Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Evaluation of automated hippocampal segmentation software in a clinically heterogeneous sample MacRae, Cassie Brieana 2013

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2013_fall_macrae_cassie.pdf [ 7.94MB ]
Metadata
JSON: 24-1.0074082.json
JSON-LD: 24-1.0074082-ld.json
RDF/XML (Pretty): 24-1.0074082-rdf.xml
RDF/JSON: 24-1.0074082-rdf.json
Turtle: 24-1.0074082-turtle.txt
N-Triples: 24-1.0074082-rdf-ntriples.txt
Original Record: 24-1.0074082-source.json
Full Text
24-1.0074082-fulltext.txt
Citation
24-1.0074082.ris

Full Text

EVALUATION OF AUTOMATED HIPPOCAMPAL SEGMENTATION SOFTWARE IN A CLINICALLY HETEROGENEOUS SAMPLE  by Cassie Brieana MacRae  B.Sc., The University of British Columbia, 2011  A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF  MASTER OF SCIENCE in THE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES (Neuroscience)  THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver)  August 2013  ? Cassie B. MacRae, 2013   ii Abstract Accurate, efficient processing of magnetic resonance images (MRI) offers potential value for both basic and clinical neuroscience; however, the current gold standard method of doing so is manual segmentation, a labor-intensive process with inherent variability. Limitations associated with manual segmentation have motivated the development of automated software programs for structural MRI processing and volumetric analysis of subcortical brain structures. The three most commonly used platforms are FSL, FreeSurfer, and SPM. As each platform uses a unique mathematical algorithm for image analysis, it is necessary to determine which program offers optimal subcortical segmentation, particularly for high field-strength imaging. The current study compared automated volumetric analysis of the human hippocampus with manual segmentation values in a clinically complex population in order to determine which method provides the most robust results in the presence of motion artifact, a common problem in MR, and visible anatomic anomalies of the hippocampi. MRI data included in the current study was obtained from residents of Vancouver B.C?s Downtown Eastside (DTES) social housing projects, the majority of whom experience co-occurring addiction/polysubstance abuse and severe mental illness. A large number of these MRI scans are characterized by abnormal hippocampal morphology, perhaps resultant of neuropsychiatric illness, addiction, infection, or interactions of these factors. Intraclass correlations (ICC) were used to determine which set of automated hippocampal volumes best correlated with manual measures in motion artifact-free MRI data with and without bilateral abnormal hippocampal cavities. FreeSurfer volumes were most correlated with manual measurements regardless of whether MRI data included hippocampal cavities or not (ICC=0.88, no cavities; 0.48 with cavities), followed by FSL/FIRST (ICC=0.64, no cavities; ICC=0.43, with cavities), and SPM8.0/AAL (ICC=0.039, with cavities; ICC=0.035, no cavities). FSL/FIRST was the most sensitive to severe motion artifact (ICC=0.62), followed by FreeSurfer 5.1 (ICC=0.85), manual segmentation (ICC=0.96), and SPM8.0/AAL (ICC=0.99). ICCs for FreeSurfer 5.1 vs. FSL/FIRST were 0.87 (no motion artifact, no hippocampal cavities), 0.87 (no motion artifact, bilateral hippocampal spaces), and 0.64 (MRI scans with severe motion artifact). These data suggest automated segmentation methods are sensitive to MRI data with compromised image quality and deviations in normal hippocampal morphology.      iii Preface This investigation was approved by The University of British Columbia (Ethics Certificate No. H08-00521), and Vancouver Coastal Health (Ethics Certificate No. V08-0265).   The identification and design of this research project was fulfilled by myself, Cassie B. MacRae, along with the guidance, leadership, and intellectual contribution of my graduate studies supervisor, Dr. Donna J. Lang, and mentors, Dr. Geoffrey N. Smith and Dr. William G. Honer.      I, Cassie B. MacRae, performed all of the manual hippocampal segmentations, hippocampal cavity measurements, and data analyses discussed in the current article. Automated segmentations with FreeSurfer 5.1, FSL/FIRST, and SPM8.0/AAL, and subjective ranking of MRI scan motion artifact were performed by an experienced biomedical imaging laboratory technician, Wayne Su. Subjective evaluation of manual and FreeSurfer 5.1 segmentation was performed by trained neuropsychiatrist/neurologist Dr. William J. Panenka.    Participant information including demographics, drug use, psychiatric diagnosis, and magnetic resonance imaging data was collected as part of the Hotel Study, a larger ongoing investigation of addiction, viral infection, and psychosis in Vancouver?s Downtown Eastside neighborhoods. Substance dependence frequencies reported in the current study were obtained with permission from a student colleague, Andrea Jones, B.Sc.    iv Table of Contents  Abstract .......................................................................................................................................... ii	 ?Preface ........................................................................................................................................... iii	 ?Table of Contents ......................................................................................................................... iv	 ?List of Tables ............................................................................................................................... vii	 ?List of Figures ............................................................................................................................... ix	 ?List of Abbreviations ................................................................................................................... xi	 ?Acknowledgements ..................................................................................................................... xii	 ?Dedication ................................................................................................................................... xiii	 ?Chapter 1: Introduction ................................................................................................................1	 ?1.1	 ? The Hippocampus: General Structure, Function, and Clinical Relevance ........................ 1	 ?1.2	 ? In Vivo Evaluation of Hippocampal Morphology: Volumetric Analysis of MRI Data ..... 2	 ?1.3	 ? Rationale for Validation of Automated Segmentation Software Using MRI Data Collected from a Subset of DTES Volunteers ............................................................................ 6	 ?1.4	 ? Overview of T1 Weighted MRI Data Acquisition ............................................................. 8	 ?1.5	 ? Automated Image Parcellation: Overview of Algorithms Used by FreeSurfer 5.1, FSL/FIRST, and SPM8.0/AAL .................................................................................................. 9	 ?1.6	 ? Overview of Specific Hypothesis Evaluated in The Current Study ................................ 13	 ?Chapter 2: Body of Thesis ...........................................................................................................15	 ?2.1	 ? Research Methods ............................................................................................................ 15	 ?2.1.1	 ? Study Participant Recruitment/Inclusion Criteria ..................................................... 15	 ?2.1.2	 ? Overview of Hotel Study Format .............................................................................. 15	 ?  v 2.1.3	 ? Current Sample Demographics ................................................................................. 15	 ?2.1.4	 ? Participant Substance Use and Dependence ............................................................. 16	 ?2.1.5	 ? Participant Neurological/Psychiatric Diagnoses ....................................................... 18	 ?2.2	 ? MRI Data ......................................................................................................................... 19	 ?2.2.1	 ? Overview of MRI Datasets ....................................................................................... 19	 ?2.2.2	 ? Image Acquisition ..................................................................................................... 19	 ?2.2.3	 ? Image Segmentation .................................................................................................. 20	 ?2.2.4	 ? Manual Segmentation ............................................................................................... 20	 ?2.2.5	 ? Manual Segmentation of MRI Data with Severe Motion Artifact ............................ 21	 ?2.2.6	 ? Automated Segmentation .......................................................................................... 21	 ?2.3	 ? Study Design, Methods, and Statistical Analyses ............................................................ 22	 ?2.3.1	 ? Statistical Analyses ................................................................................................... 22	 ?2.3.2	 ? Design 1: Subjective Comparison of Manual vs. Automated Segmentation ............ 22	 ?2.3.3	 ? Design 2: Comparison of Segmentation Methods Using MRI Data with Bilateral Abnormal Hippocampal Cavities .......................................................................................... 24	 ?2.3.4	 ? Comparison of Segmentation Methods Using MRI Data with Motion Artifact ....... 25	 ?2.3.5	 ? Comparison of FreeSurfer 5.1 and FSL/FIRST Methods: ........................................ 27	 ?Chapter 3: Results ........................................................................................................................29	 ?3.1	 ? Concurrent Validation of Manually Segmented ROIs ..................................................... 29	 ?3.2	 ? Comparison of Manual vs. Automated Segmentation Methods ...................................... 29	 ?Intraclass correlations are summarized in Table 7. ................................................................... 30	 ?3.3	 ? Comparison of Manual vs. Automated Segmentation Methods Using MRI Data with Bilateral Abnormal Hippocampal Cavities ............................................................................... 30	 ?  vi 3.4	 ? Comparison of Manual vs. Automated Segmentation Methods Using MRI Data with Motion Artifact ......................................................................................................................... 32	 ?3.5	 ? Comparison of FreeSurfer 5.1 vs. FSL/FIRST Segmentation Methods .......................... 34	 ?Chapter 4: Discussion ..................................................................................................................35	 ?4.1	 ? Overview of Study Methodology ..................................................................................... 35	 ?4.2	 ? Manual Segmentation as the Gold Standard Comparison for Volumetric Analysis of the Hippocampus ............................................................................................................................ 35	 ?4.3	 ? Automated vs. Manual Segmentation Methods: Evaluation of FreeSurfer 5.1 and FSL/FIRST ................................................................................................................................ 39	 ?4.4	 ? Hippocampal Volumes Obtained with SPM8.0/AAL are Weakly Associated with Manual Segmentation ............................................................................................................... 40	 ?4.5	 ? Manual vs. Automated Segmentation Methods: The Effect of Bilateral Hippocampal Cavities ..................................................................................................................................... 41	 ?4.6	 ? The Effect of Motion Artifact on Scan Rescan Reliability: Evaluation of Automated and Manual Methods ....................................................................................................................... 43	 ?4.7	 ? Evaluation of the Agreement Between FreeSurfer and FSL/FIRST ............................... 45	 ?4.8	 ? Discrepancy in Volumetric Analysis of the Hippocampus: Manual vs. Automated Segmentation, and Postmortem vs. MRI-Derived Measurement ............................................. 46	 ?4.9	 ? Insights, Drawbacks, and Future Directions .................................................................... 49	 ?Bibliography .................................................................................................................................51	 ?   vii List of Tables  Table 1. Summary of processing pipelines for subcortical segmentation with FreeSurfer 5.1 (automated hippocampal subfield segmentation tool), FSL/FIRST, and SPM8.0/AAL. ............. 10	 ?Table 2. Baseline Sample Demographics. Age, gender, ethnicity, self-reported years using drugs in the DTES, years of high-school education, and high-school graduation status of sample participants. Mean (M); Standard Deviation (SD); Range (R) ..................................................... 16	 ?Table 3. Baseline Participant Substance Use and Dependence Disorders. 1) Proportion of sample participants dependent on 1 or more substance(s), 2) Frequency of dependence disorders, 3) Tobacco use patterns, and 3) Proportion of substances tested positive in urine drug screens. ..... 17	 ?Table 4. Baseline Frequency of Neurological/Psychiatric Diagnoses. 1) Proportion of sample participants with BECED-determined DSM-IV psychiatric diagnosis. ....................................... 18	 ?Table 5. Summary of intraclass correlation coefficients calculated to compare values generated by manual segmentation, FreeSurfer 5.1, FSL/FIRST, and SPM8/0/AAL. Each method was compared against itself using scan data with differing degrees of motion artifact. ...................... 27	 ?Table 6. Intraclass correlation comparisons between FreeSurfer 5.1 and FSL/FIRST. ............... 28	 ?Table 7. Intraclass correlation coefficients (N=25 total hippocampal volumes per method): ...... 30	 ?Table 8. Post-hoc paired t-tests (N=25 hippocampal volumes per method). ................................ 30	 ?Table 9. Intraclass correlation coefficients (N=19 total hippocampal volumes per method): ...... 31	 ?Table 10. Post-hoc paired t-tests (N=19 hippocampal volumes per method). .............................. 32	 ?Table 11. Scan-rescan intraclass correlation coefficients. A: Comparison of scan-rescan reliability of segmentation methods using MRI data with zero vs. severe motion artifact; B:   viii Scan-rescan reliability of segmentation methods using MRI data with zero vs. mild-moderate motion artifact. .............................................................................................................................. 32	 ?Table 12. Post-hoc paired t-tests: comparison of segmentation methods using MRI data with mild-moderate motion artifact. ..................................................................................................... 33	 ?Table 13. Post-hoc paired t-tests: comparison of segmentation methods using MRI data with severe motion artifact. ................................................................................................................... 34	 ?Table 14. Intraclass correlation comparisons between FreeSurfer 5.1 and FSL/FIRST. ............. 34	 ?   ix List of Figures    Figure 1. Magnetic resonance image of the hippocampal formation (selected in red: a) coronal view, b) sagittal view of the left hippocampus, and c) axial (transverse view). ............................. 1	 ?Figure 2. Screenshot of the Mango GUI for medical image editing. The Mango toolbar is shown in the upper left-hand corner of the GUI. ....................................................................................... 4	 ?Figure 3. Software emblems for each of the freely available automated segmentation platforms evaluated in the current study. Programs can be downloaded from their respective websites: FreeSurfer: https://surfer.nmr.mgh.harvard.edu; FSL: http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/; SPM: http://www.fil.ion.ucl.ac.uk/spm/. .................................................................................................. 5	 ?Figure 4. Breakdown of MRI Datasets. ........................................................................................ 19	 ?Figure 5. Example of GUI set-up for concurrent validation of manual vs. FreeSurfer 5.1 generated ROIs. Figure depicts a desktop screenshot of 2 GUI windows displaying the same MR image. Manual and FreeSurfer 5.1 ROIs are superimposed over the hippocampal formation. Images were presented blindly and side-by-side. In this case, manually delineated ROIs are presented in the left GUI, and FreeSurfer 5.1 ROIs are presented on the right. ........................... 23	 ?Figure 6. MRI scans from 3 Hotel Study participants with hippocampal cavities (red arrows indicate abnormality). a) sagittal view of left hippocampus, b) coronal view of left hippocampus, and c) coronal view of right hippocampus. ................................................................................... 25	 ?Figure 7. Example of 3 T1-weighted MR images collected from the same Hotel Study participant with a range of motion artifacts. A) Zero motion artifact: gray-to-white matter boundaries are clear, scan resolution and pixel contrast is high, B) mild-moderate motion artifact: slight blurring   x of gray-to-white matter boundaries especially in anterior cortical regions (red arrow), pixel contrast remains high, and C), severe motion artifact: significant blurring throughout the image, resolution is poor, separate anatomical constituents are unclear in places, and pixel contrast is relatively low (red arrow indicates greyish blurring of typically black-colored ventricular areas)........................................................................................................................................................ 26	 ?Figure 8. Mean total hippocampal volumes. Segmentation methods are presented along the x-axis, and volume (mm3) is presented along the y-axis. ................................................................ 29	 ?Figure 9. Mean total hippocampal volumes. Segmentation methods are presented along the x-axis, and volume (mm3) is presented along the y-axis. ................................................................. 31	 ?Figure 10. Sagittal view of left hippocampus: a) 1.5T MR image, b) 3.0T MR image. The white matter tract separating the anterior amygdala is clearly visible on MR images collected at 3.0T (b). This tract cannot easily be distinguished on MR images collected at 1.5T (a). ..................... 36	 ?Figure 11. Sagittal view of identical MRI images collected from a Hotel Study participant with hippocampal atrophy. The ROI in the left image was segmented by manual tracing, and the ROI in the right image was segmented with FreeSurfer 5.1. Red arrows indicate areas where ROI extends beyond hippocampal boundaries. .................................................................................... 38	 ?   xi List of Abbreviations  ASM/AAM: Active Shape/Active Appearance Models CSF: Cerebral spinal fluid DTES: Downtown Eastside fMRI: Functional magnetic resonance imaging  GUI: Graphical user interface  ICBM: International Consortium for Brain Mapping ICC: Intraclass correlation coefficient  M: Mean  MNI: Montreal Neurological Institute MRI: Magnetic resonance imaging R: Range RF: Radiofrequency energy ROI: Region of interest SD: Standard deviation SOR: Single-Room Occupancy  T: Tesla   xii Acknowledgements  I offer my enduring gratitude to the faculty, staff and my fellow students at UBC. I owe particular thanks to Dr. Donna J. Lang, Dr. Geoffrey N. Smith, and Dr. William G. Honer, whose guidance and intellectual contributions have helped me succeed in bringing this project to completion.  Thanks to Mitacs Accelerate, for helping to fund this project, and special thanks are also owed to my parents, who have continued to support me through countless years of education.   xiii Dedication  Dr. Donna J. Lang ? for her faith in my abilities, her coherent responses to my endless questions, and for instilling in me an appreciation of the art of brevity.      1 Chapter 1: Introduction  1.1 The Hippocampus: General Structure, Function, and Clinical Relevance   The hippocampus, or more broadly, the hippocampal formation (Figure 1) is a critical subcortical gray matter formation in the brains of humans and other advanced vertebrates. In the human brain, the hippocampal formation is a bilaterally symmetrical sheath of cortex folded into the medial aspect of the temporal lobes in the floor of the inferior horn of the lateral ventricular system (Krebs, Weinberg, & Akesson, 2011). The name ?hippocampus? comes from the Latin roots ?coiled? and ?horse,? as early preparations of the structure were likened to the shape of a seahorse (Duvernoy, 2005).  It has long been established that hippocampal function is critical for memory consolidation and spatial navigation. Cells within the dentate gyrus are primary mediators of long-term potentiation, a principal molecular mechanism of memory storage (Bliss & L?mo, 1973). In the famous case study of anterograde amnesia (Scoville & Milner, 1957), bilateral temporal lobe lesions abolished Patient H.M.?s ability to form lasting declarative memories. Similar lesions in previously trained rats rendered them unable to navigate a Morris Water Maze (a test of spatial memory), even under circumstances where obvious spatial cues were provided (Morris, Garrud, Rawlins, & O'Keefe, 1982).   Figure 1. Magnetic resonance image of the hippocampal formation (selected in red: a) coronal view, b) sagittal view of the left hippocampus, and c) axial (transverse view).      2 The hippocampus is a region of considerable interest in psychiatric research. Magnetic resonance imaging (MRI) studies have shown that hippocampal morphology is affected by ageing and dementia (Ball, 1977), Alzheimer?s disease (Convit et al., 1997), Huntington?s disease (Spargo, Everall, & Lantos, 1993), epilepsy (Mathern et al., 1996), major depression (Bremner et al., 2000), childhood sexual abuse and posttraumatic stress disorder (Bremner et al., 2003), and other conditions. Drugs of abuse have also been shown to affect hippocampal morphology. Hippocampal volume reduction has been associated with early and late onset alcoholism (Agartz, Momenan, Rawlings, Kerich, & Hommer, 1999; Laakso et al., 2000), and chronic ecstasy (3,4 methylenedioxymethamphetamine) abuse (den Hollander et al., 2012).    Reports of hippocampal abnormalities are frequent in studies of complex psychiatric illnesses such as schizophrenia and related psychotic spectrum disorders. Volume reductions in the hippocampi have been cited in chronic and first-episode psychosis patients (Velakoulis et al., 1999), as well as in first-degree non-psychotic relatives (Seidman et al., 1999; Seidman et al., 2002). Longitudinal studies of hippocampal volume reduction in chronic and first-episode patients have not reported changes in volume over time, suggesting that abnormalities in these areas are stable fixtures of illness (Lieberman et al., 2001). Taken together, these findings suggest a central role of the hippocampus in genetic and neurodevelopmental models of psychotic illness.  Considering the extensive list of psychiatric conditions linked with temporal lobe abnormalities, reliable in vivo assessment of hippocampal structure is of value to both basic and clinical neurosciences. The ability to detect changes in hippocampal structure in a number of illnesses over time offers practical insight into mechanisms governing psychosocial and biopsychological sequelae. It is also possible that gross changes in hippocampal morphology indicate pre-morbid states of illness. Under these circumstances, routine MRI evaluation of hippocampal structure would aid in early detection of psychiatric disease.   1.2 In Vivo Evaluation of Hippocampal Morphology: Volumetric Analysis of MRI Data  Volume reduction is the most commonly and consistently reported finding in structural MRI studies of the hippocampus in psychiatric research. Reports of hippocampal volume loss in   3 schizophrenia and related psychotic spectrum disorders, in particular, are especially frequent (Harrison, 2004; Koolschijn et al., 2009). Currently, the gold standard approach to volumetric assessment of subcortical brain structures with complex gray-to-white matter boundaries is manual segmentation. Typically, experienced technicians perform manual segmentation in graphical user interface (GUI) programs capable of displaying the brain image in coronal, sagittal, and axial (longitudinal) fields of view. Anatomical regions of interest (ROI) are traced directly onto the MR image with reference to stereotypical landmarks and spatial relationships between brain structures.  One example of a non-commercial GUI program for viewing, editing, and analyzing volumetric medical images is Mango (Multi-Image Analysis GUI; University of Texas) (Figure 2). Mango provides several tools for selecting and editing ROIs. Once a neuroimage is uploaded into Mango, the ?trace? tool is used to select the ROI, and the ?edit? tool is used to erode or expand the region. Technicians may trace and then edit the ROI in all planes of view until its boundaries best match the 3 dimensional shape of the brain structure being measured. Once the rater has selected the entire ROI, Mango will calculate the ROI volume by adding the pixels traced on every slice of the volumetric image together. ROI volumes are reported in mm3 units. ROIs can be saved in a 3 dimensional format and reapplied to the image for examination at a later time.               4  Figure 2. Screenshot of the Mango GUI for medical image editing. The Mango toolbar is shown in the upper left-hand corner of the GUI.       In addition to an inherent degree of variability, manual segmentation is labor intensive and requires rater fluency in neuroanatomy. For this reason, several automated segmentation software programs have been developed for volumetric analysis of neuroimaging data. Automated segmentation methods are highly reliable compared with manual segmentation, as the exact same volume would be obtained from an MRI scan regardless of how many times it is automatically processed. In contrast, it is highly unlikely that the exact same volume would be produced by manual methods from the same MRI scan, whether the same rater or multiple raters perform the segmentation. Manual segmentation is more accurate, as the trained human rater is far more precise at identifying structural boundaries in 3-dimensional brain space compared to the software tools.  This results in far higher validity of measurement. Previous studies have !  5 shown that manual segmentation offers high validity, with inter-rater intraclass correlations have been reported as 0.954 (Doring et al., 2011), 0.92 (S?nchez-Benavides et al., 2010), and 0.89 (Morey et al., 2009). Intra-rater intraclass correlations of 0.94 (S?nchez-Benavides et al., 2010) and between 0.87-0.97 (Morey et al., 2009) have also been reported. Since it is impossible to have validity without some degree of reliability, manual segmentation is considered both a valid and reliable measure, despite a certain susceptibility to measurement variability.  Three frequently used automated imaging software packages include FSL (Functional MRI of the Brain [FMRIB] Software Library, Oxford), FreeSurfer (Massachusetts Institute of Technology, Harvard), and SPM (Statistical Parametric Mapping, Cambridge) (Figure 3). These freely available segmentation platforms are the most cited in volumetric imaging studies of the brain.   Figure 3. Software emblems for each of the freely available automated segmentation platforms evaluated in the current study. Programs can be downloaded from their respective websites: FreeSurfer: https://surfer.nmr.mgh.harvard.edu; FSL: http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/; SPM: http://www.fil.ion.ucl.ac.uk/spm/.        6 As each automated segmentation program uses a unique mathematical method for image interpolation and interpretation, the potential for variability exits among imaging findings from research groups choosing to use one automated method over another. In fact, validation studies conclude that automated segmentation output is affected by changes in MRI data preprocessing protocol (Pu et al., 2013), and is insensitive to hippocampal atrophy in some psychiatric samples (Pardoe, Pell, Abbott, & Jackson, 2009; S?nchez-Benavides et al., 2010). Contradictorily, other studies claim the use of automated segmentation methods are suitable for detecting small differences in hippocampal volume in psychiatric samples vs. controls (Doring et al., 2011), and between brain hemispheres of the same healthy individual (Lucarelli et al., 2013).  Further evaluation of automated segmentation software is necessary to ensure the most appropriate package is chosen for analyzing imaging datasets collected from a wide range of psychiatric and control subjects. The primary aim of the current study was to assess the validity of automated hippocampal segmentation tools available from Freesurfer 5.1, FLS, and SPM8.0 in a sample of individuals living in marginal housing in Vancouver?s DTES neighborhood.  1.3 Rationale for Validation of Automated Segmentation Software Using MRI Data Collected from a Subset of DTES Volunteers  Residents of Vancouver B.C?s DTES social housing projects represent a clinically unique population, as a large majority of the community struggles with co-occurring addiction/polysubstance dependence and complex psychiatric illness including schizophrenia and related psychotic spectrum disorders. A majority of the estimated 18, 500 individuals living in the DTES reside in Single Room Occupancy (SRO) housing projects funded by government social assistance programs. The Hotel Study is a large longitudinal investigation of the interactions of psychosis, addiction, and viral infection currently ongoing in the DTES. Participants are drawn from SRO facilities in the DTES and followed for up to 5 years. One aim of the Hotel Study is to investigate the relationship between structural imaging findings and mechanisms of psychiatric illness and consequences of long-term substance abuse. MRI scans collected from Hotel Study subjects commonly display moderate to severe motion artifact and global anatomical differences   7 such as severely enlarged ventricles, residual evidence of traumatic brain injury or stroke, and diffuse cortical atrophy. In addition, a large number of individuals scanned for Hotel study have large hippocampal cavities that appear as dark spaces on the images.  Volumetric analysis of large MRI datasets requires automated intervention, as manual methods are time-consuming and laborious. Currently, over 700 MRI scans have been collected from Hotel Study participants over 3 years. Recent studies have focused on validating Freesurfer and FSL in normal and Alzheimer Disease subjects (Morey et al., 2009; S?nchez-Benavides et al., 2010) and in subjects with bipolar affective disorder (Doring et al., 2011). Other studies have investigated sensitivity thresholds of these packages for the detection and lateralization of hippocampal atrophy in patients with mesial temporal lobe epilepsy (Pardoe et al., 2009). Few studies have sought to assess the validity of the Automated Anatomical Labeling (AAL) tool available with SPM, and to our knowledge, there are currently no published studies that focus on validating the ability of any automated hippocampal segmentation tool in a sample of individuals with co-morbid addiction/polysubstance dependence and complex psychiatric illness. Motion artifact and gross structural abnormalities both affect image quality and segmentation. Computational models designed to analyze volumetric images are, therefore, likely to be less effective when MRI datasets are compromised by motion distortion and anatomical anomalies. Currently, no published articles address the analytical susceptibility of automated hippocampal segmentation tools to either MRI data with moderate-severe motion artifact, or MRI data with large abnormal hippocampal cavities. Here, 2 novel analyses are proposed: 1) comparison of automated vs. manual hippocampal segmentation using MRI data with bilateral large abnormal hippocampal cavities, and 2) scan-rescan reliability of automated and manual hippocampal segmentation methods using MRI date with zero, mild-moderate, and severe motion artifact. Below, an overview of MRI data acquisition and image segmentation algorithms used by FreeSurfer 5.1, FSL/FIRST, and SPM8.0/AAL is included for completeness.          8 1.4 Overview of T1 Weighted MRI Data Acquisition   The concepts of T1 weighted MRI acquisition are summarized in Fundamental Physics of MR Imaging (Pooley, 2005). All specific information included in the following section was adapted from the content of this document.  MRI acquisition relies on the ability to manipulate hydrogen protons in biological tissues with a strong magnetic force and brief pulse sequences of radiofrequency energy (RF). Inside an MR imaging system, large electrical currents are fed through loops of wire to produce the main magnetic field. Typical MRI scanners generate magnetic forces of 1.5 Tesla (1.5T). Larger main magnetic fields (e.g. 3.0T) generate higher resolution MR images.  A standard coordinate system is used to define directionality inside an MRI system. Typically, the direction parallel to the main magnetic field is longitudinal, corresponding to the head-to-foot orientation of a patient lying horizontally, headfirst, and supine in the scanner, and is designated as ?z?. The plane perpendicular to z is termed the transverse plane or the ?x-y? plane. The x-y plane is commonly chosen such that x and y correspond to the patient?s left-right and anterior-posterior directions respectively.   When protons are exposed to large magnetic fields, most will tend to align in the direction of the magnetic force, resulting in a net magnetization aligned parallel to the magnetic field. In terms of an MRI acquisition, the net magnetization created by protons (in body fats and water) aligning with the main magnetic field is the source of the MR signal used to produce images.  RF energy, transmitted via a transmit coil (e.g. head coil for neuroimaging), is added to the system in short bursts or ?RF pulses?. When the RF frequency matches the precessional (rotational) frequency of the protons an efficient transfer of energy from the RF coil to the protons occurs ? a phenomenon know as resonance. The absorption of RF energy by the protons rotates them (and ultimately the net magnetization) away from the direction of the main magnetic field. The amount of rotation of the protons is termed the flip angle can be controlled by manipulating the magnitude and duration of the RF pulse.  Following the application of a 90O RF pulse (flip angle=90O), the longitudinal magnetization (z) is zero. When RF energy is removed from the system, protons begin to realign with the main magnetic field and the net magnetization regenerates along the longitudinal axis. This is termed   9 ?longitudinal relaxation? or ?T1 relaxation?. The rate at which the flip angle returns to 0O (protons realign parallel to the main magnetic field) differs from tissue to tissue ? that is, different tissues have unique T1 identities. T1 is defined as the time it takes for the longitudinal magnetization to regenerate to 63% of its final value after application of a 90O pulse. White matter has a very short T1 time, cerebral spinal fluid (CSF) has a relatively long T1, and gray matter has an intermediate T1. It is the T1 values of different brain tissue that provides the image contrast. In a T1 weighted image, white matter appears as very light (high intensity) pixels, CSF appears as almost black (low intensity), and gray matter appears as pixels of intermediate shades of grey (intermediate intensity).  1.5 Automated Image Parcellation: Overview of Algorithms Used by FreeSurfer 5.1, FSL/FIRST, and SPM8.0/AAL  Much in the same way 2 dimensional digital images are made up of pixels, volumetric images are made up of volumetric pixels or ?voxels?. Voxels make up the fundamental components of MRI data. Each voxel (e.g. 1mm3) is identified by its spatial position in the raw volumetric image, and its corresponding MR signal intensity. Coordinates for spatial location (e.g. x,y,z) are designated by the MRI acquisition system (scanner settings/imaging parameters), and the MR signal intensity corresponds to the T1 relaxation rate for each voxel (corresponding to white matter, gray matter, or CSF). Information describing voxel location and intensity provide the input data to be analyzed by automated segmentation programs.  Several similarities and differences exist between FreeSurfer, FSL/FIRST, and SPM8.0/AAL processing pipelines (Table 1). Following image ?stripping? (removal of non-brain tissues, e.g. skull, muscle, skin from the images), all three automated subcortical segmentation programs begin by registering raw MRI data to a brain template of known parameters (image registration). Image registration serves to spatially normalize and reduce anatomical variability between individual brains. The transformation of raw MRI data into one coordinate system is required for automated comparison and integration to occur. In addition to template registration, all three automated segmentation tools rely on manually delineated datasets (called ?training data?) from   10 which the program?s algorithm must draw the most likely ROI matches for the subcortical structures being measured.  Volumetric segmentation methods used by FreeSurfer 5.1, FSL/FIRST, and SPM8.0/AAL differ in two significant ways. First, SPM8.0/AAL and FSL/FIRST use a different registration template from FreeSurfer 5.1. Second, each platform uses an exclusive set of training data. The FIRST subcortical segmentation tool, available as part of FSL freeware is the most recently described (Patenaude, Smith, Kennedy, & Jenkinson, 2011), followed by FreeSurfer?s computational model for segmenting hippocampal subfields (Van Leemput et al., 2009), and SPM8.0?s automated anatomical labeling tool (AAL) (Tzourio-Mazoyer et al., 2002).    Table 1. Summary of processing pipelines for subcortical segmentation with FreeSurfer 5.1 (automated hippocampal subfield segmentation tool), FSL/FIRST, and SPM8.0/AAL. Automated Segmentation Program Image Registration Template ROI Segmentation Method    FreeSurfer 5.1     ICBM 452 1. Developed for segmentation of individual hippocampal subfields.  2. Bayesian modeling approach ? uses information about voxel coordinate (neuroanatomical label) and MR intensity.  3. Probabilistic atlases are used as a prior. 4.  Atlases are built from manual delineation of hippocampal subfields 5. Training data composition: 10 healthy cognitively normal individuals (5 female, 5 males; 4 older, 6 younger). 6. Only models a cuboid region surrounding the hippocampus. 7. Individual subfield volumes are added together for total hippocampal volume.     FSL/FIRST         MNI 152       1. Developed as a whole-brain subcortical segmentation tool.  2. Bayesian modeling approach supplemented by Active Shape/Active Appearance models (ASM/AAM). 3. ASM/AAM are algorithms for matching a statistical model of object shape and appearance to a new image. 4. Bayesian framework models probabilistic relationships between the shape and MR intensity of difference structures.      11 Automated Segmentation Program Image Registration Template  ROI Segmentation Method  FSL/FIRST  MNI 152 5. ASM/AAM and Bayesian models draw shape/MRI intensity information from manually labeled training datasets.  6. Training data composition: 336 MRI scans collected from normal and psychiatric subjects.    SPM8.0/AAL   MNI 152 1. Developed as an alternative to Talairach labeling system for determining regions of activation in fMRI studies.    2. AAL atlas/template was created from manual delineation of 90 (45 bilateral) ROIs from the MNI single subject brain. 3. The MNI single subject brain is an average of 27 MRI scans collected from the same cognitively normal young male adult.  4. Sulci landmarks were used to determine AAL atlas/template ROI limits. 5. Segmentation is achieved by propagating AAL template labels to regions in target image (MRI scan) that coincide with ROI spheres in the AAL atlas/template. 6. MR intensity determines whether voxels are gray matter, white matter, or CSF.            Both FSL/FIRST and SPM8.0/AAL tools register raw MRI data to templates provided by the Montreal Neurological Institute (MNI), while FreeSurfer 5.1 uses one from the International Consortium for Brain Mapping (ICBM). FSL/FIRST and SPM8.0/AAL use the MNI 152 registration template, which is the average of 152 normal structural MRI scans, while FreeSurfer 5.1 uses the ICBM 452 template, an average space constructed from 452 T1-weighted MRIs of normal young adult brains.   Choice of registration template may contribute to variability in the processing pipelines of automated segmentation tools, however, the major differences are likely to be due to the unique computational models and training datasets used by each platform to determine the anatomical identity of each voxel from raw MRI data. While FSL/FIRST and FreeSurfer 5.1 use Bayesian inference to determine voxel identity, SPM8.0/AAL employs an entirely different method based on percentage of voxel overlap with a sphere centered by a set of known coordinates.   Both FreeSurfer 5.1 and FSL/FIRST are designed within a Bayesian framework. These models capitalize on probabilistic identities of neighboring voxels before assigning each an   12 anatomical label. However, there are dissimilarities between the type of segmentation procedure (computational model) used by each program, as well as the training sets used to build their corresponding probabilistic atlases. FSL/FIRST uses a training dataset built from 15 manually segmented bilateral subcortical ROIs from 336 MR images of normal and pathological brains (including Alzheimer?s/schizophrenia patients, and individuals with prenatal cocaine exposure). In contrast, the training set used by the FreeSurfer 5.1 automated hippocampal subfield segmentation tool is specific to the hippocampus, and built from hippocampal subfield delineations performed on scans from 10 cognitively normal individuals (5 female, 5 male; 4 older, 6 younger).  The computational segmentation models used by FreeSurfer 5.1 and FSL/FIRST are very similar in that both are designed within a Bayesian framework. Differences between the two programs exist in the details: while FSL/FIRST?s Bayesian framework is supplemented by aspects of Active Shape/Active Appearance Models (ASM/AAM), FreeSurfer 5.1 simply uses a Bayesian modeling algorithm that draws information from several probabilistic atlases as a prior. Additionally, FreeSurfer 5.1?s automated hippocampal subfield segmentation tool models only a cuboid ROI (94 x 66 x 144 voxels) encompassing the hippocampus, in contrast to FSL/FIRST, which models the whole brain. Segmentation is ultimately achieved by applying these models to registered T1 MRI data in order to create a 3 dimensional mesh (defined by specific coordinates at each vertices) that covers the domain of interest. The ROI mesh is then adjusted until it best fits the boundaries of the subcortical structure being measured. This is accomplished by drawing from training datasets (built from manually segmented brain scans) that contain probabilistic information about the likelihood of a voxel?s identity given its respective MR intensity and special location relative to its neighboring voxels.  While FSL/FIRST and FreeSurfer 5.1 rely on conditional probabilities, SPM8.0/AAL uses an entirely unique method of segmentation. AAL training datasets are derived from manual delineation of the MNI single-subject brain; a compilation of 27 scans collected from past MNI post-doctorate student, Colin Homes. 45 bilateral ROIs were drawn every 2mm on axial slices, and several sulci were used as anatomical landmarks. ROIs are represented in the template as spherical regions centered by a known set of extremum coordinates with a given anatomical label. Following image registration, subcortical structures are defined by the percentage of MRI   13 voxels overlapped by template ROI spheres. SPM propagates anatomical labels to ROIs in the target image (MRI scan) based on areas of overall within the AAL template space.  The SPM8.0/AAL tool was not specifically designed for subcortical volumetric analysis. It was developed as an alternative standardized template for function MRI (fMRI) studies. Prior to the development of the AAL tool, the number of anatomical templates for activated cluster labeling in fMRI were limited. Most fMRI researchers preferred the use of the Talairach template because it offers a pre-labeled anatomical description of the human brain. SPM engineers justify the development of the AAL tool for several reasons, one being the observation that the Talairach brain displays deformations in its posterior part (likely resulting from several years of storage in formol). While a detailed discussion of fMRI methodology is beyond the scope of the current article, it is relevant to note the primary purpose of developing the AAL template was to provide a labeled brain atlas other than the Talairach template for reporting the localization of activations detected in fMRI studies.  1.6 Overview of Specific Hypothesis Evaluated in The Current Study  I. Manual segmentation is considered the gold standard method of volumetric analysis. It was predicted that manually traced ROIs would better represent true hippocampal boundaries compared with ROIs generated by FreeSurfer. II. FreeSurfer and FSL/FIRST use similar computational algorithms to achieve subcortical segmentation, however, literature suggests that hippocampal segmentation with FreeSurfer is superior to FSL/FIRST (Doring et al., 2011; Morey et al., 2009; Pardoe et al., 2009). Recently, FreeSurfer introduced a hippocampus-specific segmentation method capable of volumetric analysis of individual subfields, however, no existing studies have compared this method to FSL/FIRST or manual delineation of the whole hippocampus. It was predicted that hippocampal volumes generated with FreeSurfer?s automated hippocampal segmentation tool would be in strongest agreement with manual measurements, followed by FSL/FIRST volumes.  III. Since SPM8.0/AAL was developed for fMRI purposes, it was predicted that segmentation with the AAL tool would be least associated with manual measures.    14 IV. Regarding MRI data with large hippocampal spaces, two hypotheses were addressed: 1) it was predicted that FreeSurfer would perform more accurately than FSL/FIRST, and 2) both automated platforms would generate hippocampal volumes in lesser agreement with manual measurements when MRI data displayed bilateral hippocampal cavities.  V. Regarding MRI data with motion artifact: the Bayesian framework in which FSL/FIRST was designed relies heavily on the shape and texture (gradient of MR intensities) of subcortical constituents for voxel classification. It was predicted that FSL/FIRST segmentation algorithms would be hyper-sensitive to MRI data flaws that effect image resolution.  VI. Finally, it was predicted than FreeSurfer and FSL/FIRST volumes would be similar, yet would overestimate hippocampal volume. This hypothesis is an extension of previous literature that suggested that automated methods had a tendency to over exaggerate hippocampal boundaries.         15 Chapter 2: Body of Thesis   2.1  Research Methods   2.1.1  Study Participant Recruitment/Inclusion Criteria  The Hotel Study is a large longitudinal investigation of co-occurring psychosis, addiction, and infectious disease currently ongoing in Vancouver?s low-income DTES neighborhood (Hotel Study: Co-occurring Psychosis, Addiction, and Viral Infection; Principal Investigator: William G. Honer; Funding: Canadian Institute of Health Research). Hotel Study participants are recruited from DTES SRO social housing projects. DTES SRO residencies are affordable housing options (costing <30% of before-tax household income) intended for marginalized or disabled individuals receiving social assistance. Recruitment is staggered and ongoing. Eligibility to participate in Hotel Study includes the following criteria: 1) Residency in a SRO social housing project, 2) informed written consent, and 3) the ability to communicate in English. The sole exclusion criterion to participate in Hotel Study is the inability to give informed consent.  2.1.2 Overview of Hotel Study Format  Baseline (time of recruitment to Hotel Study) assessments were extensive. Included in the overall package of baseline and annual follow-up assessments were: brain imaging, drug use histories, education, health and trauma questionnaires, as well as full psychiatric evaluations.  2.1.3 Current Sample Demographics   55 participants, chosen at random from the larger Hotel Study database, were included in the current study. Baseline sample demographic information is given in Table 2.      16 Table 2. Baseline Sample Demographics. Age, gender, ethnicity, self-reported years using drugs in the DTES, years of high-school education, and high-school graduation status of sample participants. Mean (M); Standard Deviation (SD); Range (R)  Baseline Subject Demographics (N=55): Age (years) M=44.1; SD=10.2; R: 23-68 Gender* Female 14.5% (N=8) Male 85.5%  (N=47) Ethnicity Aboriginal: 25.5% (N=14) Caucasian: 71.0% (N=39) Other Nationality** 3.5% (N=2) Self-Reported Years Using Drugs in the DTES M=14.0; SD=9.4; R: 1-50 Years of High School Education M=10.3; SD=2.12; R: 0-12 High School Graduation Status High School Diploma: 23.5% (N=13) General Education Development (GED) Certification: 23.5% (N=13) No High School Diploma or GED:  53% (N=29) *The female-to-male ratio is reflective of the total Hotel Study sample.      **Other Nationalities included: Norwegian-African, and unknown.       2.1.4 Participant Substance Use and Dependence   DTES communities are at high risk for substance use/polysubstance dependency. Trained psychiatrists reviewed self-reported lifetime substance exposure, participant descriptions of weekly drug-use patterns, Maudsley Addiction Profile data, and drug-dependence related sections of the Mini International Neuropsychiatric Interview to determine substance dependencies. Dependency diagnoses were made according to the Best Estimate Clinical Evaluation and Diagnosis (BECED). Two psychiatrists independently assigned diagnoses. Tobacco use was assessed by survey, and laboratory drug screenings were obtained from participants? urine samples. Baseline participant substance use and dependence disorders are given in Table 3.    17  Table 3. Baseline Participant Substance Use and Dependence Disorders. 1) Proportion of sample participants dependent on 1 or more substance(s), 2) Frequency of dependence disorders, 3) Tobacco use patterns, and 3) Proportion of substances tested positive in urine drug screens.  1. Number Substance Dependencies: Proportion of Sample Percentage of Sample 1 Substance 8/55 14.5% 2 Substances 8/55 14.5% 3 Substances 15/55 27.3% 4 Substances 13/55 23.6% 5 Substances 4/55 7.2% 6 Substances 5/55 9.0% 7 Substances 2/55 3.6%  2. Dependence Disorders:  Proportion of Sample Percentage of Sample Alcohol Dependence  22/55 40% Cocaine Dependence  43/55 78.2% Crack Cocaine Dependence  39/54 72.2% Methamphetamine Dependence  20/55 36.4% Heroin Dependence  24/55 43.6% Methadone Dependence  22/55 40%  3. Tobacco Use Proportion of Sample Percentage of Sample Never/Not at All 7/55 12.7% Occasionally 10/55 18.2% Everyday  38/55 69.0%  4. Positive Urine Drug Screen Result: (N=48; 7 Participants Refused Screen) Proportion of Sample Percentage of Sample Amphetamine 13/48 27.1% Barbiturates  1/48 2.1% Benzodiazepines  10/48 20.8% Cocaine 36/48 75.0% Marijuana  26/48 54.2% Methadone  18/48 37.5% Methamphetamine 17/48 35.4% Methylene-dioxymethamphetamine (MDMA) 2/48 4.2% Opiates 17/48 35.4       18 2.1.5 Participant Neurological/Psychiatric Diagnoses  Co-occurring psychiatric diagnoses are common among Hotel Study participants. All available clinical information was considered before psychiatric diagnoses were made according to the Best Estimate Clinical Evaluation and Diagnosis (BECED). Clinical information included mental health associated hospitalization records, Mini International Neuropsychiatric Interview data, and clinical psychiatric evaluations. Two psychiatrists assigned diagnoses independently. Baseline frequency of neurological/psychiatric diagnoses is given in Table 4.    Table 4. Baseline Frequency of Neurological/Psychiatric Diagnoses. 1) Proportion of sample participants with BECED-determined DSM-IV psychiatric diagnosis. 1. Frequency of Neurological/ Psychiatric Disorders: (Some participants had more than one diagnosis) Proportion of Sample Percentage of Sample Schizophrenia 14/55 25.5%% Schizoaffective Disorder 9/55 16.4% Bipolar Disorder (Includes Bipolar I, Bipolar II, & Bipolar NOS) 2/55 3.6% Major Depressive Disorder or Depression NOS 10/55 18.2% Psychosis Not Otherwise Specified (PNOS) 3/55 5.5% Substance-Induced Psychosis Disorder 11/55 20.0% Traumatic Brain Injury  8/55 14.5% Other Psychiatric/Neurological Diagnoses* 44/55 80.0% *Other Psychiatric/Neurological Diagnoses include: cerebellar stroke, agoraphobia, panic and anxiety disorders, attention deficit disorders, post-traumatic stress disorder, personality disorders, social phobia, obsessive-compulsive disorder, seizure disorder, tardive dyskinesia, hoarding disorder, dysthymia, Huntington?s disease, anorexia/bulimia, migraine headaches, cognitive disorders, pathogenic gambling, spinal cord abscesses.         19 2.2 MRI Data  2.2.1 Overview of MRI Datasets  75 T1 weighted MRI scans collected from 55 Hotel Study Participants were included in the current analysis. All MR images were screened for motion artifact and abnormal hippocampal cavities (Figure 4). The sole exclusion criterion for MRI scan selection from the Hotel Study database was incomplete scan acquisition (participant unable to remain in scanner until scan sequence was complete).  Figure 4. Breakdown of MRI Datasets.     2.2.2 Image Acquisition   High-resolution three-dimensional (3D) T1 weighted scans were acquired on the University of British Columbia (UBC) 3T research scanner (Phillips Achieva 3.0 Tesla, software   20 version 12.2). Subjects were positioned for T1-weighted image acquisition using a three-dimensional axial inversion recovery-weighted spoiled gradient recalled (3D SPGR T1-weighted) sequence; scan duration = 07:21.1min, field of view (FOV) = 256mm, matrix = 256 x 250mm, reconstructed isotropic voxel size = 1 x 1 x 1 mm, TR/TE = 8.1 x 3.5ms, flip angle = 8.0?, SENSE-Head-8 coil, autoshim, slice thickness = 1mm no gap (total 190 slices contiguous through whole brain), and isotropic voxel dimensions.  2.2.3 Image Segmentation   Four methods were used to obtain bilateral hippocampal volumes for each of the total 75 T1-weighted MRI scans: 1) manual segmentation by rater (C.B.M), 2) automated segmentation with FreeSurfer 5.1, 3) automated segmentation with FSL/FIRST, and 4) automated segmentation with SPM8.0/AAL.   2.2.4 Manual Segmentation  Trained rater C.B.M performed all manual segmentations. The manual-segmentation protocol was adapted from (Smith et al., 2003) and General Brain Segmentation ? Methods and Utilization Manual (Version 3, 2004) available on The Center for Morphometric Analysis website (http://www.cma.mgh.harvard.edu). Bilateral anatomic regions of interest (ROI) were selected one at a time (left and then right) using Mango. All selections were made using a Mac OS X (2 X2.4 GHz Quad-Core Intel Xeon processor) and a Wacom Intuos4 digital drawing tablet and pen.  ROIs were identified and selected based on anatomical landmarks and hippocampal boundaries seen on image slices. Limits were selected with visual reference to a neuroanatomic atlas of the human brain and hippocampus (Krebs et al., 2011; Shorvon, 2000). The most anterior ROI selection was chosen at first appearance of the myelinated tract visibly separating the amygdala from hippocampus (portions of the fornix and stria terminalis [F/ST]), and the most posterior image was chosen as one slice anterior to the first appearance of the trigones of the lateral ventricles.    21 Beginning in the coronal plane, the first ROI selection was made 14 slices posterior to the first appearance of the anterior commissure; selections were continued until the first appearance of the trigones of the lateral ventricles (where the hippocampal tail could no longer be distinguished from the surrounding cortical gyri). Next, image view was switched to the sagittal plane and ROI selections were made on all slices where the hippocampus was visibly separated from amygdala by F/ST. Image view was switched back to coronal and ROI selections were made starting at the first presentation of the ROI tracing made in the sagittal view, and continued until the slice on which the very first ROI selection was made (first ROI selection- coronal plane, 14 slices posterior to the anterior commissure). Images were then examined in all planes (coronal, sagittal, and axial) and ROIs were corrected using Mango?s ?edit? tool to include more or less gray matter until the traced region best fit hippocampal boundaries. If edits were made, ROIs were re-examined several times in each plane of view to ensure boundaries remained best representative of hippocampal anatomy. Volumes were calculated by Mango: ROIs were selected using ?Select All Method?, followed by Analysis>ROI Statistics>Volume stream.   2.2.5 Manual Segmentation of MRI Data with Severe Motion Artifact  Manual segmentation was performed using the same protocol described previously, except in cases where motion artifact was so severe gray-to-white matter boundaries were incomprehensible. In these situations, a scan from the same individual with zero motion artifact was used for anatomical reference.  2.2.6 Automated Segmentation  W.S. was responsible for running all automated segmentation pipeline scripts.  I. Freesurfer 5.1: Step (1) subcortical segmentation and hippocampal subfield segmentation values were obtained using default parameters. Step (2) all subfield were merged. Step (3) visible holes in the pixelated hippocampal ROI were filled and isolated voxels were removed.    22 II. FSL/FIRST: all default parameters available from the FIRST 4.1.9/FIRST segmentation stream were used to obtain hippocampal volumes. III. SPM8/AAL: Step (1) T1 images and Montreal Neurological Institute (MNI) 152 template were co-registered with unified segmentation and normalization (default parameters). Step (2) Template space AAL hippocampal masks were mapped onto native T1 space using the transformation obtained in Step (1).  2.3 Study Design, Methods, and Statistical Analyses  2.3.1 Statistical Analyses  All statistical comparisons were performed using SPSS 21.0 software (IBM Corporation). For all analyses included in the current study, total hippocampal volumes (right volume + left volume) were used.  2.3.2 Design 1: Subjective Comparison of Manual vs. Automated Segmentation  Hypothesis IA.  Manual segmentation of the hippocampal formation is superior to automated methods. Anatomical regions of interest (ROIs) delineated by manual methods more accurately identify hippocampal boundaries then automated ROIs.  Method IA. Following the completion of all manual hippocampal segmentations, 10 scans were randomly chosen for concurrent validation of ROIs by trained neuropsychiatrist/neurologist (W.P). Images (the same MRI scan opened in 2 GUI windows) were presented blindly and side-by side. On one scan, manually selected ROIs were superimposed bilaterally. On the other scan, ROIs produced by FreeSurfer 5.1 were superimposed bilaterally. The intention was side-by-side comparison of ROIs produced by manual segmentation and FreeSurfer 5.1 (Figure 5).     23  Figure 5. Example of GUI set-up for concurrent validation of manual vs. FreeSurfer 5.1 generated ROIs. Figure depicts a desktop screenshot of 2 GUI windows displaying the same MR image. Manual and FreeSurfer 5.1 ROIs are superimposed over the hippocampal formation. Images were presented blindly and side-by-side. In this case, manually delineated ROIs are presented in the left GUI, and FreeSurfer 5.1 ROIs are presented on the right.    W.P.P. examined scan-ROI sets in each plane of view, and rated which of the two ROI sets (presented on the right or left) was most representative of hippocampal boundaries. Following 10 sets of comparisons, W.P?s choices (right or left) were reviewed with reference to a pre-made answer key listing which of the two ROI sets (right or left) was manually traced and which was obtained with FreeSurfer 5.1 for each trial.   Hypothesis IB.  The Intraclass-correlation between Manual vs. FSL/FIRST methods will be strongest, followed by Manual vs. FreeSurfer 5.1, and then Manual vs. SPM8.0/AAL.    24   Method IB. Twenty-five scans free of motion artifact and abnormal hippocampal cavities were included. Intraclass correlation coefficients were calculated for the following sets of comparisons: 1) manual segmentation vs. FreeSurfer 5.1, 2) manual segmentation vs. FSL/FIRST, 3) manual segmentation vs. SPM8.0/AAL. A repeated-measures analysis of variance (ANOVA) model with hippocampal volumes generated by manual segmentation, FreeSurfer 5.1, FSL/FIRST, and SPM8.0/AAL as within-subjects variables was used to assess whether mean total hippocampal volumes produced with any of the 4 methods differed significantly. Post-hoc paired t-tests were used to compare each method to the other.  2.3.3 Design 2: Comparison of Segmentation Methods Using MRI Data with Bilateral Abnormal Hippocampal Cavities  Hypothesis II.  Large hippocampal spaces (Figure 6) disrupt automated segmentation algorithms. Intraclass correlation comparisons between manual vs. automated methods will be lower when MRI data display bilateral abnormal hippocampal cavities.  Method II. All visible hippocampal cavities were manually measured in Mango (Multi-Image Analysis Graphic User Interface), a non-commercial GUI software for viewing, editing, and analyzing volumetric medical images by a trained rater (C.B.M). The length of each cavity was estimated by counting the number of consecutive slides on which the cavity was clearly visible. Since MRI slices are 1mm thick, cavities clearly visible on more than 3 images were classified as ? 3mm. Cavity diameter estimations were obtained by plugging cavity volumes from a single MRI slice obtained in Mango (mm3) into the equation for finding volume of a sphere. Cavities were classified as abnormal according to (Yoneoka, Kwee, Fujii, & Nakada, 2002). Criteria for   25 abnormality included the following: 1.) Cavity diameter exceeded 1mm, and 2.) Cavity length exceeded 3mm.  Figure 6. MRI scans from 3 Hotel Study participants with hippocampal cavities (red arrows indicate abnormality). a) sagittal view of left hippocampus, b) coronal view of left hippocampus, and c) coronal view of right hippocampus.    Nineteen motion artifact-free scans with bilateral abnormal hippocampal cavities were included. Intraclass correlation coefficients were calculated for the following sets of comparisons: 1) manual segmentation vs. FreeSurfer 5.1, 2) manual segmentation vs. FSL/FIRST, 3) manual segmentation vs. SPM8.0/AAL. A repeated-measures analysis of variance (ANOVA) model with hippocampal volumes generated by Manual Segmentation, FreeSurfer 5.1, FSL/FIRST, and SPM8.0/AAL as within-subjects variables was used to assess whether mean total hippocampal volumes produced by any of the 4 methods differed significantly. Post-hoc paired t-tests were used to compare each method to the other.   2.3.4 Comparison of Segmentation Methods Using MRI Data with Motion Artifact  Hypothesis III. Motion artifact disrupts automated segmentation algorithms. When MRI data display motion artifact, manual segmentation methods will measure hippocampal volume with greater consistency than automated methods.   26   Method III.  The degree of motion artifact was ranked by a trained biomedical imaging expert (W.S.) with extensive practical experience in T1 weighted MR imaging analysis. W.S ranked scans as having zero, mild-moderate, or severe motion artifact (Figure 7). Zero motion artifact was assigned to scans with clear gray-to-white matter boundaries, clear anatomical resolution, and high pixel contrast. Scans with mild-moderate motion artifact were slightly-to-moderately blurred, however, in most cases gray-to-white matter boundaries and separate anatomical constituents were adequately clear. Scans with severe motion artifact were of substantially poorer image quality, with severely blurred anatomical boundaries and low pixel contrast.   Figure 7. Example of 3 T1-weighted MR images collected from the same Hotel Study participant with a range of motion artifacts. A) Zero motion artifact: gray-to-white matter boundaries are clear, scan resolution and pixel contrast is high, B) mild-moderate motion artifact: slight blurring of gray-to-white matter boundaries especially in anterior cortical regions (red arrow), pixel contrast remains high, and C), severe motion artifact: significant blurring throughout the image, resolution is poor, separate anatomical constituents are unclear in places, and pixel contrast is relatively low (red arrow indicates greyish blurring of typically black-colored ventricular areas).     Thirty-one MRI scans collected from 11 individuals were included. Nine individuals (scanned at 3 time points) each had 1 scan with zero motion artifact, 1 scan with mild-moderate   27 motion artifact, and one scan with severe motion artifact. The remaining 2 individuals (scanned twice) had one scan with zero motion artifact, and 1 scan with mild-moderate motion artifact. Scans were grouped into categories based on their degree of motion artifact. Categories included were: zero, mild-moderate, and severe motion artifact. Intraclass correlation coefficients were calculated for the following comparisons (summarized in Table 5). A repeated-measures analysis of variance (ANOVA) model with hippocampal volumes generated by Manual Segmentation, FreeSurfer 5.1, FSL/FIRST, and SPM8.0/AAL as within-subjects variables was used to assess whether mean total hippocampal volumes produced by any of the 4 methods differed significantly. Post-hoc paired t-tests were used to compare each method to the other.   Table 5. Summary of intraclass correlation coefficients calculated to compare values generated by manual segmentation, FreeSurfer 5.1, FSL/FIRST, and SPM8/0/AAL. Each method was compared against itself using scan data with differing degrees of motion artifact. 1. Zero Motion Artifact   VS. Severe Motion Artifact a. Manual Segmentation a. Manual Segmentation b. FreeSurfer 5.1 b. FreeSurfer 5.1 c. FSL/FIRST c. FSL/FIRST d. SPM8.0/AAL d. SPM8.0/AAL    2. Zero Motion Artifact   VS. Mild-Moderate Motion Artifact a. Manual Segmentation a. Manual Segmentation b. FreeSurfer 5.1 b. FreeSurfer 5.1 c. FSL/FIRST c. FSL/FIRST d. SPM8.0/AAL d. SPM8.0/AAL  2.3.5 Comparison of FreeSurfer 5.1 and FSL/FIRST Methods:  Hypothesis IV. FreeSurfer 5.1 and FSL/FIRST use similar computational frameworks for automated hippocampal segmentation. Strong intraclass correlation exists between hippocampal volumes generated by FreeSurfer 5.1 and FSL/FIRST even when MRI data displays motion artifact and abnormal hippocampal cavities.    Method IV.    28 The following intraclass correlation coefficients (Table 6) were calculated to compare total hippocampal volumes generated FreeSurfer 5.1 vs. FSL/FIRST:   Table 6. Intraclass correlation comparisons between FreeSurfer 5.1 and FSL/FIRST.   Comparison MRI Dataset 1. Freesurfer 5.1 vs. FSL/FIRST 25 scans free of motion artifact and abnormal hippocampal cavities. 2. Freesurfer 5.1 vs. FSL/FIRST 19 motion artifact-free scans with bilateral hippocampal cavities. 3. Freesurfer 5.1 vs. FSL/FIRST 9 scans with severe motion artifact.   29 Chapter 3: Results  3.1 Concurrent Validation of Manually Segmented ROIs    Manually delineated ROIs were ranked as more accurate then ROIs obtained with FreeSurfer 5.1 in 10/10 trial comparisons.  3.2 Comparison of Manual vs. Automated Segmentation Methods   Total hippocampal volumes from 25 scans free of motion artifact and abnormal hippocampal cavities were included.  The mean (M) and standard deviation (SD) for total hippocampal volume produced with each method were as follows: 1) Manual Segmentation (M=6263.0, SD=834.7), FreeSurfer 5.1 (M=7087.0, SD=932.6), FSL/FIRST (M=7524.2, SD=1144.0), and SPM8.0/AAL (M=12114.8, SD=1031.8) (Figure 8).   Figure 8. Mean total hippocampal volumes. Segmentation methods are presented along the x-axis, and volume (mm3) is presented along the y-axis.  0	 ?2000	 ?4000	 ?6000	 ?8000	 ?10000	 ?12000	 ?14000	 ?Manual	 ? FreeSurfer	 ?5.1	 ? FSL/FIRST	 ? SMP8.0/AAL	 ?Mean	 ?Total	 ?Hippocampal	 ?Volume	 ?	 ?  30   Intraclass correlations are summarized in Table 7.    Table 7. Intraclass correlation coefficients (N=25 total hippocampal volumes per method): Comparison  Intraclass Correlation Coefficient  Manual Segmentation vs. Freesurfer 5.1 0.88 Manual Segmentation vs. FSL/FIRST 0.64 Manual Segmentation vs. SPM8.0/AAL 0.035  Repeated measures ANOVA revealed significant differences between total hippocampal volumes generated by all 4 methods (F(3,72)=449.73, p=0.000). Post-hoc t-test comparisons between each method are summarized in Table 8.    Table 8. Post-hoc paired t-tests (N=25 hippocampal volumes per method).  Paired-Sample T-test  T24 P-value Manual Segmentation vs. Freesurfer 11.1 0.000 Manual Segmentation vs. FSL/FIRST 11.1 0.000 Manual Segmentation vs. SPM8.0/AAL 27.8 0.000 Freesurfer 5.1 vs. FSL/FIRST 3.7 0.001 Freesurfer 5.1 vs. SPM8.0/AAL  23.9 0.000 FSL/FIRST vs. SPM8.0/AAL  18.2 0.000  3.3 Comparison of Manual vs. Automated Segmentation Methods Using MRI Data with Bilateral Abnormal Hippocampal Cavities  Total hippocampal volumes from 19 motion artifact free scans with bilateral abnormal hippocampal cavities were included.     31 The mean (M) and standard deviation (SD) for total hippocampal volume produced with each method were as follows: 1) Manual Segmentation (M=6411.3, SD=835.9), 2) FreeSurfer 5.1 (M=7595.8, SD=755.5), 3. FSL/FIRST (M=7922.63, SD=908.1), and 4) SPM8.0/AAL (M=12453.4, SD=963.7) (Figure 9).   Figure 9. Mean total hippocampal volumes. Segmentation methods are presented along the x-axis, and volume (mm3) is presented along the y-axis.   Intraclass correlations are summarized in Table 9.  Table 9. Intraclass correlation coefficients (N=19 total hippocampal volumes per method): Comparison  Intraclass Correlation Coefficient Manual Segmentation vs. Freesurfer 5.1 0.48 Manual Segmentation vs. FSL/FIRST 0.43 Manual Segmentation vs. SPM8.0/AAL 0.039  0	 ?2000	 ?4000	 ?6000	 ?8000	 ?10000	 ?12000	 ?14000	 ?Manual	 ? FreeSurfer	 ?5.1	 ? FSL/FIRST	 ? SMP8.0/AAL	 ?Mean	 ?Total	 ?Hippocampal	 ?Volume	 ?	 ?  32 Repeated measures ANOVA revealed significant differences between total hippocampal volumes generated by all 4 methods (F(3,54)=408.5, p=0.000). Post-hoc t-test comparisons are given in Table 10.   Table 10. Post-hoc paired t-tests (N=19 hippocampal volumes per method).  Paired-Sample T-test  T18 P-value Manual Segmentation vs. Freesurfer 7.9 0.000 Manual Segmentation vs. FSL/FIRST 9.4 0.000 Manual Segmentation vs. SPM8.0/AAL 28.2 0.000 Freesurfer 5.1 vs. FSL/FIRST 2.8 0.011 Freesurfer 5.1 vs. SPM8.0/AAL  23.2 0.000 FSL/FIRST vs. SPM8.0/AAL  19.3 0.000  3.4 Comparison of Manual vs. Automated Segmentation Methods Using MRI Data with Motion Artifact  Thirty-one scans collected from 11 individuals were included. Eleven scans were classified as having zero motion artifact, 11 scans were classified as having mild-moderate motion artifact, and 9 scans were classified as having severe motion artifact.    Scan-rescan intraclass correlation coefficients (Table 11): zero motion artifact (N=11 per method), mild-moderate motion artifact (N=11 per method), and severe motion artifact (N=9 per method).   Table 11. Scan-rescan intraclass correlation coefficients. A: Comparison of scan-rescan reliability of segmentation methods using MRI data with zero vs. severe motion artifact; B: Scan-rescan reliability of segmentation methods using MRI data with zero vs. mild-moderate motion artifact.   A. Zero vs. Severe Motion Artifact Intraclass Correlation Coefficient Manual Segmentation 0.96 Freesurfer 5.1 0.85   33 FSL/FIRST 0.62 SPM8.0/AAL 0.99  B. Zero vs. Mild-Moderate Motion Artifact Intraclass Correlation Coefficient Manual Segmentation 0.96 Freesurfer 5.1 0.86 FSL/FIRST 0.86 SPM8.0/AAL 0.99  Repeated measures ANOVA revealed significant differences between total hippocampal volumes generated by all 4 methods (F(1,10)=132.8, p=0.000) when MRI data displays mild-moderate motion artifact. Post-hoc paired t-test comparisons between methods are summarized in Table 12.   Table 12. Post-hoc paired t-tests: comparison of segmentation methods using MRI data with mild-moderate motion artifact.  Paired-Sample T-test  T10 p Manual Segmentation vs. Freesurfer 5.2 0.000 Manual Segmentation vs. FSL/FIRST 2.7 0.022 Manual Segmentation vs. SPM8.0/AAL 19.8 0.000 Freesurfer 5.1 vs. FSL/FIRST 0.27 0.789* Freesurfer 5.1 vs. SPM8.0/AAL  20.6 0.000 FSL/FIRST vs. SPM8.0/AAL  10.0 0.000 *Total hippocampal volumes generated by FreeSurfer 5.1 and FSL/FIRST were not significantly different.  Repeated measures ANOVA revealed significant differences between total hippocampal volumes generated by all 4 methods (F(3,24)=118.1, p=0.000) when MRI data displays severe motion artifact. Post-hoc paired t-test comparisons between methods are summarized in Table 13.       34  Table 13. Post-hoc paired t-tests: comparison of segmentation methods using MRI data with severe motion artifact. Paired-Sample T-test  T8 p Manual Segmentation vs. Freesurfer 0.74 0.48* Manual Segmentation vs. FSL/FIRST 2.7 0.022 Manual Segmentation vs. SPM8.0/AAL 19.8 0.000 Freesurfer 5.1 vs. FSL/FIRST 0.27 0.000 Freesurfer 5.1 vs. SPM8.0/AAL  20.6 0.000 FSL/FIRST vs. SPM8.0/AAL  10.0 0.000 *Total hippocampal volumes generated by manual segmentation and FreeSurfer 5.1 were not significantly different.  3.5 Comparison of FreeSurfer 5.1 vs. FSL/FIRST Segmentation Methods  Intraclass correlation coefficients (Table 14): 1) 25 MRI scans free of motion artifact and abnormal hippocampal cavities 2) 19 motion artifact-free scans with bilateral abnormal hippocampal cavities, 3) 9 scans displaying severe motion artifact.  Table 14. Intraclass correlation comparisons between FreeSurfer 5.1 and FSL/FIRST.  Comparison MRI Dataset Intraclass Correlation 1. Freesurfer 5.1 vs. FSL/FIRST 25 scans free of motion artifact and abnormal hippocampal cavities. 0.87 2. Freesurfer 5.1 vs. FSL/FIRST 19 motion artifact-free scans with bilateral hippocampal cavities. 0.87 3. Freesurfer 5.1 vs. FSL/FIRST 9 scans with severe motion artifact. 0.64         35 Chapter 4: Discussion   4.1 Overview of Study Methodology  In the present study evaluation of manual and automated hippocampal segmentation methods in a clinically heterogeneous sample was performed. First, manual segmentation was established as the gold standard in comparison to FreeSurfer segmentation. Second, a series of intraclass correlations were used to determine which set of automated hippocampal volumes were most correlated with manual measurements using samples with and without bilateral hippocampal cavities. Third, the scan-rescan reliability of each method using MRI data with mild to severe motion artifact was assessed. Additionally, we investigated statistical differences in mean volumes produced by each method when MRI data displayed no cavities, bilateral cavities, and mild to severe motion artifact, and evaluated similarities in the performance of the two best performing segmentation software platforms, FreeSurfer and FSL/FIRST.  4.2 Manual Segmentation as the Gold Standard Comparison for Volumetric Analysis of the Hippocampus  Due to its complex three-dimensional shape and ambiguous gray-to-white matter boundaries, accurate volumetric analysis of the hippocampal formation is difficult irrespective of the segmentation method used. In order to confirm manual tracing was the more accurate hippocampal segmentation protocol in this sample, example ROIs obtained with FreeSurfer and manual methods were blindly presented to a neurologist for evaluation. Ten out of 10 manually selected samples of hippocampal ROIs were rated as more accurate than FreeSurfer segmentation of the same ROI set. Our rationale for using FreeSurfer ROIs for comparison with manual segmentation was based on several validation studies (Doring et al., 2011; Morey et al., 2009; Pardoe et al., 2009) that cite hippocampal segmentation with FreeSurfer as superior to other automated methods.    36 The accuracy of manual segmentation depends on the rater?s ability to visualize variation within and between brain structures, while automated methods rely on their respective algorithms to do so. One particular issue with manual segmentation is the difficultly in establishing the amygdalo-hippocampal boundary. For this reason, previous studies with lower resolution data (e.g. MRI scans collected with 1.5T magnet) have grouped the entire amygdala and hippocampus together (Yoshida et al., 2009), or divided the hippocampus into anterior and posterior regions (O'Driscoll et al., 2001). Advances in MRI acquisition technology now allow for ascertainment of ultra high-resolution images of the brain with much greater detail than before. In fact, on virtually all MRI scans manually segmented in the current study, the white matter tract separating the anterior hippocampus from the amygdala was obviously discernable (Figure 10). The ability to clearly distinguish this tract provided an accurate and consistent reference for choosing where to draw the first segmentation lines at the boundary of the anterior hippocampus in the current study.    Figure 10. Sagittal view of left hippocampus: a) 1.5T MR image, b) 3.0T MR image. The white matter tract separating the anterior amygdala is clearly visible on MR images collected at 3.0T (b). This tract cannot easily be distinguished on MR images collected at 1.5T (a).    37   Visual inspection of manually derived ROIs compared with those obtained by automated methods, clearly showed that in most cases manual segmentation more accurately represented hippocampal boundaries (Figure 11). Automated modalities, while suitable for certain MRI datasets, are currently not capable of establishing object boundaries with as much accuracy as the human visual system. Subcortical structures, in general, are not easily distinguished from one another, since a large degree of variability in histological composition corresponds to within-structure fluctuations in MRI pixel intensity. For example, the significantly higher myelin content of the pallidum makes its T1-weighted MR intensity higher than the caudate, despite the fact both structures are composed of gray matter. As previously described, the hippocampus is composed of several different subfields that are not easily distinguished, even on ultra high-resolution images. While manual raters benefit from the ability to visualize fine anatomical detail on higher resolution images, automated segmentation programs require more advanced computational algorithms to model sharp edges and intensity gradients within the same subcortical structure. Accurate automated volumetric analysis of high field MR images cannot be achieved by measuring voxel intensity alone (e.g. gray matter, white matter,   38 CSF); rather, segmentation algorithms must model extensive variation in the shape and MR intensity gradients of the same subcortical structure.             Figure 11. Sagittal view of identical MRI images collected from a Hotel Study participant with hippocampal atrophy. The ROI in the left image was segmented by manual tracing, and the ROI in the right image was segmented with FreeSurfer 5.1. Red arrows indicate areas where ROI extends beyond hippocampal boundaries.       39  The current sample, which included subjects with abnormal hippocampal morphology, introduced a large degree of individual-based complexity that was not correctly identified by the software segmentation paradigms. Individual morphological variability is likely to exist in a number of patient populations, however, manual evaluation of random MRI subsamples from patient studies may help to ensure ROIs generated by automated means are sufficiently congruent with manual standards. Further efforts to refine computational algorithms are required before fully automated hippocampal segmentation is equivalent to the gold standard.    4.3 Automated vs. Manual Segmentation Methods: Evaluation of FreeSurfer 5.1 and FSL/FIRST  The hypothesis that FreeSurfer would produce hippocampal volumes that correlated most strongly to those obtained using manual methods was supported. In this sample of Hotel Study participants, intraclass correlations between manual segmentation and FreeSurfer 5.1 were strongest, followed by manual vs. FSL/FIRST. This was true for MRI data with and without bilateral hippocampal cavities, although intraclass correlation comparisons were lower when MRI data included cavities (this result is discussed in more detail in section 4.5).  These results are similar to previous validation studies that cite FreeSurfer performance as superior to FSL/FIRST. In fact, the intraclass correlations between manual vs. FreeSurfer and FSL/FIRST segmentation (MRI data without hippocampal cavities) observed in the current study were similar to those observed by other groups using FreeSurfer?s whole-brain segmentation tool. Doring et al. (2011) reported intraclass correlation values of 0.846 (right) and 0.859 (left) and 0.746 (right) and 0.654 (left) between manual segmentation vs. FreeSurfer and manual segmentation vs. FSL/FIRST respectively. Here, we chose to use FreeSurfer?s recently described tool for automated segmentation of the hippocampus at the subfield level, and observed a robust agreement between total hippocampal volumes generated by this method and manual measurements   40 in this sample. While the current study cannot directly speak for the validity of FreeSurfer?s whole-brain segmentation tool, overall these data, along with findings from previous investigations, support the hypothesis that FreeSurfer subcortical segmentation toolboxes outperform FSL/FIRST.  Considering the FreeSurfer subfield segmentation tool (evaluated here) was created to model the hippocampus, specifically its subregions, while FSL/FIRST was designed to model all major subcortical constituents (not just the hippocampus), these results are not surprising. Based on the current results, namely that FreeSurfer?s subfield segmentation tool generated hippocampal volumes that are in robust agreement with manual measurement, it appears this method meets its primary objective: modeling hippocampal regions with acceptable accuracy. It should be noted that other research groups have not assessed the validity of this tool, and therefore, it remains unknown whether this processing pipeline is appropriate for use in other samples. Nonetheless, the current results suggest that hippocampal volumes obtained with FreeSurfer?s subfield segmentation tool adequately agree with manual segmentation when MRI data is not compromised by morphological anomalies.     4.4 Hippocampal Volumes Obtained with SPM8.0/AAL are Weakly Associated with Manual Segmentation  Our hypothesis that SPM8.0/AAL would produce volumes least correlated with the manual measures was supported by the current data, regardless of whether MRI data included hippocampal cavities or not. In fact, the intraclass correlation between manual and SPM8.0/AAL derived hippocampal volume was close to zero in both cases.  In addition, SPM8.0/AAL mean hippocampal volumes differed significantly from all other segmentation methods (FreeSurfer, FSL/FIRST, manual segmentation) under all circumstances (MRI data with/without hippocampal cavities, MRI data with motion artifact).  Considering the AAL template is not specifically designed to achieve the most accurate subcortical segmentation, this result is not surprising. Moreover, SPM8.0/AAL   41 is based on the manual delineation of a single brain, thus reducing the amount of structural information available in segmentation training data. Furthermore, manually delineated hippocampal ROIs built into SPM8.0/AAL include the uncus (this was not the case for the current study since manual segmentations did not include uncal regions as part of the hippocampus), and are purposely extended beyond the gray matter layer to account for the low spatial resolution of fMRI data. The current data suggests hippocampal segmentation with SPM8.0/AAL is the least congruent with manual measures. Based on the current data, use of this software for volumetric analysis of the hippocampus is not recommended over alternative automated methods such as FreeSurfer 5.1 and FSL/FIRST.   4.5 Manual vs. Automated Segmentation Methods: The Effect of Bilateral Hippocampal Cavities  When MRI data included hippocampal cavities the intraclass relationships between manual segmentation vs. FreeSurfer and manual segmentation vs. FSL/FIRST were decreased, but intraclass correlations between manual segmentation and SPM8.0/AAL were not notably affected. Overall, these data support the predicted decrease in the performance of automated segmentation methods when MRI data displayed morphological anomalies. The intraclass comparison between manual segmentation and FreeSurfer was reduced by a greater magnitude than manual vs. FSL/FIRST when MRI data included bilateral hippocampal cavities vs. when MRI data did not include cavities.  Variation in the composition of training datasets likely accounts for the majority of inconsistency in segmentation results obtained with automated software. Segmentation algorithms rely on structural information in training datasets to best match ROIs in the target image. Theoretically, any morphological variation (e.g. hippocampal cavities) excluded from training datasets would not be adequately modeled by automated segmentation algorithms. It is estimated that only 21-25% of the non-senior aged population have normal hippocampal spaces (Yoneoka et al., 2002). Under these criteria   42 for normalcy, training datasets used by FSL/FIRST may contain up to 80 MRI scans with hippocampal cavities (since a total of 336 MRI scans were included). However, since training datasets used by SPM8.0/AAL and FreeSurfer 5.1 were based on MRI scans collected from 1 and 10 individuals respectively, it may be the case that normal cavity variants were not included in these training dataset samples. It should also be noted that the MRI samples chosen for the current study contained hippocampal spaces that surpassed the criteria for normalcy described by Yoneoka et al. (2002).  In the current study, MRI scans were purposely chosen to reflect morphological variation in hippocampal regions. It was predicted that automated analysis would be compromised by large deviations in normal anatomical morphology. While FreeSurfer and FSL/FIRST performance appears to be affected, SPM8.0/AAL performance was unaltered. Recall that FreeSurfer and FSL/FIRST rely on computational frameworks centered by Bayesian probability. This model does not account for pathological spaces within the given ROI and this results in mislabeling of both internal and external boundaries. In contrast, SPM8.0/AAL relies on spatial warping into a semi-fixed framework that aligns both volumes based on sulcal landmarks and pre-determined subcortical regional masks. This model assumes all pixels within the predetermined subcortical mask are part of the pre-defined ROI.  Since abnormal hippocampal cavities are displayed as large dark spaces in regions that would otherwise be identifies as gray matter, it is, perhaps, unsurprising that modeling algorithms are susceptible to failure under these circumstances. FreeSurfer and FSL/FIRST, in particular, rely specifically on statistical information pertaining to the spatial relationship and MR intensities between voxels and their neighbors. The current data suggest FreeSurfer and FSL/FIRST segmentation algorithms are not designed to compensate for large dark clusters of black pixels in areas that usually contain intermediate intensities (gray matter, white matter, or a combination of both).  As previously mentioned, the intraclass relationship between manual segmentation and SPM8.0/AAL was poor, irrespective of whether or not MRI date displayed large deviations from normal hippocampal anatomy or not. Again, segmentation with this tool is not recommended over FreeSurfer or FSL/FIRST. Manual   43 evaluation of MRI data prior to processing with automated methods is, however, always recommended in order to identify factors that are likely to affect automated segmentation (e.g. large hippocampal spaces).   4.6 The Effect of Motion Artifact on Scan Rescan Reliability: Evaluation of Automated and Manual Methods  With respect to scan-rescan reliability, FreeSurfer outperformed FSL/FIRST when MRI data displayed severe motion artifact, however, FSL/FIRST performance was comparable to FreeSurfer when MRI data displayed only mild-moderate motion artifact. SPM8.0/AAL produced the most robust scan-rescan reliability irrespective of motion artifact severity with an intraclass correlation coefficient close to 1.0. Interestingly, when MRI data had severe motion artifact, mean volumes obtained with FreeSurfer and FSL/FIRST exhibited a statistically significant difference. The difference in mean volumes generated with manual versus FreeSurfer segmentation was non-significant when MRI data contained severe motion artifact.  The purpose of analyzing MRI data with a range of motion artifact severity was to evaluate the agreement between automated segmentation software and manual methods when MR image quality was non-optimal. To our knowledge, no existing studies have sought to determine the scan-rescan reliability of any segmentation method (e.g. automated or manual) using MRI data with motion artifact, although a recent study describes the scan-rescan reliability of FreeSurfer and FSL/FIRST in unflawed MRI data from healthy young adult subjects. MRI data with severe motion artifact appears to disrupt segmentation with FSL/FIRST, but not FreeSurfer 5.1, although FSL/FIRST performance does not appear to be impacted by mild-moderate motion artifact. In contrast to MRI data with abnormal hippocampal morphology, slight variation in the Bayesian computational frameworks of FSL/FIRST and FreeSurfer 5.1 are more likely to affect image parcellation than the composition of training datasets (scans with motion artifact are typically excluded from training datasets). The computational approach of FreeSurfer and FSL/FIRST are very   44 similar. While FreeSurfer uses a Bayesian modeling approach that capitalizes on likelihood distributions that predict how a voxel will be labeled given its set of coordinates and MR intensity, FSL/FIRST relies on Active Shape/Active Appearance Model (ASM/AAM) principles to conduct image segmentation. The sensitivity of FSL/FIRST to motion artifact may be explained by the heavy reliance of ASM/AAM on information about the relationship between a voxel?s intensity and the intensities of its neighboring voxels.  FSL/FIRST is likely more sensitive to blurring due to motion artifact since this results in poorer image resolution, and thus, a loss of information pertaining to the spatial relationship between voxels with different MR intensities.  In contrast, the FreeSurfer pipeline chosen for evaluation in the current study only models the hippocampus (all voxels outside a cuboid region of space surrounding the hippocampus are assigned an intensity value of 0). FreeSurfer algorithms are, therefore, only affected by blurring in temporal lobe areas, and not the remainder of the image.  In the current study, SPM8.0/AAL demonstrated the most robust scan rescan reliability of the 4 methods tested. In fact, we observed an intraclass correlation close to 1.0, regardless of whether SPM8.0/AAL volumes were obtained from scans with severe or mild-moderate motion artifact. This result leads us to question the sensitivity of the SPM8.0/AAL tool. A recent study of scan-rescan reliability of FreeSurfer and FSL/FIRST reported intraclass indices as low as 0.92 (FreeSurfer) and 0.80 (FSL/FIRST) for right hippocampal volumes segmented from scans collected from the same individual 1 hour apart (Morey et al., 2010). This study was conducted in a sample of healthy young adults, and excluded MRI data with motion artifact. In contrast, the participants included in the current study were scanned once per year and the MRI data had a range of both motion artifact and anatomical abnormality. As SPM8.0/AAL appears to be unaffected by gross adjustments in MR image quality, it is unlikely that this tool was able to identify slight changes in hippocampal volume, or identify volume reduction in a patients.  Given the current results, it is concluded that FreeSurfer?s method for subfield segmentation is superior to FSL/FISRT in the presence of severe motion artifact. While these data do not confirm whether SPM8.0/AAL is an effective tool for identifying slight   45 changes in hippocampal volume, the apparent insensitivity of this tool to gross adjustments to MR image quality suggests the validity of studies that have used this tool for volumetric analysis are questionable.   4.7 Evaluation of the Agreement Between FreeSurfer and FSL/FIRST    Regardless of whether MRI data displayed bilateral hippocampal spaces, the intraclass relationship between FreeSurfer 5.1 and FSL/FIRST segmentation was robust. The only exception was when MRI data displayed severe motion artifact. Under these circumstances, the relationship between FreeSurfer and FSL/FIRST was compromised.  Given that both FreeSurfer and FSL/FIRST use similar Bayesian computational frameworks, this result is neither surprising nor in disagreement with our a priori hypothesis. Remarkably, none of the studies that simultaneously compared the performance of FreeSurfer and FSL/FIRST to manual methods reported the correlation between the results obtained by the two automated programs (Doring et al., 2011; Lucarelli et al., 2013; Morey et al., 2009; Pardoe et al., 2009). For this reason it is difficult to conclude whether other studies observed stronger agreement between FreeSurfer and FSL/FIRST compared with manual methods, or whether they observed similar intraclass relationships to those described here. In a recent study focused on comparing the performance of several automated methods, where manual segmentation was not included (Seixas et al., 2010), the researchers reported intraclass indices between FreeSurfer and FSL/FIRST as 0.92 for the left hippocampus and 0.70 for the right hippocampus in a sample of cognitively normal subjects scanned at 3.0T (intraclass indices were lower for 1.5T MRI data). In the current study, similar correlations between FreeSurfer and FSL/FIRST volumes were observed whether MRI displayed bilateral hippocampal spaces or not (ICC=0.87 in both cases). This relationshipwas worsened when MRI data displays severe motion artifact (ICC=0.64). It should be mentioned, however, that although intraclass indices between FreeSurfer 5.1 and FSL/FIRST were relatively strong, mean volumes obtained with   46 FSL/FIRST were always significantly larger than FreeSurfer, except when MRI data display severe motion artifact.     4.8 Discrepancy in Volumetric Analysis of the Hippocampus: Manual vs. Automated Segmentation, and Postmortem vs. MRI-Derived Measurement    In the current study, all automated methods produced significantly larger hippocampal volumes compared with manual segmentation, this was true when MRI data did and did not include hippocampal cavities. SPM8.0/AAL volumes were significantly larger than those produced with both FreeSurfer and FSL/FIRST. All methods produced volumes that were significantly different from one another with the exception of Manual vs. FreeSurfer segmentation of 9 MRI scans with severe motion artifact.   These results are similar to previous validation studies, where FreeSurfer was found to overestimate hippocampal volume in a sample of probable Alzheimer?s patients (S?nchez-Benavides et al., 2010) and chronic major depression disorder (Tae, Kim, Lee, Nam, & Kim, 2008). Similarly, FSL/FIRST and FreeSurfer v3.05 tended to label more tissue as hippocampus compared with manual methods in a sample of individuals with temporal lobe epilepsy (Pardoe et al., 2009). Few studies have focused on validating hippocampal segmentation with SPM8.0/AAL, although a recent study comparing image registration techniques reported that differences in atrophy patterns identified with the AAL template were largely dependent on the choice of registration protocol (Pu et al., 2013). Another research group concluded that the volumetric error across a sample of healthy subjects was 73% for ROIs derived with the AAL tool, although in this study manual segmentation was not considered the gold standard (Rodionov et al., 2009).  Despite an abundance of literature discussing automated segmentation in neuroimaging analyses, there is little mention of whether the subcortical volumes generated via automated means are reflective of true anatomical dimensions. MRI-derived hippocampal volumes in the current study were similar to published values for all 4 segmentation methods. Specifically, manually derived hippocampal volumes were similar to those reported in MRI studies of Alzheimer?s (Bobinski et al., 1999; S?nchez-  47 Benavides et al., 2010) and affective bipolar patients (Doring et al., 2011). Hippocampal volumes obtained from FreeSurfer 5.1 and FSL/FIRST were also similar to those reported in literature (Doring et al., 2011; S?nchez-Benavides et al., 2010), as were those generated with the AAL atlas (Rodionov et al., 2009). Other validation studies failed to report mean values for either manual or automated measures (Bartzokis et al., 1993; Morey et al., 2009). Nonetheless, mean total hippocampal volumes generated by all 4 methods were at least 2 times greater than values reported in relevant postmortem studies (Bobinski et al., 1999; Falkai & Bogerts, 1986). Volume estimates generated with SPM8.0/AAL were the least consistent with postmortem studies, and at least 1.5 times greater than any other methods used in the current study.  It is possible that fixative solutions used to preserve postmortem samples shrink hippocampal tissues to a fraction of in vivo volumes (Stowell, 1941). Other explanations for the discrepancy in postmortem versus MRI derived volumes might be differences in population sample characteristics, and/or in the measurement protocol used (Keller & Roberts, 2009). Postmortem studies are subject to available data (e.g. usually lower sample sizes than MRI studies), and post mortem samples are typically from elderly/Alzheimer?s patients. Differences in the measurement protocol used in either postmortem or MRI study vary depending on the author?s definition of hippocampal boundaries.  While each of these explanations likely account for small percentage of measurement incongruity, the question of why MRI-derived hippocampal volumes are at least twice the size of reported postmortem values remains unanswered. It is possible that MRI and/or MRI analysis tools inflate the true anatomical dimensions of the total brain, thereby inflating the boundaries of the hippocampus. Studies of fresh whole brain volume in a large sample of individuals are rare. Recently, a study of intelligence and brain size recorded whole brain (cerebrum and cerebellum) weights from 100 recently deceased patients. Although, the authors failed to report whole brain volume, they did report cerebrum volume and weight. Cerebrum volume was approximately 94% of cerebrum weight for both right and non-right handed males and females, if the same method was applied to the whole brain (cerebrum and cerebellum), postmortem whole brain weights   48 could be approximated as 1173ml and 1295ml in females and males respectively. These estimates are consistent with manual assessment of the whole brain volume from MR images (Allen, Damasio, & Grabowski, 2002), but not brain volumes generated with FreeSurfer and SPM (reported as 1448-1463mm and 1744-1750mm respectively) (Nordenskj?ld et al., 2013).  Although whole brain volume appears to be inflated by automated MR imaging analysis tools, an approximate increase in overall volume between 15.2-29.4% does not account for hippocampal volume estimates 2 to 3 times the size of postmortem values. Moreover, manual assessment of total brain volume is similar to postmortem estimates, suggesting some other variable must account for manual overestimations both in this study and in others (Bobinski et al., 1999; Doring et al., 2011; S?nchez-Benavides et al., 2010). The most parsimonious explanation for differences in whole-brain volumes derived with automated methods is variation in registration template dimensions. Variations in template dimensions may account for differences in FreeSurfer and SPM estimates of whole brain volumes, however, it would not account for any disparities in total brain volume estimates with FSL/FIRST and SPM8.0/AAL since both platforms use the MNI152 template for registration. With respect to the manually derived volumes obtained in the current study, it is possible that Mango v2.5 inflates anatomical dimensions of the MRI image, or that ROIs were overly inclusive at the most anterior and posterior boundaries where the hippocampus is not easily distinguishable from surrounding tissue.  Structural MRI-based studies typically focus on group differences (e.g. patients vs. controls) rather than obtaining the most accurate measurement of true anatomical volumes. Nonetheless, the existence of a standardized set of anatomical dimensions of the human brain would greatly benefit future studies that involve volumetric analysis of subcortical structures. Current brain models developed by the software creators have not included their own standardized region-of-interest volumes in the public domain.  This is likely to arise from the difficulty involved in determining true standard volumes. Though theoretically viable, the development of such an atlas in reality is much more difficult and would require both age and gender-dependent brain maps across the lifespan to be   49 created. Not only does an infinite degree of structural variability exist within the human population, a surprising amount of discrepancy regarding true anatomical proportions exits in literature. One possible method for alleviating some uncertainty in future volumetric studies of the hippocampus is the development of a standard set of anatomical dimensions based on ROI volumes obtained from previous studies that have used manual segmentation methods, and the ROI volumes included in data training sets of automated segmentation platforms. While this procedure may not entirely elucidate the inconsistency between MRI-derived and postmortem volumes of the hippocampus and other subcortical brain structures, it would provide researchers with expectations about the results obtained by automated (or manual) means. As previously mentioned, several earlier validation studies failed to report mean hippocampal volumes obtained by either automated or manual measurements. Furthermore, information about manually derived ROI volumes included in the data training sets of automated segmentation platforms is also not readily available. It may prove beneficial to the greater field of structural neuroimaging if future authors submit raw volume measurements to a larger database. While the generation of standardized set of anatomical dimensions may not currently be feasible, the development of a peer-based database of reference volumes would, at the very least, provide guidelines for further structural imaging research groups.   4.9 Insights, Drawbacks, and Future Directions  Despite the inflation of true ROI dimensions, volumetric analysis of MR images provides the best possible estimate of subcortical brain volumes in vivo. While manual segmentation is still considered the most accurate method of volumetric analysis, it is laborious (segmentation of 1 hippocampus in the current study took between 2-4 hours) and subject to a large degree of intra and inter-rater variation. Similarly, substantial variability also exists between segmentation results obtained by automated means, and at present, ?fully automated? segmentation algorithms are incapable of consistently replicating manually derived results. In terms of the most appropriate approach to data analysis, MRI researchers are faced with several methodological obstacles: while manual   50 segmentation methods are inadequate for large datasets, the use of fully automated programs increases the likelihood of false negatives (or positives). Moreover, automated platforms are not designed for MRI data with compromised image quality, and regardless of segmentation method used, it?s unlikely the MRI-derived subcortical volumes are reflective of true anatomical dimensions.  In conclusion, volumetric analysis of hippocampal regions with FreeSurfer?s subfield segmentation tool is comparable to manual methods, however, manual tracing remains more accurate. Two alternative segmentation procedures are described for use in future volumetric studies of the hippocampus (or other subcortical structures in the human brain). First, ROIs obtained by automated programs like FreeSurfer and FSL/FIRST can be uploaded and sequentially edited by trained raters using GUI programs like Mango. This would theoretically reduce the amount of time it takes to manually trace an ROI from scratch, yet ensure the ROI best represents the boundaries of the subcortical structure of interest. Second, FreeSurfer offers a step-by-step method for developing unique sets of training data (probabilistic atlases describing the location and MR intensity of different subcortical structures) to be used with their whole-brain subcortical segmentation tool. It is likely that building unique segmentation atlases from a subset of MRI scans from the dataset of interest will be better suited to modeling individual variation within the group. This is especially useful in MRI studies where deviations from normalcy are expected (e.g. psychiatric samples). While respect to analysis of structural imaging data from Hotel Study, future volumetric studies may benefit from the construction of such an atlas.            51 Bibliography 1. Agartz, Ingrid, Momenan, Reza, Rawlings, Robert R, Kerich, Michael J, & Hommer, Daniel W. (1999). Hippocampal volume in patients with alcohol dependence. Archives of General Psychiatry, 56(4), 356.  2. Allen, John S, Damasio, Hanna, & Grabowski, Thomas J. (2002). Normal neuroanatomical variation in the human brain: An MRI-volumetric study. American journal of physical anthropology, 118(4), 341-358.  3. Ball, MJ. (1977). Neuronal loss, neurofibrillary tangles and granulovacuolar degeneration in the hippocampus with ageing and dementia. Acta neuropathologica, 37(2), 111-118.  4. Bartzokis, George, Mintz, Jim, Marx, Peter, Osborn, Diane, Gutkind, Daniel, Chiang, Frances, . . . Marder, Stephen R. (1993). Reliability of in vivo volume measures of hippocampus and other brain structures using MRI. Magnetic resonance imaging, 11(7), 993-1006.  5. Bliss, Tim VP, & L?mo, Terje. (1973). Long-lasting potentiation of synaptic transmission in the dentate area of the anaesthetized rabbit following stimulation of the perforant path. The Journal of physiology, 232(2), 331-356.  6. Bobinski, M, De Leon, MJ, Wegiel, J, Desanti, S, Convit, A, Saint Louis, LA, & Rusinek, H. (1999). The histological validation of< i> post mortem</i> magnetic resonance imaging-determined hippocampal volume in Alzheimer's disease. Neuroscience, 95(3), 721-725.  7. Bremner, J Douglas, Narayan, Meena, Anderson, Eric R, Staib, Lawrence H, Miller, Helen L, & Charney, Dennis S. (2000). Hippocampal volume reduction in major depression. American Journal of Psychiatry, 157(1), 115-118.  8. Bremner, J Douglas, Vythilingam, Meena, Vermetten, Eric, Southwick, Steven M, McGlashan, Thomas, Nazeer, Ahsan, . . . Garg, Pradeep K. (2003). MRI and PET study of deficits in hippocampal structure and function in women with childhood sexual abuse and posttraumatic stress disorder. American Journal of Psychiatry, 160(5), 924-932.  9. Convit, A, De Leon, MJ, Tarshish, C, De Santi, S, Tsui, W, Rusinek, H, & George, A. (1997). Specific hippocampal volume reductions in individuals at risk for Alzheimer?s disease. Neurobiology of aging, 18(2), 131-138.  10. den Hollander, Bj?rnar, Schouw, Marieke, Groot, Paul, Huisman, Henk, Caan, Matthan, Barkhof, Frederik, & Reneman, Liesbeth. (2012). Preliminary evidence of hippocampal damage in chronic users of ecstasy. Journal of Neurology, Neurosurgery & Psychiatry, 83(1), 83-85.  11. Doring, Thomas M, Kubo, Tadeu TA, Cruz, L Celso H, Juruena, Mario F, Fainberg, Jiosef, Domingues, Romeu C, & Gasparetto, Emerson L. (2011). Evaluation of hippocampal volume based on MR imaging in patients with bipolar affective disorder applying manual and automatic segmentation techniques. Journal of Magnetic Resonance Imaging, 33(3), 565-572.  12. Duvernoy, H.M. (2005). The Human Hippocampus: Functional Anatomy, Vascularization and Serial Sections with MRI: Springer.   52 13. Falkai, Peter, & Bogerts, Bernhard. (1986). Cell loss in the hippocampus of schizophrenics. European archives of psychiatry and neurological sciences, 236(3), 154-161.  14. Harrison, Paul J. (2004). The hippocampus in schizophrenia: a review of the neuropathological evidence and its pathophysiological implications. Psychopharmacology, 174(1), 151-162.  15. Keller, Simon S, & Roberts, Neil. (2009). Measurement of brain volume using MRI: software, techniques, choices and prerequisites. Journal of Anthropological Sciences, 87, 127-151.  16. Koolschijn, PC, van Haren, NE, Cahn, W, Schnack, HG, Janssen, J, Klumpers, F, . . . Kahn, RS. (2009). Hippocampal volume change in schizophrenia. J Clin Psychiatry.  17. Krebs, C., Weinberg, J., & Akesson, E.J. (2011). Neuroscience: Wolters Kluwer/Lippincott Williams & Wilkins Health. 18. Laakso, MP, Vaurio, O, Savolainen, L, Repo, E, Soininen, H, Aronen, HJ, & Tiihonen, J. (2000). A volumetric MRI study of the hippocampus in type 1 and 2 alcoholism. Behavioural brain research, 109(2), 177-186.  19. Lieberman, Jeffrey A, Chakos, Miranda, Wu, Huwei, Alvir, Jose, Hoffman, Elaine, Robinson, Delbert, & Bilder, Robert. (2001). Longitudinal study of brain morphology in first episode schizophrenia. Biological Psychiatry.  20. Lucarelli, RT, Peshock, RM, McColl, R, Hulsey, K, Ayers, C, Whittemore, AR, & King, KS. (2013). MR Imaging of Hippocampal Asymmetry at 3T in a Multiethnic, Population-Based Sample: Results from the Dallas Heart Study. American Journal of Neuroradiology, 34(4), 752-757.  21. Mathern, Gary W, Babb, Thomas L, Leite, Joao P, Pretorius, K, Yeoman, Kristin M, & Kuhlman, Paula A. (1996). The pathogenic and progressive features of chronic human hippocampal epilepsy. Epilepsy research, 26(1), 151.  22. Morey, Rajendra A, Petty, Christopher M, Xu, Yuan, Pannu Hayes, Jasmeet, Wagner II, H Ryan, Lewis, Darrell V, . . . McCarthy, Gregory. (2009). A comparison of automated segmentation and manual tracing for quantifying hippocampal and amygdala volumes. Neuroimage, 45(3), 855-866.  23. Morey, Rajendra A, Selgrade, Elizabeth S, Wagner, Henry Ryan, Huettel, Scott A, Wang, Lihong, & McCarthy, Gregory. (2010). Scan?rescan reliability of subcortical brain volumes derived from automated segmentation. Human brain mapping, 31(11), 1751-1762.  24. Morris, RGM, Garrud, Paul, Rawlins, JNP, & O'Keefe, John. (1982). Place navigation impaired in rats with hippocampal lesions. Nature, 297(5868), 681-683.  25. Nordenskj?ld, Richard, Malmberg, Filip, Larsson, Elna-Marie, Simmons, Andrew, Brooks, Samantha J, Lind, Lars, . . . Kullberg, Joel. (2013). Intracranial volume estimated with commonly used methods could introduce bias in studies including brain volume measurements. Neuroimage.  26. O'Driscoll, Gillian A, Florencio, Patrik S, Gagnon, Danny, Wolff, Anne-Lise V, Benkelfat, Chawki, Mikula, Lynn, . . . Evans, Alan C. (2001). Amygdala?  53 hippocampal volume and verbal memory in first-degree relatives of schizophrenic patients. Psychiatry Research: Neuroimaging, 107(2), 75-85.  27. Pardoe, Heath R, Pell, Gaby S, Abbott, David F, & Jackson, Graeme D. (2009). Hippocampal volume assessment in temporal lobe epilepsy: How good is automated segmentation? Epilepsia, 50(12), 2586-2592.  28. Patenaude, Brian, Smith, Stephen M, Kennedy, David N, & Jenkinson, Mark. (2011). A Bayesian model of shape and appearance for subcortical brain segmentation. Neuroimage, 56(3), 907-922.  29. Pooley, Robert A. (2005). Fundamental Physics of MR Imaging1. Radiographics, 25(4), 1087-1099.  30. Pu, Fang, Xu, Liqiang, Li, Deyu, Fan, Yubo, Niu, Haijun, & Li, Shuyu. (2013). Comparison of two nonlinear registration techniques to investigate brain atrophy patterns in normal aging. Journal of Neuroradiology.  31. Rodionov, Roman, Chupin, Marie, Williams, Elaine, Hammers, Alexander, Kesavadas, Chandrasekharan, & Lemieux, Louis. (2009). Evaluation of atlas-based segmentation of hippocampi in healthy humans. Magnetic resonance imaging, 27(8), 1104-1109.  32. S?nchez-Benavides, Gonzalo, G?mez-Ans?n, Beatriz, Sainz, Aitor, Vives, Yolanda, Delfino, Manuel, & Pe?a-Casanova, Jordi. (2010). Manual validation of FreeSurfer's automated hippocampal segmentation in normal aging, mild cognitive impairment, and Alzheimer Disease subjects. Psychiatry Research: Neuroimaging, 181(3), 219-225.  33. Scoville, William Beecher, & Milner, Brenda. (1957). Loss of recent memory after bilateral hippocampal lesions. Journal of neurology, neurosurgery, and psychiatry, 20(1), 11.  34. Seidman, Larry J, Faraone, Stephen V, Goldstein, Jill M, Goodman, Julie M, Kremen, William S, Toomey, Rosemary, . . . Caviness, Verne S. (1999). Thalamic and amygdala?hippocampal volume reductions in first-degree relatives of patients with schizophrenia: an MRI-based morphometric analysis. Biological Psychiatry, 46(7), 941-954.  35. Seidman, Larry J, Faraone, Stephen V, Goldstein, Jill M, Kremen, William S, Horton, Nicholas J, Makris, Nikos, . . . Tsuang, Ming T. (2002). Left hippocampal volume as a vulnerability indicator for schizophrenia: a magnetic resonance imaging morphometric study of nonpsychotic first-degree relatives. Archives of General Psychiatry, 59(9), 839.  36. Seixas, Fl?vio Luiz, Saade, D?bora Christina Muchaluat, Conci, Aura, de Souza, Andrea Silveira, Tovar-Moll, Fernanda, & Bramatti, Ivanei. (2010). Anatomical brain MRI segmentation methods: volumetric assessment of the hippocampus. IWSSIP, 17, 17-19.  37. Shorvon, Simon. (2000). The Human Hippocampus. Functional Anatomy, Vascularization and Serial Sections with MRI. By HENRI M. DUVERNOY. (Pp. viii+213; 255 figures; ?95.50/$159 hardback; ISBN 3 540 63205 0.) Berlin: Springer. 1998. Journal of Anatomy, 197(3), 513-518. doi: 10.1046/j.1469-7580.2000.197305133.x   54 38. Smith, Geoffrey N, Lang, Donna J, Kopala, Lili C, Lapointe, Jocelyne S, Falkai, Peter, & Honer, William G. (2003). Developmental abnormalities of the hippocampus in first-episode schizophrenia. Biological Psychiatry, 53(7), 555-561.  39. Spargo, E, Everall, IP, & Lantos, PL. (1993). Neuronal loss in the hippocampus in Huntington's disease: a comparison with HIV infection. Journal of Neurology, Neurosurgery & Psychiatry, 56(5), 487-491.  40. Stowell, Robert E. (1941). Effect on tissue volume of various methods of fixation, dehydration, and embedding. Biotechnic & Histochemistry, 16(2), 67-83.  41. Tae, Woo Suk, Kim, Sam Soo, Lee, Kang Uk, Nam, Eui-Cheol, & Kim, Keun Woo. (2008). Validation of hippocampal volumes measured using a manual method and two automated methods (FreeSurfer and IBASPM) in chronic major depressive disorder. Neuroradiology, 50(7), 569-581.  42. Tzourio-Mazoyer, N, Landeau, B, Papathanassiou, D, Crivello, F, Etard, O, Delcroix, N, . . . Joliot, M. (2002). Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain. Neuroimage, 15(1), 273-289.  43. Van Leemput, Koen, Bakkour, Akram, Benner, Thomas, Wiggins, Graham, Wald, Lawrence L, Augustinack, Jean, . . . Fischl, Bruce. (2009). Automated segmentation of hippocampal subfields from ultra-high resolution in vivo MRI. Hippocampus, 19(6), 549-557.  44. Velakoulis, Dennis, Pantelis, Christos, McGorry, Patrick D, Dudgeon, Paul, Brewer, Warrick, Cook, Mark, . . . Murrie, Vanessa. (1999). Hippocampal volume in first-episode psychoses and chronic schizophrenia: a high-resolution magnetic resonance imaging study. Archives of General Psychiatry, 56(2), 133.  45. Yoneoka, Yuichiro, Kwee, Ingrid L, Fujii, Yukihiko, & Nakada, Tsutomu. (2002). Criteria for Normalcy of Cavities Observed Within the Adult Hippocampus: High-Resolution Magnetic Resonance Imaging Study on a 3.0-T System. Journal of Neuroimaging, 12(3), 231-235.  46. Yoshida, Takeshi, McCarley, Robert W, Nakamura, Motoaki, Lee, KangUk, Koo, Min-Seong, Bouix, Sylvain, . . . Niznikiewicz, Margaret A. (2009). A prospective longitudinal volumetric MRI study of superior temporal gyrus gray matter and amygdala?hippocampal complex in chronic schizophrenia. Schizophrenia Research, 113(1), 84-94.    

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

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.24.1-0074082/manifest

Comment

Related Items