Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Telomere length measurements using fluorescence microscopy Poon, Steven Sui Sang 1997

You don't seem to have a PDF reader installed, try download the pdf

Item Metadata

Download

Media
ubc_1998-272273.pdf [ 27.89MB ]
Metadata
JSON: 1.0064855.json
JSON-LD: 1.0064855+ld.json
RDF/XML (Pretty): 1.0064855.xml
RDF/JSON: 1.0064855+rdf.json
Turtle: 1.0064855+rdf-turtle.txt
N-Triples: 1.0064855+rdf-ntriples.txt
Original Record: 1.0064855 +original-record.json
Full Text
1.0064855.txt
Citation
1.0064855.ris

Full Text

Telomere Length Measurements Using Fluorescence Microscopy STEVEN SUI SANG POON, P.Eng. M.A.Sc, The University of Britisch Columbia, 1989 B.A.Sc, The University of Britisch Columbia, 1985 A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY in THE FACULTY OF GRADUATE STUDIES Department of Electrical Engineering We accept this thesis as conforming to the required standard THE UNIVERSITY OF BRITISH COLUMBIA December 1997 © Steven Sui Sang Poon, 1997 In presenting this thesis in partial fulfilment of the requirements for an advanced degree at the University of British Columbia, I agree that the Library shall make it freely available for reference and study. I further agree that permission for extensive copying of this thesis for scholarly purposes may be granted by the head of my department or by his or her representatives. It is understood that copying or publication of this thesis for financial gain shall not be allowed without my written permission. Department of <5^2Q^72«tX <5v The University of British Columbia Vancouver, Canada Date DE-6 (2/88) TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY ii Abstract We describe a system to estimate the lengths of telomere repeat DNA sequences in individual chromosomes of cells. Conventional systems based on gel electrophoresis of digested DNA can only determine the telomere length distribution of a population of cells but cannot provide estimation of telomere lengths of individual chromosomes using a limited number of cells. Our method is based on analyzing microscopy images of metaphase chromosomes prepared using fluorescence in-situ hybridization technology. In order to obtain reliable and reproducible images for measurements, we have developed and optimized an image acquisition system specifically for this purpose using commercially available components. In this study, two types of images are used. The first image highlights the chromosome regions. The second highlights only the telomere regions and consists of a set of multi-focus plane images. We first perform segmentation on the acquired multi-focus plane images in order to determine the region that each telomere occupies. For each telomere region determined, the total integrated fluorescence intensity (IFI) is obtained. This calculated IFI value is normalized for spatial unevenness in illumination in the field of view as well as the day to day variation in the illumination intensity. The acquired chromosome image is then segmented using novel and simple algorithms developed for such purpose. The location of the segmented telomere and chromosome boundaries are then overlaid onto the chromosome image. Different colour boundaries are used to highlight and identify different TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY iii telomeres within a chromosome as well as the number of telomeres detected within each chromosome. The resulting image and the calculated telomere IFI values for each chromosome are then presented on the computer screen for user verification and editing as required. The algorithms described in this thesis are presently being used on a daily basis to collect data on telomeres and to study the role telomeres play in the biology of cells. Our method of analysis has made it possible to generate reliable estimates of the length of telomeres in individual chromosome arms using a limited number of cells. In addition, our automation of the analysis has significantly reduced the user interaction time for editing and verification of the results. To date, at least two different biological studies of telomeres have been completed on the system we developed. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY iv Table of Contents Abstract ii Table of Contents iv List of Tables x List of Figures x Acknowledgments xiii Chapter 1.Introduction 1 1.1. Document Introduction and Overview 1 1.2. Telomeres and Their Function in The Cell 5 1.3. Telomere Quantification Methods 9 1.3.1. Conventional Technique: Southern Analysis 9 1.3.2. New Technique: Quantitative FISH and Image Analysis 10 1.4. Objectives 12 Chapter 2.Background 15 2.1. Imaging Systems2.1.1. Overview2.1.2. Wide-field Microscopes 20 2.1.3. Confocal Microscopes 4 2.2. Image Analysis 28 TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY v 2.2.1. Overview 28 2.2.2. Image Pre-Processing 30 2.2.2.1. Background Subtraction 32.2.2.2. Flat-field Compensation 1 2.2.2.3. Wavelength Compensation 32 2.2.2.4. Photobleaching Compensation2.2.3. Focus and Three Dimensional Reconstruction 33 2.2.4. Image Segmentation 35 2.2.4.1. Overview2.2.4.2. Thresholding or Clustering 36 2.2.4.3. Edge Detection 38 2.2.4.4. Region Extraction 9 Chapter 3.Imaging System 41 3.1. Overview 43.2. Image Acquisition Hardware 42 3.2.1. Overview3.2.2. Fluorescence Microscope 43 3.2.2.1. Illumination Source 5 3.2.2.2. Excitation and Emission Filters 46 3.2.2.3. Objective Lens 47 3.2.3. Focussing Mechanism 9 3.2.4. High Resolution Camera 50 3.2.5. Computing System 2 3.3. System Temporal Stability and Aberrations 53 TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY vi 3.3.1. Overview 53 3.3.2. Temporal Fluctuations in Illumination 53 3.3.3. Photobleaching Effects 56 3.3.4. Uneven Illuminated Field of View 58 3.3.5. Flat-Field Compensation Results 61 3.4. Image Acquisition Software 62 3.4.1. Overview 63.4.2. Image Exposure and Photometric Range Selection 63 3.4.3. Image Pre-Processing: Sensor Defects 64 3.5. Image Analysis System 65 3.5.1. Image Analysis Hardware 63.5.2. Image Analysis Software 5 Chapter 4.Acquisition System Characteristics 70 4.1. Background 74.2. Our Derivation of the System OTF 75 4.3. Theoretical OTF and PSF Results 80 4.4. Initial OTF/PSF Comparative Study 5 4.5. Comparison With Experimental PSF 89 Chapter 5.Telomere Segmentation and Integrated Fluorescence Intensity Measurements 94 5.1. Overview 95.2. Theory in IFI Quantification 96 5.3. Segmentation and IFI Quantification Algorithm 100 5.4. Number of Focus Planes Required 108 TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY vii 5.5. Algorithm Evaluation And Validation 112 5.5.1. Overview 115.5.2. Spatial Resolution of IFI Segmentation 114 5.5.3. Simulated Objects 115 5.5.4. Fluorescent Beads 121 5.5.5. Plasmids 122 5.5.6. Summary of Algorithm Validation Results 124 5.6. Human Telomeres Results 126 5.6.1. Number of Focus Planes Required 125.6.2. Telomere Distribution in Cells 130 5.7. Chapter Summary 131 Chapter 6.Segmentation of Chromosomes 133 6.1. Overview 136.2. Rank Difference Filter 136 6.3. Comparison of Edge Detectors 144 6.4. First Approximation to Edges: Thresholding 154 6.5. Second Approximation to Edges: Texture Detection 157 6.6. Third Approximation to Edges: Region Refinement and Labeling 163 6.7. Feature Extraction and Artifact Removal 166 6.8. Associate Telomere with Chromosome 166.9. Segmentation Performance 168 Chapter T.Conclusion and Future Suggestions 172 7.1. Overview 17TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY viii 7.2. System Performance 173 7.2.1. Imaging system 6 7.2.2. System Characteristics 177 7.2.3. Telomere IFI Value7.2.4. Chromosome Segmentation 178 7.3. Current Biological Studies 179 7.4. Future Suggestions and Applications 182 Chapter 8.Bibliography 184 TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY ix List of Tables 5.1. IFI values of typical human telomeres at different focus levels 119 5.2. IFI values of simulated test objects calculated at different focus level sampling spacings 120 5.3. IFI values of different size beads 121 5.4. IFI values of different size plasmids 123 5.5. Normalized IFI values of objects at different focus positions 125 5.6. IFI values of typical human telomeres at different focus levels 126 5.7. IFI values of human telomeres calculated at different focus level sampling spacings 129 TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY x List of Figures 1.1. Diagram of telomere location in a metaphase chromosome 6 1.2. Telomere lengths during chromosome duplication prior to cell division 8 1.3. Fluorescence image of telomeres and chromosomes 13 2.1. Block diagram of a typical imaging system 17 2.2. Model of a microscope system 22 2.3. Principle of confocal microscopy 5 2.4. Process for cell analysis 29 3.1. Block diagram of the imaging system 43 3.2. Block diagram of the excitation and emission filter system 48 3.3. Illumination variations over time 55 3.4. Photobleaching effects on telomere fluorescence 56 3.5. Photobleaching effects on bead fluorescence 7 3.6. Illumination variation over the field of view 59 3.7. Intensities of the central ros of pixels of a homogenous sample 62 3.8. Flat field compensation for spatial illumination variations 62 3.9. Telomere analysis program 67 4.1. Relationship of the in-focus and defocus distances in the object and image side of the objective lens 73 4.2. Theoretical OTFs of the system in the xy-plane 81 4.3. Theoretical PSFs of the system in the xy-plane 82 4.4. Theoretical PSF of the system in the xz-plane 82 TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY xi 4.5. Theoretical PSF distribution of the system at various z-spacings 83 4.6. In-focus system response for various objectives 88 4.7. Typical images of four 0. lm beads acquired at different focus spacings 90 4.8. Comparison of experimental and theoretical PSF distributions as a function of z-focus position 91 5.1. Errors in calculating the object IFI 102 5.2. Application of the average difference filter 105 5.3. Process of segmenting a typical telomere image 106 5.4. Simulated test objects for spatial resolution estimates 115 5.5. Simulated test object values and shapes 118 5.6. Normalized IFI values at varying focus of different objects 119 5.7. IFI distribution of different size beads 122 5.8. IFI Distribution of Different Size Plasmids 124 5.9. IFI of typical human telomeres at different focus levels 127 5.10. Telomere IFI distribution in a cell 131 6.1. Central pixel of rank difference filter 139 6.2. Shape of rank difference filter region 140 6.3. Size of rank difference filter region 141 6.4. Effect of varying upper and lower rank numbers 143 6.5. Effect of additive noise and varying upper and lower rank numbers.... 144 6.6. Performance of edge filters on test object 145 6.7. Performance of edge detectors in presence of noise added to the test object 146 6.8. Performance of edge detectors on the peppers image 148 TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY xii 6.9. Edge filter performance of peppers image with additive Gaussian noise 150 6.10. Edge filter performance of peppers image with additive uniform noise 151 6.11. Edge filter performance on chromosome image 153 6.12. Chromosome image at various thresholds 155 6.13. Second approximation to edges: texture detection 159 6.14. Rank difference of the average difference image 162 6.15. Third approximation to edges: region refinement and labeling 165 6.16. Chromosome and telomere segmentation results 168 6.17. Comparison of our segmentation algorithm to the Canny filter 169 7.1. Telomere lengths of individual chromosomes in a cell 181 TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY Xlll Acknowledgments I am most grateful to my supervisory committee for their guidance, advice, support, and patience throughout this work. In particular, I am in debt to Dr. Rabab Ward, my supervisor, for her time and sound advice during our discussions, particularly with respect to the technical issues. She was also very instrumental in providing alternative methods to solve various problems encountered throughout the project. I am also very grateful to Dr. Branko Palcic, who is not only a member of my supervisory committee but also an excellent mentor and visionary. Without him, I would not be able to apply my imaging methods for use in an new and interesting biological research applications. Lastly, I am most appreciative to Dr. Peter Lansdorp, another member of my supervisory committee, for his support, facilities, and resources for the biological aspects of this project. His eagerness in the telomere project, his vision into the directions of how the system would be used, his understanding of the impact as well as the limitations of automating some of the processes in the biological studies, and the regular feedback from him and his group during the development of the project were key in making this project successful. In addition, I would like to thank all the staff and student members of the B.C. Cancer Research Centre who have assisted me in this thesis. Particularly, I would like to thank the group involve with telomere research in the Terry Fox Laboratory for Hematology/Oncology (Dr. Peter M. Lansdorp, Dr. Uwe, M. Martens, Dr. J. Mark J.M. Zijlmans, and Liz Chavez). I would like to TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY xiv give special thanks to Dr. Uwe M. Martens for his help in preparing the microscope slides, acquisition of some of the images used in this thesis, and most importantly, offering constructive criticism and feedback on the use of the telomere image analysis software during its development. I would also like to thank Mr. Hector Huang for modifying some of the code in the acquisition program to incorporate the z-drive control and automatic exposure time setting. In addition. I would like to thank Mr. Bong Choun Ming for processing the test images using the Canny and Difference of Gaussian edge filters. This work is supported by scholarships from the Science Council of British Columbia and grants from the National Institutes of Health, the Medical Research Council of Canada, the National Cancer Institute of Canada, the German Science Foundation (U.M.M.) and the Dutch Cancer Society (J.M.J.M.Z.), and equipment from Xillix Technologies Corp. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 1 Chapter 1. Introduction 1.1. Document Introduction and Overview This thesis describes a new method for determining the length of telomeres in a cell. Telomeres are nucleo-protein complexes containing specific DNA repeat sequences found at the ends of each of the 46 chromosomes in a human cell. These repeat sequences represent approximately one ten thousandths of the total DNA in the cell. Due to their minute size, the task of quantifying the number of repeat sequences at individual chromosome ends has not been accomplished before. The method we developed has made it possible to estimate the length of individual telomeres and to perform new studies in telomere biology. All cells, including human, are known to divide. Each time cells divide, their number doubles. In most somatic cells, the maximum number of times the cell-division cycle can repeat itself is estimated to be 70 to 100 times. After that, cells die or become senescent: they consume less food and their membranes deteriorate; a process that is very similar to aging. To study the basis of cellular aging and mortality, current studies focus on a small area at each tip of each chromosome which is called the telomere (from telos = end and meros = part). A telomere consists of repetitive sequences of base pairs which do not appear to code for any traits and associated proteins. The repetitive DNA sequences appear to protect the ends of chromosomes and provide a cap. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 2 Each time a cell divides, some of the repeats are lost and the telomeres in the daughter cells become shorter. Finally, when the telomeres reach a critical length, the cell stops dividing (typically after 50 to 100 replication rounds). Some researchers theorize that when the telomere reaches its critical length, certain genes become active and produce proteins that trigger tissue deterioration associated with aging. While almost every cell in the human body exhibits telomere loss, a few such as sperm, egg, and cancer cells do not. Such cells are characterized by their ability to divide not just 100 times but thousands. Researchers are using what they know so far about telomeres and other cellular mechanisms to attack the diseases that keep the very old from becoming even older. Other researchers are studying how to block the telomerase RNA enzyme (which helps to lengthen telomeres) in cancer cells, leading to withering of telomeres and the death of the no longer so prolific cells. In his recent new novel, "Holyfire", science fiction author Bruce Sterling popularized telomeres. He describes a procedure by which his heroine, a very rich 95 year old woman, gets a complete cellular makeover in which the telomeres within every cell of her body were lengthened. As a result, she was transformed into a healthy 20 year old young woman. Although this phenomena is not likely to occur in the near future, this thesis will give researchers a tool which can help better study the role of telomeres in biology and perhaps, get them closer to this goal. Alternatively, new facts about telomeres uncovered by such studies may rejuvenate enthusiasm in biological sciences. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 3 Our method for determining telomere length is based on applying image analysis techniques to microscope images of chromosomes prepared using fluorescence in-situ hybridization (FISH). Our method provides more detailed information about telomere length and is considerably more sensitive than the conventional method based on gel electrophoresis. In addition, our technique enables studies on telomeres of single chromosomes and on telomere length distributions within cells. Based on this thesis work, telomere length variations within individual chromosomes and amongst different chromosomes in cells, can now, for the first time, be conveniently determined. Information about telomere lengths is used in various studies to investigate the role of telomeres in the biology (aging and malignant transformation) of cells from human and other species. Besides describing the importance of this research, the present chapter gives an overview of the thesis. It describes what telomeres are and how the technology described in this thesis can help researchers to further understand the role of telomeres in basic cell biology and molecular genetics. It introduces our technique for detecting individual telomeres and measuring their lengths using fluorescence microscope imaging and image analysis systems. The objectives of this research are also presented in this chapter. Chapter 2 presents the background of imaging systems and image analysis algorithms for cell analysis. The accuracy of identifying and quantifying the chromosomes and the telomeres is highly dependent on the focus position of objects in the microscope images. Hence, background on three dimensional (xy-spatial and z-focus plane) image analysis techniques are also covered in this chapter. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 4 Chapter 3 describes the image acquisition system that we developed and used in this research. As there are no commercially available systems for this work, the rationale, behind our selection of components is described with emphasis on the key components in the system. Image calibration of the acquired images is also covered in this section. Chapter 4 presents a new and more accurate method for analyzing and formulating the characteristics of the hardware system. We derive the characteristics of the system and compare them with the experimental results. Chapter 5 describes the theory and algorithms used to quantify telomere lengths/fluorescence in the three-dimensional plane. The theoretical analysis leads to a new and simplified method for accurate telomere quantification. Our new method directly extracts information from the multi-focus plane images of the telomeres without having to perform 3D image reconstruction. Chapter 6 outlines a novel chromosome segmentation process which yields a good approximation to the border of chromosomes (including those which are touching). For this purpose, we developed a new algorithm called the Rank Difference Filter. This filter can act as an edge detection filter as well as a selective morphologic dilating/erosion filter. The results from the telomere segmentation are then linked to the chromosome segmentation results and presented to the user for verification and editing. Finally, Chapter 7 summarizes the conclusions of this research and outlines possible areas for future investigation. The recent advances in the fluorescence in-situ hybridization (FISH) technology has made it possible to quantitatively stain specific gene sequences such as those in telomeres. Emphasis of this thesis will concentrate on the TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 5 engineering aspects of the project. A brief description of the biological aspects is presented for better comprehension of the project. Specifically, the description of the telomeres and the FISH technology are presented in this Chapter. 1.2. Telomeres and Their Function in The Cell Telomeres contain proteins and specific DNA sequences found at the natural ends of chromosomes in eukaryotic (nucleated) cells (Figure 1.1). DNA is composed of a unique sequence of base-pairs which enables differentiation of individuals within a single species. Each base-pair can be one of the thymine (T), adenine (A), cystosine (C), or guanine (G) nucleic acids, paired with its counterparts (A with T, and C with G). Specifically, telomeric DNA consists of highly repetitive base-pair sequences which are unlike the rest of the DNA sequences in the chromosome. (Benbow, 1992; Wilson et al., 1993) Depending on the species, a single telomeric DNA contains around 20 to 15,000 base-pairs consisting of repetitive 6 to 8 base-pair sequences called repeats. Within a single species, however, the length of the telomeric sequence is closely controlled. In humans for example, the repetitive sequence is 6 base-pairs long and consists of TTAGGG. The length of the human telomere ranges from around 1000 to 15,000 base-pairs (1/10000 of the total DNA). In total, there are 46 chromosomes in a human cell. Since each chromosome has a telomere at each end, there are 92 telomeres in each cell. As chromosomes are most often studied after duplication but before separation (metaphase chromosomes), a total of 184 (twice the number) telomeres can be observed in TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 6 typical chromosome slide preparations of single cells. (Benbow, 1992; Wilson et al., 1993) Figure 1.1. Diagram of telomere location in a metaphase chromosome. Telomeres are located at the ends of chromosomes. In condensed chromosomes, they may appear to only occupy a portion of the chromosome tip. Telomeres are known to play and are postulated to have a number of roles in the function of the cell (Benbow, 1992; Wilson et al., 1993). The basic function of the telomere is to cap and protect the ends of chromosomes. This function is mediated by specific repeat sequences of the telomere which are not present in the rest of the DNA. Hence, the cell is able to differentiate between the ends of a chromosome and a break within. By detecting the break (telomere repeat sequence not present at the chromosome end), the cell can appropriately initiate its repair mechanism to repair the damage. Telomeres also appear to play a role in the nuclear architecture of the cell. Specific TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 7 proteins interact with the telomere and attach each chromosome to particular sites on proteins of the nuclear matrix. The specific localization of each chromosome within the cell may determine how each chromosome functions. Furthermore, during meiosis (a special form of cell division giving rise to sperm or egg cells), the telomeres attach themselves to the nuclear envelope, resulting in a stage that promotes proper chromosome pairing. Telomeres may play a role in gene regulation as well. It has been postulated that the'length of the telomere may determine if particular genes in that chromosome are expressed or depressed. Telomeres also play an important role in normal cell division. As shown in Figure 1.2, an RNA primer first attaches itself within the first 200 base-pairs at the 3' (G-riched) end of a single strand of chromosome. After the primer has been attached, the rest of the chromosome is transcribed or replicated joining each nucleic acid with its complement (A with T, and C with G) by DNA polymerase enzymes which can only duplicate DNA in the 5' to 3' direction of the newly formed strand. At the end of the process, the RNA primer is cleaved or removed. As a result, a portion of the parent 3' telomere strand may not be fully replicated in the daughter cell. An enzyme, telomerase, can increase the 3' length of telomeres but is generally not expressed in normal somatic (non-germ line) cells. It has been recently shown that telomere lengths shorten with age until the telomeres reach a certain length which prevents the cell from dividing (Harley et al. 1990; Allsopp et al. 1992; Levy et al. 1992; Vaziri et al. 1994). The relationship of telomere lengths and its relationship with cancer TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 8 and with age is an active research area (Hastie et al. 1990; de Lange 1995; Lansdorp et al. 1996; Lundblad and Wright, 1996). (CCCTAA), (a) Parent Chromosome Strands (TTAGGG) n 1 3' RNA Primer (GGGATT) n 0 (AATCCC) n (b) Replicated Strands <^ 3 5' (c) Two Sister Chromosomes <^ -|- Telomere Replication Loss 5' Figure 1.2. Telomere length during chromosome duplication prior to cell division, a) An RNA primer attaches anywhere within the first 50 to 200 base-pairs of the 3' end. b) Once the primer is attached, the chromosome replicates, c) At the end of the replication process, the RNA primer is removed. The resulting sister chromosomes would then have shorter telomeres than the parent. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 1.3. Telomere Quantification Methods 9 1.3.1. Conventional Technique: Southern Analysis The conventional technique for measuring the length of telomere repeat sequences is based on a method called Southern analysis (Allshire et al. 1988; de Lange et al. 1990). In this technique, DNA is first digested by enzymes to obtain segments containing the telomeric repeats. The resulting fragments are deposited onto a gel and the process of electrophoresis is initiated. The smaller DNA fragments move further in the gel than the larger ones. The distance these fragments move in the gel, under the force of the electric field, has a logarithmic relationship to the size of the DNA fragments (telomeres). A radiolabelled probe specific for the TTAGGG telomeric repeat sequence is then allowed to hybridize (bind) to the telomeric repeats in the (blotted) gel. After washing off the excess, the presence of radiolabelled probe can be visualized on an X-ray plate to record the final destination of the telomeric repeat segments onto film. The distribution of telomere lengths within a sample can then be obtained. The size distribution of telomere segments can be determined from the spot signal intensities and their respective distance from the starting point. There are a number of drawbacks to this technique. First, at least 100,000 cells are required to give enough DNA to obtain a good representation of the telomeric lengths within a sample. Second, the size of the telomere may be over-estimated. The digestive enzymes used typically cleaves at the sub-telomeric region resulting in telomeric segments with additional base-pairs that do not have the telomeric repeat sequence. Hence the apparent telomere size is TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 10 increased and over-estimated. Even if the same enzyme is used, the amount of non-telomere repeats within the segments may vary from chromosome to chromosome. Lastly, the size and number of short telomeres in the sample are under-represented. Short telomeres of similar lengths are more dispersed in the gel compared to the longer telomeres because of the logarithmic size relationship. Hence, a larger number of short telomeres are required to generate the same or similar signal intensities as those generated by long telomeres. Due to the limited dynamic range of the film, the transducers and the logarithmic representation of telomere size, longer telomeres are favored as their signals are stronger and the short telomeres become under-represented. The system described below overcomes these difficulties. Instead of around 100,000 cells, less than 30 cells are required to measure the telomere length distribution. The size of the telomeres obtained is not under or over estimated and the obtained telomere signals are not biased in their lengths. Although our system cannot analyze cells which do not divide (e.g. senescent cells), many other biological studies on telomeres can still be carried out using the system we developed. 1.3.2. New Technique: Quantitative FISH and Image Analysis A new technique based on the fluorescence in-situ hybridization (FISH) protocol has been developed to measure telomere lengths (Lansdorp et al. 1996). FISH technology relies on probes (nucleic acid sequences) that can hybridize (bind base-pairs) to specific sites in denatured chromosomes. By attaching a fluorescence label onto the probe, the location of specific sites in TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 11 the chromosome can be identified under a fluorescence microscope. Although the FISH technique has been around for the past 10-15 years, it is not until recently that probes have been developed which hybridize to telomeres with sufficient efficiency such that quantitative analysis of FISH images can be considered. Synthetic peptide nucleic acid (PNA) oglionucleotide probes are used for this purpose. It was found that such probes can hybridize under conditions that do not favor DNA to DNA and DNA to RNA hybridization (Nielsen et al. 1991; Egholm et al. 1993). As a result, these new probes do not need to compete with the surrounding DNA or RNA to bind to the telomeric repeat sites. In our study, two types of images of the cell are used. The first one is an image (or images at different focal planes) of the telomeres only. From this image(s), the fluorescence intensity of every telomere is determined. However, to determine which telomere belongs to which chromosome, we use a second image which is an image of the chromosomes without the telomeres. To obtain the first image, a fluorescent probe (PNA sequence of (CCCTAA)3) is used for the detection of the TTAGGG repeats in the telomeres. Since, there are no other sequences present in the cell which compete with the binding of this probe, the resultant image acquired highlights only the telomere signals. Because in general, the length of a telomere is correlated very well with the number of fluorescently labeled probes attached to it, the measured fluorescence intensity of the telomere provide information about the length of the telomere. By combining multiple measurements, random variations in TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 12 fluorescence measurements are filtered out and reliable estimates of telomere length can be obtained. A different coloured fluorescence label can also be used to highlight an entire chromosome. As a result, a second image of only chromosomes (without telomeres) is then obtained. The telomere image can then be superimposed onto the chromosome image to help determine which telomere belongs to which chromosome (Figure 1.3). By identifying each chromosome in a cell, telomere length distribution of each chromosome in a particular cell type may be obtained. Initially, FITC and PI probes were used for marking the telomere and chromosome respectively. Currently, CY3 and DAPI probes are used instead of FITC and PI probes respectively. The currently used probes are better for this analysis because i) there is less spectral overlap between the probes used, ii) the CY3 probe photobleaches less than the FITC probe and iii) the DAPI probe highlights the bands within the chromosome which facilitates karyotyping (identification and differentiation of chromosome types in the cell). 1.4. Objectives The hypothesis of this thesis is that the length of individual telomeres in a particular cell type can be determined from fluorescence microscope images of the telomeres in a limited set of metaphase chromosome spreads of that cell type. In order to test this hypothesis, we had to adapt existing microscope techniques and develop new image analysis algorithms. Because the hypothesis is based on new DNA detection methods, the corresponding tasks have not been accomplished before. As a result, comparisons with existing TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY Figure 1.3. Fluorescence image of telomeres and chromosomes. The pseudo colour image is generated by superimposing the red, CY3 labeled telomere image onto the blue DAPI labeled chromosome image. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 14 techniques are not straight forward. Conventional systems using Southern analysis can only determine the distribution of telomere lengths in a population of cells and not the telomere lengths of individual chromosomes or cells. The goal of this thesis is to prove our hypothesis by developing a fluorescence microscope imaging and analysis system to quantify the telomere fluorescence and thereby the telomere lengths in individual chromosomes. We will use the break-through offered by the PNA-FISH technology (Lansdorp et al., 1996) to quantitatively highlight individual telomeres of the cell. Since the telomeres and their attached fluorescence probes occupy a 3-dimensional space, we shall use 3-dimensional images in our analysis. To achieve our goal, the tasks needed are: 1. to build the fluorescence microscope imaging and analysis system; 2. to develop programs for acquiring images of telomeres and chromosomes; 3. to calibrate the acquired telomere images for illumination unevenness, optical aberrations, camera defects, and photobleaching effects; 4. to detect and segment telomeres from the acquired 3-dimensional (x,y,z) telomere images; 5. to obtain a reliable estimate of the length of every telomere from the telomere images by measuring the telomere's integrated fluorescence intensities (IFI) in multiple metaphase cells; 6. to detect and segment chromosomes from the chromosomes image; and 7. to associate every telomere (and its calculated telomere IFI value) with the corresponding segmented chromosome in the images and present the results to the user for editing and verification. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 15 Chapter 2. Background 2.1. Imaging Systems 2.1.1. Overview The application of machine vision and robotics to microscopy began in the early 1950s with the introduction of blood cell analyzers (Young and Roberts, 1951; Walton, 1952). Since then, many applications of machine vision to other areas of cell biology have been developed. These include applications in cervical cell screening (e.g. Bengtsson et al., 1979; Tucker, 1979; Shoemaker et al., 1982, Palcic et al., 1992) and in the analysis of chromosomes (e.g. Preston, 1976; Philips and Lundsteen, 1985). Currently, there are a number of commercially available imaging systems for general cell analysis, such as those from Becton Dickinson Incorporated and Oncor Instrument Systems. Current imaging systems specifically developed for the analysis of chromosomes include those developed by Applied Imaging Inc., Biological Detection Systems Inc., Vysis Inc., and Perceptive Scientific Instruments Inc. All of these systems have the capability to manipulate and work with multi-spectral images. The applications they are designed for include spot counting, comparative genomic hybridization and karyotyping, chromosome probe mapping, and intensity and morphometric measurements. In general, all these systems require human interaction to verify and correct the results generated. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 16 Although there are many capabilities built into these imaging systems, these systems do not support new areas of chromosome research as in the topic of this project. As these systems are application specific, proprietary and are not "open systems", it is difficult to integrate new algorithms, modify existing algorithms and incorporate new components into them to perform our research. The source code of these programs is huge and difficult to obtain. Even when the source code is made available, there is little technical support provided. Thus, many researchers resort to develop their own in-house systems for addressing new research application areas (e.g. Lockett et al. 1990; Poulin et al., 1989, Nederlof et al., 1992). Hence, we decided to design and develop our own system for telomere image acquisition and analysis. This would give us the ease and flexibility to implement algorithms which need to be tailored to our specific area of research. In designing the acquisition system, a good understanding of the properties, characteristics, and tradeoffs of the diverse selection of components is required. Using this knowledge, the appropriate components can then be selected such that the integration of these components are optimal for the intended telomere application (Chapter 3). An overview of the different components of the acquisition system is described below. Although many image cytometry (cell measurement) systems have been developed, they are all very similar in design and operation. The basic system consists of an illumination source, microscope optics, a mechanical stage, a camera, digitizing circuitry, image memory, a display monitor, and processors (Figure 2.1). The stage which holds the samples is capable of moving objects in TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 17 the x and y directions (in a plane parallel to the detector's field of view). In some systems, a motorized z direction is provided for automated focussing purposes. Light is first transmitted from the illumination source to the sample. An image of the sample is then acquired by the camera detector. The detected image is transformed into a digital image by the digitizing circuitry. The resulting digital image is then stored in computer memory from where it can be displayed on the monitor and/or processed and analyzed by the computer. This technology has brought accuracy, uniformity, reproducibility, and a control level of quality to microscope analysis. Camera Illumination Source Digitizing Circuitry Microscope Optics Microscope Stage Display Monitor Image Memory ~7K Computer/ Processors! Figure 2.1. Block diagram of a typical imaging system. The major differences amongst these imaging systems are the optics and transducers (detector) employed and the method used in scanning. The most TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 18 important component in the optical system is the objective lens of the microscope. This lens determines the magnification of the sample (the object) and the spatial resolution of the image produced. These lenses are not perfect and generally introduce distortions, aberrations, and shading effects. Another important factor in the optical system is the arrangements of the optics and components in the microscope. The most common optics arrangement for looking at samples are found in widefield microscopes. Confocal microscopes may also be used, particularly in situations where out-of-focus blur is a problem. In imaging systems such as the ones described above, the quality of the output image is dependent not only on the type of optics used but also on the transducer and electronic (digitizing) circuitry. The function of the transducer is to convert the optical image to an electronic form. Most systems use a two dimensional detector such as those found in tube cameras or a two dimensional array charge coupled device (CCD) cameras (e.g. Tucker, 1979; Jaggi et al., 1987, 1988, 1990). Video cameras scan and sample the sensor to obtain a signal in an analog (video) format which then requires sampling by a computer to convert it to a digital form. Very few systems use linear detectors such as CCD or diode arrays (e.g. Bengtsson et al., 1979; Jaggi et al. 1985, 1986, Palcic et al., 1987; Tucker et al., 1987). In these systems, a digitized image is obtained by moving the sample across the sensor and then piecing together the image lines in the computer. The advantage of linear scanners is that they, in general, have more sensor elements than a single row or a single column in a two dimensional detector and hence provide a wider field of view. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 19 Another advantage is that the elements of the sensor are digitized and transmitted as digital data. The disadvantages of such a system, for the purpose of acquiring high resolution 2-dimensional images, are i) the requirements for a precise mechanical scanning system to move the linear sensor across the image and ii) the increase in time for acquisition. Systems which use a combination of the linear and matrix (video) detectors (e.g. Graham and Norgen, 1980) also exist. Here, objects detected by the linear array are moved into the field of view of the two dimensional detector where a higher resolution image is acquired through a higher magnification lens. There are also systems which use a one element detector such as a photomultiplier or a photodiode (e.g. Ingram and Preston, 1970; Shoemaker et al., 1982). In these systems, a mirror is used to deflect the laser spot to scan the object and the corresponding signal from each spot is assembled in the computer to create the image. Most cameras today are built for the television broadcast community where the image detected by the transducer is converted to an analog signal. This signal, representing the image, is later digitized (at a pixel grid which is generally different from that of the sensor) for machine analysis. As a result, some information is lost due to the indirect image sampling and related errors. Better noise performance cameras directly sample and digitize the sensor image into digital data for machine analysis. They are however more expensive and do not conform to a fixed standard of data transmission. Regardless of the components used in the system, distortions introduced by the optics, TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 20 transducer and digitizing circuitry should be compensated for, to help simplify the analysis process and obtain more accurate results. 2.1.2. Wide-field Microscopes The wide-field microscope is the basic instrument used to obtain a magnified view of the sample. The basic components of the wide-field microscope are i) the illumination source, ii) the objectives, and iii) the viewing apparatus. The illumination source supplies the medium (light) which is modified and transported from one component to the next. The source consists of some light generation mechanism and a focussing media to direct the light to the sample as either coherent or incoherent illumination. The light measured can arise from light which is i) transmitted through the sample (as in bright-field microscopy), ii) reflected off the sample (as in reflectance microscopy), or iii) absorbed by the sample and re-emitted at a different wavelength (as in fluorescence microscopy, the mode used in our study). The transmitted, reflected or re-emitted light is focussed by the objectives and possibly other lenses to produce a magnified image at the viewing surface such as the oculars and/or a transducer. The transducer converts the light to an electronic signal which can then be digitized and stored in a computer for further analysis. For the wide-field microscope, the limiting resolution depends mainly on the optics of the system. The limiting resolution of a system defines the distance by which two points can be resolved. Let the xy-plane be perpendicular to the direction of the focal axis z. It can be shown from Raleigh's criterion that the x and y resolution of the microscope is proportional TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 21 to the wavelength of light (k) used and inversely proportional to the numerical aperture (NA) of the system. In the z direction, Francon's criterion states that the resolution is proportional to the wavelength and the index of refraction (n) of the medium and inversely proportional to the square of the numerical aperture. (Inoue, 1986). That is, dv = ^± (2-1) NA dz = 1^ (2'2) NA2 where dv is the limiting resolution in the x-y plane, and dz is the limiting resolution in the z direction and is often referred to as the depth of field of microscopes. Sheppard (1988) has also characterized the depth of field of microscopes that have high numerical aperture objectives in an air medium. He has derived a number of different expressions based on diffraction optics. These include i) a divergent beam, ii) a divergent beam in (the more realistic) aplanatic systems where the beam is focussed to a plane wave, and iii) a paraxial approximation to the aplanatic system. These expressions result in depth of field values which are closed to Francon's criterion (stated above) for low numerical aperture (< 0.6) objectives and is 50 to 75% of Francon's criterion for numerical aperture approaching 1.0. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 22 Using equations 2-1 and 2-2, the limit of resolution in the x (or y) and z plane for a wavelength of 600nm and for a numerical aperture of 1.4 with an oil medium index of refraction of 1.5 (conditions which are used in our research), are 0.26 and 0.46 microns respectively. The telomeres and chromosomes in the samples we used generally lie at different focal planes spanning a depth which can be thicker than the resolution limit in the z direction. As a result, out-of-focus blur from planes above and below the focal plane are present in the acquired image. Thus, more than one image plane (as shown later in Chapter 5) are required to obtained a more accurate representation of the telomere length. Sample Image M s(x,y,z) Microscope Characteristics h(x,y,z) Observed ^ Image o(x,y,z) Figure 2.2. Model of a microscope system. The image of the sample is modified by the microscope to result in the observed image. We need to characterize the system in order to understand its behaviour. The general characterization is given below, while the details specific to our system will be elaborated upon in Chapter 4. The image formed at the transducer can be characterized using a simplified model of the microscope TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 23 (Figure 2.2). In this model, the characteristics of the light source, lenses, and transducer are combined into a "black box" with the sample image as the system input and the observed image as the system output. The analysis of such system is carried in the three spatial dimensions to correspond to the three dimensional nature of objects. Thus, in the mathematical formulation, the sample is described as s(x,y,z). This sample image is modified by the objectives and other lenses of the system and its characteristics can be represented by a transfer function in the form of a three dimensional point spread function (PSF), h(x,y,z). The observed image, however, is a two dimensional one, o(x,y). Albeit, a three dimensional image, o(x,y,z) can be obtained by acquiring and storing several xy-plane images at varying focal levels (z-direction of the microscope). The resulting three dimensional image can be described as the 3-dimensional convolution of the system PSF with the sample, i.e.: where ® denotes the convolution operator. In the Fourier domain, the observed image spectrum, 0(u,v,w), is the multiplication of the microscopes optical (contrast) transfer function (OTF), H(u,v,w), with the image spectrum of the sample, S(u,v,w). o(x,y,z) = h{x,y,z)®s{x,y,z) (2-3) 0(u,v,w) = U(u,v,w) • S(u,v,u>) (2-4) The OTF of the microscope can be determined theoretically. Hopkins (1955) has derived the OTF function for a general optical system with a circular TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 24 aperture and incoherent illumination. Later, Stokseth (1969) has introduced an approximation to the equation for large amounts of defocus. In addition, Goodman (1968), Castleman (1979), and others have given a general description on this topic and defined the function for different shape apertures. Many assumptions of ideality are used in the derivations, but these do not hold for high numerical aperture objectives (e.g. the approximation of sinB with 9 when 9 is large). For image reconstruction, some researchers use the experimentally determined three dimensional OTF (Agard et al. 1989), while others use the theoretical OTF (Erhardt et al., 1985). One experimental method to determine the OTF is to use small fluorescent beads as point sources of light and apply Fourier transforms to the observed varying focal depth images (Hiraoka et al. 1990). 2.1.3. Confocal Microscopes The confocal microscope is an alternate choice of microscope for capturing fluorescence images used in our study. It can produce higher resolution images than the wide-field microscopes but has other drawbacks. The confocal microscope is originally proposed by Minsky (1957). The principle of its operation (Figure 2.3) is that the illumination source is focused to a single point on the sample (the illumination pinhole) and the light from this point is focussed and detected by the transducer through the detector pinhole. Only a portion of the information (mostly in-focus) at the illuminated point is allowed to pass through the pinhole while most of the out-of-focus information is blocked. Consequently, the resulting image is of high lateral and axial TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 25 resolution. A three dimensional image can be generated by moving the illumination and detection pinholes using mirrors and/or moving the sample using the microscope stage to scan the object. The speed in which an image is acquired is governed by how fast the scanning takes place and how long a spot must be illuminated in order to generate enough photons for detection. Thus an intense illumination source such as a laser or mercury arc lamp is often used. illumination condensor object objective detector pinhole lens lens pinhole Figure 2.3. Principle of confocal microscopy. A point source of light is used to illuminate the sample. The out-of-focus image is blocked and only the in-focus image is observed through the detector pinhole. In confocal microscopes, it is critical that the illumination and detection light paths behave similarly such that a focused image is seen at the detector pinhole. There can be variations in lenses as one lens is likely to behave differently than another. To overcome these variations, the objective and TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 26 condenser lenses are generally the same lens. In this instance (and also in wide-field fluorescence microscopy), a dichroic mirror is used to deflect the illumination to the sample and the emitted light is passed through to the detector. In biological work, the confocal microscope is usually used for fluorescence imaging where light at a particular wavelength is used to excite the sample. The components which are excited by the light will fluoresce (emit) light at a different wavelength. The spectral selection of specific components is particularly useful for extracting and identifying the objects of interest. As in the case of wide-field microscope, the PSF or OTF of the confocal microscope can also be characterized. The characteristics of the confocal microscope has been described by (Inoue, 1986; Pluta, 1988; van der Voort et al., 1988). In the ideal case where the pinhole is infinitely small, the system PSF can be described by the square of the PSF of the wide-field microscope (Wilson and Sheppard, 1988) and is given by o{x,y,z) = h2{x,y,z)® s(x,y,z) > (2-5) As a result of the squaring function, the PSF of the confocal system is much sharper than that of the wide-field microscope and thus is of higher spatial and axial resolution. In practice, however, the confocal microscope is rarely used in its ideal state. Generally, the pinhole has a certain diameter and is adjustable to compensate for the faint light reflected or emitted from the sample. Hence the PSF or OTF of the system should be experimentally determined for each given setup. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 27 Because of the time it takes to scan a single point over an area of the sample, faster confocal microscopes are desired. This has been realized using i) a tandem scanning microscope (Inoue, 1989), ii) the linear scanning microscope (Wilson and Hewlett, 1990), or acousto-optic scanning microscope (Goldstein et al., 1990). In the tandem scanning microscope, a large number of spatially separated pinholes on a disk (Nipkow disk) are used such that more than a single point can be imaged without interference from neighbouring points at a given time. As the disk is rotated, a different set of points are imaged and hence the entire image can be composed at a faster rate. In the linear scanning microscope, a slit shaped illumination and the corresponding line rather than a point detector is used. This results in high resolution in one direction (e.g. x) similar to that of confocal and lower resolution in the others (e.g. y and z) compared to that of confocal. The resolution of this system is still better than that of the wide-field microscope. The reason why we chose the wide-field microscope in this study is because confocal microscopes often have a more intense illumination source which can cause higher sample photobleaching. In addition, the acquisition time in confocal microscopes is significantly increased if high spatial resolution is to be maintained. Long exposure times, up to 10s, are typically required to capture the weak telomere signals. To reduce the acquisition time in confocal microscopes, a sub-region of the image is typically used for object focussing before the entire image is captured. As a result, this sub-region will exhibit higher photobleaching effects because it is exposed to more light during the focussing process than the rest of the image. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 2.2. Image Analysis 28 2.2.1. Overview A typical process for analyzing cells and chromosomes involves the following six steps: i) acquisition of images, ii) pre-processing the acquired images, iii) segmentation of the cells in the scene, iv) post-processing the segmented regions, v) extraction and quantification of features, and vi) classification of the segmented objects (Liedtke et al., 1987; Poon et al., 1989a, 1992c, 1993a) (Figure 2.4). The first two steps, acquiring and pre-processing the images, are critical since high quality input images do simplify and reduce the amount of processing required in the later stages of the analysis. Quality of the sample preparation and acquisition system is thus very important. The next step segments or defines the regions of interest in the image, such as the objects from the background. Post-processing (either automatically or interactively) of the defined regions is required to fine-tune the mask of each region. Features are then calculated based on the defined region boundaries. The objects in the scene are then classified based on the values of the features and/or the segmented results. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 29 Acquisition Pre-processing Segmentation Post-processing Feature Extraction Object Classification Figure 2.4. Process for cell analysis. In our analysis, we perform a similar process as described above. However, the specific algorithms and methods used are tailored for our telomere and chromosome segmentation and the extraction of the IFI value. These algorithms will be discussed in Chapters 5 and 6. A general description of some basic pre-processing and segmentation algorithms is given below. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 30 2.2.2. Image Pre-Processing Image acquisition and pre-processing are key steps in the analysis process for obtaining consistent and repeatable results. Images of the same sample may be different depending on i) the consistency in preparing the sample within the same batch and amongst batches, ii) the location of the object on the slide, iii) the location of the object in the microscope's field of view (x, y, and focus), and iv) the time the image is acquired (illumination stability, aging and photobleaching effects). Sample preparation differences can be compensated for by calibrating the cells of interest to similar stained objects/cells with known characteristics on the same slide (Palcic et al., 1992). Pre-processing techniques have been developed over the years, to correct for differences in illumination, sensor, and optics aberrations at different x, y locations (Poulin et al., 1994). Amongst these techniques are background subtraction methods (Castleman, 1979), flat-field compensation methods (Poulin et al., 1994), wavelength compensation methods (Castleman, 1993), and photobleaching methods (Rigaut et al. 1990, 1991). These methods are briefly described below. Some of these methods or variations of them are used as applicable in our algorithm (Chapter 3). 2.2.2.1. Background Subtraction One method of correcting for the shading and aberration effects is to perform background subtraction. In this method, the local difference in the field of view (represented by the bright background image, B(x,y)) is subtracted from the acquired image, I(x,y), resulting in the corrected image, C(x,y): TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY C(x,y) = l{x,y)-B(x,y) + K 31 (2-6) A constant value, K, equal to the average value of the bright image is added such that the range of grey levels in the corrected image is similar to that of the acquired image. As no multiplications or divisions are used in this method, truncation errors are minimal. 2.2.2.2. Flat-field Compensation A more accurate method than the background subtraction correction is to use flat-field compensation. This compensation attempts to scale the acquired image depending on the local bright background image. This method is based on the conversion of the light transmittance (which is detected by the sensor) to optical density values by taking the logarithms of the pixel values. The optical density values can then be added and subtracted to simulate the physical properties of the system. The conversion operation is determined as follows: This equation can be converted to the following flat-field compensation equation by removing the logarithms and subtracting a dark background image, D(x,y), from the measured images (i.e. I(x,y) = F(x,y) - D(x,y) and B(x,y) \ogC{x,y) = logl{x,y) - logB(x,z/) + log(k) (2-7) = B'(x,y) + D(x,y): C(x,i/) = k V{x,y)-D{x,y) B'{x,y)-D{x,y) (2-8) TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 32 As division is used in this technique, truncation errors are likely to distort the distribution of (discretely binned) grey levels. 2.2.2.3. Wavelength Compensation Multiple probes are often used in microscopy imaging. This generally requires acquiring images at different wavelengths to highlight different features of the cell or tissue. Depending on the probe used and their spectral characteristics, there can be an overlap in the spectral properties of the probes used. Castleman (1993) has developed a method of compensating for the overlap effects and isolating the signals from each probe using matrix algebra. In this method, the spectral response of each probe at each of the observed wavelengths can be represented by a vector in matrix C. For example, probe X may have normalized responses of 0.1, 0.7, and 0.2 at wavelengths Xlf X2, and X3, respectively. The response of each probe (vector R) can be calculated from the observed spectral images (vector I) as follows: R = IC1 (2-9) 2.2.2.4. Photobleaching Compensation . Photobleaching of the sample, particularly in acquiring fluorescence images, can have a considerable effect on the intensity of images taken over time. Each time the sample is illuminated, the amount of fluorophores which can be excited is reduced resulting in a less intense emission. Hence, the intensity of the image may vary depending on the time required to focus the TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 33 image before acquisition takes place and/or the time the image is acquired in the 3D (multiple focus plane) image acquisition cycle. Anti-photobleaching agents are generally being used in sample preparation to reduce the effects of photobleaching. However, there is usually some reminance photobleaching effect (of approximately a few percent) over minutes of exposure. Rigaut et al. (1990, 1991) have proposed methods to apply reconstruction techniques to correct images in confocal microscopy for effects of optics and photobleaching. These methods typically use an exponential time decay function to simulate the photobleaching effects. 2.2.3. Focus and Three Dimensional Reconstruction Generally, it is assumed that the image taken contains objects which are all in-focus or near focus. This assumption is not necessary true. As we have shown earlier, images of sufficient detail and clarity are required in quantitative microscopy to obtain consistency in object classification and discrimination (Poon et al, 1987, 1989b, Spadinger et al., 1989, 1990, Poon and Palcic, 1991). To see the details, these images are often taken with higher magnification and numerical aperture objective lenses. These lenses have a lower depth of focus and as a result, the details of the entire object can not be captured in one focus plane. In addition, objects in a scene and even the details in the objects themselves are not all at the same focal plane. Therefore, we have shown that these objects must be individually focussed before they are analyzed (Poon et al., 1989b, 1991, 1992a,b,c). For correct and consistent object segmentation and feature calculation, an objective method in focussing all objects of interest TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 34 in the sample is necessary. In some situations, 3D (x,y and z-focus) image reconstruction is initially performed. Several methods have been developed to solve the three dimensional problem of reconstructing the in-focus image from the observed images which contains both in-focus and out-of-focus information. Castleman (1979) have proposed methods of using i) inverse filtering, ii) simultaneous equations, or iii) nearest neighbour approximation. Agard's group (1984, 1989, 1990) later modified some of Castleman's work and introduced the solution of iterative constrained deconvolution with a non-negativity constraint. Carrington's group (1987, 1989, 1995) analyzed the inverse and iterative constrained deconvolution techniques and proposed the constrained least squares technique as a solution. Holmes' group (1989a, 1989b, 1991) maximized the log-likelihood function of the system with respect to the object's optical density for object reconstruction. Due to limitations in the detector and optics, higher resolution images are desired for detailed image analysis. Bertero et al. (1987, 1989, 1990) and Sheppard (1988) proposed methods for obtaining "super" resolution from images taken from a confocal microscope. In this system, a two dimensional CCD detector is used instead of a single element photomultiplier to detect the image created by a single point of illumination. As the point illumination is scanned over the image, a 2-dimensional image response rather than a single value is obtained for each point illumination. This extra information at each pixel is used to help reconstruct an image with higher resolution. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 35 2.2.4. Image Segmentation 2.2.4.1. Overview The most difficult step in automating the analysis process is to define the regions which belong to the object. Unlike feature extraction where only mathematical computation over the defined region will suffice, segmentation also requires prior knowledge of the geometrical, morphological, and topological properties of the objects in the scene as well as a heuristic approach for analyzing the problem. Since object classification is based on feature values which are derived from the segmented regions, segmentation is crucial for the correct interpretation of the objects in the scene. Many segmentation techniques have been developed over the past several decades (e.g. Davis, 1975; Fu and Mui, 1981, Mantaz, 1987, MacAulay et al., 1988). These methods can be categorized into three different classes: i) characteristic feature thresholding or clustering, ii) edge detection, and iii) region extraction. A single algorithm generally cannot segment a particular scene and hence a combination of segmentation processes is often used. Generally these processes perform well in some applications but may fail in others. Detailed descriptions of the different classes of known segmentation algorithms are discussed in the following sections. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 36 2.2.4.2. Thresholding or Clustering Thresholding is a common technique used in segmenting regions in a scene. The process assigns distinct labels to areas based on some properties of the image. A property may be a characteristic feature such as the image grey levels or may be of local nature such as the gradient or Laplacian of the grey levels. In all cases, a specified range of values of a given property is used to define the pixels in the image which belong to the same region. Often, a histogram of an image property is used to determine the thresholds for each region. These histograms are generally smoothed to remove noise. Care must be taken, however, to avoid smoothing out small but valid minima or maxima. A thresholding technique which can be applied to grey level histograms is the mode method. This type of histogram gives an indication of the number of pixels which have the same grey level in the image. Each peak (mode) of the histogram represents areas where large number of pixels have a similar intensity level. A boundary is then placed at the valley between peaks to separate the regions. The rationale for choosing such points is to minimize the probability of misclassifying each region. Since the number of pixels at the valley compared to the peaks is relatively small, misplacement of the threshold from the exact location has relatively little noticeable effect on the resulting image. For example, Poon et al. (1992c, 1993a) used this technique to segment the blood cells from the background of the image. A different technique is used for thresholding gradient histograms. Since these histograms represent the sum of the magnitude of gradients at a TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 37 given grey level, the boundary is placed at the highest point in the histogram. This point signifies the location of the largest intensity differences (the edges) of the image. This method works well with some images but fails in others. For images where there are many similar intensity pixels with a small gradient, their sum may overmask the sum generated at the edge of rare objects and hence a wrong threshold level is generated. Clustering extends the technique of thresholding to the multi dimensional space. This technique is used when a single feature yields poor discrimination regions while distinct regions can be seen in histograms of two or more characteristic features (e.g. cluster plots). Any feature which is useful for segmenting a region, such as the grey levels of images seen through different spectral filters, gradients, texture features, etc., can be used. Poon et al. (1992c, 1993a) used the green and blue image components of the image to separate the nucleated cells from the red blood cells. Algorithms for cluster analysis have been available for locating the decision boundary between regions in a multi-dimensional space (Amadasun and King, 1988; Umesh, 1988). To reduce the amount of computations required in the analysis, the smallest number of features which can discriminate the regions is employed. Thresholding and clustering techniques are global operators which use some aggregate properties of different features. These features are very dependent on the type of regions which are segmented in the image. Although the segmented regions are closed, some images may require smoothing to eliminate the noisy boundaries. Since no spatial information is used in the selection of the threshold, the resulting regions may not be contiguous. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 38 2.2.4.3. Edge Detection Edge detection algorithms use the information of edge points to determine the boundary between objects. The edge points are located where there is an abrupt change in grey levels in the image. In this technique, the elements which are candidates to belong to an edge are first extracted and then combined to form the boundary. The extraction of edge pixels requires a measure which corresponds to the change in grey value of the pixel with its surrounding. Various methods, such as the gradient, Sobel, Kirsch, and Prewitt operators (Rosenfeld and Kak, 1982; Young and Fu, 1986), have been developed for this purpose. These operators can be implemented as a series of image convolutions where the weights in the convolution kernel are different for each filter. The resulting value of the convolution at a pixel gives an indication of the strength of the changes around the pixel. The edge points are then extracted by thresholding the processed image. Marr (1982) has developed a Laplacian of a Gaussian edge filter. In this method, the zero-crossings of the filter correspond to the edges of the structures which have a space constant greater than (or a lower spatial frequency than) a selected value used in the Gaussian blurring process. Canny (1983, 1986) developed a good edge detector which convolves the noisy image with a spatial function (representative of the result) and then finds the maxima values in the resulting convolution. There are several problems with edge detection techniques. First, the transition from one region of the image to the other sometimes occurs over TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 39 several pixels and is then not abrupt enough. Second, the contours produced from thresholding edge information are generally more than one pixel wide and not necessarily closed. Hence, some post-processing using thinning and contour-closing algorithms are required. Another problem is that the texture of some regions is significant enough to be thresholded and interpreted as edge points, resulting in erroneous image segmentation. Nevertheless, the results from the edge detection techniques can be used in conjunction with other methods in determining particular regions. 2.2.4.4. Region Extraction Another segmentation approach is to group pixels with similar properties, such as grey levels, texture, color information, etc., into regions. These region extraction techniques can be separated into three categories: region merging, region splitting, and a combination of region merging and splitting (Ohlander et al. 1978; Garbay et al., 1986). In region merging or growing techniques, the image is initially divided into many small regions such as a pixel or a small neighbourhood of pixels. Various properties that reflect the characteristics of the object are computed for each region. The characteristics of each region are compared with its neighbouring regions. If the properties of the adjacent regions are similar, these regions are combined or merged into one. This process is iterated by recomputing the object membership properties for each enlarged region and merging the regions which have similar characteristics. The segmentation is TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 40 completed when all adjacent regions have significantly different properties such that no merge can further be made. The region splitting or dividing techniques begin with the entire image instead of many small regions. A predicate describing the various properties of the object is evaluated from the entire region. An example is to determine if all pixels in the region have grey levels which do not differ by a certain amount. If the predicate is not satisfied, the region is divided into smaller regions and the predicate for each of the sub-regions is recomputed. The process continues until the predicates for all regions are satisfied. The split and merge technique uses a combination of region merging and splitting to obtain regions of similar properties. Regions are merged when adjacent regions have similar properties and are split when the predicate describing the property is not satisfied. Liedtke et al. (1987) used this technique on microscope images of blood cells to extract the primitives used in his segmentation method. Region extraction techniques utilize the local properties of the image directly. Although they produce closed and contiguous regions, the drawback is that these algorithms are computationally intensive. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 41 Chapter 3. Imaging System 3.1. Overview This chapter describes in detail the imaging system that we developed for this project. As mentioned earlier in Chapter 2, there is no commercially available system that is capable of performing the required task, nor is there any system where modifications can easily be made. Hence, we built the imaging system for this project by selecting and integrating basic commercial components and developing the algorithms and software (as described later in this chapter). This imaging system performs two basic functions. First, it acquires images and stores them into files. Second, it analyzes the acquired images and generates telomere length information for each detected chromosome. The first function requires developing image acquisition hardware and software while the second function primarily involves developing the analysis software. Rather than combining both functions into one program and one system (as is done in most commercial systems), we developed separate software programs for the acquisition and for the analysis. The acquisition software can only operate with the hardware of the acquisition system. The analysis software, on the other hand, do not depend on the dedicated acquisition hardware and a darkened room to operate. Since, the analysis systems are less costly to build and operate than the acquisition system, the separation into two software programs results in a more TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 42 economical and efficient use of our hardware resources. Given the high demand for the use of the systems for biological studies, we built one acquisition system and multiple analysis systems and operated them independently. After the images are acquired, the images are transferred from the acquisition system to the analysis system via a computer network. The telomere/chromosome image analysis can then be performed in the analysis system. This chapter first describes the components of our acquisition system. It then outlines the algorithms we developed for acquiring the multi-focus plane images. Lastly, this chapter describes our analysis system and the algorithms we developed for pre-processing the acquired images. 3.2. Image Acquisition Hardware 3.2.1. Overview A block diagram of our image acquisition system is shown in Figure 3.1. The major components of this system are i) the microscope, ii) the focussing mechanism, iii) the camera, and iv) the computing system. The motorized focussing mechanism varies the focus of the objects placed on the microscope. This allows the acquisition of a series of 2-dimensional images of the object where the focus position of each image with respect to that of another in the series is computer controlled and known. A 3-dimensional image representation of the object is thus obtained. The camera transforms the image in the microscope into digital form and this image can be displayed in the TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 43 computing system. The computing system controls the entire process and stores the multi-focus plane images for later processing. Camera CCD Sensor] Digitizer luminatioiii j Microscope! Source \~j Optics ! j : ! X | Microscope! Microscope ! Stage j Image ! ! Display ! Memory \~ j 1 Monitor j Computer / Processor Computing System Focussing Mechanism Figure 3.1. Block diagram of the imaging system. The major components are the microscope, focussing mechanism, camera, and the computing system. 3.2.2. Fluorescence Microscope The fluorescence microscope is the key component of the imaging system. It transforms and magnifies the telomeres and chromosomes for visualization. We chose the widefield microscope over the confocal microscope for this application to avoid photobleaching (fading of the fluorescence probe TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 44 caused by the illumination source and the long acquisition time) as well as cost considerations. Fluorescence microscopes from brand-name manufacturers (e.g. Leitz, Nikon, Olympus, and Zeiss) all have comparable performance. We used the Zeiss Axioplan fluorescence microscope for this project because of its availability in the laboratory. The major components of the fluorescence microscope are i) the illumination source, ii) the excitation and emission filters, iii) the objectives, and iv) the microscope stage. An important restriction in the selection process of each of these components is that the component must be compatible for use with our chosen Zeiss microscope. Before discussing the selection of components, the basic operation of the fluorescence microscope is described. A fluorescence microscope generally has a number of slots for filters and dichroic mirror block assemblies. First, a selected wavelength of light from the illumination source is passed through an excitation filter. The selected light is reflected off the dichroic mirror to the sample via the objective lens. The sites within the sample which are excited by the selected wavelength of light will emit light at a higher wavelength (lower energy). The emitted light signals are focussed and magnified by the objective lens of the microscope. The emission filter then allows only a selected wavelength of the emitted light from the sample to pass through and blocks other wavelengths of the light including the reflected light at the excitation wavelength. The resulting image of the selected sites are then visualized through the dichroic mirror and onto the oculars of the microscope or onto a camera for display on a monitor. Thus, the choice of excitation, dichroic, and TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 45 emission filters are important for determining which fluorescence labeled probe is desired to be seen in the image of the object. The objective lens and illumination source are also important in quantitative imaging because they govern the image magnification, distortions, stability, and intensity. 3.2.2.1. Illumination Source The choice of the illumination source is dependent on the fluorescence spectral characteristics of the sample. An ideal illumination source is one which i) gives even (uniformly distributed) illumination, ii) has sufficient intensity in the desired excitation wavelength, and iii) does not fluctuate over time. By careful adjustment and alignment of the illumination source for Koehler illumination, a fairly even illumination around the center field of view (5% variation) can be obtained. Typically, fluorescence microscopy ft illumination sources are based on either mercury or xenon. Mercury (200W, Zeiss) tends to have a number of intensity peaks including the wavelengths at 405nm and 546nm. These two intensity peaks can be used to excite the DAPI and CY3 probes used in this project, respectively. Xenon (150W, Zeiss) tends to be less intense at these wavelengths but has better temporal stability than that of mercury. In addition, the lifetime of xenon bulbs is longer than that of mercury (500 hours compared to 200 hours). However, the xenon lamp is not suitable for this project because it does not give sufficient intensity for CY3 excitation. Fortunately, a hybrid mercury/xenon lamp (200W, OptiQuip distributed by Zeiss) is commercially available. This hybrid lamp fluctuates less than the mercury lamp in time and is more intense than the xenon lamp at TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 46 the wavelengths of interest. Its spectral characteristics is very similar to that of the mercury lamp but has a much longer bulb lifetime (2000 hours). Hence, we chose this lamp for this project. The results and analysis of the temporal stability of the hybrid lamp used are discussed later in this Chapter. 3.2.2.2. Excitation and Emission Filters As mentioned earlier, the choice of the excitation and emission filters and the dichroic mirror plays an important role in what is seen in the resulting image. Filters and dichroic block assemblies are then selected to match the properties of the fluorescence probes used in the experiment. For example, one probe highlights the entire object while other probes, with different fluorescence spectral characteristics, highlight specific sites within the object. As the block assemblies are interchanged throughout the experiment, the alignment of these blocks with each other and with its previous position in the imaging path may vary. Consequently, the problem of image registration of multiple probe images results (up to 10 pixels shift). Fortunately, new types of filters have been developed in the last five years to minimize the image registration problem. We used one type of these filters in this project. As in conventional systems, the excitation filter in this filter system selects and allows a specific wavelength of light to pass through to the object. Instead of allowing only one wavelength to pass,.the emission filter we used allows multiple bands of wavelengths of light to pass. Hence, with multiple excitation filters and only one emission filter, objects labeled with multiple fluorescence probes can be independently imaged. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 47 The spectral characteristics of the filters used in this project are shown in Figure 3.2. In our filter system, the mechanical selection of excitation is performed only in the illumination path using a filter wheel (Pacific Scientific Inc.) which has openings for 8 different excitation filters. A single multi-spectrum dichroic mirror and emission filter assembly is then used to image all probes in the experiment. The excitation and emission filters are selected in conjunction with the appropriately matched probes to minimize spectral crosstalk in the observed image. Since no optical components are moved in the imaging path, the shift between multi-spectrum images is small and is less than 2 pixels. There are 2 drawbacks in using this type of filter system which can be compensated for with adequate intensity illumination and appropriate selection of filters and probes. First, the amount of light that is allowed to pass through, at a selected wavelength, is diminished by approximately 10-50% compared to single bandpass emission filters. Second, more noise is present as undesired light from other wavelengths, although minimal in most cases, is allowed to pass through. 3.2.2.3. Objective Lens The objective lens is perhaps the most important component in the microscope. It plays an important role in determining i) the spatial resolution, ii) the chromatic response, iii) the chromatic and spherical aberrations, and iv) the intensity of light in the system. Spatial resolution in the objective lens is dependent on the wavelength as well as the numerical aperture of the objective lens (Section 2.1.2). Spherical aberration relates to consistency in the size of TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 48 Camera Figure 3.2. Block diagram of the excitation and emission filter system. An 8-position filter wheel is used to select the excitation wavelength in the illumination path. A double band pass dichroic and emission filter is used in the imaging path. the object at different points in the field of view. Chromatic aberration relates to the size of an object when seen under different wavelengths. The object size varies because the focal point changes with different wavelengths. For high quality objective lenses, such as the ones we used, spherical and chromatic aberrations are relatively small in the central field of view where the objects lie TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 49 and thus can be ignored. Lastly, the amount of light loss through the objective lens is of particular importance in fluorescence as the signal intensities are low. The process of chromosome imaging typically requires high magnifications and high numerical aperture objectives. We used the fluorescence 63x magnification objective lens with a numerical aperture of 1.4 (Plan Apochromat 63x/1.4, Zeiss) for this project. This objective lens was found to exhibit the best all around performance over other (Zeiss) 63x and lOOx magnification objectives for imaging chromosomes and telomeres. 3.2.3. Focussing Mechanism We incorporated an automated focussing mechanism into the system for two reasons. First, the telomeres and chromosomes do not all fall at the same focus plane. Second, the focus depth of the objective is smaller than the size of the objects being studied (i.e. the entire object is not captured by the lens). The motor (ZSS 43-200-1.2, Phyptron, Germany) and controller (MAC4000, Marzhauser, Germany) used can move the focus position in step size increments of 0.1 um. However, the backlash or hysteresis effects inherent in the mechanics of the microscope (for moving the focus position in a direction that is opposite from its previous movement) can range up to 0.3 um. A simple technique is then employed to obtain the multiple focus position 2-dimensional images such that the spacing in the z-direction is consistent. In our technique, images are acquired by first moving the objects so they are out-of-focus and then acquiring multiple images as the focus position is stepped in equal TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 50 intervals in the opposite direction passing through the focus point of the object. As a result, the acquired multi-focus plane images are evenly spaced from each other. The exceptions are in the first couple of images where the focus spacing can be different because of the backlash/hysteresis effects. Hence, these first couple of images in the series are not used in the analysis. 3.2.4. High Resolution Camera The quality of the image acquired is highly dependent on the camera used. Hence, we based our selection of the camera on a number of key considerations which are desirable for quantitative fluorescence microscopy. These considerations resulted in selecting a camera which have the following requirements: i) high spatial resolution and large field of view, ii) sufficient photometric resolution, high sensitivity, and large dynamic range, and iii) multi-spectral image acquisition capability and relatively fast readout rates (Jaggi et al., 1993; Pontifex et al., 1994, Poon and Hunter, 1994, Vrolijk et al. 1994). We chose the Microlmager MI1400-12 digital camera (Xillix Technologies Corp.) for this project as it meets these requirements. Other cameras which employ the same CCD sensor and meet these requirements are available from other manufacturers (e.g. Photometries Ltd. or Princeton Instruments Inc.) but are more expensive. For the first requirement stated above, it is desirable to have a detector which can sample at a rate which is at least twice the resolution of the rest of the system such that aliasing effects do not occur. This translates to a sensor with a pixel spacing of less than 7 microns. Given such a small pixel size, the TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 51 detector should have approximately 1000x1000 pixels so that it can capture the entire cell (metaphase chromosomes) with sufficient detail and resolution in one image. Square and 100% fill factor pixels detectors are used in the Xillix camera. This eliminates the need for geometrical pixel compensations and increases the pixel sensitivity as the entire pixel region (rather than a portion) is sensitive to light. The second requirement dictates the use of high dynamic range cameras with sufficient photometric resolutions and sensitivity. In fluorescence microscopy, signals may range over several orders of magnitudes in intensities (0.0001 - 1 lux). Typically, variable exposure time (less than 10s to avoid significant photobleaching effects) cameras combined with a dynamic range of 10 or more true bits (>1024 grey levels or >60dB signal to dark-noise ratio) are used in fluorescence microscopy imaging. Generally, a photometric resolution of only 8 true bits (256 grey levels) of information will suffice especially when fluorescence images are typically fairly noisy. Hence, the eight most significant bits of high resolution cameras (e.g. Xillix 12 bits or 4096 levels camera) are typically used to represent the image stored in the computer. Alternatively, for less intense images, a sub-region of the sensor's dynamic range may be used to obtain the same photometric resolution data (256 levels) without the need to increase the exposure and acquisition times. In this instance, a selected portion of the 4096 grey levels of the camera is linearly mapped to 256 grey levels. The remainder of the 4096 grey levels are then set to 0 or 255 depending on whether the respective grey level is below or above the selected portion. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 52 Lastly, the camera must be able to capture images over a broad spectrum of visible light and acquired images at sufficient rates. Although the filters in the fluorescence microscope are used to select the wavelength of light for imaging, the sensor must also be sensitive to light in this range. Most CCD detectors are made from silicon and these exhibit a higher quantum efficiency in the red region than that in the blue. The favoured red region corresponds to the emission wavelength of the CY3 probe used for labeling telomeres in this project. Readout rates is an issue in megapixel cameras as each new object needs to be re-focused. Too fast a readout rate (30MHz) can pose a strain on the sensor and degrade the signal-to-noise performance. As a compromise, we use the binning and the 8MHz readout rate features of the Xillix camera for focussing purposes. In the binning mode, several pixels are combined together before they are readout. This increases the sensitivity and reduces the total number of pixels in the image. As a result, shorter readout and exposure times are required to obtain a similar intensity image in the normal (not binned) mode. 3.2.5. Computing System Our computing system is similar to the one which was developed for general imaging by Xillix Technology Corp. (Jaggi et al., 1991). This system was chosen because of its availability, our familiarity with the system, and the availability of the source code to facilitate modifications as required. The computing system consists of i) a host 486-based personal computer, ii) 1 gigabyte mass storage disk space, iii) high resolution image acquisition and TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 53 display card (1280x1024x24 bits), and iv) a corresponding high resolution monitor. 3.3. System Temporal Stability and Aberrations 3.3.1. Overview It is important to ensure that the images can be and are calibrated such that the results from within an experiment and from different experiments can be compared in a meaningful way. In order to determine what corrections and compensations are required on the acquired data and generated results, we performed a number of experiments on the system. These experiments and their results are described in the following sections. The analysis of the data from these experiments is then used to justify what pre-processing algorithms are required for data correction and compensation. 3.3.2. Temporal Fluctuations in Illumination The aim here is to ensure that accurate measurements of telomeres are made irrespective of when these images are acquired. The illumination source is the major cause for temporal variations in the acquired image. The light intensity emitted by the lamp will vary due to the variations in the power supplied. In addition, the light intensity will change over time as the light bulb ages. Although we have selected a hybrid mercury-xenon lamp which exhibits good temporal stability properties, we would still need to characterize its TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 54 stability such that appropriate measures for image normalization and corrections can be developed and implemented. For characterizing the fluctuations in the illumination source, we chose a coloured plastic (acrylic) slide as the test object. This plastic exhibits similar fluorescence characteristics as telomeres by giving off a faint red fluorescence when it is illuminated with green light. Similar filter settings were used to acquire images of the plastic and those of telomeres. Images of the plastic sample were then acquired at different time intervals to determine the stability of the illumination source over time. The images were acquired in binning mode to reduce the number of pixels to process and store, that is, every 2x2 pixel in the image was combined into one pixel. Hence, the acquired image was reduced to 640x512 pixels in size. For each acquired image, we calculated the average of the measured fluorescence intensity in the central 512 x 512 pixel region. The averaging helps to smooth out the spatial non-uniformities and noise over the region. If we assume that the fluorescence emission of the sample is directly proportional to the intensity of excitation, the average value then gives an indication of the amount of light emitted by the illumination source. The distribution of the mean scene intensity as a function of time is shown in Figure 3.3. It can be seen that the lamp intensity can change drastically over time. This can be explained by changes that have been made to the illumination source during the course of its use. The changes may include centering/focussing the light bulb, replacing filters in the light path, etc. It is also noticed that the lamp does remain constant over several hours of TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 55 use at a time. Thus, the illumination can be assumed to be constant for each acquisition session which takes 2 to 3 hours to complete as long as no changes are made to the system during the experiment. Figure 3.3. Illumination variations over time. The illumination remains constant during the duration of the (2-3 hour) experiment. The illumination level can vary in between experiments as the optics may be moved or changed during such time (represented by the discontinuous horizontal lines). Based on the above experiment, a calibration method for variation in illumination was developed. First, an average of 10 fluorescence intensity measurements of the acrylic test sample was made at the beginning of each "telomere" experiment. The averaging helps to compensate for the temporal noise in image acquisition. This average intensity value signifies the amount of luminance of the light source for each experiment. Hence, the data from each experiment can then be scaled appropriately by comparing the average acrylic fluorescence values and appropriately scaling the telomere results amongst experiments. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 56 3.3.3. Photobleaching Effects Photobleaching (fading) is another important aspect to consider in quantitative fluorescence microscopy. Anti-bleaching agents are often used to minimize the fading effects of the probes used. If a sample was significantly photobleached (i.e. the sample fluorescence emitted was reduced as a function of light exposure), a method will need to be developed to determine the amount or the stage of photobleaching in this sample so that the results generated could be correlated or compared with another sample. The distribution of the normalized maximum intensity of telomeres over exposure time is shown in Figure 3.4. A similar distribution for 0.1 um fluorescence beads is shown in Figure 3.5. 0.00 200 4.00 6.00 8.00 10.00 Elapsed Time (minutes) Figure 3.4. Photobleaching effects on telomere fluorescence. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 57 200 §J 140 CU - 120 £, ioo en en c CD 80 60 40 20 4. 0 0.00 1.00 2.00 3.00 4.00 5.00 6.00 Elapsed Time (minutes) 7.00 8.00 9.00 Figure 3.5. Photobleaching effects on bead fluorescence. It can be seen from Figure 3.4 that the fluorescence decay is approximately 1% over minutes of light exposure. The fluorescence decay of beads is higher at approximately 3% per minute. In our experiments, we decided not to compensate for photobleaching effects. The reason is that the variation due to photobleaching (approximated by the line in Figure 3.4) is much less than the variation in acquiring an image (represented by the difference between consecutive sample points in Figure 3.4). The latter variation can be up to 10% from the expected (line) value. The estimated temporal variance in the intensities of the acquired images due to photobleaching is 1% as it takes 2-3 minutes (with excitation light exposure) to TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 58 setup and focus the telomere image and approximately another 1-2 minute to capture the series of multiple plane images. 3.3.4. Uneven Illuminated Field of View We next investigate the unevenness of sample illumination. It can be seen from Figure 3.6 that there is a variation of illumination over the field of view. The bright spots in the image corresponds to the image of the arc of the bulb. The, ring-like contours of equal intensities show that the intensity decrease away from the bright spot. The intensity difference between the brightest and dimmest spot in the field of view is approximately 10%. To correct for this spatial variation, we chose the flat-field compensation method (mentioned in Section 2.2.2.2). As shown in Chapter 2, the derivation for the flat-field compensation involves taking the logarithm of the transmittance to obtain the optical density values. However, the integrated fluorescence intensity (IFI) value in fluorescence microscopy is proportional to the fluorescence intensity and not the logarithm of the intensity. Even though no logarithm conversion is required, this compensation is still valid. Our reason for this can be explained by the unevenness in illumination in the field of view as shown below. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 59 Figure 3.6. Illumination variation over the field of view. The image shows a contour map of the variations in light intensities. There is a two grey level difference between two adjacent regions. The brightest spot in the image is represented by the grey intensity region near the center. The light intensities decrease towards the edges of the image. If the illumination at the center of the field of view generates an object fluorescence of intensity Ic - D (where D is the fixed offset value which is independent of the level of illumination), then a different intensity in illumination scaled by a factor of s, will generate a scaled object fluorescence of intensity slc -D= IE -D. If the scaled illumination is not in the center of the field of view, the fluorescence intensity at the off-centered position is still a scaled version of the center intensity. Hence, each measured fluorescence value I(x,y) TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 60 in the image can be scaled by the appropriate scaling factor s(x,y) (corresponding to illumination differences in the field of view) at that pixel to generate a compensated fluorescence value C(x,y). This is expressed as: C(x,„)-D = M^ ,3-!) s[x,y) The scale factor s(x,y) at each point can be obtained from an homogeneous fluorescence material placed in the microscope. If the fluorescence response of the background material at each point is B(x,y) - D and the fluorescence at the center of the field of view is Bc - D (constant k), the scale factor at each point is given by: »,«,„, = ^E£=«(£fz£ ,3.2| By inserting this scale factor into equation (3-1), the following result which is similar to the known flat-field compensation formula (Section 2.2.2.2) is obtained: Clx,y) = k I{X,y) ° +D (3-3) 1 ,y> B(x,y)-D The above flat-field compensation method involves division operations which will generate real (rather than integer) numbers. Since images are stored in integer numbers (8 bits), truncation or quantization errors will arise. Hence, this flat-field compensation image is not carried at this stage. Instead, we TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 61 performed flat-field compensation later and only on those pixels which are used in calculating the IFI value of the telomeres. The results and discussions of this compensation method is presented below. 3.3.5. Flat-Field Compensation Results The fluorescence acrylic sample that is used earlier for the temporal fluctuations in illumination experiment is used for flat-field compensation measurements. In this experiment, images are acquired at different illumination levels (inserting neutral density filters in the illumination path). The brightest image is used as the background reference image and the other images are processed using the flat-field compensation method of Chapter 3.3.4. A cross-section of the intensity distribution in the central row of pixels of the field of view, before and after the flat-field compensation algorithm is shown in Figure 3.7 and 3.8 respectively. Using flat-field compensation, it can be seen (Figure 3.8) that a linear response can be obtained over the field of view (standard deviation of less than 1 grey level) for different levels of intensity of illumination. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 62 250 _ 200 m I 150 cu S £ 100 cu ^ 50 200 217.9 +.6.1 150.6 +4.3 106.4 +3.4 43.2 .+ 1.9 400 600 800 Pixel Position 1000 1200 1400 Figure 3.7. Intensities of the central row of pixels of a homogenous sample. 250 ~ 200 v> > cu - 150 cu 5 100 '35 e co £ 50 228.0 +.0.0 157.6 JiO.7 111.3 Hi 0.7 45.2 +0.8 +-0 200 400 600 800 Pixel Position 1000 1200 1400 Figure 3.8. Flat field compensation for spatial illumination variations. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 3.4. Image Acquisition Software 63 3.4.1. Overview The "SSM" program which we used in our acquisition system was developed by Xillix Technologies Corp. for acquiring, storing, and manipulating the images acquired by the Microlmager camera. This program runs under the Microsoft DOS operating system. It has the capabilities to control all camera functions but does not have all the functions which we need for this project. Hence, we modified the program to allow for i) automatic camera exposure time and photometric range selection, ii) integration of multiple focus plane acquisition, and iii) corrections for pixel defects. 3.4.2. Image Exposure and Photometric Range Selection We implemented automatic exposure time and photometric range selection functions into the program to control the camera. These functions maximize the dynamic range of the image such that the grey level in the acquired image span over most of the 256 levels. The automation of these functions also simplifies the setup process for image acquisition and reduces the time required for adjustment. The linearity of the camera as a function of exposure time and photometric response makes the automation algorithm simple and easy to implement. In the program, the photometric range is implemented by a lookup table which maps the camera's 0-4095 grey levels scale to the computer memory's 0-255 grey levels. In this algorithm, all grey levels below the chosen minimum value in the camera is set to 0 in the TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 64 computer memory. Similarly, all grey levels above the chosen maximum value in the camera are set to 255 in the computer. The range in-between the minimum and maximum values is then linearly mapped to the closest integer in the range of 0-255. In addition, the range between the chosen minimum and maximum levels is set to a multiple of 256. This avoids a bias in mapping more levels in the camera to a given level (between and not including 0 to 255) in the computer. The resulting grey level histogram of the image would then be smooth and would not have peaks at the bias locations. The image exposure time and the chosen values for the minimum and maximum levels are based on information in the image. Initially, an image is acquired with an exposure time of Is and with a linear mapping of the 0-4096 range to the 0-255 range. The intensity distribution of the image is evaluated and the exposure time is adjusted (down to a minimum of the 30ms limit and up to a maximum of the 10s limit of the camera). The aim is to obtain an image which i) has the brightest pixel with an intensity value of greater than 230 grey levels, ii) has less than 10 pixels set to a grey level of 255 (i.e. to avoid image saturation), and iii) has the dark region set to within the grey levels of 0-50. If the selected exposure time does not give the desired image intensity distribution, the minimum and maximum values for the mapping function are adjusted accordingly. 3.4.3. Image Pre-Processing: Sensor Defects Megapixel CCD cameras often have a number of inherently defective pixels. These pixel defects are largely caused by the local accumulation of dark TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 65 current. As the size of the accumulated dark current is related to the temperature of the sensor, some cameras employ cooling to reduce the size and number of such defects. These defects typically appear as bright spots or lines in the image. They are not a problem at low integration times. However, at long integration times (more than 3s and longer for cameras where the CCD is cooled), the defective pixels may become saturated even when no light is present. As the information in these pixels are lost, the only method is to find an approximate value for the pixel. We used the average of the surrounding non-defective pixel intensities as the replacement intensity value. As these defects only pose a problem at long integration times, the defective pixels in the image are replaced only if an integration time of greater than 2 s is used to acquire the image. 3.5. Image Analysis System 3.5.1. Image Analysis Hardware The image analysis hardware consists of just a computing system. The requirements for this computing system are less stringent than those of the acquisition system. We chose both Pentium and 486-based personal computers for this purpose. The Pentium based machines are strictly used for data analysis. The slower 486-based systems can be used for image analysis when it is not being used for image acquisition. We also chose commercially available displays (e.g. NEC 3D) and display cards (e.g. ATI VGA Wonder) TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 66 capable of handling a minimum resolution of 640x400x8 bits. Typically, a resolution of 1024x768 is used. 3.5.2. Image Analysis Software The analysis software first reads the image files acquired from the acquisitions system into the computer memory. It then performs the segmentation and generates the results for interactive verification and editing. We developed this program to operate in the Microsoft Windows Version 3.1 operating system environment. The results of the telomere and chromosome analysis is shown in Figure 3.9. The key user-interaction features which were required as defined in consultation with the users are implemented into the analysis program. These factors are: i) displaying an enhanced view of the banding structure in the chromosome for ease in karyotyping (chromosome classification), ii) displaying the relative telomere positions on the chromosome image, iii) marking and labeling each telomere in a chromosome, iv) displaying the segmentation results, v) marking and labeling each chromosome, vi) displaying the results of the quantification (chromosome number, number assigned, telomere IFI for each chromosome, chromosome IFI and area feature values), vii) sorting the chromosome list based on the selected feature, and viii) editing capabilities. The resulting implementation is shown in Figure 3.9. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 67 pie fcdit applications Options! Erocess Window Steven's linage Analysis N:\STEV\IELOX.IMU \Aiti x 4JZ x U HFP| W f/li Labelled Object Histagrarr Options Fdit Objects 1 3 n Si:? - 6 =s 3 n IFI 06 «in Mas Court M-dr Mm in =Td Dev 9: 576 52 i:o 273 -:' 14 J7 73 538 46 :-6 257 55 1 33 27 7i 558 39 24 256 46 1 36 3; 84 452 3' :i6 247 51 103 83 1 : 114 3 18 86 21 47 52 2A i£4i 3 ! *"•• =| Cliruinusumt: Statibtiu&i h -Fit it Savr Sort Print L:.l C I* TI T; T3 T4 T- Tot.l =>a:inT»l. C D 3 jr.u Aiaa Currr-iwn ± |.| 8 II 11 Oil ' '639 11" 582 -- 135 :7 • 2:0 232 0 917 0 79 4 81:0 67 •4 34 2" 7 0 0 0 217 0 00 9419 99 ; • so : m 251 216 0 1232 1 43 4 72=7 ft ft •445 145 • 24 120 129 1.1 583 0 85 4 •3 3 3ft 7! •426 _ 2* : 365 115 99 1.1 r96 .' 71 4 31554 •I: 929 • 341 225 • :>: 0 983 1 64 4 5510 'i •035 • •1: 4:3 ;43 8 1242 0 61 4 10742 87 2027 0 0 0 0 0 00 0 5 l 00 22 : 2" 2 231 654 542 0 1645 0 38 4 44-36 '5 979 207 31.1. II 8 0 4 0 Lll.l 2 49'tl 9 i '058 •: 25: 55 1:7 0 0 577 246 3 2226 67 542 147 '6. I.I IJ 0 30: 0 00 2 •64r .'. 4:-• J •::; •392 150 155 Ij 1601 4 25 4 1 0421 95 "919 ft 35: 257 :-9 20 Ij 37; 2 79 4 2324 73 65 331 294 =9 132 Ii 95: 2 71 4 : 0 92 "335 • 33; 381 2;s 256 0 1252 1 32 4 10329 »• 2028 i ft 1 95 1 1 254 0 614 II 72 4 42:3 11 i o 147 21; 3:1 0 Ij ft 01 1.09 3 5421 80 0 32 15: 297 614 0 Ij lljft: 0 74 3 6129 80 " 129 407 341 VS ;sfi Ii ' 394 1"6 4 7,3 70 114 "558 _ 15: 1": '1: 0 3:0 0 458 0 II 15; 1114-0 00 ' 1 'ii : 25 12216 1 1 3 21.19 ' 14 243 261 = 6 19- Ij 701 1 82 4 7417 79 632 _ :-: 39: 30 1:8 20 3 35 l"92 1 40 5 3276 T2 7:7 : (. 02: 317 138 112 1 07 lift: 2 02 6 9127 104 " 825 " i 212 252 0 0 ci 1 39 3 2930 99 847 579 4i • • ' 10" I 132; .' Ill 4 6227 ••I'l 1,4113 20 55; 271 i;4 0 j i o i :• 4 51 3 9929 68 •910 3 0 234 "62 0 0 0 404 0 00 2 "829 95 632 : i 164 29: 3;4 268 0 1062 0 76 4 •:: i o 83 •41" - _ 224 0 0 c 2724 0 00 155 13" 226 1 77 • II .•111 . 85 II 86 0 71 4 [•,'•11 II 91) "44" 24; 22: 177 229 Ij ss: 1 •? 4 79:3 33 iW 15: " U: 2:4 .j: IJ 81 : 0 47 4 ::: 8" 956 46; 564 1:8 295 0 1511 2"2 4 6533 80 "256 - ~ 164 Ij 0 0 ft 4 0 00 11 98 78 ; Ci 264 ' 27 259 233 I'l 004 0 79 4 8124 98 "975 :9 1 7: 0 0 Ij 173 I] 00 25 35 :9 : 597 4I-.4 2*6 253 I53r 2 22 4 II 4 B" 676 J 12; '1; 0 0 I] 242 0 00 2 •634 97 533 y 17; 14 -::: 493 0 1312 0 29 4 31 3 3 77 933 : • 474 in' 24' Ii 145 = 1 82 4 13550 ?2 2515 u 227 •6: 1=4 IT; 0 =•53 1 08 4 7577 1 OU •832 '_ J" 3D: 294 528 :89 1 1433 8 72 4 'Ii 1 111 734 41 is: 95 322 ;60 0 967 0 40 4 1 2420 73 2657 " 47 141 15 2;: 3 35 27 907 0 39 5 I3027 32 2535 _ 48 67: 51: 149 189 0 152-2 3 02 4 G729 • • '517 Figure 3.9 Telomere ^Analysis Program. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 68 The resulting image displayed shows the borders of the telomeres overlaid onto the processed chromosome image. To accomplish this, the image of the chromosome is first thresholded such that the background pixels are set to a mid-intensity grey level of 128. This grey level allows the chromosome borders and both dark and bright intensity bands in the chromosomes to be seen on a grey background. The image is then inverted (linear map of 0-255 to 255-0). Contrast stretching is then performed on the chromosome to enhance the details of its banding structures. The image of the telomere objects are next processed (as described later in Chapter 5). To ensure that the telomeres lie at the ends of the chromosomes in the superimposed telomere-chromosome image, a pattern matching algorithm is then used to determine the placement of the detected telomere borders onto the chromosome image. In this matching algorithm, the telomere image is first shifted, pixel by pixel, from the chromosome image. For each pixel shift location, the number of detected telomeres that are within the borders of the detected chromosomes is determined. The pixel shift value (x,y) which generates the highest number of telomere/chromosome object matches becomes the pixel shift value for aligning the two images. The border of the telomeres are then superimposed onto the enhanced chromosome image. A different color is assigned to each of the telomeres in the chromosome for ease of user recognition (purple, red, blue, cyan, orange, and green are assigned to telomere number 1, 2, 3, 4, 5, and 6 or more). Although there are only 4 telomeres in each chromosome, 2 additional colors are used to facilitate the TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 69 editing of the detected telomere results and to identify other non-telomeric probes (e.g. centromere of chromosome 17). Finally, the border of the chromosomes are overlaid on the enhanced chromosome image with the highlighted telomere borders. Again, a different color is used to indicate the number of "telomere" objects in the chromosome (cyan, green, and yellow are assigned to 3 or less, 4 and 5 or more "telomere" objects, respectively). Each chromosome objects are then labeled with a unique number. A corresponding list of chromosome and telomere IFI can be displayed in a window beside the image. An enlarged sub-region of the image can also be displayed to aid in visualizing the details of the image. The user can then edit the information displayed. The editing capabilities include i) the joining of chromosomes (which are improperly segmented), ii) splitting "telomere" objects (touching telomeres), iii) reassigning the telomere number in the chromosome (pair sister telomeres: 1 with 2 and 3 with 4), iv) ranking the chromosomes based on its size or IFI value, v) assigning a chromosome number to each chromosome (for karyotyping purposes), and vi) adding a comment to the data. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 70 Chapter 4. Acquisition System Characteristics 4.1. Background An understanding of the characteristics of the acquisition system would give insight into how images are formed and how telomere fluorescence measurements could be evaluated. This system characteristics can be represented mathematically. As mentioned in Chapter 2 (equation 2-3), the input object is modified by the transfer function of the system to result in an output image. In our system, the image is affected by the components in both the microscope and the camera. If we assume linearity and space invariance in the system, the overall optical transfer function, OTFsys, will consist of the continuous transfer function of the microscope, OTFopt, multiplied by the transfer function of the discretizing camera, OTFcam. OTFsys(u,y) = OTFopt(u,y) x OTFcam(u,v) (4-1) The transfer function of the microscope can be derived based on geometric and Fourier optics theory. If light is considered to be composed of many point rays, the incoherent illumination used in the fluorescence microscope would then be composed of point light rays where the phase of each point varies independently of one another. The incoherent PSF is then represented by the power spectrum of the pupil function of the system (Goodman 1988). Thus, if the pupil function p(x, y) of the system is TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 71 represented by the system's aperture, the pupil function at the image plane (at d distance from the aperture of the objective lens) is then represented by p(Xdx, Xdy), where X is the wavelength of light. The PSFopt of the microscope is then the Fourier transform of the pupil function and is given by: PSFopt(x,t/)= F\ p(Xdx, Xdy) } (4-2) The normalized incoherent OTFopt can then be determined from the normalized auto-correlation of the pupil function and is given by: OTFAu,v) optK""U) 2?p(0,0) J Jp{Xdx, Xdy)p(Xdx - u,Xdy - v)dxdy (4-3) —oo—oo 00 00 J" j p2 (Xdx,Xdy)dxdy Hopkins (1955) was the first to theoretically determine the OTF of an incoherent illuminated system for both the out-of-focus and in-focus cases. In his derivation, the out-of-focus pupil function for a circular aperture system of radius A (as shown in Figure 4.1) observed at plane d2 when the object is at distance da and the in-focus pupil function is at plane d{ (i.e. 1 = — d0 dt f where/is the focal length of the lens and rij and n2 are the refractive indices in the object and image sides of the lens) is given by: TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 72 p(r) = red 2A exp jkw (4-4) where rect(x) - 1 if x < 1 and = 0 otherwise tu = -d, - 8Z cosa + (d,2 + 2d,8z + 522 cosa2)^ = defocus error X a = arctan d,. dt = location of the focussed image of object at distance dQ = d2 =location of the focussed image of object at distance d1 di + 5z = d2 Notice that if 5Z = 0, then w = 0 and the pupil function p(r) becomes the system's aperture function. By substituting the pupil function into equation 4-3 and the spatial frequencies (u,v) with (s), the resulting OTFopt becomes: OTFopl(s) = —cos 1 as fMH + zl-ir1 sin(2np) n=l 2n [J2n-l(a)- ^2n+l(°0 sin %a — as v2 j sin[(2n + l)p] Z(-D n=0 where a = 2kws (3 = cos -/c = spatial cutoff frequency. o i Van(a) - J2n-2(a)] 2n + l L J 2Vu2 (4-5) s = fe TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 73 Refractive Index n n In-focus Object _JL +_2 =JL d +d f 1 2 Image / Detector Plane Figure 4.1. Relationship of the in-focus and defocus distances in the object and image side of the objective lens. This derivation takes into account both the x,y spatial (s) and defocus error (w) effects. The resulting OTF is not a simple function and takes a long time to compute. Stokseth (1969) later produced an approximation of this OTF which takes less time to compute. Stokseth's approximation for a defocus error of vu is given by: OTFop( (s, w) « (l - 0.69s + 0.076s2 + 0.043s3) jinc 4ktu r i s S 1 - J V 2 ) 2 where TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY J,(x) jinc(x) = 2 x for s < 2 74 (4-6) J1(x) = the first order Bessel function of x. Castleman (1979) found a more accurate version of Stokseth's approximation for small values of defocus error to. For Castleman's method, the polynomial in Stokseth's approximation is replaced by an in-focus OTF term (—(2(3 - sin2(3)) multiplied by a defocus effect (jincfc) term which is a K function of the spatial frequencies, u and v, and the defocus value, u>). Hence, at a focus plane (out-of-focus distance of amount 82) from the in-focus image at distance di in Figure 4.1) and with a lens system having an aperture of radius A, the OTF is given by: OTFop( (q,w) « -(2p - sin2p) jinc where q2 = u2 + v2 4kw fa fc 2A fc (4-7) The above OTFopt represents the transfer function for general defocus optic systems. For the special case of concern which is the microscope system, Erhardt [1985] and others used a similar version of the above formula. They expressed the theoretical OTF in terms of the parameters of the microscope system. For example, Erhardt's derivation of the system OTF is based on Stokseth's method and is given by: TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 75 2TI • h(p) -fc-dz where (4-8) f(p) = 1 - 1.38/? + 0.03p2 + 0.344p3 h(p) = NA-p{l-p) 2 U2 +V 2 P = L2 NA = numerical aperture of the objective. In Erhardt's derivation, the variables are based on the parameter dz which is on the object side of the lens. The defocus error, here, is assumed to be the out-of-focus distance dz. Because J1[dJ/dz or jinc(dj is symmetrical about dz=0, the OTF values at the positive defocus distance dz is the same as that at the negative defocus distance -dz (i.e. OTFopt [p,dz) = OTFop( (p,-dz )). This symmetry characteristic, however, is not true in reality. In what follows, we find the OTF which shows its asymmetric nature. 4.2. Our Derivation of the System OTF In our system, as with most microscopes including Erhardt's, the location of the sensor from the objective lens, distance d^, is fixed and the sample is moved in relationship to the objectives (Figure 4.1). When the system is in-focus, an object point source in the object plane d1 produces a point source in the imaging (sensor) plane d^. When the object point source is moved by a distance dz to the out-of-focus plane d^, a point source at distance d{ results and the response observed at the sensor plane d^ (given by equation 4-9) is a blurred image. By moving the object point source to many different TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 76 positions around the distance d1 and along the optic axis, and by determining the OTF values at the detector plane for each of these object positions, the 3-dimensional OTF of the system can be obtained. It can be deduced from dx+d + n0 1 (refer to Figure 4.1) that the variable dj, and thus 8Z and dt f cos a vary as the defocus amount dz varies. These variables are used to calculate the defocus error, iv (equation 4-4). We now consider Castleman's approximation for the OTF of the microscope for small defocus amounts (equation 4-7): OTFop( [u,v,tu) « -(2P - sin2p) jinc 8% X w-ylu2 + V2 fc 4u2 + v2 fc (4-9) where w, the defocus error, is defined in equation 4-4. For positive and negative defocus amounts, ±dz, the calculated values for the defocus error, w, are not the same. Hence, the PSF is asymmetric (i.e. has different values at equal positive and negative defocus distances dz from the focal point). Thus, we shall use equation (4-9) to represent the OTF of the microscope. Our next step is to define expressions for the d{, 5Z, and cos a terms and use these terms to evaluate the defocus error, w, and then OTFopt of equation 4-9. To find the relationship between the in-focus and defocus distances 8Z, we write the relationship between the object plane distance and that of the image plane (as indicated in Figure 4.1) as determined from Lens law. The relationship of the in-focus position is: TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 77 1 _ Mj n2 f d2 ^ 10^ and that of the out-of-focus is 1 f d„ d- dx + dz n2 d2 - (4-11) where the distance d^ is generally fixed in a microscope system. The two equations in (4-10 and 4-11) are then equated (dropping the focal length term). The terms in the equation are then rearranged such that the change of focus in the image side is expressed in terms of i) the change in the object side (dz), ii) the refractive indexes of the media on both side of the objectives (n1( n^), iii) the distance of the sensor from the objectives (d^), and iv) the magnification (M) of the objectives. This expression is given by: d0 • M -n. -d, 5Z = 2 L_E (4-12) Tl2 d0 + (M n. +n0)-d, M 12/2 where M = ^ dx There remain two other variables in the general equation (4-4), the aperture A and angle a, which need to be expressed in the more commonly known parameters of the microscope system. First, the aperture, A, of the system can be derived from geometry and equated to the numerical aperture, NA, of the objective lens. This is given by: TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 78 r\ • sin 9 = nx • = NA (4-13) ^d2 + A2 In the derivation of the OTF (equation 4-9), it was assumed that A « dv Hence, equation (4-13) can be approximated by: nl- — = NA (4-14) dx Rearranging equation (4-14) for A and expressing the distance d1 in terms of the distance of the imaging plane from the lens d^ results in the following expression for A: d9 • NA A = ^ (4-15) Second, the term cos a, can also be derived from geometry and is given by: d, cosa = , (4-16) Vd,a + A2 where di = d^ - 5Z. The values for 8Z and A can be evaluated from equations (4-12) and (4-15) respectively. The expressions defined for d;, 5Z, and COS a, can then be used in equation 4-9 to evaluate OTFopt Now that the OTF of the microscope is determined, the next step in determining the OTF of the system is to derive the transfer function of the TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 79 camera. CCD cameras are commonly used for quantitative microscopy imaging. The photons (light), which reaches the CCD and have sufficient energy, will generate electrons in the CCD. In the active sensing area of the CCD, the electron (or the corresponding hole) charges will be trapped in the potential wells (pixels). In the non-sensing area, the electrons are absorbed by the sensor and does not contribute to the charges in the potential well. Hence, the camera samples and integrates the light which only falls in the active sensing area of each pixel. If the area of the sensing element is rectangular, and if the response to light over the entire area of the pixel dose not vary, then the resulting point spread function, PSF, is the convolution of the sampling function, sampQ, and the rectQ function. This can be written as: PSFcam(x,y,z) _1_ Px rect Px samp (*}] 1 T- rect p. V y_ P« samp vAJ. PxPy rect vPxy rect ry J i j where (4-17) Px > Py = s^ze °f pixel in the x and y direction respectively Ax , Ay = sample spacing in the x and y direction respectively. For our camera, the pixel size and the spacing between any 2 pixels are essentially the same in both the x and y directions and may be replaced by p (i.e. p = px = py = Ax = Ay). If we assume that the x and y directions are independent of each other, then the OTF of the camera which is just the Fourier transform of the PSF and is given by: TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 80 (4-18) Previously, only the transfer function of the microscope and the sampling properties of the camera have been used for image reconstruction purposes (Erhardt et al. 1985) and the integration process at each sample point is ignored. If our addition of the pixel sensing area is taken into account in the derivation (as shown above), the resulting theoretical transfer function (given by equation 4-1) should better resemble the experimentally determined function (i.e. the desired OTF ). The images of our theoretical system OTFs are shown in Figure 4.2. The calculation for the OTFs are based on the derivations in Section 4.2. These OTFs are evaluated with an objective lens having a magnification of 63x and a numerical aperture of 1.4 and located at a distance of 150mm from the camera detector, an oil medium with a refractive index of 1.515 between the objective lens and the sample, an emission wavelength of 570nm, z-spacings of 0.1 um, and a detector pixel size of 6.8x6.8um2. A point source object is positioned at focus plane z=0. Starting at an out-of-focus position of -1.6um, a 32x32 pixel image of the OTF distribution is generated. The out-of-focus position is subsequently increased by 0.1 um and another image of the OTF is generated. The process of generating OTF images at 0.1 um z-spacing is then repeated until a total of 32 OTF images are generated ranging from -1.6um to 1.5um (Figure 4.3. Theoretical OTF and PSF Results TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 81 4.2). The corresponding calculated images of the system PSFs are shown in Figure 4.3. Each of the 32 images of the PSF is obtained from the inverse Fourier transform of the corresponding OTF image in Figure 4.2. The system PSF in the x (or y), z-plane is displayed in Figure 4.4. Lastly, a plot of the system (radial xy) PSF distribution at varying z-spacings (focus levels) is shown in Figure 4.5. Figure 4.2. Theoretical OTFs of the system in the xy-plane. Each OTF image is evaluated at 0.1 um in the z-direction from its neighbouring images. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 82 Figure 4.3. Theoretical PSFs of the system in the xy-plane. Each PSF image is evaluated at 0.1 um in the z-direction from its neighbouring images. Figure 4.4. Theoretical PSF of the system in the xz-plane. Image (b) is a logarithmic scaled version of image (a). The pixel spacing in the z-direction is 0. lum. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 0.00 5.00 10.00 15.00 Pixels (0.1um) J(a) c CO •o CO "3 E ho O Z 6 8 Pixels (O.lum) J(b) Figure 4.5. Theoretical PSF distribution of the system at various z-spacings. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 84 It can be seen from the images of the system OTFs and PSFs that their responses appear circular symmetric in the xy-plane; this is as expected since the active sensing element and pixel areas have the same x and y dimensions. The intensity of the 3D PSF is highest at the central point (in the xyz-space). The intensity decreases as the distance from the central point is increased in the x, y directions. From Figure 4.4, it can be seen that the width of the spot in the x (or y) direction is half that of the z direction. Hence, as estimated in Section 2.1.2, the x (or y) direction is twice the resolution of the z direction. Within approximately 0.6um (the location corresponding to the cutoff frequency) in the z-direction from the central point, the intensity also decreases with increasing distance from the central point. At above +0.6um in the z-direction, however, alternating dark and bright rings (intensities) begins to appear in the OTF (Figure 4.2) and PSF (Figure 4.3) images. This is caused by the oscillatory effects of the Bessel function in the equation of the theoretical OTF. It can also be seen from the images that the PSFs and OTFs at equal distances in the positive and negative z-directions (about the central point) are similar (particularly within ±0.6um range). At approximately ±0.6um, slight difference in the PSFs in the positive and negative z-directions begins to become apparent in the image. It appears that the negative z-direction appears to decrease in magnitude slightly faster than in the positive direction. The asymmetries of the PSF in the z-direction is the result of the non-linear mapping of the positive and negative z-movements as discussed in Section 4.2. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 85 It can be seen from Figure 4.5 that the asymmetries between positive and negative focus amounts are present even at small defocus amounts (< 0.6um). An interesting point to note from the theoretical PSF image (Figure 4.3) is that the sum of intensity values over the 32x32 pixel (x,y) area for each focus plane is the same. That is, the total intensity information can be obtained from either the "sharp" in-focus image or the "very blurred" out of focus image (e.g. at 1.5um) where the point source is hardly seen. The intensity sum variations are in the order of 2%. These variations result from truncation errors (256 grey level images) and because the 32x32 pixel area is not large enough to encompass the entire function. Although the intensity information is present in the out-of-focus image, it is very difficult to extract such information since the boundary of the point source cannot be easily detected. When there are other objects nearby, the segmentation problem at out-of-focus planes becomes more difficult. As seen later in Chapter 5, we have developed a segmentation and intensity extraction method to estimate the total intensity information from the spot. 4.4. Initial OTF/PSF Comparative Study An initial study was carried out to compare the theoretical PSFsys with the experimental PSFsys of the system. Two different experimental approaches in obtaining the PSFsys of the system were taken (Poon et al. 1993b). Our first approach utilized small fluorescent beads to simulate point sources of light. The resulting 3D images acquired will then represent the PSF of the system. In this method, 0.15um fluoresbrite beads (Polyscience Inc. Warrington, PA) were TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 86 used. These beads were excited with blue light and emitted light in the green (515nm) spectrum. These beads were smaller than the resolving resolution of the microscope, but yet large enough to enable detection by the sensors. Sets of 50 images taken at O.lum z-spacing were acquired. Because the fluoresbrite beads photobleaches, the images acquired were corrected (restored), to compensate for the decrease in signal intensity as a function of time. The correction was based on the assumption that the fractional reduction in intensity in sequential images is a constant (Lockett et al., 1994). After the correction, the highest intensity pixel in the stack of images was chosen as the center of the bead. Each xy-plane image in the z-stack was then radially transformed to produce the radial distribution of intensity levels about the center point. The radial distribution was later used for comparison with other theoretical and experimental PSF of the system. Our second approach for determining the PSF uses a step edge to derive the line spread function, LSF, of the system (Tatian, 1965; Castleman, 1979). The LSF can be obtained by differentiating the image of a step image response (since the derivative of a step function has an infinite value at the point of the edge and zero elsewhere). The LSF is then rotated to give the PSF (i.e. PSF(-N/x2 + y2 ) = LSF(x)). The test target slide for the step edge consists of chrome evaporated onto glass (Edmund Scientific Company) to produce an opaque area on the slide. Like the bead images, sets of 50 images were acquired at 0.1 um z-spacings. From each xy-plane image, the central point of each line perpendicular to the edge was determined. The central point is the location where the intensity is half way between the bright and dark regions of TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 87 the line profile. Each line in the xy-plane image was then aligned with others at their corresponding center points. The respective values along the aligned lines were summed and averaged. The resulting step function was then median filtered, by a filter of 5 pixels in width, to remove sharp peaks and valleys in the function. The step function was then differentiated using a difference kernel of [-1, 1], to generate the PSF of the system. The in-focus (z=0) system responses for various objective lenses (20x/0.70 NA, 40x/0.85 NA, 60x/1.40 NA, and 100x/1.40 NA, Nikon) are shown in Figure 4.6. For each objective lens used, a comparison is made amongst i) the experimental PSF from bead images, ii) the experimental PSF from step edge images, iii) the theoretical PSF based on Erhardt, and iv) the theoretical PSF we derived. Both the bead and step edge images are acquired through a 515nm band pass filter. It can be seen that the measured and theoretical system responses are very similar in shape. For all objective lenses, the PSF obtained from the step edge has the widest response, followed by the bead PSF, our theoretical PSF, and then Erhardt's PSF. The results for this sequence can be explained by the following. The step edge is not infinitesimally thin but has some thickness which is larger than the size of the beads. In turn, the beads although small are of finite size, and hence they are only approximate point sources. Our PSF is wider than Erhardt's because we took into consideration the integrating effects of the camera sensor in our derivation. Hence, our PSF is more representative of the behaviour of the system as it is closer to the experimental results. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 88 20x/0.70Objective Lens 0.0 0.5 1.0 Distance from center (microns) 40x/0.85 Objective Lens 0.0 0.5 1.0 Distance from center (microns) 0.0 0.5 1.0 Distance from center (microns) 0.0 0.5 1.0 Distance from center (microns) OurTheoretical PSF Erhardt's PSF PSF from Beads PSF from Step Edge Figure 4.6. In-focus system response for various objectives. It can also be seen that our theoretical PSF is very similar to the experimental bead results. This is especially true for the 40x and lOOx TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 89 magnification objective lenses. As the beads are smaller than the resolution of the microscope and hence can be considered as point sources, our theoretical PSF gives a very good approximation to the characteristics of the microscope system. Of the two experimental approached carried, the step edge images are easier to acquire. They can be performed in brightfield microscopy. Thus, they require less exposure time for image acquisition and hence less noise is generated in the images. Due to the thickness of the chrome in the step edge and the jagged looking edge at high magnifications, however, the system response generated is not as good as the response of the beads. Hence, the fluorescent beads at a similar wavelength to that of the CY3 telomere probe are used for further comparison (mainly the values of the PSF in the z-direction) on the system PSF. 4.5. Comparison With Experimental PSF In the following experiment, we again use beads as test objects for comparison with the theoretical PSF. This time, however, the beads are acquired under the similar conditions as those used for telomere acquisition. That is, the beads are acquired under the same magnification and spectral wavelength used for acquiring telomere images. We used the 63x magnification objective with a numerical aperture of 1.4 and the green excitation and the hybrid red (570nm) emission filters. To simulate point sources of light, beads of 0.1 and 0.2urn in diameter are used as test objects. The images acquired from these beads will then be an approximation to the PSF of the system. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 90 Smaller size beads are more representative of point sources but are harder to manufacture consistently and have a weak fluorescence signal which could not be reliably detected by our imaging system. Multi-focus plane images of a number of 0.1 um beads at z-spacings of O.lum are shown in Figure 4.7. It can be seen from the bead images that their responses resemble that of the theoretical results (Figure 4.3). They all have circular symmetric responses in the xy-plane. Unlike the theoretical results, the images of the beads occupy over approximately twice the distance in the z-plane. This difference can be more easily seen in the intensity profile of the bead over different focus levels (Figure 4.8). Figure 4.7. Typical images of four O.lum beads acquired at different focus spacings. Each image is spaced 0.1 um in the z-direction from its neighbouring images. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 91 100 cu en <d 80 w 60 c •a 40 cu N E 20 O Z 0 -0.5 0 0.5 Z-Focus Position (microns) Figure 4.8. Comparison of experimental and theoretical PSF distributions as a function of z-focus position. A typical 0.1 um bead is used to represent the experimental PSF. The intensity distribution of the bead is broader than that of the theoretical PSF. This is caused by the additional blurring effect introduced when the beads are imaged through materials with different refractive index. In preparing the sample, the beads are first placed onto the glass slide (which has a refractive index of 1.5. A mounting medium solution is then placed over the beads. This solution contains the anti-photobleaching agent, Vectorshield, and has a refractive index of 1.45. The thickness of this solution can vary from 50 to lOOum. A glass coverslip of refractive index of 1.5 is then placed over the solution. Immersion oil of refractive index of 1.5 is then placed in between TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 92 the coverslip and the objective lens. Because of the difference in the refractive index between the mounting medium and the rest of the material in the imaging path, light from a point source object is refracted when it hits the refractive surface interface. Thus, from the objective lens, it appears that the light rays that passes through originate from a number of different point sources at different focus planes. This blurring effect has been recognized by a number of investigators (Vander-Voort and Brakenhoff 1990; Sheppard and Cogswell, 1991, Carlsson, 1991, Visser et al., 1991, 1992; and Hell et al. 1993, 1995). We did not take into account the mounting medium in the derivation of the theoretical PSF because the thickness of the mounting medium is variable and is not fixed from sample to sample. Hence, the amount of refraction and the resulting blurring will be different from image to image. In addition, it is difficult to predict the reflection of the fluorescence signals from the glass surface on which they lie as the distance of the object from the surface may vary over 0.5um. This phenomena is seen in Figure 4.8 where another intensity peak is observed in the negative focus position. This secondary peak is derived from the additional light which is partially reflected from the bottom glass surface. Since the intensity and distribution of the reflection signals are governed by how far the object is and what reflective index is at the mounting medium and glass interface, a generalization for the theoretical PSF would be difficult to generate. In summary, if the mounting medium has the same refractive index as the other material in the imaging path, then our theoretical PSF gives a good TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 93 approximation to the response function of the system (Figure 4.6). This, however, is not the case with our telomere study. Nevertheless, we used our theoretical PSF to generate images of simulated objects such that we can evaluate the robustness of our IFI algorithm (Section 5.5) to objects of varying intensities and shapes. As shown later in Section 5.2, the IFI value of the object calculated from the image generated using our theoretical PSF will be similar to that calculated from the image obtained using the true PSF of the system. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 94 Chapter 5. Telomere Segmentation and Integrated Fluorescence Intensity Measurements 5.1. Overview This chapter describes the method we developed to calculate the integrated fluorescence intensity (IFI) of telomeres in fluorescence microscope images. The IFI value is a measure of the total amount of fluorescence emitted from the object and is correlated to the length of the telomere. For our study, we use a number of images acquired at different focus planes. The IFI values are first determined for each of the multiple-focus plane images. These IFI values are then combined in a number of different ways and the results are compared to determine the best combination scheme to evaluate the IFI value of the object. Traditionally, most of the research in the quantification of FISH images involve spot counting, relative distance measurements, and event enumerations. With the introduction of more accurate and efficient quantification probes, and better microscope systems and anti-photobleaching agents, new biological studies to determine the fluorescence intensity of the probes are being carried out. Typically, only one image for each probe/spectral wavelength at the best focus level is analyzed to extract the relevant information. This "best" focus level is often chosen manually. That is, the TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 95 sharpest looking image is visually selected as the focus is manually varied. A different method of locating the "best" focus location is to automate the "best" focus selection process. In this case, the "best" focus level is chosen by minimizing or maximizing a particular feature. However, as we have shown in an earlier study (Poon et al. 1992a,b), the optimal focus level for one feature may not be the optimum one for other features. Furthermore, the z-distance between optimal focus level of different features within one cell differs from cell to cell. In that study, we have also found that a tighter feature distribution can be obtained if more than one image around the "best" focus position is acquired and the feature values are accumulated over these planes. We have also shown in the study that a tighter distribution of feature values can be obtained if the feature calculations are performed on processed images where out-of-focus blur are first removed. As mentioned earlier, most quantitative FISH image research use only one image, the best focussed image, at each wavelength of light. In some cases, more than one image acquired at different focal planes are used. For multiple focus plane analysis, the tendency has been to first reconstruct the 3D image. In this instance, the blur of the multi-focus plane images is removed before extracting the features. Unlike our present research, in the traditional applications of FISH imaging, the values of the fluorescence intensities do not have to be accurately known. In those applications, the resolution in the quantification can be coarse but within limits which enable detection of the location of the fluorescence spots or objects. A high degree of accuracy, however, is required in our present work in determining the IFI since the TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 96 amount of fluorescence detected is used to estimate the length of telomeres and the distribution of their lengths in a chromosome. Recently, a study which required much higher degree of accuracy than those of traditional applications and which determine the IFI has been investigated (Lansdorp et al., 1996). This chapter first describes the theoretical analysis for calculating the IFI of an object. It then discusses our method for calculating and determining the IFI values of the objects. Our telomere segmentation algorithm and our method for determining the number of focal planes required for the IFI quantification are also described. Finally, a number of methods to validate our results are discussed. 5.2. Theory in IFI Quantification In fluorescence microscopy, the IFI of an object is the total fluorescence detected from the object, o(x,y,z). Hence, the IFI of a fluorescence object, is the total intensity of the object and can be represented by the integral of intensity at each point of the object (in the x,y,z plane) and is given by: (5-1) As mentioned in Section 2.2.3, the original object can be reconstructed from the observed image, i(x,y,z), by convolving the latter with a reconstruction filter, g(x,y,z), i.e. o{x,y,z) = i[x,y,z) ® g(x,y,z), and thus, TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 97 IFI = jjji(x,y,z)® g(x,y,z)dxdydz (5-2) x,y,z where ® represents the convolution operator. Equation (5-2) can be simplified by replacing the convolution operator with its multiplication counterpart as follows: IFI - |JJ jjji(x,y,z) • g(u - x, v - y, u> - z) • dudvdwdxdydz x,y,zu,v,w - jjji(x,y,z) • jjjg[u - x,v- y,w-z)dudvdu> \-dxdydz x,y,z \u,v,w J = jjji(x,y,z) • K • dxdydz x,y,z = K • jjji(x,y,z) • dxdydz (5-3) x,y,z where K is the integral over the entire range of the reconstruction filter. This integral value, K, is a constant and does not vary with the image. Since we are only interested in the relative or normalized IFI value of objects, it is not necessary to determine the value of K as it will be factored out in the normalization process. It is shown in equation (5-3) that the IFI of the object is proportional to the IFI calculated from the observed image. Hence to calculate the relative IFI of the object, it is not necessary to first reconstruct the observed image i(x,y,z) to obtain o(x,y,z). The IFI of the object can also be calculated from a transformed image s(x,y,z) obtained by convolving the observed image i(x,y,z) with some filter, p(x,y,z), (whether it is reconstructive or not). To show this, we define IFIp as: TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 98 IFIp = K • jjj s(x,y,z)dxdydz x,y,z = K • 111 i(x,y,z) <8> p[x,y,z)dxdydz x,y,z From a similar analysis as above, the result of the convolution in equation (5-4) is a constant factor multiplied by the IFI of the object. IFIp = KpK • jjji(x,y,z) • dxdydz = Kp-IFI (5-5) x,y,z This means that if the IFI is difficult to calculate from the observed image, the observed image can be transformed into another image where the IFI calculations can be more readily determined. The above analysis shows that the IFI of an object can be calculated from the observed 3D image of the object according to equation (5-3). In our system, the 3D image is obtained by acquiring and combining a series of different focus level images as seen on the image detector plane. The image at the detector plane, i(x,y,zd), contains either an in-focus or out-of-focus object. When the object is in-focus, most of the signal intensities will be localized in a small region. When the object is out-of-focus, the observed intensities are dispersed over a larger region as the image is altered by the out-of-focus PSF. It can be seen from the previous chapter that the magnitudes of the central frequencies of the OTF of an image at any z location are the same. That is, OTF(0,0,z) = l Vz (5-6) TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 99 This then implies that the total intensity of the PSF in the xy-plane is the same for all z-focus values: As shown in equation 5-7, the total intensity observed in the detector plane (assuming that the image is large enough to capture the intensity variations) is the same for in-focus or out-of-focus objects. This can also be explained from a geometric perspective. It is known that the intensity of light observed at a distance from the point source varies inversely with the square of the distance. If we place an aperture in the light path, only the light that enters the aperture of the objective lens is allowed to pass through. If we look at the amount of light which hits the image plane when the latter is at a distance located close to the aperture, the light will be dispersed over an area similar to the size of the aperture. If the image plane is moved further away, the light observed will be less intense but would cover a larger area which increases by the square of the distance from the point source. The total light observed in both instances is the same. In our system, we assumed that the changes in focus is small compared to the distance of the object to the objective. Hence, the amount of light which passes through the aperture can be assumed to be invariant with respect to small focus changes. This analysis is valid if we also assumed that the light losses in the optic path (e.g. filters, objectives, cover slip, oil, mounting medium, etc.) are invariant with focus changes. (5-7) TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 100 The calculation of the IFI can then be simplified to calculations performed on only one image plane (since similar intensity sums are present in other focus planes). Equation (5-3) becomes: where Z is a constant representing the number of Z-focus planes of the observed object. 5.3. Segmentation and IFI Quantification Algorithm From the above analysis, the IFI of the object can be simply obtained if their observed intensities are summed over the entire region in which they occupied regardless of which focus plane the image is taken. The difficulty lies in determining the region in which to perform the integration/summation. Telomeres are generally not isolated from one another such that the signal intensities of one do not interfere with that of another. When they are in-focus, most of the telomere signal intensities are concentrated in a spot while the rest of the signal is dispersed in the surrounding which may contain other telomeres. When they are out-of-focus, the telomere signal intensities are spread over a larger region and hence is more likely to interfere with the signals of other telomeres making it more difficult to determine the IFI of each telomere. (5-8) TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 101 Thus, the major problem in accurately quantifying the IFI of telomeres lies in their segmentation, i.e. determining the exact boundaries of each telomere. Most telomeres are relatively easy to detect since they appear as bright spots. Approximate locations of these spots can be found by thresholding or edge detection methods (e.g. Section 2.2.4). However, a problem arises when finding the exact location of the borders. Figure 5.1 shows the (two possible) borders B1 and B2 of a cross section of a telomere. If the estimated telomere borders are closer to their centre intensity (Bj), the IFI value will be under-estimated. Conversely, if too much background intensity is included in the estimated border (B2), the IFI value will be over-estimated. There is also the problem of segmenting telomeres which are close to each other and determining which pixels belong to which telomere. In addition, not all telomeres lie in the same focus plane. Hence, segmentation in 3D space (which is rarely done in biological imaging) may be required. The relationship between the true IFI of the object and that calculated from the segmented object from a given 2D image assuming the level for the background intensity is Bgnd (Figure 5.1) can be represented by the following equation: = ^Reg.Bgnd + eBgnd,ReS + e0ut + QE (5"9) where g.Bgnd ~ the calculated value of the IFI of the segmented region calculated at the given background (Bgnd) level and over the defined region (Reg) 8Bend Reg =the error in selecting the background level TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 102 sDut = the error consisting of IFI calculated outside the defined volume QE = the error in quantizing the signal into discrete values. Cross section of telomere Figure 5.1. Errors in calculating the object IFI. The inherent noise in the system (optics and illumination aberrations, camera noise, sample preparation noise, etc.) and neighbouring telomeres make it difficult to define the "true" background level and segmented region for the IFI calculation. It can be seen from equation (5-9) that there is a compromise in TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 103 selecting the segmentation area and background level such that the total of all IFI errors is small. It can also be seen that it is easier to estimate the true boundary of the telomere if it is in-focus rather than out-of-focus. As a result, the errors of the IFI calculation are smaller for in-focus objects. Not all telomeres are at the same focus level and have the same shape. Thus, segmentation over different focus planes are required. The first step of our segmentation algorithm is to find the location of each telomere in the x, y, as well as the z-directions since telomeres have varying lengths and distributions in all directions. This is accomplished by first searching and recording the locations of the different local bright spots (center parts of telomeres) in each of the different focus plane (x,y) images. These spot locations are then compared to their corresponding spot locations in other images of different neighbouring focus planes. For each spot location, the image at the focus (z) plane which has the brightest spot intensity is considered to be the z-plane containing the center of the spot. Once the center of a telomere spot is found, the extent of that spot in 3D space is then determined by using the algorithm described later in Section 5.4. The segmentation algorithm we used to find the spot in each of the multi-focus level xy planes is similar to that of the Laplacian filter (Russ, 1990). For each pixel, the average intensity value of its surrounding pixels is subtracted from its intensity, I(x,y) to generate an edge image, E(x,y), that is: TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 104 1 1 1 E[x,y) = I(x,y)--YYl{x-i,y-j) (5-10) We named this operation the Average Difference filter and its operation (in one dimensional space) is illustrated in Figure 5.2. If the resulting E(x,y) value is above a threshold level, the pixel is considered to be a telomere pixel. Otherwise, it is considered as a background or boundary pixel. The algorithm just described eliminates noise by using a threshold above the noise level and detects intensity peaks. At the center portion of a telomere, the average value of the surrounding pixels is generally less intense and hence the value E(x,y) is positive and large. At the edges of the telomere, the average value of the surrounding pixels is generally the same as the pixel value since on average, half of the surrounding pixels have lower intensities than the central pixel and the other half have higher intensities. As a result, the value E(x,y) is small and near zero. Similarly, at the background region, the average value of the surrounding pixels is similar to that of the pixel. Hence, by using thresholding, the noise pixels in the background and the edge pixels of the telomeres are removed. Telomeres which are close to each other can also be separated using this technique. The reason is that the valley in between two nearby telomeres is lower in intensity than the average surrounding and thus can be removed by thresholding. An example of the use of our segmentation method on a typical telomere image is shown in Figure 5.3. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 105 Figure 5.2. Application of the average difference filter. The Average Difference filter is first applied to the simulated signals of the telomere spots. A threshold above the Average Difference operation then defines the centers of spots (shaded in the diagram) detected by the algorithm. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 106 (0 (d) Figure 5.3. Process of segmenting a typical telomere image. A typical fluorescence image of the telomeres (a) is processed with the Average-Difference filter to generate image (b). A threshold level is then selected from the histogram of the processed image and this threshold is applied to 5.3b to generate a binary image (c) of the telomeres. The resulting bright spots in 5.3c are first labeled and then dilated to generate the final mask image for the telomeres. The boundaries of the segmentation results (d) are then generated. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 107 The threshold level for the segmentation is chosen such that most of the background noise are eliminated by thresholding. If the distribution of noise is assumed to be Gaussian, the peak value will correspond to the central point in the Gaussian distribution. Since most of the pixels in a telomere image belong to background pixels, the maximum peak of the histogram corresponds to the background level. The value at 5% of the peak value corresponds to the point at 1.7a of the Gaussian distribution. Hence, all points less than 1.7a value from the peak corresponds to approximately 95% of the points in the Gaussian distribution. The search for the point corresponding to 5% of peak value is made at the left of the peak (lower intensity values) in the histogram. This side of the histogram peak is used (instead of the higher intensity side) because the interference from the telomere signals which are located at higher intensity values would be less. The 1.7a point to the right of the peak can then be calculated. The distance of this point from the peak is then added to the peak point to obtain the threshold level which eliminates approximately 95% of the background pixels. This level seems to be the optimal for removing background noise pixels and also preserving the relevant telomere peaks. Objects which are too small and whose intensities are similar to the background level (those objects pointed out by the arrows in Figure 5.3c) are classified as artifacts and rejected from further analysis. The above algorithm removes the edges of telomeres. To recover these edges, we first need to dilate (Figure 5.3d). However, this will combine or fuse multiple telomeres into one. To overcome this problem, we first perform labeling. In the labeling process, each continuous connected object is given a TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MCROSCOPY 108 unique number and then the size of each telomere mask is dilated by one to recover the lost edges of that telomere. Note that if dilation was performed before the labeling, objects which are closed together may be connected and considered as one object by the labeling process (those objects pointed out by the arrows in Figure 5.3d). 5.4. Number of Focus Planes Required It was shown in Section 5.2 that only a single image from a single focus plane is required to represent the IFI value of the object (see Section 5.8). However, from Section 5.3, it can be seen that if the objects are out-of-focus, more errors are introduced in the IFI value due to segmentation errors (i.e. defining the region for the IFI calculation and the intensity of the background). Hence, the best focus image should be used for determining the IFI value of the object. In a given image, however, not all objects are at their corresponding best focus. Telomeres can be located at different focus levels and they can have varying lengths and distributions in the z-direction. In humans, chromosomes on a microscope slide are typically 0.5um thick in the z-direction (the approximate size of the chromosome tip). Telomeres range in size from 0.1 to 0.3um in diameter and can lie anywhere within the tip of the chromosome. Hence, to accurately quantify the telomeres, images from multiple focus planes should be examined. To experimentally determine the number of planes required for calculating the telomere IFI values, approximately 20 images at 0.1 um spacing TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 109 are acquired. One of the central images in this series contains the "best" focus image. The first and last few images in this 20 image series are blurry but are still discernible images of the telomeres. A z-spacing of O.lum is used because it is the approximate limit of the step size of the z-drive motor. It also corresponds to the size of the smallest telomere, and corresponds to at most 20% of the sampling resolution in the z-direction of the objective lens i.e. we are over sampling. Three different methods for calculating the IFI of a telomere are performed. The results of these are then compared to determine which is the best method to use. One method uses the image plane which appears to be at the best focus to calculate the IFI of each telomere in the image. Not all telomeres in the image are in-focus in a single image plane. Hence, the calculated IFI values for out-of-focus telomeres is lower in value from the corresponding IFI calculated at the best focus position for that telomere. The second method utilizes a number of different focus plane images and selects the best focus (highest evaluated IFI value) image to calculate the IFI value for each telomere in the image. In the last method of IFI calculation, the IFI value of each telomere in the image is determined from the sum of the corresponding telomere IFI values calculated at each of the multi-focus plane images. The methods are elaborated upon below. For each image acquired, the telomere segmentation and the IFI calculation measurements are performed. Each telomere in each image is then matched with each of its corresponding telomere in the other multi-focus plane TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 110 images and is given a unique label. Since the sample is only moved in the z-direction (no other optical component is moved during the acquisition sequence), there is little registration error. Hence, no correction in the x-y shift is required in the multiple plane images. There can be, however, some differences in the location of the x-y center of a telomere spot in different z-plane images. This is because telomeres can be irregularly shaped and can lie in any orientation in 3-dimensional space. Thus, the center of each telomere spot in a particular focus plane image may be different from its neighbouring planes. To match telomeres at different focus planes, the telomere center at a given focus plane should correspond to a point not necessarily the center but inside the area of the corresponding telomere in an adjacent focus-plane image. Two nearby telomeres can also be segmented by our algorithm. As a telomere becomes out-of-focus, it gets blurred and gradually fades into the background. Hence, when two nearby telomeres are in-focus, they may be detected as two individual objects by our segmentation algorithm. However, when they are out-of-focus, they get blurred and thus may be detected as a single telomere. To overcome this problem, the area of each telomere is calculated as well as its IFI value for each focus plane. If the area of a detected telomere in one image is greater than twice that of a corresponding telomere in another focus-plane image, then two (or more) telomeres must be present. In this instance, the area corresponding to the telomere in the image with the highest IFI value (in-focus image) will define the area from which the IFI is recalculated in the other image plane (out-of-focus image). This would reduce the area with which the IFI should be calculated. Thus the IFI calculated TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MCROSCOPY 111 would be smaller than the "actual" value. Nevertheless, it would be better to segment nearby telomeres and estimate their IFI values than not to include this data in the total IFI calculation for multiple focus plane images. The IFI values calculated for a telomere from all images at different focus planes are then summed to obtain the summed IFI value of that telomere. The above method is not practical to perform in most situations because of the large number of images required for each cell. Handling these images would then require large disk storage space and long image acquisition and processing times. To reduce the amount of data and acquisition time, the number of planes acquired needs to be reduced. To accomplish this, we may keep the same O.lum spacing between images but only acquire the central images of the 2um span. This method must accommodate for a large variation in telomere sizes and z-focus positions. Another method of reducing the data is to acquire images at higher sampling spacings (e.g. 0.2um, 0.3um, ...) but over the 2um span. In this case, it then becomes a matter of determining the highest sampling spacing that do not significantly increase the error in the IFI calculation of the telomere. Theoretically, if we sample at half the resolution of the objective lens, the images acquired should be representative of the object. This corresponds to a sampling spacing of no more than 0.3um. This spacing is experimentally explored below. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MCROSCOPY 112 5.5. Algorithm Evaluation And Validation 5.5.1. Overview In this section, we first estimate the spatial resolution yielded by our algorithm. This evaluation gives an indication of how far apart two telomeres have to be before they are classified as two distinct objects by our telomere segmentation algorithm. We then test the validity of our IFI quantification method. For this, we use test objects with known IFI values. A theoretical analysis is not easy to perform because of the non-linearities in the segmentation step (the IFI values are based on the area defined by the segmentation result). We use 3 different methods outlined below to validate our results. The results and discussions of the methods are covered below. One method for validation is to use simulation, i.e. to construct simulated objects with known IFI values and shapes and convolve these objects with the PSF of the system. The various IFI quantification algorithms are then applied to the generated images. The theoretical PSF of the system although similar to that of the microscope system used, may not be an accurate representation. Nevertheless, this PSF is sufficient for validation purposes since the function used for the system PSF needs only be spatially invariant as shown in equation 5-5. Using this method, we must choose objects which are representative of the shapes and distributions of telomeres. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MCROSCOPY 113 Another method for validating our quantification method is to use different sizes of small fluorescence beads as test objects. The images of these beads are acquired and our IFI quantifying process is applied. The measured IFI values are then compared with the estimated IFI of the beads based on the assumption that the IFI of the bead is proportional to its volume. A precaution in using this method is that the telomeres unlike beads are generally not spherical in shape. Another approach for validation is to use telomere objects of known lengths. Plasmids (circular pieces of DNA) can be used for this purpose. DNA with telomere sequences (TTAGGG repeats) of known lengths and up to 3200 base-pairs can be inserted into a plasmid. A precaution of this approach is that these telomere lengths are much smaller than those in human chromosomes. As a result, the fluorescence signals are weaker, harder to detect, and may not span over a similar number of focus planes. Another precaution is due to other variations in the system (such as biological and sample preparation variations) which may interfere with the accuracy of the validation process. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MCROSCOPY 114 5.5.2. Spatial Resolution of IFI Segmentation Simulated test objects are used to estimate the spatial resolution of using our algorithm to separate two nearby telomeres (Figure 5.4). We use the theoretical PSF of the system and convolve it with our simulated test objects to generate the images of the test objects. The test objects consists of two point sources which are incrementally moved apart from one another in steps corresponding to one pixel spacing of the detector. Our algorithm is then applied to the generated objects to determine how far the test objects have to be away from each other before they are treated as being two objects, It can be seen that one can just visually distinguish 2 points in the simulated images when the points sources are 4 pixels (the approximate resolution of our microscope and camera system) away from each other. At 5 pixels away from each other, the human eye and our algorithm can both clearly separate the two objects. This distance corresponds to approximately 0.54um in the object plane at 63x magnification. This distance is also twice the theoretical resolution of just the microscope. Thus, it is estimated that our segmentation algorithm can separate objects that are 5 or more pixels (> 0.54um) apart. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 115 Figure 5.4. Simulated test objects for spatial resolution estimates. The top row shows the characteristics of the test objects. The middle row shows the calculated in-focus image. The last row shows the borders of our telomere segmentation algorithm. Our algorithm separates the two point sources when they are at least 4 pixels apart. 5.5.3. Simulated Objects To validate our telomere IFI quantification algorithm, we used different simulated test objects where the relative IFI value of each is known. The simulated objects have varying shapes and intensity distributions (to simulate varying shape and intensity of telomeres) but the same IFI value. These objects are used to test the robustness of our IFI algorithm to see if similar IFI values are generated. The simulated objects consist of i) a single point source whose dimensions are 1 pixel (object #1 in Figure 5.5), and ii) point sources whose dimensions are greater than 1 pixel in the x directions (objects #2, #3, and #4 TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 116 in Figure 5.5) and greater than 1 pixel in the x and y directions (object #5 in Figure 5.5) and greater than 1 pixel in the z direction (objects #6, #7, #8, and #9 in Figure 5.5). Simulated images were generated by convolving test objects with the theoretical PSF of the system as described in Chapter 5.3. While the spatial distribution of each simulated object is different from the others, its IFI value is the same as the others (i.e. the sum of pixel intensity values for each object is the same as the sum of pixel intensity values for other objects). The values and shapes of the simulated test objects (in the x-z plane) are shown in Figure 5.5. The results of the IFI calculations (using equation 5-9) at each differently focus image are shown in Table 5.1. A plot of the IFI variations for some of these simulated objects as a function of focus is shown in Figure 5.6. The total IFI estimation of each object is then determined from the sum of the IFI values over different focus images (sum of row values in Table 5.1) and are shown in Table 5.2. First, the single point source image is chosen as a reference for comparing IFI values. After normalization, the IFI of the in-focus image of this single point object is set to 100%. From the results (Table 5.1 and Figure 5.6), it can be seen that the IFI value calculated (at the best focus plane) of each simulated object can vary by 9% amongst objects. The variation of the IFI of an object over ±0.1 um from its best focus plane can be as high as 10% (e.g. object #6). The variation are higher at higher de-focus amounts. At ±0.2um de-focus (a typical z-distance for metaphase telomeres prepared on a slide), the calculated IFI can be high (approximately 30%). From the theory (equation 5.7), the IFI calculation at any focus plane {zl of z2) should be the same. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 117 However, due to the difficulty in defining the exact border of the telomere objects in different images as the objects become out-of-focus, the calculated IFI values are different. Thus, the results suggest that the focus plane from which the image of the object is captured plays an important role in the accuracy of measuring the IFI value of the object. If the IFI value obtained at each focus plane (calculated using equation 5.9) is summed to give the total IFI value for the object (e.g. column 1 of Table 5.2), then the variation in the total IFI value is only 3%. Thus, summing individual IFI values calculated from a series of images captured at different focus planes generates more precise results than the IFI value calculated from an image captured at only a single focus plane. Instead of summing planar IFI values of images taken at all 0.1 um z-spacings, the summed IFI values are calculated from images captured at larger sample spacings (e.g. 0.2, 0.3, and 0.4um z-focus apart). As less z-focus planes are used in the summation, the resulting IFI value will be proportionately smaller (e.g. only half of the O.lum z-spacing images are used in the 0.2urn sample spacing calculation and hence the generated IFI sum is half that of the O.lum sample spacing calculation). Hence, the summed values are multiplied by the sampling frequency (e.g. 2, 3, and 4, respectively) to result in a similar magnitude in total IFI as that of the O.lum spacing. The results of different variations of the sampling just described are summarized in Table 5.2. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 118 Object #1 1.00 Object #2 0.50 0.50 Object #3 0.3333 0.3334 0.3333 Object #4 0.25 0.25 0.25 0.25 Object #5 0.13 0.25 0.06 0.22 0.15 0.19 Object #6 0.50 0.50 Object #7 0.3333 0.3334 0.3333 Object #8 0.25 0.25 0.25 0.25 Object #9 0.20 0.20 0.20 0.20 0.20 Figure 5.5. Simulated test object values and shapes. For all objects, the horizontal direction represent the extent of luminance in the x-direction. For object #5, the vertical direction represent the extent of luminance in the y-direction. For objects #6, #7, #8, and #9, the vertical direction represent the extent of luminance in the z-direction. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 119 Obj. # Normalize Calculated IFI Percentage at Sub-Micron Z-Spacing -0.5 -0.4 -0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 1 15 40 61 81 95 100 96 81 62 42 15 2 16 35 61 80 93 98 94 81 62 37 18 3 18 38 63 80 94 98 90 77 60 40 19 4 20 40 60 77 90 95 91 78 61 42 22 5 20 40 60 78 91 95 91 79 61 42 22 6 0 24 51 71 88 98 97 88 72 53 25 7 16 40 60 79 92 96 92 80 62 42 17 8 26 50 70 85 92 93 85 70 51 26 11 9 19 42 59 75 86 91 87 76 60 42 20 Table 5.1. IFI values of typical human telomeres at different focus levels. The bold face values represent the highest IFI values over the range of focus levels for each telomere. -0.6 -0.4 -0.2 0 0.2 0.4 0.6 Z-focus plane (microns) Figure 5.6. Normalized IFI values at varying focus of different objects. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MCROSCOPY 120 It can be seen that if the IFI calculation were summed over every other 0.2 or 0.3 urn z-spaced images, the summed IFI value can vary by 5% amongst objects. The variation in the sum increased to 9% when every other 0.4 z-spaced images are summed. This suggests that images can be spaced at up to 0.3 um apart from one another without significantly increasing the error in the calculation of the summed IFI value of the simulated object. Obj. # Normalized Calculated IFI Percentage Summed Over Different Z-Spacings (1:0) (2:0) (2:1) (3:0) (3:1) (3:2) (4:0) (4:1) (4:2) (4:3) 1 100 100 100 98 101 101 94 100 105 100 2 98 96 100 96 99 99 94 100 99 101 3 98 97 100 96 100 99 91 100 103 100 4 98 97 100 94 100 100 90 100 103 101 5 99 97 101 95 101 101 91 100 103 101 6 97 97 97 96 97 98 93 93 101 101 7 98 98 98 95 100 99 92 98 104 98 8 97 97 98 98 97 98 97 99 98 97 9 97 97 96 98 97 98 96 96 101 96 Table 5.2. IFI values of simulated test objects calculated at different focus level sampling spacings. The summed value is obtained by including images spaced at 1, 2, 3, or 4 (O.lum) z-focus spacings. The first number in parenthesis (in the column heading) represents the z-focus spacing used to generate the summed value. The second number in parenthesis represent different starting point or offset (for each focus spacing) in which the sum is calculated. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 121 5.5.4. Fluorescent Beads Images of different size fluorescence beads (0.1, 0.2, 0.5, and 1.0 um) are acquired and the IFI value at each focus plane and the total IFI value of all focus planes are generated for each bead. If we assume that the fluorescence intensity of the bead is related to the 3-dimensional size of the bead, then the IFI of the bead would be proportional to the cube of its 1-dimensional size (diameter). The results of this experiment and the normalized expected theoretical IFI values are summarized in Table 5.3 and plotted in Figure 5.7. Diameter of Bead (um) Best Focus IFI (Mean 8B Standard Deviation Best Focus Normalized IFI Sum Z-Planes IFI (Mean 85 Standard Deviation Sum Z-Planes Normalized IFI Theor. IFI 0.1 9.24 ± 2.64 1.00 ± 29% 110±31 1.00 ±28% 1 0.2 99.1 + 13.3 10.7 ± 13% 968 ±130 8.8 ± 13% 8 0.5 1162 ±37 126 + 3% 12503 ± 530 113 + 4% 125 1.0 9344 ± 203 1011 ±2% 146189 ± 3224 1329 ± 2% 1000 Table 5.3 IFI values of different size beads. It can be seen from the graph (Figure 5.7) that the calculated IFI values for the beads correspond closely to the theoretical IFI values. It can also be TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MCROSCOPY 122 seen that the IFI value calculated at the best focus plane gives slightly more accurate results than the summed value for all focus planes in the cases of the larger size beads (0.5 and 1.0 um) The results show that our IFI algorithm gives a good estimate of the IFI of the bead. T5 10 00 T3 0) N "(5 E 10000 1000 100 10 • Best Focus Sum of Z-Planes Estimated Size 10 100 Estimated IFI (normalized to 0.1 micron bead) 1000 Figure 5.7. IFI distribution of different size beads. 5.5.5. Plasmids Similar to the analysis for the beads, images of different size telomeres within plasmids (150, 400, 800, and 1600 base-pairs) are acquired and the total IFI value is generated for each plasmid. If we assume that the TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 123 fluorescence intensity of the plasmid is related to the number of telomere base-pairs present (like our assumption for telomeres in chromosomes), then the IFI of the plasmid would be proportional to the number of base-pairs in the plasmid. The results of this experiment and the normalized expected theoretical IFI values are summarized in Table 5.4 and plotted in Figure 5.8. Column 1 Column 2 Column 3 Column 4 Column 5 Column 6 Size of Plasmid (base pairs) Normalized to 150 base pairs Best Focus IFI (Mean 8s Standard Deviation Best Focus Normalized IFI Sum Z-Planes IFI (Mean 8s Standard Deviation Sum Z-Planes Normalized IFI 150 1.0 4.17 ± 1.31 1.00 ±31% 48.0± 18 1.00 ±37% 400 2.7 11.2 ± 3.4 2.69 ± 29% 119 ±40 2.47 ± 33% 800 5.3 24.3 ± 4.6 5.83 ± 19% 221 ±41 4.60 ± 19% 1600 10.7 44.7 ± 5.7 10.7 ± 13% 464 ± 64 9.67 ± 13% Table 5.4 IFI values of different size plasmids. Again, as with the results with the beads, the plasmid results also show good correlation between the calculated and the expected theoretical values (compare columns 4 or 6 with column 2 of Table 5.4). That is, our IFI algorithm gives a good estimate of the length of telomeres. The best focus IFI values are better matched to the theoretical values than the summed value for all focus planes. The variance of the calculated values are higher for the TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 124 smaller plasmids. This suggests that there are larger variation in the size of the telomere repeat sequence in smaller plasmids. o -I 1 1 1 1 1 1 1 1 0 200 400 600 800 1000 1200 1400 1600 Size of Plasmid (base pairs) Figure 5.8. IFI Distribution of Different Size Plasmids. 5.5.6. Summary of Algorithm Validation Results The previous three methods used to evaluate the IFI value of the object showed that the IFI of the object should be evaluated using images captured at different focus planes. The IFI value evaluated using the IFI value from the best focus plane image for the object produced similar results to the IFI value evaluated using the sum of IFI values from images at different focus planes. For the simulated objects, the best focus method results in 3% variation in IFI value amongst different objects while the summed IFI values (up to sample spacing of 0.3um) from multiple focus plane method results in 5% variation. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 125 For the beads and plasmids, both methods produce similar results. As there are more variability in the smaller size beads and plasmids compared to the larger objects, the variations in their sizes are also higher. If only one image near the best focus plane for all objects is used, the focus of individual objects in the image can be ±0.2um from the best focus position for that object (as shown in Figure 5.6 for the simulated objects, Table 5.5 for the beads and plasmids, and later in Table 5.6 for the telomeres). The IFI value evaluated from a single image plane could then vary by 20% from the IFI value estimated from the best focus image. Object Z-focus position (microns) -0.50 -0.40 -0.30 -0.20 -0.10 0.00 0.10 0.20 0.30 0.40 0.50 0.1 (im bead 0.77 0.85 0.98 0.98 0.98 1.00 0.89 0.85 0.79 0.53 0.43 0.2\m\ bead 0.33 0.52 0.69 0.84 0.96 1.00 0.95 0.87 0.76 0.60 0.48 0.5(xm bead 0.42 0.57 0.72 0.85 0.93 1.00 0.97 0.89 0.80 0.69 0.57 1 .Oum bead 0.69 0.79 0.87 0.93 0.98 1.00 0.99 0.95 0.88 0.78 0.66 1 50bp plasmid 0.48 0.64 0.66 0.68 0.85 1.00 0.82 0.70 0.61 0.59 0.54 400bp plasmid 0.67 0.77 0.86 0.97 0.98 1.00 0.95 0.76 0.66 0.48 0.35 800bp plasmid 0.48 0.62 0.78 0.91 0.97 1.00 0.89 0.67 0.57 0.47 0.33 1600bp plasmid 0.63 0.70 0.83 0.88 0.94 1.00 0.93 0.84 0.59 0.41 0.27 Table 5.5 Normalized IFI values of objects at different focus positions. These different methods of IFI calculation is next applied to telomere images to see if similar results are obtained. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 126 5.6. Human Telomeres Results 5.6.1. Number of Focus Planes Required A similar type of analysis as that performed on the simulated objects is made on the telomere images (using equation 5-9 for each focus plane). Multi-focus plane images of telomeres are acquired at O.lum z-spacing. There were 125 telomeres analyzed from the images. The IFI for each focus plane for each telomere is generated. The results of some of the 125 telomere IFI's are shown in Table 5.6. Telomere Number Z-Position (microns) -0.9 -0.7 -0.5 -0.3 -0.1 0.1 0.3 0.5 0.7 0.9 1 63.3 127.3 212.1 324.8 383.6 381.2 307.1 235.2 158.5 97.4 2 3.0 21.8 42.5 79.4 111.2 125.2 110.5 93.8 70.0 42.7 3 7.7 35.3 73.7 139.6 202.0 238.3 231.7 201.3 156.0 104.9 4 28.8 67.0 113.3 196.7 261.9 289.6 267.7 229.0 184.4 120.1 5 0.4 18.9 41.6 90.2 149.4 183.5 195.6 170.0 136.5 95.5 6 19.2 53.2 97.8 162.6 234.7 256.5 246.0 213.4 170.4 113.0 7 23.6 63.1 107.4 183.1 254.6 279.0 231.8 183.3 134.9 71.4 8 16.8 69.5 137.3 239.2 337.3 374.2 331.4 267.4 197.9 122.4 9 14.1 40.1 79.5 139.9 220.6 266.8 267.9 236.0 193.0 135.2 10 51.3 116.6 195.3 301.0 350.6 334.9 270.8 187.2 118.9 65.6 Table 5.6. IFI values of typical human telomeres at different focus levels. The bold face values represent the highest IFI values over the range of focus levels for each telomere. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 127 It can be seen in Table 5.6 that the position of the in-focus plane (highest IFI value) for each telomere in a metaphase chromosome sample varies by at least 0.4u,m. In addition, there is at least a three fold difference between the lowest and highest value amongst the telomeres in the example. -1 -0.5 0 0.5 1 Distance from In-Focus Position (microns) Figure 5.9. IFI of typical human telomeres at different focus levels. The IFI values at different focus planes are then normalized so that they can be compared with those of other telomeres. To normalize the IFI values, the highest IFI value for each telomere is set to 1.00. The IFI values at other focus planes are then divided by the highest IFI value for that telomere to give TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 128 the normalized results. A plot of the IFI variations for the average of 125 normalized telomere IFI values as a function of focus is shown in Figure 5.9. It can be seen from Figure 5.9 that if a single focus plane is used to represent the IFI value of the telomere, there can be at least 10% error in the calculated IFI if the telomere is 0.2um out-of-focus. That is, if a single image is used to represent the best focus image for all telomeres, there can be some telomeres which are 0.2um out-of-focus (Table 5.6). Hence, the calculated telomere IFI value could be 10% less than their in-focus value. Similar to the simulated objects, the total IFI is calculated as the sum of values at three differently spaced z-intervals (0.2, 0.3, and 0.4um). The results of different variations of the sampling for the telomeres are summarized in Table 5.7. It can be seen from Table 5.7 that the standard deviation of the total IFI values for the 125 telomeres is approximately 18% of the average IFI value. This implies that if the total IFI is not calculated over multiple focus planes, there can be 18% difference between the multi-focus plane IFI value and the best focus IFI value. The difference would be higher if the comparison is made with the IFI value calculated from the image where the telomere is 0.2um out-of-focus (i.e. if only one image plane is used in the analysis as in most quantitative cytometry studies). It can also be seen that summing images at every 0.2 or 0.3um z-spacing produce similar results with error of approximately 1.5% of the total IFI value. The error more than doubles to 3.8% when the summing step size is increased to 0.4um. Hence, a 0.3um step TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MCROSCOPY 129 size as predicted in Section 5.4 is used for acquiring multi-plane telomere images. The total IFI values for each telomere can then be obtained from these images. Alternatively, as shown in the 3 IFI validation experiments, the best focus image (selected as being the focus plane that gives the highest IFI value for that object over O.lum z-spaced images) could be used to calculate the representative IFI of the object. Z-Spacing (um) Offset (um) Average IFI Standard Deviation Average CV 0.1 0.0 10.863 1.802 0.0000 0.2 0.0 10.856 1.802 0.0148 ±0.0086 0.2 0.1 10.867 1.812 0.3 0.0 10.843 1.816 0.0157 ±0.0088 0.3 0.1 10.875 1.802 0.3 0.2 10.870 1.810 0.4 0.0 10.992 1.887 0.0382 ±0.0111 0.4 0.1 10.894 1.858 0.4 0.2 10.720 1.774 0.4 0.3 10.840 1.834 Table 5.7. IFI values of human telomeres calculated at different focus level sampling spacings. For each z-spacing, there can be a number of different starting points or offsets from which the sum is generated. The summed IFI value is generated by summing the IFI values from images at different z-spacings for each telomere. The average of 125 summed IFI values is calculated. For each z-spacing category, the average coefficient of variation (CV) is calculated (i.e. CV is the difference between the standard deviation and its mean standard deviation determined from within the z-spacing group divided by the mean standard deviation). TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 130 5.6.2. Telomere Distribution in Cells The distribution of telomere IFI values in a cell is shown in Figure 5.10. It can be seen that the plot is asymmetric about its peak and resembles a Poisson distribution rather than a Gaussian distribution. An explanation for this is that there are variations in telomere lengths in a given cell as shown previously using the Southern analysis (Allshire et al. 1988). Even if all telomeres in the cell have the same telomere length to begin with, after a number of cell division, the majority of the telomere lengths will be at around the same value. There are however, others which do not have a similar rate of reduction in telomere lengths. Thus, a spread of long telomeres can be observed. The number of very short telomeres compared to the majority of telomeres in the cell is small because the cell has reached its critical state and tends not to divide. If the normal cell happens to divide, only those cells which have longer telomeres will tend to survive. By analyzing less than 30 metaphases samples, a statistical interpretation of the IFI distribution of the cell population can be obtained. Previously, using the Southern analysis, approximately 100,000 cells were required to give similar results as ours. In addition, our IFI analysis can give information of the IFI distribution within each individual cells. If karyotyping (the identification/classification of each chromosome within a cell) is performed on the chromosomes, the IFI of each of the 24 different types of chromosomes in the cell and in a population of cells can be obtained. We have performed TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 131 these studies on mouse telomeres (Zijlmans et al.; 1997) and the p-arm of chromosome 17 (Martens et al; 1997). 35 cu 30 u 2 25 l 20 1 15 c CU 3 O" cu LL 10 0 r- OO co C^J CO CM CO CO CsJ co CO LO CO CO to r---CO CO CO CO Lf) CD Lf) CD IFI Value (grey levels) Figure 5.10. Telomere IFI distribution in a cell. 5.7. Chapter Summary In this Chapter, we described what the IFI of an object is and how it can be theoretically evaluated. We have shown that although objects occupy a 3D space, only one image obtained at the focus plane is sufficient to determine the IFI of the object. We next described the difficulty in calculating the IFI value as telomeres are in close vicinity of one another and hence their signal intensities overlap. The errors in IFI calculation as a result of segmentation are then discussed. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 132 We then introduced our telomere segmentation algorithm. This algorithm is capable of separating images of point sources whose locations are 4 pixels apart from one another. Our algorithm also segments objects over noisy and varying (not flat or similar in intensity) background intensities. Three methods of evaluating the IFI value of an object are introduced. The first calculates the IFI value of every objects within the image using the same focus plane image. The latter two methods use images captured from multi-focus planes. One method selects the best focus (highest IFI value) image for every telomere to evaluate the IFI of the object. The other method evaluates the IFI value of the object as the sum of the corresponding IFI values for the object determined from images at equally spaced multi-focus planes. We used simulated objects, beads, and plasmids to evaluate and validate our IFI segmentation and quantification methods. We then applied our IFI algorithm on telomeres and compared the results. We have shown that our IFI quantification algorithm can be used to estimate the lengths of telomeres (verified by the beads and plasmid experiments). We have also shown that images from more than one focus plane are required to reduce the errors in the IFI calculation. The best focus IFI value or the sum of IFI values evaluated from multi-focus plane images give similar acceptable results. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 133 Chapter 6. Segmentation of Chromosomes 6.1. Overview This chapter describes how we segment fluorescence microscopy images of metaphase (chromosomes which have duplicated but have not separated in a dividing cell) chromosomes. The objective here is to determine the regions occupied by each chromosome in the image. By segmenting the chromosomes, the telomere IFI values (obtained from Chapter 5) can be linked to specific segmented chromosome objects. The user can then classify the chromosome type and the length of the telomeres of each chromosome can be obtained. The variability in the chromosome texture (intensity) within individual chromosomes and amongst different chromosomes make it difficult to find the exact borders of each chromosome. In addition, the high noise levels associated with low light level fluorescence images pose another difficulty for segmentation. Yet another segmentation difficulty lies in defining the boundaries of touching and overlapping chromosomes. Although one can select metaphases where all the chromosomes are isolated from one another, such images are rare to find. Hence, one typically scans a slide to find a metaphase where most of the chromosomes are isolated from one another and some of which are touching or overlapping. Images of these are then acquired and used for our analysis. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 134 We have applied several known edge detection filters to segment chromosomes. The edge detection filters tried included the traditional difference of Gaussians (DoG) filter (Marr 1982) and the Canny filter (1983). These edge detectors generate incomplete or discontinuous edge boundaries as well as edges within and outside the chromosome region. Furthermore, additional processing (which can be quite computationally involved) are required to select and fine tune the edge pixels to form a continuous edge around each object. Thus, we developed a method which remedies the boundary discontinuity problem and give the required chromosome boundaries. Our method employs only integer operations and hence is faster to compute. This method involves a novel filter called the Rank Difference filter which can be used as an edge detector or a morphological filter. Due to the difficulties in chromosome segmentation mentioned above, there are no single or combination of segmentation techniques which can correctly segment all chromosomes. Ji (1994) developed a method to segment chromosome images from brightfield microscopy (which has better contrast and less noise than images from fluorescence microscopy, the microscopy mode used in our study). His method uses an iterative rule-based approach to obtain the required number of segmented chromosomes per image. We did not use Ji's method because it can take a long time to compute (as many iterations may be required) and his method is not designed for chromosomes which has been stained to highlight their banding structures (such as the fluorescence DAPI stain that is used in our study). We require this banding structure information for karyotyping (identify the type of chromosome in a cell) such TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 135 that telomere length information of a particular chromosome in a cell can be obtained. Commercial chromosome analysis systems (Applied Imaging Inc., Biological Detection Systems Inc., Vysis Inc., and Oncor Instrument Systems) tend to use a simple semi-automated segmentation algorithm to generate an initial estimate of the chromosome borders. The algorithms typically used for this step consists of first interactively or automatically thresholding the image and then extracting the chromosomes borders from the thresholded image. Once the initial borders are found, researchers then manually verify and correct the segmentation results. By using our method, most of the chromosomes are correctly segmented. Hence, less user interaction is required in the manual verification process resulting in a less tedious and a more economical overall interactive analysis. Our chromosome segmentation algorithm consists of a combination of different segmentation methods since a single technique does not produce good results. Each method or step in the sequence improves on the results obtained by the previous step. Thresholding is first used to define the first approximation of the regions occupied by chromosomes. Texture information in the segmented region is then used to generate the second approximation of the chromosome region. In this step, we first detect the local high intensity pixels. We then use our Rank Difference filter as a morphological operator to merge detected pixels into different chromosome regions and at the same time separate touching chromosomes. Standard dilation and erosion morphological filters were not used in this step because they do not perform as well as our TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 136 filter for both merging and separating the different chromosome regions. The final step first finds the edges of the chromosome regions using our Rank Difference filter again, but this time it functions as an edge detector. Our Rank Difference filter is simple, employing only integer operations, to implement and generates results comparable to or better than the more complex edge detectors (difference of Gaussian and Canny). The edges are then used to refine the borders for each detected chromosome region. Feature values are extracted from each chromosome region and these extracted values are used to reject objects whose signal intensities are too weak and whose sizes are too small to be chromosomes. Finally, the segmented telomere image is used to confirm that the segmented regions are indeed chromosomes with telomeres located at their ends. As the Rank Difference filter is used in two stages of our segmentation algorithm, we first describe the formulation and usefulness of this filter (Section 6.2). We then compare the edge detection properties of our Rank Difference filter with those of traditional edge detectors (Section 6.3). Finally, we describe the details of our chromosome segmentation algorithm as outlined above (Section 6.4-6.8). 6.2. Rank Difference Filter Our Rank Difference filter is formed from a difference of two rank filters. Rank filters are widely used in image processing. A popular and commonly used rank filter, the median filter, removes random noise and preserves edges in images (Huang, 1979, Russ and Russ, 1986). Another commonly used rank TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 137 filter, the maximum or minimum rank filter, is used to obtain smooth background regions (Bright and Steel, 1986). By combining the maximum and minimum rank filtered images and calculating the difference, another image containing the maximum local contrast in the original image (Russ, 1990) is generated. From this latter filter, we obtain the Rank Difference filter by extending the choice of rank filters in the difference to include those in between the minimum and maximum ranked values. As a result (as shown later in this section), the choice of rank filters to use in the difference determines the filter's tolerance to random noise in the image. It is also shown later in this section that our Rank Difference filter can be used as a selective morphologic filter. This filter has properties which out-performs those of other filters for use in chromosome segmentation. In formulating the Rank Difference filter, we first define a region S(i(x,y)) in the image i(x,y) containing a total of v pixels and a pixel (x,y) which may or may not be inside S(i(x,y)). The magnitudes of the pixels in S(i(x,y)) are ordered such that the smallest value is denoted by R^Sfifcy))] and the largest value is denoted by R^Sfifcy))]. The rank filter R^S(i(x,y))] is the filter whose output is the ith smallest value of the pixels in S(i(x,y)). Our Rank Difference filter Ru [S(i(x,y))\ simply combines the difference of two rank filters, an upper rank Ru[S(i(x,y))] and a lower one R^Sfayfl, into one filter as follows: K,,[S(i(x,y))} = Ru[S(i(x,y))} - RlSfifcy))] where 1<l< u< v (6-1) TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 138 The output value of the Rank Difference filter is thus the difference between the uth smallest value and the Ith smallest value of the pixels in S(i(x,y)). This value is assigned to the corresponding pixel (x,y) in the filtered image. The Rank Difference filter image is then generated by performing the above operation on all pixels in the image. The behaviour of this filter is governed by a number of parameters: i) the location of pixel (x,y) in the region S(i(x,y)), ii) the shape and size of S(i(x,y)), and iii) the values of the upper and lower rank numbers u and I. First, the location of pixel (x,y) with respect to the region S(i(x,y)) determines the location of the resulting edge in the generated edge image with respect to those in the original image. This phenomena is illustrated in Figure 6.1. Pixel (x,y) can be located anywhere within or outside S(i(x,y)). This filter treats all locations in S(i(x,y)) equally since only the magnitudes of the intensities in S(i(x,y)) determine the outcome. For the generated edges to be well registered with respect to the original image, (x,y) should be located as close to the center of S(i(x,y)) as possible (Figure 6.1b). In this case, the magnitudes of the pixels in the vicinity surrounding the pixel (x,y) determines the fate of that pixel and hence, there is no pixel shift in the resulting edge. If the pixel (x,y) is away from the central point of the region S(i(x,y)) (i.e. outside the region), then an image shift occurs in the result (Figure 6.1c). In this instance, pixel (x,y) is replaced by the result of the Rank difference operation calculated over a region which does not include pixel (x,y). TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 139 Figure 6.1. Central pixel of rank difference filter. Second, the shape of region S(i(x,y)) determines which spatial orientation of edges are emphasized more than others. This region can assume any shape. With a square shaped region, the filter emphasizes those edges which are at an angle more than those edges which are perpendicular or parallel to the square edges. This phenomena is illustrated in Figure 6.2 where circular and rectangular test objects located at different angles from the horizontal axis are used. Hence, a square shaped region would be useful for detecting objects in the image which lie in one general orientation (Figure 6.2b). With a circular shaped region, edges in all direction will be equally emphasized (Figure 6.2c). TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 140 That is the detected edges of the circle and diagonal lines have similar widths. Since the chromosomes in our analysis can lie in any orientation, a circular region is preferred. (a) (b) (c) Figure 6.2. Shape of rank difference filter region. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 141 Figure 6.3. Size of rank difference filter region. The original image is in row (a). Operations with the 3x3, 5x5, and 7x7 filter regions are shown in rows (b-d), respectively. Column (i) shows the filter region. Column (ii) is the original image. Column (iii) and (iv) have added uniform noise of standard deviation of 0.5 and 5.0 added, respectively. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 142 The size of the region s(x,y) is found to determine the spatial resolution of the details/edges in the image. For example, in a 3x3 square pixel region, edges which span over 3 or less pixels wide can be detected. Small region sizes are preferred over larger ones because they require less time to process. With a smaller region size, however, it is more difficult to obtain a circular looking region in a rectangular pixel grid. For example, in a 3x3 pixel rectangular area, all 9 pixels in the square region would be used to approximate a circular shape. Hence, there is a compromise as to which filter size should be used in a given application. Figure 6.3 shows the results of applying different size filter regions to the test image. With noise free images as shown in Figure 6.3ii, the edges in the resulting image tend to be thick and similar in width to that of the filter's size. With noisy images, the larger size filters tends to do a better job of selecting the edges (Figure 6.3iii and iv). For example, even when uniform noise of standard deviation of 5.0 is added to the test image such that the objects cannot be visually identify amongst the noise in the image, our 7x7 Rank Difference filter is able to identify the edges (Figure 6.3d,iv) The upper and lower rank numbers, u and I, determines the filter's tolerance to noise in the image and the thickness of the edge in the resulting image. By choosing the highest number (v) for the upper rank and the lowest number (1) for the lower rank number, the filter generates ah image consisting of the largest local intensity differences (Figure 6.4). This special case of our Rank Difference filter becomes the filter used by Russ (1990). As the upper and lower rank numbers are moved away from their extreme values, the edges generated by the filter are thinner and less intense (Figure 6.4c). There is a point when the edges become too thin that they disappear resulting in a TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 143 discontinuous border around the object (Figure 6.4d). If the maximum and middle rank numbers are used as the upper and lower rank numbers in the Rank Difference filter, the outer edge of the object (increased by half the width of the filter) results (Figure 6.4e). Similarly, if the middle and lowest rank numbers are used, the inner edge of the object results (Figure 6.4f). This is true even in the presence of noise as shown in Figure 6.5. Figure 6.4. Effect of varying upper and lower rank numbers. The original image is shown in (a). A circular 7x7 Rank Difference is applied (as shown in Figure 6.3di). The maximum rank of 37 and minimum rank of 1 is used in (b). The upper and lower rank numbers are (29,9), (22, 16), (37,30) and (8,1) for (c), (d), (e), and (f), respectively. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 144 Figure 6.5. Effect of additive noise and varying upper and lower rank numbers. Uniform noise of standard deviation of 1.0 is added to the original image and the result is shown in (a). A 3x3 Rank Difference is applied. The maximum rank of 9 and minimum rank of 1 is used in (b). The upper and lower rank numbers are (7,3), (6,4), (9,5) and (5,1) for (c), (d), (e), and (f), respectively. 6.3. Comparison of Edge Detectors In this section, the performance of our Rank Difference filter is compared with other edge detectors: the difference or Laplacian of Gaussians and the Canny. A variety of different test images are used in the comparison. The first test image consists of rectangular and circular shaped objects. As TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 145 explained in section 6.2 (Figure 6.2), the rectangular objects are oriented at varying angles from the horizontal axis of the pixel grid such that the filter's performance to edge orientation can be tested. The results of the edge detection filters are shown in Figure 6.6. It can be seen that the DoG filter performs the best in defining the edges of the test object. Our Rank Difference filter performs similarly with edges which are slightly wider (2 to 3 pixels in width) than that of the DoG result. The Canny filter did not perform as well as the others in defining the corners of the square and rectangular objects. i \ ^ i—i i' \^ r P ^ ! ^,-1 j • •v_y i 5 Figure 6.6. Performance of edge filters on test object. The original image is shown in (a). The results of the 3x3 median and then Rank Difference filter, R(8,2), is shown in (b). The results of the DoG (a = 16) and Canny (a = 3) filters are shown in (c) and (d), respectively. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 146 Figure 6.7. Performance of edge detectors in presence of noise added to the test object. The top row represent the original test objects. From left to right, the images contains additive Gaussian noise of 0.5 and 1.0 standard deviation, and additive uniform noise of 1.0 and 7.0, respectively. The second row is the results of applying the median and our Rank Difference filter (7x7 circular R(31,7), 7x7 circular R(31,7), 3x3 square R(9,l), 7x7 circular (37,1). The third and fourth rows are the results of the DoG (a = 16, 16, 14, 16) and Canny (a = 3, 4, 4, 5) filters, respectively. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 147 Random noise is then added to this image to test the filter's tolerance level to noise. The results are shown in Figure 6.7. To filter some of the noise in the images, we first applied a similar size median filter before we applied our Rank Difference filter. The use of the median filter significantly reduce the number of false edges introduce by our filter. From the results shown in Figure 6.7, it can be seen that the DoG filter performs the worst in defining the edges of objects. The Canny filter perform the best on images with additive Gaussian noise. On the other hand, our Rank Difference filter performs the best in situations where uniform noise is added to the original image. The corners of the objects are generally more preserved using our filter. Our filter was also able to detect most of the edges in the noisy uniform image where it is even difficult for the eye to distinguish all the edges (Figure 6.7, top 2, right images). Images of real objects are then used to evaluate the performance of the edge filters. First, an image of peppers is used to compare the various edge filters (Figure 6.8). The DoG filter again has the worst performance. The location of the edges do not exactly correspond to the location of the edges in the image but are in the vicinity of the true edges. The Canny filter performs the best. Sharp, single pixel wide borders in the resulting image correspond closely to the true edges of the image. Our Rank Difference filter performs almost as good as the Canny filter. The edges are wider and more fuzzy and they also match the location of the true edges in the original image. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 148 Figure 6.8. Performance of edge detectors on the peppers image. The original image is shown in (a). The results of our 3x3 Rank Difference (R(9,l)), the DoG (CT =20), and the Canny (a = 1) filters are shown in (b), (c), and (d), respectively. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 149 The effects of the filters performance to additive Gaussian noise are then examined (Figure 6.9). A degradation in defining most of the edges and their true locations is evident compared to the filters' performance on the original image. More false edges (as a result of the added noise) are also detected. This is especially true for our Rank Difference filter. Again, the Canny filter perform the best in defining the major edges of the image. Our Rank Difference filter, however, does a better job of defining the true location of the borders as seen in the border between the long pepper and the light coloured pepper. It is also observed that more noise in the image typically requires more pre-filtering. That is the noise parameter needs to be increased in the DoG and Canny filters and the kernel size needs to be increased in our Rank Difference filter to obtain favourable segmentation results. The effects in the performance of the filters to additive uniform noise is also examined (Figure 6.10). A similar result is seen as that for the case of additive Gaussian noise. The exception is that our Rank Difference filter performs better with additive uniform noise than Gaussian noise. In this instance, more pronounced edge pixels correspond closely to the exact positions of the edges in the original image. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 150 Figure 6.9. Edge filter performance of peppers image with additive Gaussian noise. Gaussian noise of standard deviation of 0.5 is added to the original image to result in (a). The results of the median and our 7x7 circular Rank Difference (R(31,7)), the DoG (a =24), and the Canny [a = 3) filters are shown in (b), (c), and (d), respectively. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 151 Figure 6.10. Edge filter performance of peppers image with additive uniform noise. Uniform noise of standard deviation of 3 is added to the original image to result in (a). The results of the median and our 7x7 circular Rank Difference (R(37,l)), the DoG (a =25), and the Canny (a = 6) filters are shown in (b), (c), and (d), respectively. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 152 An image of metaphase chromosomes is then used to compare the edge filters (Figure 6.11). Also shown in this figure are the filters performance on the Gaussian and uniform noise added to the original image. Like the other examples, the Canny filter performs the best overall in terms of detecting the chromosome edges. Our Rank Difference filter gives thicker boundaries and also has better localization of the true edges of the chromosomes in the image. Our Rank Difference filter also perform well on images with additive uniform noise and not as well on images with additive Gaussian noise. With all three filters, as more noise is added to the image, more processing time is required. Larger filter sizes are generally required to smooth out the noise in the image. As the filter size increases, the processing time also increases in a relationship of approximately the square of the size of the filter. The operations in the Rank Difference filter are based mainly on the sorting of integers. On the other hand, both the DoG and Canny filter utilizes floating point arithmetic which takes much longer time to process. Although the Canny filter generates the best edges of the three edge detectors tried, the results are not good enough for our purpose as many touching chromosomes are not properly segmented and many edges are not connected (discontinuous object boundaries). Thus, we have to develop a better technique which overcomes these problems. Our technique contains a number of segmentation steps and includes the use of our Rank Difference filter in two of these steps. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 153 Figure 6.11. Edge filter performance on chromosome image. The original images are shown in the top left. The top right image have Gaussian noise with a standard deviation of 0.3 added. The second row are the corresponding results of the Rank Difference filter (3x3 square R(7,3) and 7x7 circular R(31,7)). The third row are the corresponding results of the DoG filter (a =5 and 11). Finally, the last row are the results of the Canny filter (a =1 and 3). TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 154 6.4. First Approximation to Edges: Thresholding In this and the following two sections, we describe the details of each of the three steps of our segmentation algorithm. The first step in our segmentation algorithm generates an approximate chromosome region or mask by determining which pixels belong to the background. These pixels are then excluded from further analysis. Thresholding is chosen for this step because it is simple to implement. Thresholding performs well on most isolated chromosomes but behaves poorly in segmenting touching and almost touching chromosomes (Figure 6.12f,g,h). The chromosome regions are faded into the background as the threshold level is increased. It can be seen that no one threshold can be used to segment all chromosomes, especially those which are touching. Even adaptive thresholding does not separate nearby chromosomes as the intensity levels of the overlapping and touching pixels are much similar in values to those pixels within the chromosomes. This similarity in intensity levels can be seen in Figure 6.12 where the original image is thresholded to generate thresholded images at various intensity levels. As seen in the thresholded images, no single threshold can be used to separate all chromosomes. As we are only interested in an approximation of the chromosome region in this step, we chose a conservative threshold level such that the background region constitute a large portion but more importantly no chromosome region is eliminated from the thresholding process. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 155 Figure 6.12. Chromosome image at various thresholds. The original image (a) is thresholded at 90 to 150 in increments of 10 grey levels to generate binary images (b) to (h). In the binary images, gray levels above the threshold are represented by white. Image (i) shows the result of setting all values of the original image which are below the threshold of 110 grey level to a value of 110. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 156 To obtain the value of the threshold, a histogram of intensity levels of the chromosome image is first generated. Traditional thresholding techniques generally search for the valley between the peaks in the histogram or for a location at a certain distance from a defined location of the maximum peak. These traditional techniques can not be used for the chromosome images because of the following reasons. First, even if smoothing is applied to remove the noise in the image, the histogram may contain more than 2 peaks. Second, the histogram may contain only one peak, as the chromosome pixels are spread almost uniformly over the intensity scale. Lastly, the maximum peak in the histogram do not always correspond to the mode of the background pixels, but can represent different chromosome regions (i.e. dark or bright bands in the chromosome). Hence, we use a different technique in obtaining the threshold level. In our method, we first define the range in which the chromosome and background pixels lie. Most artifact and extreme noise pixels are first rejected from this range. The minimum level is then the level where approximately 0.2% of the total number of pixels in the image have lower intensities. Similarly, the maximum level is the level where approximately 0.2% of the total number of pixels in the image have higher intensities. The selected minimum and maximum intensity levels then define the range of intensities corresponding to approximately 99.6% of the total pixels in the image. A threshold level, T, is then set at 3/5 of the intensity range from the minimum intensity level. This threshold level was found to work best for images acquired under a number of different situations. This level corresponds mostly to those background pixels which are near the borders of chromosomes. The pixels away from the chromosomes generally have intensity values below TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 157 this level. As there are relatively few pixels in the image at or near the threshold level, an error of approximately 10% in the exact location of the threshold level would result in approximately 1 pixel shift in the borders of the segmentation result (Figure 6.10c,d,e). This shift is not significant since the purpose of this step is to remove the pixels in the background which are away from the chromosomes. Those background pixels which are close to the chromosomes would be segmented in later steps. The first approximation image to the chromosome regions i2(x,y) is then the thresholded version of the chromosome image i(x,y) and is given by the following: Once the threshold value is found, all pixels in the image which are below the threshold are set to the threshold (background) level (Figure 6.10). Note that this first approximation of the segmentation may contain a few erroneous pixels both within and outside the chromosome. 6.5. Second Approximation to Edges: Texture Detection The second step in our segmentation algorithm refines the chromosome regions obtained from the first approximation. In this step, we search in the previously defined regions for texture information which is characteristic of the banding structures and the texture in the chromosomes of fluorescence images. This search process is divided into two parts. The first part finds points in the hix>y) i[x,y) if if i(x,y)>T i(x,y)<T (6-2) T TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 158 approximated region with high local intensities. The second part then expands these points into their neighbours to define the chromosome regions. To find the local high intensity points, we use the Average Difference filter described earlier in Section 5.3. This time, instead of using a 3x3 region, the value of the locally averaged image at (x,y) is the average value of the intensities over a 5x5 square neighbourhood region center about (x,y). This larger filter region smoothes out more noise and texture that are present in the chromosome images. We then impose a non-negative constraint on the difference image. The entire process is formulated as follows: •^2 2 (a) j[x,y) = il(x,y)-—- £ ^i^x - m,y - n) m=-2 n=-2 (b) if j(x,y) < 0, replace by j(x,y) = 0 (6-3) TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 159 Figure 6.13. Second approximation to edges: texture detection. The thresholded image is generated from the first approximation step and is shown in (a). A 5x5 average filter is then applied to result in (b). The average difference is then generated by subtracting image (b) from (a) to result in image (c). The background has a value of 0 . Darker regions represents negative values while brighter regions represent positive values. The positive values of image (c) is then shown in (d). TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 160 As a result of this operation, the background regions, which are close to the chromosomes, and most of the chromosome edges are set to 0 because its resulting intensity is below the local average of the first approximate segmentation (6.13). The isolated background regions where variation in intensity can result in positive difference values are not considered because these points have been rejected by thresholding in the previous step. Detected points (those set above 0) are found scattered further inside the chromosome as the image intensity difference begins to fluctuate between positive and negative values (Figure 6.13d). This fluctuation is due to the banding structures of the chromosome and the random noise in areas where the chromosome intensities are similar. These cluster of detected points are usually less than 2 pixels wide. Most of the points in-between touching chromosomes are also eliminated, since they have negative difference values. The second part in the texture detection algorithm is to expand and connect the detected texture points into chromosome regions. Morphological filters (dilation filter and combination of erosion and dilation filters) are the obvious choice of algorithms to use. For example, in a 3x3 dilation filter, the value of the pixel is set to 255 if a selected number of the pixel's neighbourhood has a value of 255. Otherwise, the pixel value is set to 0 as most of the neighbours have a value of 0. Hence, the gaps in between detected chromosome points which are mostly 2 or less pixels wide are filled and set to a value of 255. This filter also fills the region between touching chromosomes. If the erosion filter (the filter where pixel values are set to 0 if some of its neighbours are 0 and to 255 otherwise) is used prior to the dilation filter for the TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 161 purpose of separating chromosomes, the chromosome itself (clustered points mentioned above) may become separated by the process. We overcome this challenge of joining chromosome points and still be capable of separating touching chromosomes by using our Rank Difference filter which we developed and described earlier in Section 6.2. As the gaps to be filled are 2 or less pixels wide, a 3x3 neighbourhood region S(j(x,y)) is chosen. Larger neighbourhoods will increase the processing time with no significant improvement on the results. An upper and lower rank number of 7 and 1, respectively is chosen. The Rank Difference image is then binarized by setting all negative values to 0 and all others to 255. The resulting image i2(x,y) (Figure 6.14b) is the second approximation to the chromosome region and is given by: a) b) i2(x,y) = R71[S0(x,y)))] if i2(x,y) > 0, replace i2(x,y) with 255 otherwise replace i2(x,y) with 0 (6-4) TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 162 Figure 6.14. Rank difference of the average difference image. The positive portion of average difference image is shown in (a). The thresholded Rank Difference operation on (a) is shown in (b). This selection of rank numbers and binarizing dictate that if there are 3 or more pixels in the 3x3 neighbourhood which have a value greater than the lowest value in the region, the value of the pixel is set to 255. Otherwise, the pixel value is set to 0. Hence, this filter behaves like a selective dilation and erosion filter. It is identical to the dilation filter if the lowest value is always 0 (which is not the case in our images) and all other values are binarized to 255. Points with similar edge magnitudes in a neighbourhood are eroded and set to zero (e.g. edges of chromosomes and some areas in between chromosomes). Conversely, varying magnitude points are dilated and set to 255 (e.g. areas within the chromosome). The result is a more refined mask of the chromosome region. It resembles a skeleton outline of chromosomes as only the interior of TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 163 the chromosome is present and most of the edge pixels are eliminated. Almost all background regions and regions in between touching chromosomes are removed by this Rank Difference filter operation. 6.6. Third Approximation to Edges: Region Refinement and Labeling The final step in determining the chromosome region is to refine the results of the previous approximation to obtain a better estimate of the chromosome border and to label or distinguish the region of one chromosome from another. At this stage, we can use either the Difference of Gaussian, Canny or Rank Difference filter since for binary pictures (Figure 6.6), they all produce good results. We chose to use our Rank Difference filter because of the following reasons. First, the algorithm is already available within the program. Second, the algorithm uses only integer operations and is less complex and hence it is faster to compute. Last and most importantly, the Rank Difference filter gives thick edges at the appropriate locations such that some of the remaining touching chromosomes which are not segmented in previous approximations can be separated. In this instance, the Rank Difference Filter is used as an edge detector instead of a selective dilation filter. The purpose of this filter operation is to determine the borders of the chromosome regions. This filter operates over a 3x3 neighbourhood using 9 and 1 as the upper and lower rank numbers, respectively. This filter generates the boundary image b(x,y) which is defined as follows: b(x,y)=R9A[S(i2(x,y))\ (6-5) TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 164 Since the input image is binary, the resulting image is also binary (values of 0 and 255). The filter generates thick edges (approximately 3 pixels wide) around the boundaries of the chromosome. Pixels in between touching chromosomes which have 4 or less continuous pixels now become an edge pixel and are set to 255. A logical arithmetic operation is then employed to separate the touching chromosomes. In this operation, a new chromosome region, m(x,y) is generated based on the logical AND (•) of the previous chromosome approximation region, i2(x,y), with the logical NOT (—)of the newly calculated boundary image, b(x,y) as follows: m[x,y) = i2[x,y)»b[x,y) (6-6) The resulting image, m(x,y) then contains regions defined by the second approximation image i2(x,y) less those boundary pixels which lie in both i2(x,y) and b(x,y). Although the regions found are smaller than the actual regions of the chromosomes, they are mostly distinct and isolated from one another. Each object is next labeled such that each isolated region is given a distinct number. The size of the region of each labeled object is then increased such that it is representative of the size of the chromosomes. This increasing process is accomplished by dilating each labeled region twice using a 3x3 dilation filter. In this dilation process, the center pixel in the 3x3 region is set to 255 if any of the pixels in the region has a value of 255. Otherwise, the center pixel is set to 0. As different label numbers are used in the dilation process, regions which touch one another after the dilation are kept distinct with different label numbers. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 165 Figure 6.15. Third approximation to edges: region refinement and labeling. A 3x3 R(9,l) Rank Difference filter is applied to the result of the second approximation step (Figure 6.14b) and is shown in (a). An inverse of image (a) is generated and shown in (b). A logical AND is performed on the image from the second approximation step (Figure 6.14b) and the image in (b) to give image (c). Objects which touches the edges of the image are deleted. The image is then labeled and dilated. A border is then placed around each labeled chromosome to result in image (d). TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 166 6.7. Feature Extraction and Artifact Removal In all previous steps, no a-priori morphological information about the chromosome is used. In this step, we utilize features of the "chromosome" objects as parameters for rejecting artifacts from the detected objects from the third approximation image. Chromosome objects are generally more intense than artifact objects which have intensities similar to that of the background. Hence, the IFI over the defined area of the object can help in object discrimination. The area is calculated from the total number of pixels within the labeled object. The IFI is calculated by first summing the intensities of all pixels in the detected object and then subtracting the average background intensity multiplied by the number of pixels in the detected region. The background is calculated by determining the average intensity of the pixels which lie just outside the object region. Finally, the decision for rejection is then to eliminate objects which are dim and have an average of 5 or less gray levels above the background intensity level. 6.8. Associate Telomere with Chromosome To further refine both the telomere and chromosome regions, both the telomere and chromosome images are used. Other chromosome segmentation algorithms do not have our added advantage of having corresponding telomere images which can help in defining the ends of the chromosomes. As there are no corresponding reference points in the telomere and chromosome images for image registration, a probabilistic matching of the two images is first performed. The telomere image is shifted at 2 pixel steps in both the x and y TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 167 directions from the chromosome image. At each shift position, the number of telomeres that are located within the chromosome regions are determined. The shift number which corresponds to the largest number of telomeres found becomes the chosen shift value. Once the shift values have been found, the number of telomeres within each labeled chromosome can then be calculated. As there can only be four telomeres in a chromosome, telomeres or chromosomes which do not follow this rule are highlighted accordingly such that they can be easily seen during the manual editing and verification stage. Those telomeres which are more than 2 pixels away from any chromosome mask are treated as artifacts and are rejected from further analysis. An example of the results of overlaying the telomere borders onto the chromosome image and segmented results are shown in Figure 6.16. In this image, the details within the chromosomes are enhanced by contrast stretching such that they can be more readily seen. The background is also set to a gray colour (instead of the normal black) for visual enhancement of the details. It can be seen from the image that some of the telomeres are lying partially or just outside the chromosome border (e.g. chromosomes #6 and #7 in Figure 6.16). These telomeres are properly associated with the correct chromosome. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 168 Figure 6.16. Chromosome and telomere segmentation results. The borders of the chromosomes and telomeres are shown in black. The background of the original image is changed from black to gray and the intensities within the chromosome are contrast enhanced and inverted to help visualize the details within. Chromosomes which lie on the boundary of the image are not segmented in the algorithm but their telomere results which do not lie on the image boundary is shown. 6.9. Segmentation Performance Our segmentation method described is compared to the best of the different edge filters previously described, the Canny filter, for chromosome TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 169 segmentation. Sample chromosome segmentation results from the Canny and our method are shown in Figure 6.17. I (a) (b) (c) Figure 6.17. Comparison of our segmentation algorithm to the Canny filter. The original image is shown in (a). The results of the Canny filter is shown in (b). The results of our algorithm is shown in (c). In our algorithm, all chromosomes have closed boundaries and chromosomes which touch the edge of the image are discarded. It can be seen that our method gives superior results compared to the Canny filter. The borders in our method encompasses the outer edge of the chromosome. The larger mask would help in associating individual telomeres, which can fall outside the chromosome region, to the corresponding chromosome. The borders shown are continuous and they properly describe the region occupied by the chromosome. Even the background in between the arms of chromosomes are marked as not being part of the chromosome. In the Canny image however, some of the chromosomes (in particularly the smaller ones) do not have continuous boundaries. The edges within the chromosomes are also not properly joined such that these background regions are properly classified. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 170 It is difficult to accurately quantify the performance of our algorithm. There are a number of reasons for this. First, metaphase chromosomes are manually selected. Hence, the performance of our algorithms is largely dependent on which metaphases are selected for the analysis. Second, we do not have access to commercial chromosome segmentation packages from which we can compare our results. Lastly, there are cases where visually, it is difficult to distinguish which portion of a touching or an overlapping chromosome belongs to which of the two chromosome. Thus, we performed a qualitative evaluation of our segmentation algorithm. From the hundreds of metaphase chromosomes analyzed, there are a number of different types of errors observed. Most of these errors can be subsequently corrected by interactively editing the generated results. First, errors may arise when some touching or overlapping chromosomes are not properly separated. In these instances, the intensities at the borders resemble those within the chromosomes. Thus, we included a utility to force regions to split by drawing cutting lines in the image before the segmentation is performed. Second, chromosomes may be improperly separated. This usually occurs at the boundary of chromosomes where the observed intensities in the overlapping region are more intense than those observed in non-overlapping regions. This split can also occur in an arm of the chromosome where a substantially wide dark band is present. These split regions can be joined together during the interactive editing phase. Third, telomeres may be assigned to the wrong chromosome as they are more in the vicinity of another neighbouring chromosome or are outside the borders of the telomere search region. Finally, telomeres may not be properly paired and ordered within the TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 171 chromosome. With the editing features, proper pairing and labeling of each of the telomeres in the p and q arms of the chromosome can be made. Although there may be some errors in segmentation using our segmentation algorithm, there are considerable regions which are correctly segmented. As our segmentation algorithm is used to pre-process the acquired images before manual editing, substantial savings in time have resulted as a majority of chromosomes are correctly segmented. With the interactive editing features, nearly all chromosomes can be separated and analyzed. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 172 Chapter 7. Conclusion and Future Suggestions 7.1. Overview We have accomplished the objectives set forth for this project. Based on the work described in this thesis, our original hypothesis, which postulates that the length of individual telomeres in a cell can be determined from digital images of fluorescence in situ hybridization prepared cells, is accepted. This conclusion is based on our studies described in Section 7.3. In order to perform these studies, we had to develop the hardware system, algorithms, and software to allow for reliable measurements. Conventional systems using the Southern analysis can only determine the average length of telomeres of a population of cells, but can not determine the length of individual telomeres belonging to every chromosome in a cell. A summary of our work and the performance of our system are discussed in Section 7.2. Our system is currently being used in the Terry Fox Laboratory at the B.C. Cancer Research Centre and in the Netherlands on a routine basis to study the behaviour and role of telomeres in cells. Two such studies are described in Section 7.3. There are also plans to utilize our analysis system and extend the telomere studies into Germany and the United Kingdom. Improvements to the system and other areas of development are discussed in Section 7.4. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 173 7.2. System Performance With the telomere analysis system that we developed, the distribution of telomere lengths of the cells under study can be generated (Lansdorp et al., 1997). The advantage of our system is that significantly fewer cells (less than 30 cells) are required to obtain results compared to the conventional Southern analysis which requires analysis of approximately 100,000 cells. This makes it possible to carry biological studies when only a limited number of cells are available for analysis. In addition, telomere length studies can now be carried out on individual cells as well as individual chromosomes in every cell. No other method is currently available to determine the length of individual telomeres. Hence, no direct method for verifying the accuracy of our algorithms for telomere length measurements is available. For this reason, we resorted to indirect methods to validate our fluorescence measurements. For this purpose, we used objects of known fluorescence intensities which resemble telomeres. These objects included i) simulated objects of different shapes and sizes, ii) fluorescence beads of known size and relative fluorescence intensities, and iii) plasmids with known telomere insert lengths (which are typically an order of magnitude less in length than the telomeres in the cells). Our algorithm estimated the integrated fluorescence intensity (IFI) of simulated objects of varying shapes and sizes to within ±3%. The estimated mean IFI values correlated well (correlation coefficient of 0.99) with the size of the fluorescence beads and with the length of telomere insert in plasmids (Martens et al., 1997). The standard deviation in the estimation ranged from 2% for the lum beads to 13% for the 0.2um beads to 29% for the O.lum beads. The TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 174 standard deviation was larger for the smaller beads because it is more difficult to fabricate them. The standard deviation for telomere inserts in plasmids was around 20% of the mean estimated IFI value. This variance is most likely due to the variable efficiency of the hybridization procedure (binding of the probe). We observed a similar variation in hybridization on chromosomes after hybridization with probes specific for centromere repeat sequences of invariable length. Although the variation appears to be large, we observed that by averaging the results of 10 or more cells, a good indication of the telomere length on a particular chromosome arm in a population of cells can be obtained. That is, with results from 10 or more cells (>40 telomeres), the Wilcoxon-rank-sum test showed a significance level of less than 0.05 in differentiating between telomere lengths of chromosome groups. It is important to note that there are no other methods which can produce similar, let alone better results. In our analysis of each cell, we need to capture multi-focus plane images containing only telomere signals and a single image containing only chromosome signals. The images of telomeres are used to evaluate the telomere IFI values which give an estimate of the telomere lengths in the cell. The corresponding image of the chromosome is required to identify the regions occupied by the chromosomes. By identifying and classifying each chromosome in the image and associating it to its corresponding telomere, an estimate of the length of every telomere of each chromosome in the cell is obtained. From analyzing a number of cells, the telomere fluorescence distribution for each type of chromosome in the cell is realized. Our algorithm TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 175 for telomere segmentation, telomere fluorescence measurement, and chromosome segmentation for each metaphase chromosome sample takes less than 1 minute to perform on a 100MHz Pentium-based microcomputer. In accomplishing our major goal for the project, we have built and uniquely characterized a fluorescence microscopy imaging system for capturing images of metaphase chromosomes and multi-focus plane images of telomeres. The characterization of the system was then used to generate simulated images of the different shape and size test objects. In order to compare data acquired at different intervals in time and space, we have developed image processing techniques to compensate for the spatial and temporal distortions introduced by the acquisition system. The temporal distortions is a result of the decay in fluorescence intensity of the probe attached to the object. We have also developed novel techniques to analyze telomeres and chromosomes. In these developments, we have introduced algorithms to first segment multi-focus plane images of telomeres and then extract the integrated fluorescence intensity (IFI) values for each detected telomere. The IFI value is proportional to the telomere length. In addition, we have developed algorithms to segment chromosomes including those which are just touching. The segmentation results and the calculated telomere IFI values are then presented to the user for verification and editing. The automation of the telomere and chromosome extraction and IFI calculation process has simplified the user verification and editing process. On average, over 90% of the chromosomes are segmented properly. The success rate in segmentation is dependent on the metaphase sample which typically contains a TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 176 few overlapping chromosomes. The improperly segmented chromosomes can be corrected within 5 minutes by the user. In addition, the user does not need to perform the tedious task of defining the exact border for every telomere which is required to obtain a consistent telomere IFI value. 7.2.1. Imaging system We have successfully developed and built an imaging system for fluorescence microscopy. This system is capable of acquiring a large range of signal intensities (0.00001 to 100 lux) from very faint to strong signals. The system can also acquire images at different focus planes spaced at 0.1 um or more from each other. The critical elements of our system, which we paid special attention to during the component evaluation and selection process are the illumination source, fluorescence excitation and emission filters, objective lens and high resolution integration camera. The components were selected such that the quality of the captured image is sufficiently high and thus very little pre processing is required to correct or compensate for the aberrations in the images. The pre-processing and other functional algorithms which we determined to be essential to incorporate into the analysis included faulty pixel correction, flat-field correction, automatic selection of integration time and lookup table selection, and multi-focus plane image acquisition. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 177 7.2.2. System Characteristics We developed a new method for characterizing the response of our microscope imaging system. We based our method on Castleman's derivation and incorporated the contribution of the response function of the image detector. Our results are more representative of the system behaviour than those using Erhardt's method. Our theoretical PSF, however, is only an estimate and does not properly characterize the system response of our system. This is because the solution used to fix the chromosome onto the microscope slide (and to prevent photobleaching of the fluorescent probe) has a refractive index which is different from that of the cover slip for the slide and the immersion oil. As the location (z-direction) of the telomere varies within the solution, the extent of the blurring effects caused by the difference in refractive indices can not be predicted for each telomere. Hence, our theoretical PSF function is only used for generating simulated objects to test the telomere IFI algorithms. 7.2.3. Telomere IFI Value We first performed an analysis to gain an understanding of what the IFI value represents and how this value can be theoretically calculated. The IFI value was found to be proportional to the sum of all light intensities that originate from the object. We determined that this value needs not be summed over a 3D space but can be obtained from a single image plane as long as a sufficiently large region is used. The practicalities in the sample preparation and system such as close proximity of telomeres, quantization limit and errors TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 178 in the detector, noise in the image and problems in segmentation limited us to such an analysis. We thus resorted to analyze multiple focus plane images to estimate the telomere IFI value. We have developed a telomere and IFI quantification algorithm which can segment objects which are spaced at distances greater than 0.54um from each other in the system. We compared our IFI algorithm with simulated test objects as well as with experimental results using beads and plasmids where the relative fluorescence intensities are presumably known. Our results correlated well with the results of these experiments. We observed that better results can be obtained if the IFI value is chosen from the best focussed image (i.e. image with the highest IFI value) for each object in the set of multi-focus plane images. A 20% reduction in the IFI value from the best focus value can result if only a single image is used in the analysis since an image can contain objects which are ±0.2um in z-focus away from the best focus image. Alternatively, results within our acceptable accuracy limit can also be obtained if the sum of the IFI values from a stack of images spaced no more than 0.3um from each other is calculated. 7.2.4. Chromosome Segmentation We have developed a chromosome segmentation algorithm which is successful in determining the regions which belong to chromosomes. The resulting automation in the segmentation eliminates the lengthy time required by the user to manually define the borders of each chromosome in the image. Once the image is segmented, karyotyping (chromosome type identification) can TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 179 be performed. In conjunction with the telomere IFI results obtained earlier, the length of every telomere in each chromosome in the cell is obtained. For our segmentation algorithm, we first introduced the Rank Difference filter which is used in each of the two steps of our segmentation algorithm. This Rank Difference filter can act as an edge detector or a morphological filter. As an edge detector, our Rank Difference filter gives better localization of the edges than the Difference of Gaussians or Canny filters. It also performs better than the other filters on images with additive uniform noise. , For chromosome images, our segmentation algorithm outperforms other edge detectors in defining continuous regions and in separating touching chromosomes. Since our algorithm uses only integer operations, it performs faster than the Difference of Gaussians or Canny filters which use floating point arithmetic. As a vast majority of chromosomes (typically >90%) are properly separated by our algorithm, less user interaction is required to correct and edit those chromosomes which are improperly segmented by the algorithm. Hence, greater productivity in the analysis is obtained. 7.3. Current Biological Studies The system is currently being used on a daily basis to analyze telomeres at the Terry Fox Laboratory of the B.C. Cancer Research Centre. A dedicated system is used for acquiring the telomere and chromosome images. At least two other processing systems are being used to analyze the acquired images using our analysis software. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 180 Two biological studies have been completed to-date using our system. The first study (Zijlmans et al. 1997) investigates the telomere length distribution in mice. Previous studies in this area have shown that mice have long telomeres which do not appear to shorten as they age. This phenomena contradicts the concept that telomeres shorten with age. Using our analysis system, we have found that there is a large variation in the telomere lengths in mouse cells. We also observed that there are specific chromosomes in bone marrow and skin fibroblast cells in individual mice which have similar telomere lengths. We also observed the presence of very short telomeres which may be the critical link in limiting the cell replication process (aging process) in mice. The second study (Martens et al. 1997) investigates the telomere length distributions in human cells. The results of the analysis of one of the metaphase samples is shown in Figure 7.1. Our image analysis generates the borders of the telomeres from the telomere (figure labeled CY3) images and the borders of the chromosomes from the chromosome (figure labeled DAPI) image. The segmentation results are superimposed onto the processed chromosome image (figure labeled Image Analysis). The "X" chromosome is highlighted in the example to illustrate its telomere IFI values on the p and q chromosome arms. A Pseudo-Colour image is also generated from the chromosome and telomere image. This pseudo-Colour image is then used to generate the Karyogram image which sorts and identifies the different chromosome types (chromosomes #1 to #22 and chromosomes X and Y). The respective telomere IFI values of each chromosome arm is also generated (figure labeled with Telomere Karyogram) to facilitate in the biological analysis and interpretation. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 181 DAPI a * W C <% * t*. If/* . Image ^ Analysis, ^ Cy-3 0 Pseudo-colour Karyogram jbr S| HX 44 R« # CV* />/ p2 ql q2 X 324 266 166 197 93 33 Telomere Karyogram p-arm q-arm h A |L L L g " 3 4 4 5 5 6 6 7 7 8 8 9 9 10 10 11 11 12 112 2 3 : LJl 12 131314 14 15 15 16 16 17 17 1818 19 19 20 20 21 21 22 22 X Y """l|B'U"|F"'P Chromosomes Figure 7.1. Telomere lengths of individual chromosomes in a cell. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 182 In this second study, we found that for an individual, the telomere lengths in a specific chromosome from a certain tissues are very similar to those of other tissues. However, the telomere lengths do vary from an individual to another. We also noted from a study of 11 unrelated individuals that the telomeres on the arms of chromosome 17p are consistently among the shorter telomeres in the cell. Studies in this area may give an insight into why cancer cells frequently lose the ends of chromosome 17p. Using our telomere analysis system, other experiments and investigations can now be performed to study and determine the role telomeres play in the aging process and in patients with cancer or genetic disorders. 7.4. Future Suggestions and Applications Although the system is currently used on a daily basis, there are a number of improvements which can be made to the system. In terms of the hardware, a cooled integrating CCD camera would be useful in obtaining better quality images. With such a camera, less faulty pixels would be present. Hence, the accuracy of the IFI algorithm would be improved since it is no longer necessary to estimate the value of the faulty pixel by taking the average of its surroundings. Another hardware component which would benefit the system and improve the calculated IFI value is a more accurate and repeatable z-focussing mechanism. We recently acquired a piezo-electric motion controller which adjust the position of the objective lens. This system has less backlash and is more accurate than the existing mechanical motor attached to the focussing knob of the microscope. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 183 One method of improving the throughput of analyzing specimens is to automate the karyotyping process. This process is currently the most time consuming step as each chromosome image is manually sorted by a cytogenetics technician. The automated process developed should be such that the sorted chromosomes is easier to link to the corresponding telomere IFI value generated by the current program. Finally, research into segmenting telomeres in interphase nuclei (a circular shaped nucleus where chromosomes are clumped and are indistinguishable from one another) would significantly increase the number of samples which can be analyzed. In addition, the time to acquire images would be reduced as it is no longer necessary to only scan and select metaphase chromosomes from the slide. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 184 Chapter 8. Bibliography Agard D., Hiraoka Y., and Sedat J. (1989a) Three-dimensioal microscopy: image processing for high resolution subcellular imaging, in New Methods in Microscopy and Low Light Imaging, ed., James Wampler, SPIE 24-30. Agard D., Hiraoka Y., Shaw P., and Sedat J. (1989b) Fluorescence microscopy in three dimensions, in Methods in Cell Biology, ed. Lansing Taylor, 353-377. Agard D. (1984) Optical sectioning microscopy: cellular architecture in three dimensions. Ann. Rev. Biophys. Bioeng. 13:191-219. Allshire R.C., Gosden J.R., Cross S.H., Cranston G., Rout D., Sugawara N., Szostak J.W., Fantes P.A., and Hastie N.D. (1988). Telomeric repeat from T. thermophilia cross-hybridizes with human telomeres. Nature 332:656-659. Allsopp R., Vaziri H., Patterson C. Goldstein S. Younglai E., Futcher B., Greider C.W., and Harley C.B. (1992) Telomere length predicts replicative capacity of human fibroblasts. Proc. Natl. Acad. Sci. USA 89:10114-10118. Amadasun M., and King R.A. (1988) Low-level segmentation of multispectral images via agglomerative clustering of uniform neighbourhoods. Pattern Recognition, 21(4):261-268. Benbow R.M. [1992] Chromosome structures. Sci. Progress Oxford. 76:425-450. Bengtsson E., Eriksson O., Holmquist J., and Stenkvist B. (1979) Implementation and evaluation of a diode array scanner for digitizing microscopic images, in The Automation of Cancer Cytology and Cell Image Analysis, eds. Pressmann N.J., and Weid S.L., 269-289. Bertero M., Boccacci P., Brakenhoff G.J., Malfanti F., and van der Voort H.T.M. (1990) Three dimensional image restoration and super-resolution in fluorescence confocal microscopy. J. Microscopy 157(l):3-20. Bertero M., Boccacci P., Defrise M., de Mol C, and Pike E.R. (1989) Super-resolution in confocal scanning microscopy: II. the incoherent case. Inverse Problems 5:441-461. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 185 Bertero M., Brianzi P., and Pike E.R. (1987) Super-resolution in confocal scanning microscopy. Inverse Problems 3:195-212. Bright D.S. and Steel E.B. (1986) Brightfield image correction with various image processing tools, in Microbeam Analysis, eds. Romig A.D. Chambers W.F. San Francisco Press. 517-520. Canny J.F. (1986) A computational approach to edge detection. IEEE Trans, of Pattern Analysis and Machine Intelligence. 8(6):679-698. Canny J.F. (1983) Finding edges and lines in images. Technical Report TR-720, MIT, Cambridge, Massachusetts. Carlsson K. (1991) The influence of specimen refractive index, detector signal integration, and non-uniform scan speed on the imaging properties in confocal microscopy. Journal of Microscopy 163(2):167-178. Carrington W.A., Lynch R.M., Moore E.D.W., Isenberg G., Fogarty K.E., and Fay F.S. [1995] Science 268:1483-1487. Carrington W., Fogarty K., Lifschitz L., and Fay F. (1989) Three-dimensional imaging on confocal and wide-field microscopes, in The Handbook of Biological Confocal Microscopy, ed., James Pawley, IMR Press, Madison, 137-146. Carrington W., and Fogarty K. (1987) Three-dimensional distribution in living cells by deconvolution of optical sections using light microscopy, in Proc. 13th Annual North-East Bioengineering Conference, ed., K. Foster, 108-111. Castleman K.R. (1993) Color compensation for digitized FISH images. Bioimaging 1:159-165. Castleman K.R. (1979) Digital Image Processing. Prentice Hall, New York. de Lange T., Shiue L., Myers R.M., Cos D.R., Naylor S.L., Killery A.M., and Varmus H.E. (1990) Structure and variability of human chromosome ends. Mol. Cell. Biol. 10:518-527. Davis L.S. (1975) A survey of edge detection techniques. Computer Graphics and Image Processing, 4:248-270. de Lange T. [1995] Telomere Dynamics and Genome Instability in Human Cancer, in Telomeres, eds. Blackburn E. H., and Greider C. W. Cold Spring Harbour Laboratory Press, New York, 265-293. Egholm M., Buchardt O., Christensen L., Behrens C, Freier S., Driver D.A., Berg R.H., Kim S.K., Norden B. and Nielsen P.E. (1993) PNA hybridizes TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 186 to complimentary oligonucleotides obeying the Watson-Crick hydrogen bonding rules. Nature 365: 566-568. Erhardt A., Zinser G., Komitowski D., and Bille J. (1985) Reconstructing 3-D light microscopic images by digital image processing. Applied Optics 24(2): 194-200. Fu K.S., and Mui J.K. (1981) A survey of image segmentation. Pattern Recognition, 13:3-16. Garbay C, Chassery J.M., and Brugal G. (1986) An iterative region-growing process for cell image segmentation based on local color similarity and global shape criteria. Analytical and Quantitative Cytology and Histology, 8:25-34. Goldstein S.R., Hubin T., Rosenthal S., and Washburn C. (1990) A confocal video-rate laser-beam scanning reflected-light microscope with no moving parts. Journal of Microscopy 157(l):29-38. Goodman J.W. (1968) Introduction to Fourier Optics. McGraw-Hill, San Francisco. Graham M.D., and Norgren P.E. (1980) The Diff3 Analyzer: A Parallel/Serial Golay Image Processor, Real-Time Medical Image Processing, eds., Onoe M., Preston K. Jr., and Rosenfeld A., Plenum, New York. Harley C.B., Futcher A.B., and Greider C.W. (1990). Telomeres shorten during aging of human fibroblasts. Nature 345:458-460. Hastie N.D., Dempster M., Dunlop M.G., Thompson A.M., Green D.K., and Allshire R.C. (1990). Telomere reduction in human colorectal carcinoma and with ageing. Nature 346:866-868. Hell S., and Stelzer E.H.K. (1995) Lens aberrations in confocal fluorescence microscopy, in Handbook of Biological Confocal Microscopy, ed. Pawley J.B. Plenum Press, New York, 347-354. Hell S., Reiner G., Cremer C, and Stelzer E.H.K. (1993) Aberrations in confocal fluorescence microscopy induced by mismatches in refractive index. Journal of Microscopy 169(3):391-405. Hiraoka Y., Sedat J.W., and Agard D.A. (1990) Determination of three-dimensional imaging properties of a light microscope system: partial confocal behaviour in epifluorescence microscopy. Biophy. J. 57:325-333. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 187 Holmes T.J. and Liu Y.H. (1991) Acceleration of maximum-likelihood image restoration for fluorescence microscopy and other non-coherent imagery. Journal of Optical Society of America 8(6):893-907. Holmes T.J. and Liu Y.H. (1989a) Richardson-Lucy / maximum likelihood image restoration algorithm for fluorescence microscopy: further testing. Applied Optics 28(22):4930-4938. Holmes T.J. (1989b) Expectation-maximization restoration of band-limited, truncated point-process intensities with application in microscopy. Journal of Optical Society of America 6(7): 1006-1014. Hopkins H.H. (1955) The frequency response of a defocused optical system. Proc. R.Soc. London A231:91-103. Huang T.S. (1979) A fast two-dimensional median filtering algorithm. IEEE Trans. ASSP 27:13-18. Ingram M., and Preston K.Jr. (1970) Automatic Analysis of Blood Cells. Scientific America, 223:72-82. Inoue S. (1989) Foundations of confocal scanned imaging in light microscopy, in Confocal Microscopy Handbook, ed., James Pawley, 1-13. Inoue S. (1986) Video Microscopy. Plenum, New York. Jaggi B. Pontifex B., Swanson J., Poon S.S.S. [1993] Performance Evaluation of a 12-Bit, 8Mpel/s Digital Camera. SPIE Proceedings, Cameras Scanners, and Image Acquisition Systems, 1901:99-108. Jaggi B., Poon S.S.S., Pontifex B., Fengler J.J.P., Marquis J., and Palcic B. (1991) A quantitative microscope for image cytometry. SPIE Proceedings, Camera and Input Scanner Systems, 1448:89-97. Jaggi B., Poon S.S.S., Pontifex B., Fengler J.J.P., and Palcic B. (1990) Evaluation of a quantitative microscope for image cytometry in Advances in Analyt Cell Pathol., eds., G. Burger et al. Elsevier Science Publishers B.V., Amsterdam, 31-32. Jaggi B., Poon S.S.S., MacAulay C, and Palcic B. (1988) Imaging System for Morphometric Assessment of Conventionally and Fluorescently Stained Cells. Cytometry, 9:6. Jaggi B., Poon S., and Palcic B. (1987) Optical memory disks in image data base management for cytometry. Applied Optics 26(16):3325-3329. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 188 Jaggi B., and Palcic B. (1985) The Design and Development of an Optical Scanner for Cell Biology. IEEE Proceedings, Engineering in Medicine and Biology, 2:980-985. Jaggi B., Poon S.S.S., and Palcic B. (1986) Implementation and Evaluation of the Dmips Cell Analyzer. IEEE Proceedings, Engineering in Medicine and Biology, 3:906-911. Ji L. (1994) Fully automatic chromosome segmentation. Cytometry 17:196-208. Lansdorp P.M., Poon S., Chavez E., Dragowska V., Zijlmans M., Bryan T., Reddel R., Egholm M., Bacchetti S., and Martens U. (1997) Telomeres in the hematopoietic system. CIBA Foundation Symposium No:211. Telomeres and Telomerase. John-Wiley & Sons Ltd. Chichester U.K. (in press). Lansdorp P.M., Verwoerd N.P., van de Rijke F.M., Dragowska V., Little M.T., Dirks R.W., Raap A.K., and Tanke H.J. [1996] Heterogeneity in telomere length of human chromosomes. Human Molecular Genetics 5(5):685-691. Levy M.Z., Allsopp R.C., Futcher A.B., Greider C.W., and Harle C.B. [1992] Telomere End-replication Problem and Cell Aging. 225:954-960. Liedtke C.E., Gahn T., Kappei F., and Aeikens B. (1987) Segmentation of Microscopic cell scenes. Analytical and Quantitative Cytology and Histology, 9:197-211. Lockett S.J., and Herman B. (1994) Automatic detection of clustered, fluorescent-stained nuclei by digital image-based cytometry. Cytometry 17:1-12. Lockett S.J., O'Rand M., Rinehart C, Kaufman D., Jacobson K, and Herman B. (1990) Automated measurement of DNA ploidy using fluorescence microscopy in Optical Microscopy for Biology. Herman B. and Jacobson K. ed., Wiley-Liss, New York, 603-613. Lundblad V. and Wright W.E. (1996) Telomeres and Telomerase: Meeting Review. A simple picture becomes complex. Cell 87:369-375. MacAulay C, and Palcic B. (1988) A comparison of some quick and simple threshold selection methods for stained cells. Analytical and Quantitative Cytology, 10:134-138. Mantas J. (1987) Methodologies in pattern recognition and image analysis - a brief survey. Pattern Recognition, 20(1): 1-6. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 189 Marr, D. (1982) Vision, W.H. Freeman and Company, New York. Martens U.M., Zijlmans J.M.J.M., Poon S.S.S., Dragowska V., Yui J., Chavez E.A., Ward R.K., and Lansdorp P.M. (1997) Short telomeres on the p-arm of human chromosome 17. Nature Genetics (in press). Minsky M (1957) Microscopy Apparatus, U.S. Patent #3013467. Nederlof P.M., van der Flier S., Raap A.K., and Tanke H.J. (1992a) Quantification of inter- and intra-nuclear variation of fluorescence in situ hybridization signals. Cytometry 13:831-838. Nederlof P.M., van der Flier S., Vrolijk J., Tanke H.J., and Raap A.K. (1992b) Fluorescence ratio measurements of double-labeled probes for multiple in situ hybridization by digital imaging microscopy. Cytometry 13:839-845. Nederlof P.M., van der Flier S., Verwoerd N.P., Vrolijk J., Raap A.K., and Tanke H.J. (1992c) Quantification of fluorescence in situ hybridization signals by image cytometry. Cytometry 13:846-852. Nielsen P.E., Egholm M., Berg R.H. and Buchardt O. (1991) Sequence-selective recognition of DNA by strand displacement with a thymine-substituted polyamide. Science 244:1497-1500. Ohlander R., Price K., and Reddy R. (1978) Picture segmentation using a recursive region splitting method. Computer Graphics and Image Processing, 8:313-333. Palcic B., MacAulay C, Sclien S., Treurniet W., Tezcan H., and Anderson G. (1992) Comparison of three different methods for automated classification of cervical cells. Analytical Cellular Pathology. 4:429-441. Palcic B., Jaggi B. and Nordin J.A. (1987) Dynamic Microscope Image Processing Scanner. U. S. Patent #4,700,298. Philips J. and Lundsteen C. (1985) Semiautomated Chromosome Analysis: A Clinical Test, Clinical Genetics, 27:140-146. Pluta M. (1988) Advanced Light Microscopy, Principles and Basic Properties. Elsevier, New York. Pontifex B. Poon S., and Jaggi B. [1994] Design of a High Resolution Programmable Digital Camera. SPIE Proceedings, Image Acquisition and Scientific Imaging Systems, 2173:130-139. Poon S.S.S., and Hunter D.B. [1994] Electronic cameras to meet the needs of microscopy specialists. Advance Imaging. 9(7):64-67, 84. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 190 Poon S.S.S., Ward R.K., and Palcic B. (1993a) Automated image detection and segmentation in blood smears. Yearbook of Medical Informatics, 271-279. Poon S.S.S., Lockett S.J., and Ward R.K. (1993b) Characterization of a 3D microscope imaging system. SPIE Proceedings, Biomedical Image Processing and Biomedical Visualization, 1905:121-128. Poon S.S.S., Ward R.K., and Palcic B. (1992a) Feature extraction from three-dimensional images in quantitative microscopy. Microns and Microscopica Acta, 23(4):481-489. Poon S.S.S., Ward R.K., and Palcic B. (1992b) Analysis of three-dimensional images in quantitative microscopy. SPIE Proceedings, Biomedical Image Processing and Three-Dimensional Microscopy, 1660:178-185. Poon S.S.S., Ward R.K., and Palcic B. (1992c) Automated image detection and segmentation in blood smears. Cytometry 13:766-774. Poon S.S.S., and Palcic B. (1991) Importance of focus in image cytometry. Cytometry Supplement 5:40 (abstract: conference presentation). Poon S.S.S. (1989a) Algorithms for detecting and segmenting nucleated blood cells. Thesis disertation, Univ. of British Columbia. Poon S.S.S., Jaggi B., Spadinger I., and Palcic B. (1989b) Focussing methods used in the cell analyzer. IEEE Proceedings, Engineering in Medicine and Biology, 11(2):754-756. Poon S.S.S., Jaggi B., and Palcic B. (1987) Cell Recognition Algorithm for the Cell Analyzer. IEEE Proceedings, Engineering in Medicine and Biology, 3:1455-1456. Poulin N., Harrison A., and Palcic B. (1994) Quantitative precision of an automated image cytometric system for the measurement of DNA content and distribution in cells labeled with fluorescent nucleic acid stains. Cytometry 16:227-235. Poulin N., Poon S.S.S., Kandla G., and Palcic B. (1989) Automatic detection of metaphase chromosomes. IEEE Proceesings, Engineering in Medicine and Biology, 11(2):350-351. Preston K.Jr. (1976) Digital picture analysis in cytology, in Digital Picture Analysis, ed. Rosenfeld A., Springer Verlay, New York 209-293. Rigaut J.P. (1991a) A new technology is born: confocal image cytometry. Anal. Cell. Pathol. 3(2): 137-141. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 191 Rigaut J.P., and Vassy J. (1991b) High-resolution three-dimensional images from confocal scanning laser microscopy: quantitative study and mathematical correction of the effects from bleaching and fluorescence attenuation in depth. Cytometry (submitted). Rigaut J.P., Vassy J., Herlin P., Duigou F., Masson E., Briane D., Foucrier J., Carvajal-Gonzalez S., Downs A.M., and Mandard A.M. (1991c) Three-dimensional DNA image cytometry by confocal scanning laser microscopy in thick tissue blocks. Cytometry 12:511-524. Rosenfeld A., and Kak A.C. (1982) Digital Picture Processing. Second Edition, Volume 2. Academic Press, New York. Russ J.C. (1990) Computer Assisted Microscopy: The Measurement and Analysis of Images. Plenum Press, New York. Russ J.C. and Russ J.S. (1986) Image processing for the location and isolation of features, in Microbeam Analysis, eds., Romig A.D. Chambers W.F. San Francisco Press. 501. Sheppard C.J.R. (1988a) Super-resolution in confocal imaging. Optik 80(2):53-54. Sheppard C.J.R. (1988b) Depth of field in optical microscopy. J. Microscopy 149(l):73-75. Sheppard C.J.R. (1988c) Aberrations in high aperture conventional and confocal imaging systems. Applied Optics 27(22) :4782-4786. Sheppard C.J.R. and Cogswell C.J. (1991) Effects of aberrating layers and tube length on confocal imaging properties. Optik 87:34-38. Shoemaker R.L., Bartels P.H., Hillman D.W., Jones J., Kessler D., Shack R.V., and Vukobratovich D. (1982) An ultrafast laser scanner microscope for digital image analysis. IEEE Transactions on Biomedical Engineering, 29(2):82-91. Spadinger I., Poon S.S.S., and Palcic B. (1990) Effect of focus on cell detection and recognition by the cell analyzer. Cytometry 11:460-467. Spadinger I., Poon S.S.S. and Palcic B. (1989) Automated detection and recognition of live cells in tissue culture using image cytometry. Cytometry 10:375-381. Stokseth P.A. (1969) Properties of a defocused optical system. J. Optical Society of America 59(10): 1314-1321. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 192 Tatian B. (1965) Method for obtaining the transfer function from the edge response function. J. Optical Society of America 55(8): 1014-1019. Tucker J.H., Husain O.A.N., Watts K., Farrow S., Bayley R., and Stark M.H. (1987) Automated densitometry of cell populations in a continuous-motion imaging cell scanner. Applied Optics, 26(10):3315-3324. Tucker J.H. (1979) An image analysis system for cervical cytology automation using nuclear DNA content. Journal of Histochemistry and Cytometry 21(l):613-620. Umesh R.M. (1988) A technique for cluster formation. Pattern Recognition, 21(4):393-400. Van der Voort H.T.M., Brankenhoff G.J., and Janssen G.C.A.M. (1988) Determination of the 3-dimensional optical properties of a confocal scanning laser microscope Optik 78(2):48-53. Vaziri H, Dragowska W., Allsopp, R.C., Thomas T.E., Harley C.B., and Lansdorp P. M. [1994] Evidence for a mitotic clock in human hematopoietic stem cells: loss of telomeric DNA with age. Proc. Natl. Acad. Sci. USA, Cell Biology, 91:9857-9860. Visser T.D., Brakenhoff G.J. and Groen F.C.A. (1991) The one-point fluorescence response in confocal microscopy. Optik 87:39-40. Visser T.D., Oud J.L., and Brakenhoff G.J. (1992) Refractive index and axial distance measurements in 3-D microscopy. Optik 90:17-19. Vrolijk J., Sloos W.C.R., Verwoerd N.P., and Tanke H.J. (1994) Applicability of a noncooled video-rate CCD camera for detection of fluorescence in situ hybridization signals. Cytometry 15:2-11. Walton W.H. (1952) Automatic counting of microscope particles. Nature 169: Weinstein M., and Castleman K.R. (1971) Reconstructing 3-D specimens from 2-D section images. Proc. Soc. Photo-Opt. Instrum. Eng. 26:131-138. Wilson E., Larosche T., and Gasser M. (1993) Telomeres and the functional architecture of the nucleus. Trends Cell Biol. 3:128-134. Wilson T. and Sheppard C.J.R. (1984) Theory and practice of scanning optical microscopy. Academic Press, London. Wilson T., and Hewlett S.J. (1990) Imaging in scanning microscopes with slit-shaped detectors. J. Microscopy 160(2): 115-139. Young J.Z., and Roberts F. (1951) A flying spot microscope. Nature 167:231. TELOMERE LENGTH MEASUREMENTS USING FLUORESCENCE MICROSCOPY 193 Young T.Y., and Fu K.S., Eds. (1986) Handbook of Pattern Recognition and Image Processing. Academic Press, London. Zijlmans J.M.J.M., Martens U.M., Poon S.S.S., Raap A.K., Tanke H.J., Ward R.K., and Lansdorp P.M. (1997) Telomeres in the mouse have large inter-chromosomal variations in the number of T2AG3 repeats. Proc. of National Academy of Science USA, Genetics 94:7423-7428. 

Cite

Citation Scheme:

    

Usage Statistics

Country Views Downloads
China 25 26
United States 14 1
Russia 3 0
United Kingdom 2 0
Iran 1 1
Canada 1 0
Japan 1 0
City Views Downloads
Beijing 21 0
Ashburn 6 0
Unknown 4 1
Shenzhen 4 26
Mountain View 3 0
Plano 2 0
Ryazan 2 0
Sunnyvale 1 0
Wilmington 1 0
Vancouver 1 0
Seattle 1 0
Tokyo 1 0

{[{ mDataHeader[type] }]} {[{ month[type] }]} {[{ tData[type] }]}
Download Stats

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0064855/manifest

Comment

Related Items