Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Adaptive ultrasound imaging to improve the visualization of spine and associated structures Zhuang, Bo 2018

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

Item Metadata

Download

Media
24-ubc_2018_november_zhuang_bo.pdf [ 57.15MB ]
Metadata
JSON: 24-1.0372823.json
JSON-LD: 24-1.0372823-ld.json
RDF/XML (Pretty): 24-1.0372823-rdf.xml
RDF/JSON: 24-1.0372823-rdf.json
Turtle: 24-1.0372823-turtle.txt
N-Triples: 24-1.0372823-rdf-ntriples.txt
Original Record: 24-1.0372823-source.json
Full Text
24-1.0372823-fulltext.txt
Citation
24-1.0372823.ris

Full Text

Adaptive Ultrasound Imaging to Improve theVisualization of Spine and Associated StructuresbyBo ZhuangM. S., University of Washington, 2007M. S., Tsinghua Uniersity, 2002B. S., Bejing Uniersity of Aero and Astro, 1999A THESIS SUBMITTED IN PARTIAL FULFILLMENTOF THE REQUIREMENTS FOR THE DEGREE OFDoctor of PhilosophyinTHE FACULTY OF GRADUATE AND POSTDOCTORALSTUDIES(Electrical and Computer Engineering)The University of British Columbia(Vancouver)October 2018c© Bo Zhuang, 2018The following individuals certify that they have read, and recommend to the Fac-ulty of Graduate and Postdoctoral Studies for acceptance, the dissertation entitled:Adaptive ultrasound imaging to improve the visualization of spine and associ-ated structuressubmitted by: Bo Zhuang in partial fulfillment of the requirements for the degreeof Doctor of Philosophyin Electrical and computing engineeringExamining Committee:Purang Abolmaesumi, Electrical and computer engineeringCo-supervisorRobert Rohling, Electrical and computer engineeringCo-supervisorPeyman servati, Electrical and computer engineeringSupervisory Committee MemberRoger Tam, Department of RadiologySupervisory Committee MemberTim Salcudean, Electrical and computer engineeringSupervisory Committee MemberEdmond Cretu, Electrical and computer engineeringSupervisory Committee Memberlaher ismail, Department of Anesthesiology, Pharmacology TherapeuticsExam chairiiAbstractVisualizing vertebrae or other bone structures clearly in ultrasound imaging is im-portant for many clinical applications such as ultrasound-guided spinal needle in-jections and scoliosis detection. Another growing research topic is fusing ultra-sound with other imaging modalities to get the benefit from each modality. In suchapproaches, tissue with strong interfaces, such as bones, are typically extractedand used as the feature for registration. Among those applications, the spine is ofparticular interest in this thesis. Although such ultrasound applications are promis-ing, clear visualization of spine structures in ultrasound imaging is difficult dueto factors such as specular reflection, off-axis energy and reverberation artifacts.The received channel ultrasound data from the spine are often tilted even after de-lay correction, resulting in signal cancellation during the beamforming process.Conventional beamformers are not designed to tackle this issue. In this thesis,we propose three beamforming methods dedicated to improve the visualization ofspine structures. These methods include an adaptive beamforming method whichutilizes the accumulated phase change across the receive aperture as the beam-forming weight. Then, we propose a log-Gabor based directional filtering methodto regulate the tilted channel data back to the beamforming direction to avoid bonesignal cancellation. Finally, we present a closed-loop beamforming method whichfeeds back the location of the spine to the beamforming process so that backscat-tered bone signals can be aligned prior-to the beamforming. Field II simulation,phantom and in vivo results confirm significant contrast improvement of spinalstructures compared with the conventional delay-and-sum beamforming and otheradaptive beamforming methods.iiiLay SummaryClear visualization of vertebrae structures are important for many clinical appli-cations such as ultrasound-guided spinal needle injections and scoliosis detection.However, it is challenging to image bones using ultrasound since the soft tissue-bone surfaces act like mirrors which could reflect ultrasound beams to a differentdirection. Therefore, this thesis proposes several methods to regulate the bone sig-nals back to the desired direction. As a result, the contrast, sharpness, sensitivityand specificity of vertebrae structures as significantly enhanced.ivPrefaceThis thesis is primarily based on four manuscripts, resulting from the collaborationbetween multiple researchers. All publications have been modified to make the the-sis coherent. Ethical approval for conducting this research has been provided bythe Clinical Research Ethics Board, certificate numbers: H11-02594, H09-01423,H07-01691, and SCOMP-003-07.A study described in Chapter 2 and 3 have been published in:• Bo Zhuang, Robert Rohling and Purang Abolmaesumi, Accumulated anglefactor-based beamforming to improve the visualization of spinal structuresin ultrasound images, IEEE transactions on ultrasonics, ferroelectrics, andfrequency control, 65(2):210-222, 2018.• Bo Zhuang, Robert Rohling and Purang Abolmaesumi, Phase-factor basedbeamforming to improve the visualization of hyper-echoic targets, MedicalImaging 2017: Ultrasonic Imaging and Tomography 10139, 1013910.The contribution of the author was in developing, implementing, and evaluat-ing the accumulated angle factor-based beamforming method. Dr. Rohling andDr. Abolmaesumi made suggestions in improving the methodology. Dr. Rohlinghelped with the purchse of the spine phantom. All co-authors contributed to theediting of the manuscript. Dr. Svetoslav Nikolov contributed in the discussion ofField II specular reflection simulation strategy. Dr. Jorgen Jensen helped with thediscussion on artifacts related to bone ultrasound imaging.vA version of Chapters 4 has been published in:• Bo Zhuang, Robert Rohling and Purang Abolmaesumi, Directional log-Gaborfiltering on the pre-beamformed channel data to enhance hyper echoic struc-tures, Ultrasonics Symposium (IUS), 2017 IEEE International, 1-4The contribution of the author was in developing, implementing, and evalu-ating the proposed directional Log-Gabor filtering method. Dr. Rohling and Dr.Abolmaesumi helped with their valuable suggestions in improving the methodol-ogy. Dr. Rohling helped with the purchse of the spine phantom. Dr. Abolmaesumiadded the region-of-interest analysis method. All co-authors contributed to theediting of the manuscript.A version of Chapters 5 has been accepted for publication in:• Bo Zhuang, Robert Rohling and Purang Abolmaesumi, Tensor-based off-axis noise suppression in pre-beamformed data to enhance the visualizationof hyper echoic structures, Ultrasonics Symposium (IUS), 2016 IEEE Inter-national, 1-4The contribution of the author was in developing, implementing, and evalu-ating the tensor-based off-axis noise suppression method. Dr. Rohling and Dr.Abolmaesumi helped with their valuable suggestions in improving the methodol-ogy. All co-authors contributed to the editing of the manuscript.viTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiLay Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiList of Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxvAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxviDedications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxvii1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Ultrasound basics . . . . . . . . . . . . . . . . . . . . . . . . . . 41.2 Ultrasound-guided lumbar spine needle injections . . . . . . . . . 51.3 Specular reflection . . . . . . . . . . . . . . . . . . . . . . . . . 81.4 Beamforming . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91.5 Adaptive beamforming methods . . . . . . . . . . . . . . . . . . 121.6 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14vii2 Simulation and experimental setup . . . . . . . . . . . . . . . . . . . 162.1 Field II simulation . . . . . . . . . . . . . . . . . . . . . . . . . . 162.1.1 Specular reflection Field II simulation methods . . . . . . 172.1.2 Specular reflection Field II simulation results . . . . . . . 182.2 Ultrasound experiments: vertebrae phantom studies . . . . . . . . 202.3 Ultrasound experiments: in vivo volunteer studies . . . . . . . . . 242.3.1 Quantitative analysis for proposed beamforming methods . 242.3.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . 263 Accumulated angle factor based beamforming . . . . . . . . . . . . 273.1 Sine wave demonstration . . . . . . . . . . . . . . . . . . . . . . 273.2 Equation and flow chart for AAF . . . . . . . . . . . . . . . . . . 303.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313.3.1 Field II simulation results . . . . . . . . . . . . . . . . . 313.3.2 Ultrasound experiment: vertebrae phantom results . . . . 323.3.3 Ultrasound experiment: vertebrae in vivo results . . . . . 333.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 374 Directional filtering and structure tensor-based masking . . . . . . . 494.1 Theoretical background and algorithm overview . . . . . . . . . . 504.1.1 Log-Gabor directional filtering . . . . . . . . . . . . . . . 514.1.2 Tensor-based masking . . . . . . . . . . . . . . . . . . . 534.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 564.2.1 Field II simulation results . . . . . . . . . . . . . . . . . 564.2.2 The phantom experiments results . . . . . . . . . . . . . 584.2.3 The in vivo results . . . . . . . . . . . . . . . . . . . . . 594.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 675 Region of interest based closed-loop beamforming for spinal ultra-sound imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 705.1 Algorithm overview . . . . . . . . . . . . . . . . . . . . . . . . . 715.1.1 Bone location detection . . . . . . . . . . . . . . . . . . . 72viii5.1.2 Tensor-based filtering . . . . . . . . . . . . . . . . . . . . 735.1.3 Data alignment . . . . . . . . . . . . . . . . . . . . . . . 745.1.4 Simulation and experiments . . . . . . . . . . . . . . . . 755.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 765.2.1 Tensor-based filtering results . . . . . . . . . . . . . . . . 765.2.2 Simulation results . . . . . . . . . . . . . . . . . . . . . 775.2.3 Phantom results . . . . . . . . . . . . . . . . . . . . . . . 795.2.4 In vivo results . . . . . . . . . . . . . . . . . . . . . . . . 815.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 835.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 856 Conclusions and future work . . . . . . . . . . . . . . . . . . . . . . 906.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 906.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 946.2.1 Algorithm acceleration . . . . . . . . . . . . . . . . . . . 946.2.2 Robustness verification . . . . . . . . . . . . . . . . . . . 956.2.3 Translation to clinical practice . . . . . . . . . . . . . . . 956.2.4 Combination with artificial intelligence . . . . . . . . . . 96Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98ixList of TablesTable 2.1 The parameters for the Field II simulation. . . . . . . . . . . . 22Table 2.2 The system parameters for the vertebrae phantom and volunteerstudies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25xList of FiguresFigure 1.1 Ultrasound front-end and back-end processing flow chart. Beam-forming is one of the most important processing steps to in-crease the ultrasound signal to noise ratio. . . . . . . . . . . . 5Figure 1.2 (a) shows an example image from the planning stage of anepidural injection. The physician uses ultrasound to locate andmark each lumbar spine level and then decide the needle inser-tion path. (b) shows an actual needle insertion plot where theneedle has to penetrate through the gap between vertebrae andreach the epidural space. . . . . . . . . . . . . . . . . . . . . 7Figure 1.3 (a) shows common spine ultrasound imaging planes includethe para-sagittal view (shown as the purple slice) and the trans-verse view (shown as the yellow slice). An example of ultra-sound para-sagittal view image is shown in (b) and an exampleultrasound transverse view image is shown in (c). Accordingto the para-sagittal view image, the bone surface (laminae) arethe saw-tooth shape hyper-echoic wave. The tissue underneaththe bone surface is dark since most acoustic energy is reflectedback from the bone surface. The bone surface is blurry due toreasons mentioned in the previous section. . . . . . . . . . . . 8xiFigure 1.4 Three common interaction scenarios between the ultrasoundwave and the vertebrae surface. The transducer head is dis-played on the top of the figure and the thick solid lines shownin the bottom represent the vertebrae surface. The arrows showthe center of the ultrasound beam direction and the curvesaround the arrow display the actual beam shape. (a) The ul-trasound beam is perpendicular to the bone surface. (b) Theultrasound beam interacts with the bone surface in an angle of90 - θi degrees. (c) The center of the beam does not interactwith the bone but the side lobes do (the off-axis interference). 10Figure 1.5 The normalized delay-compensated channel data from threeultrasound bone interaction scenarios (in vivo data sets). (a)The ultrasound beam is perpendicular to the bone surface. (b)The ultrasound beam is non-perpendicular to the bone surface(θi is around 10 degrees). (c) The center of the beam does notinteract with the bone but the side lobes do interact with thebones (off axis interference). . . . . . . . . . . . . . . . . . . 11Figure 1.6 Delay-and-Sum (DAS) Beamforming mechanism for 128 re-ceive channels. The number of active receive elements/chan-nels is called the receive aperture (in this example: 128, fromchannel 0 to channel 127). The red arrow lines show the pulsetravel time difference between each individual channel to thecenter channel (channel 63). This difference reduces with theincrease of depth and eventually becomes close to zero (thecurves become flatter as the depth increases). After summa-tion, the resulting pulse shown on the right has a higher SNRsince the signal is correlated while the noise is not. . . . . . . 12Figure 2.1 The simulation experiment setup. The transducer is located onthe top and the mimicked ‘bone’ points are shown as the bluelines and located between 2 to 5 cm in depth. . . . . . . . . . 19xiiFigure 2.2 The tilting in receive channel data caused by translation. (a)and (b) show the received channel data before and after timedelay correction when the ‘bone’ points are located in the cen-ter of the transmit aperture. (c) and (d) show the received chan-nel data before and after time delay correction when the ‘bone’points are shifted 1 cm to the right of the transmit aperture. . . 20Figure 2.3 The tilting in receive channel data caused by rotation (non-specular reflection). (a) and (b) show the received channel databefore and after time delay correction when the ‘bone’ pointsare rotated with an angle of 0.1 radians counter clockwise (5.7degrees). (c) and (d) show the received channel data before andafter time delay correction when the ‘bone’ points are rotatedwith an angle of 0.2 radians counter clockwise (11.5 degrees). 21Figure 2.4 (a) and (b) show the channel data when the vertebrae surface isperpendicular to the ultrasound beam without and with delaycorrection. According to the figure, the channel data are par-allel to the horizontal direction after delay correction. (c) and(d) show the channel data when the ultrasound beam interactsthe vertebrae surface with an angle of 70 degrees (rotate 20degrees counterclockwise). The delay-corrected channel datatilt towards left. (e) and (f) show the channel data when theultrasound beam is 0.3 cm left of a vertebrae surface. As aresult, the channel data after delay correction also tilt towardsleft. This tilting follows the same trend as the Field II simula-tion results. . . . . . . . . . . . . . . . . . . . . . . . . . . . 23Figure 2.5 The manually outlined bone surface region and background.The region shown in white is considered as the bone surfacesignal and the background region surrounding the bones isshown in black. The CR is compared based on the mean in-tensities between these regions following Equation (2.1). . . . 26xiiiFigure 3.1 Demonstration of simplified bone-reflected channel data (af-ter delay correction) using a 2D sine wave with different tilt-ing levels. The sine wave has 4 full wavelengths along thedepth direction (y direction) and 192 samples along the chan-nel direction (x direction). Signals from perpendicular specularreflections have small or no tilting (a) while signals from non-perpendicular specular reflections and off-axis interference havelarge tilting angles (b) and (c). The left column shows the 2Dsine wave with k = 0, 0.05 and 0.1. The middle column showsthe phase change for the black horizontal line shown in the firstcolumn. Notice that the phase change is wrapped around. Theunwrapped phase change is shown in the right column. As thetilting angle increases, the unwrapped phase change gets larger. 38Figure 3.2 (a) The relation between the weight and the slope value, k,for each adaptive beamforming method in a demonstration 2Dsine wave study. The 2D sine wave represents the signal com-ing from the specular reflection targets (bones). The tilt of thesine wave is controlled by changing the slope value, k. Low kvalues mimic the received data from perpendicular specular re-flections. Medium k values represent non-perpendicular spec-ular reflections and off-axis interference. Diffuse scatteringis mimicked by much higher k values. The proposed methodbrings a continuous weight reduction as k increases. (b) Theestimated AAF values from an in vivo data set. The mean AAFvalue for overall vertebrae is around 28. Signals from non-perpendicular reflection and off-axis interference have AAFvalues around 50 to 100. The soft tissue has a mean AAF valuelarger than 120. Therefore, the proposed accumulate angle fac-tor method can successfully suppress signals from soft tissueand off-axis interference. . . . . . . . . . . . . . . . . . . . . 39xivFigure 3.3 For AAF beamforming, the time delay corrected and dynami-cally apodized channel data at a given sample depth are Hilberttransformed to estimate the angle variation across the aperture.These angle values are unwrapped. The accumulated anglechange is then calculated by the difference between maximumand minimum of the unwrapped angles across the aperture.The weight is one over this accumulated angle change. . . . . 40Figure 3.4 Field II specular reflection simulation results. This figure showsthe weight changes for each beamforming method when thebone points are shifted right from 0 to 4 cm away from thecenter of the transmit beam (lateral translation experiment).(a), (b) and (c) correspond to different simulations where the‘bone’ points are located in different depth (3 cm, 4 cm and 5cm). 4 cm is the transmit focus. All weights are normalizedto the maximum weight at the same depth and showed in dBscale. The proposed AAF method is shown in black. Com-pared with other beamforming methods, the AAF method hasa gradual reduction in weight as the ‘bone’ moves away fromthe center of the transmit beam. . . . . . . . . . . . . . . . . 41Figure 3.5 Field II specular reflection simulation results. This figure showsthe weight changes for each beamforming method when thebone points are rotated counter clock-wisely from 0 to 30 de-grees (rotation experiment). (a), (b) and (c) correspond to dif-ferent simulations where the ‘bone’ points are located in dif-ferent depth (3 cm, 4 cm and 5 cm). 4 cm is the transmit focus.All weights are normalized to the maximum weight at the samedepth and showed in dB scale. The proposed AAF method isshown in black. According to the figures, most beamformingmethods show similar behaviour when the tilting angle of the‘bone’ points increases. Wiener shows high weight values (lowsuppression) which matches the sine wave simulation results. . 42xvFigure 3.6 The beamforming results for the vertebrae phantom study: (a)DAS, (b) Wiener, (c) PCF, (d) CF, (e) GCF and (f) AAF method.The proposed AAF method shows the highest contrast with nolarge distortion in the vertebrae surface compared with othermethods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43Figure 3.7 Vertebrae in vivo results. The sagittal view of vertebrae forsix different beamforming methods: (a) DAS (b) Wiener, (c)PCF, (d) CF, (e) GCF and (f) the proposed AAF method. Thelaminae are the hyper-echoic wave shapes in the middle of theimage. The skin layers are the lines in the near field. Betweenthem are the muscle / fat tissue layers. The region under thelaminae show dark image since most of ultrasound signals arereflected back in the vertebrae surface. The dynamic range forall beamforming methods was set to 60 dB due to lower signalto noise ratio compared with the phantom. The CR betweenthe laminae and the surrounding soft tissue for DAS, Wiener,PCF, CF, GCF and AAF is 0.40, 0.56, 0.86, 0.73, 0.71 and0.93, respectively. The AAF method generates a high contrastlaminae surface against soft tissue. . . . . . . . . . . . . . . . 44Figure 3.8 Vertebrae in vivo results. The transverse view of vertebrae forsix different beamforming methods: (a) DAS (b) Wiener, (c)PCF, (d) CF, (e) GCF and (f) the proposed AAF method. Theskin layers are the lines in the near field. The hyper-echoiclines in the image represent different vertebrae structures suchas the spinous process, laminae and transverse process. Be-tween them are the muscle / fat tissue layer. Same as the sagit-tal view, the dynamic range for all beamforming methods wasset to 60 dB. The CR between the laminae and the surroundingsoft tissue for DAS, Wiener, PCF, CF, GCF and AAF is 0.49,0.62, 0.82, 0.75, 0.74 and 0.90, respectively. The AAF methodgenerates a high contrast laminae surface against soft tissue.The ROI is defined as the red box shown in (f). . . . . . . . . 45xviFigure 3.9 The results from several other in vivo studies. Since PCF showsbetter results compared with CF, GCF and Wiener beamform-ing, only PCF is shown with DAS and the proposed AAF re-sults. According to the results, the AAF shows the highest CR. 46Figure 3.10 The mean and variance of CR in 12 in vivo data sets. Themean CR for the DAS is 0.49, Wiener is 0.64, PCF is 0.82, CFis 0.77, GCF is 0.76 and AAF is 0.91. The t test results provethe significant difference between the proposed AAF methodwith other beamforming methods (p < 0.001). . . . . . . . . 47Figure 3.11 The ROC curve for the in vivo data set in Figure 3.8. Thesensitivity is defined as the percentage of pixels classified asbone pixels correctly inside the bone ROI. The specificity isdefined as the percentage of pixels classified as backgroundpixels (not bone pixels) correctly outside the bone ROI. Thearea under the ROC curve is 0.755 for Wiener, 0.863 for PCF,0.833 for CF, 0.8765 for GCF and 0.938 for AAF. . . . . . . . 48Figure 4.1 The processing chain of the proposed DFTM method. The pre-beamformed channel data from each transmit event are firstdelay compensated. Then, a region of interest (ROI) alongthe depth direction is selected. To avoid the near field strongreflections, the ROI is chosen starting with 1 cm below thetransducer surface and all the way to 15 cm which covers mostof the spinal structures. The ROI is then filtered using log-Gabor directional filtering. The noise generated from the di-rectional filter along with reverberation/soft tissue artifacts arethen masked out using the tensor-based method. Finally, thefiltered and masked channel data within the ROI are summedalong the channel direction to generate a single scan line justas regular DAS beamforming. Notice that the procedure is re-peated for each transmit event to generate the whole image. . . 52xviiFigure 4.2 Two example frequency response of the proposed log-Gabordirectional filter. (a) has a wide pass band both in the angleand frequency direction and (b) has a narrow pass band. In (a),the ω0 is set to 0.6, σ f is set to 1.6, A is set to 1. In (b), the ω0is set to 0.6, σ f is set to 0.18, A is set to 8. ω0 is normalizedspatial frequency. Selecting it as 0.6 is to ensure the pass bandcan cover most spatial frequency range. σ f and A are chosenbased on the angle and frequency pass band. The 2D filterkernel size is set as 15 so two full wavelengths of the inputchannel data can be covered. . . . . . . . . . . . . . . . . . . 53Figure 4.3 Field II specular reflector simulation results. (a) shows the re-ceived channel data collected when the specular reflector isperpendicular to the transmit ultrasound pulse. (b) shows thesame channel data as (a) but with the proposed DFTM method.(c) shows the channel data collected when the specular reflec-tor is tilted 15 degrees counterclockwise. (d) shows the samechannel data as (c) but with the proposed DFTM method. (e)and (f) show the peak beamformed channel data changing withtilting angle with and without the DFTM method. The signalfrom the bones is 2 to 6 dB higher than DAS beamforming ifusing the wide-band directional filter and 3 to 10 dB strongerif using the narrow-band filter. . . . . . . . . . . . . . . . . . 57Figure 4.4 The phantom results comparison between the DAS and DFTMmethod on a vertebrae phantom. (a), (c) and (e) are DASresults. (b), (d) and (f) are the corresponding results usingDFTM. The quantitative result shows that the CR for vertebraestructures is 0.86 in the DAS result and 0.98 in the proposedDFTM result. . . . . . . . . . . . . . . . . . . . . . . . . . . 58xviiiFigure 4.5 An example of in vivo transverse view image. The DAS beam-formed image is shown in (a). The dynamic range was set to 50dB. The channel data used to beamform the red line in (a) areshown in (c). Delay correction has been applied on the channeldata. The middle bright region in the channel data correspondsto the lamina surface. Log-Gabor directional filtered channeldata are shown in (d). The tensor-based masked channel dataare shown in (e). . . . . . . . . . . . . . . . . . . . . . . . . 61Figure 4.6 Examples of in vivo transverse view images. (a), (c) and (e)are DAS results. (b), (d) and (f) are DFTM outputs. The di-rectional log-Gabor filtered results show clear suppression onthe soft tissue, off-axis interference and reverberation artifacts.On the other hand, strong signals along the horizontal directionare maintained. . . . . . . . . . . . . . . . . . . . . . . . . . 62Figure 4.7 An example of in vivo para-sagittal view image. The DASbeamformed image is shown in (a). The dynamic range wasset to 50 dB. The channel data used to beamform the red linein (a) are shown in (c). Delay correction has been applied onthe channel data. The middle bright region in the channel datacorresponds to the lamina surface. Log-Gabor directional fil-tered channel data are shown in (d). The tensor-based maskedchannel data are shown in (e). . . . . . . . . . . . . . . . . . 63Figure 4.8 Examples of in vivo para-sagittal view images. (a), (c) and (e)are DAS results. (b), (d) and (f) are DFTM outputs. The di-rectional log-Gabor filtered results show clear suppression onthe soft tissue, off-axis interference and reverberation artifacts.On the other hand, strong signals along the horizontal direc-tion are maintained, resulting high contrast in bone surfaces.(c) and (d) show the anterior complex disappears in the DFTMresult (red boxes). (d) and (e) show the soft tissue segmentsget falsely enhanced in the DFTM result (red arrows). . . . . . 64xixFigure 4.9 An example of transverse view image filtered using two differ-ent pass-band as shown in Figure 4.2. The DAS beamformedoutput is shown in (a). The DFTM method using wide-banddirectional filtering is shown in (b). The DFTM method usingnarrow-band directional filtering is shown in (c). The origi-nal channel data for constructing the red line in (a) are shownin (d). The corresponding narrow-band and wide-band direc-tional filtered output for input (d) are shown in (e) and (f), re-spectively. The tensor-masked narrow-band directional filteredchannel data are shown in (g). The tensor-masked wide-banddirectional filtered channel data are shown in (h). . . . . . . . 68Figure 4.10 An example of para-sagittal view image filtered using two dif-ferent pass-bands as shown in Figure 4.2. The DAS beam-formed output is shown in (a). The DFTM method using wide-band directional filtering is shown in (b). The DFTM methodusing narrow-band directional filtering is shown in (c). Theoriginal channel data for constructing the red line in (a) areshown in (d). The corresponding narrow-band and wide-banddirectional filtered output for input (d) are shown in (e) and (f),respectively. The tensor-masked narrow-band directional fil-tered channel data are shown in (g). The tensor-masked wide-band directional filtered channel data are shown in (h). . . . . 69xxFigure 5.1 The flow chart of the proposed closed-loop method. The pro-cessing sequence of each processing block is marked with num-bers. First, the channel data are delay-compensated. Direc-tional filtering is then applied to align the channel data fol-lowed by beamforming. These steps are repeated to generateall the scan lines in an ultrasound image. Then, a rectangu-lar ROI containing bones is selected in the beamformed im-age. The bone locations are automatically detected by locatingthe first intensity peak from deep to shallow direction alongeach beamformed line within the ROI (bottom to top). Mean-while, a tensor-based image filtering is applied on the delaycorrected channel data. This filter suppresses soft tissue clutterand only maintains strong signals in the channel data. Cross-correlation is then performed on filtered channel data based onidentified bone locations. After cross-correlation alignment,another beamforming is performed. This process is repeateduntil all scan lines of an ultrasound image is generated. To fur-ther improve the bone visualization, the closed-loop process(steps 4, 6 and 7) can be repeated until the improvement ofbone is minimal. . . . . . . . . . . . . . . . . . . . . . . . . 72Figure 5.2 Point and disk phantom experiment results. (a) shows the pre-beamformed channel data received after a single transmit event.Dynamic receive beam delay has been applied on the data basedon relative location to the receive aperture. (b) shows the samepre-beamformed data but after structure tensor based filtering.(c) and (d) are the normalized beamformed results for the pointand disk phantom using DAS and tensor-based filtering, re-spectively. The CR of the point target and disk are both en-hanced. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76xxiFigure 5.3 Simulation results. (a) shows an example channel data re-ceived from simulated bones in Field II. In this case, the ‘bone’points are rotated counterclockwise for 10 degree from the hor-izontal position. As a result, the received channel data aftertime delay correction are tilted. This can be seen from the sig-nals around the 4 centimeter depth. Notice the signal from agrating lobe is shown in the weak lines on the top left cornerof the image. (b) shows the channel data after tensor-basedfiltering. The main-lobe signal from bones are maintainedwhile other noise such as from grating lobe are suppressedsince they are relatively weak. (c) shows the result after thecross-correlation alignment. The signals are aligned back tothe horizontal direction. As a result, the positive and negativedata will not be cancelled out during the beamforming process.The peak amplitude will not drop as the tilting increases. . . . 78Figure 5.4 The beamformed output amplitude change in dB as the tiltingincreases from 0 to 15 degrees. The signal drops quickly in theDAS beamforming while it maintains the same in the proposedclosed-loop beamforming method until it reaches the searchboundary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79xxiiFigure 5.5 A transverse view of a vertebrae phantom. (a) shows the DASbeamformed result. (b) shows the result from the proposedclosed-loop beamforming method. (c) shows the beamformedresult prior to scan conversion using the directional filteringmethod (DFTM) introduced in the previous chapter. DFTMclears out the noise under the bone surface which simplifiesthe bone detection procedure. The automatically detected bonesurface points are shown as the red line. They are the firstintensity peaks from the bottom to the top of each scan line.The contrast of the bones are greatly improved in (c) comparedwith the DAS result (a). Notice that the bone segment thathave been fed back to the beamforming process is especiallysharpened. The shape of the bone itself matches the result fromthe DAS result. There is no detectable distortion generatedfrom the proposed closed-loop method. . . . . . . . . . . . . 80Figure 5.6 A para-sagittal view of the vertebrae. The laminae are shownas hyper-echoic wave shapes in the middle of the image in(a)-(f). The DAS results are (a) and (b). The red line in (b)marks the position of the scan line which is generated by thechannel data in (g). The result from directional filtering is (c).The same result as (c) but without scan conversion is (d). Theautomatically-detected bone surface is shown as the red line in(e). In (g), signals from the lamina surface are high amplitudeechos at the depth of 4 cm. (h) shows the tensor-filtered (g).The cross-correlation-aligned channel data are (i). The finalclosed-loop beamforming result is (f). Compared with (c) and(f), not only the contrast and sharpness of the bone surface areimproved, the soft-tissue clutter is also suppressed. The redrectangular box shown in (d) represents the ROI box used forcontrast and ROC analysis. . . . . . . . . . . . . . . . . . . . 86xxiiiFigure 5.7 A transverse view of the vertebrae. The DAS results are (a) and(b). The spinous process is shown as the hyper-echoic regioninside the blue rectangular box. The result from directional fil-tering is (c). The same result as (c) but without scan conversionis (d). The automatically-detected bone surface is shown as thered line in (e). (g) shows the channel data for the spinous pro-cess along the scan line listed with the red line shown in (b).(h) shows the tensor-filtered (g). The cross-correlation alignedchannel data are (i). The final closed-loop beamforming resultis (f). Compared with (c) and (f), not only the contrast andsharpness of the bone surface are improved, structures such asthe spinous process can be clearly visualized. . . . . . . . . . 87Figure 5.8 More in vivo results for comparison for DAS, directional filter-ing and closed-loop beamforming. In all cases, both the direc-tional filtering and closed-loop beamforming increase the con-trast for bones. The soft tissue clutter is greatly reduced. Com-pared with the directional filtered result, the closed-loop beam-forming method further suppressed soft tissue noise throughtensor-based filtering. The bone surface is also sharper andthinner after the channel data alignment. One important im-provement is the suppression of strong soft tissue ligamentsabove the bones. This is very clear in the difference between(b) and (c). . . . . . . . . . . . . . . . . . . . . . . . . . . . 88Figure 5.9 Multiple iteration in closed-loop beamforming. The result fromDAS is shown in (a), closed-loop beamforming with one iter-ation is shown in (b) and closed-loop beamforming with twoiterations plus vertical enhancement is shown in (c). Com-pared with (b), the bone surface shown in red boxes are morecontinuous and smooth in (c). . . . . . . . . . . . . . . . . . 89Figure 5.10 The ROC curve for the in vivo data set in Figure 5.6. Thearea under the ROC curve is 0.93 for the AAF beamformingmethod, 0.97 for the directional filtering method and 0.99 forthe proposed closed-loop beamforming method. . . . . . . . . 89xxivList of AbbreviationsAAF Accumulated Angle Factor based beamformingADC Analog to Digital ConversionCF Coherence Factor beamformingCPU Central Processing UnitCR Contrast RatioCT Computed TomographyDAS Delay and Sum with beamformingDFTM Directional Filtering with Tensor based MaskingDTB Dual Tensor BasisFPGA Field Programmable Gate ArrayGCF Geberalized Coherence FactorGPU Graphic Processing UnitLNA Low Noise AmplificationLUT Look Up TableMRI Magnetic Resonance ImagingPCF Phase Coherence Factor beamformingPSF Point Spread FunctionRF Radio FrequencyROI Region Of InterestSNR Signal to Noise RatioST Structure TensorTGC Time Gain CompensationxxvAcknowledgmentsThe author would like to thank Dr. Purang Abolmaesumi and Dr. Robert Rohlingon their continuous support and guidance on my Ph.D. study. It has been a greatpleasure working with them.Special thanks are owed to my wife, parents and kids, whose have supported methroughout my years of education.xxviDedicationsTo my family, nothing can be done without your continuous support and help.xxviiChapter 1IntroductionUltrasound imaging is widely used for diagnosis [1] and interventional guidance [2].Compared with other imaging modalities, such as magnetic resonance imaging(MRI) and computed tomography (CT), ultrasound has inherent advantages of be-ing non-invasive and real time. A key disadvantage is that ultrasound has tradition-ally been unable to clearly depict bone and surrounding structures. Recently, withimprovements in ultrasound image beamforming and ultrasound imaging hard-ware, there has been increasing interest in using ultrasound to image bone andligament tissue, such as the spine. Example applications include ultrasound-guidedspinal needle injections and scoliosis detection [3, 4]. Another growing researchtopic is fusing ultrasound with other imaging modalities to get the benefit fromeach modality [5]. In such approaches, tissue with strong interfaces, such as bones,are typically extracted and used as the feature for registration [6–8]. Among thoseapplications, the spine is of particular interest in this thesis [9–11].Although such ultrasound applications are promising, it is still challenging toclearly image spine structures and locate the bone surface accurately [12]. Thebone surface is typically weakened in ultrasound images, therefore affecting theaccuracy in extracting the bone locations [13]. Many reasons contribute to thedegradation. Other than general reasons shared with soft-tissue ultrasound imag-ing such as limited transmit focusing (especially when away from the transmitfocus) [14] and speed-of-sound variation [15], there are several artifacts that areunique when imaging bones. Typically the ultrasound beam has higher energy1along the center of the transmit beam than other off-axis directions/regions. How-ever, due to the large impedance mis-match between bone and soft tissue, strongoff-axis/sidelobe artifacts can be generated when weak ultrasound side lobes in-teract with bony surfaces. In addition, the same reason causes a portion of theultrasound pulses bounces back and forth between bone layers, resulting in rever-beration artifacts under the bone surface [16]. Another challenge for bone ultra-sound imaging comes from the specular reflection between the ultrasound beamand bone interface [17]. When the ultrasound waves are not perpendicular to bonesurfaces, the reflected ultrasound signals could be tilted in the receive aperture,which weakens the bone signal. All these phenomenon/artifacts could result inpoor contrast for the bones.Previously, researchers have suggested modifications in various stages of theultrasound imaging pipeline to suppress these artifacts and improve the bone vi-sualization. These methods include transducer geometry changes, beamformingmodifications and image post processing [12, 18]. Mauldin et al. proposed usinga single element piston transducer with a physical lens and mechanically movedby a motion stage to generate the output image. Due to the lack of array struc-ture, the piston transducer does not create grating lobes which is the main reasoncontributing to the off-axis interference in bone ultrasound imaging [12]. Phantomresults showed high contrast enhancement on areas adjacent to bones and clearbone visualization. However, employing a piston transducer also means the trans-ducer geometry has to be changed and advantages from array transducers, suchas dynamic transmit and receive beamforming, cannot be achieved, resulting ina degraded resolution. Hacihaliloglu et al. suggested using phase-based imageprocessing methods to enhance the bone surface from standard B-mode imaging.The vertebrae surface can be significantly highlighted and extracted after process-ing [18]. One potential limitation of this approach is that image processing is in thelater stage of the ultrasound processing chain. Loss of signal to noise ratio in earlyprocessing steps (such as beamforming) is hard to recover in the later processingstage. For example, strong artifacts are often treated as signals in traditional sig-nal and image processing stages. Therefore, my research focuses on beamformingmodifications.Beamforming is one of the most important processing steps for medical ul-2trasound imaging [19]. The performance of the beamformer plays a key role inthe image quality. Delay and sum beamforming (DAS) is the most commonlyused beamforming method [20]. It demonstrates high image quality especiallywhen imaging soft tissues [21]. In DAS beamforming, the weight for each channelwithin the receive aperture is pre-designed (called apodization, typically a Hanningwindow). On the other hand, adaptive beamformers dynamically apply the weightbased on the data statistics to improve the resolution [22, 23]. The computationalcomplexity is therefore increased. Examples include minimum variance basedbeamforming [24, 25] and coherence based weighting [26, 27]. Several promis-ing adaptive beamforming methods are of special interest due to their computationsimplicity. These algorithms are Wiener beamforming [28, 29], phase coherencefactor beamforming (PCF) [30], coherence factor beamforming (CF) [31] and gen-eralized coherence factor beamforming (GCF) [32]. All these methods showedpromising results for contrast and resolution enhancement on general soft tissueultrasound imaging [33, 34]. However, most of these methods are not specializedfor improving the contrast of hyper-echoic targets like bones. The effect on contrastimprovement is also not clear given the complex ultrasound bone surface interac-tions mentioned above. For example, most adaptive ultrasound beamforming al-gorithms only change the apodization weights based on signal coherence [35, 36].The data used for beamforming are not dynamically adjusted. This means oncethe bone signals are tilted and the coherence between channel data drops, the datacannot be recovered. Therefore, the signal cancellation issue is not addressed bycurrent adaptive beamforming methods.My goal is to significantly improve the visualization of hyper-echoic struc-tures. Therefore, the performance of these adaptive beamforming methods onhyper-echoic structures is first evaluated. Based on the evaluation result, a seriesof beamforming methods are proposed which are dedicated to improve the visu-alizations of bones. My study is focused on spinal applications since it is one ofthe most commonly used ultrasound applications involving bones [3, 9, 37, 38],although the techniques are designed to be general. The detail on spinal ultrasoundimaging will be introduced later.This thesis is organized based on the following order: following a brief intro-duction on ultrasound, specular reflection and beamforming algorithms (Chapter31), a Field II simulation method for simulating specular reflection, phantom and invivo experiment set-up and performance evaluation strategy are discussed (Chapter2). An adaptive beamforming method, accumulated-angle-factor based beamform-ing, is then proposed to improve the contrast of bones. The result is compared withother beamforming methods (Chapter 3). Next, the contrast of bones is furtherimproved by directional filtering and tensor-based masking on the channel data(Chapter 4). A closed-loop beamforming method is then described where the posi-tion of bones are fed back to the beamforming process to help with the channel dataalignment (Chapter 5). In the closed-loop beamforming, structural-tensor based fil-tering is employed to clear the soft tissue clutters and highlight the hyper-echoicstructures. Directional filtering proposed in Chapter 4 is used to benefit the bonesurface detection. Finally, the advantages and disadvantages for each proposedmethod are discussed and a plan for future work is proposed (Chapter 6).1.1 Ultrasound basicsMedical ultrasound has a frequency range between 1 - 20 MHz. Higher frequencymeans a shorter pulse wavelength and therefore, higher resolution with a trade-offof higher attenuation. The ultrasound pulses are generated through an ultrasoundtransducer array which converts electrical signals into mechanical waves (and viceversa). The transducer array is composed of multiple transducer elements (typ-ically 128 or 192) so the focus and steering can be controlled by arranging thedelay for each transducer element. The generated pulses are transmitted to tissueand then, reflected and picked up by the same transducer. To enhance the signal-to-noise ratio (SNR), the data received by each transducer element (channel data) arebeamformed dynamically based on travel time difference between each elementand a point of interest. The beamformed radio-frequency (RF) data are then passedthrough a series of front-end and back-end processing steps to generate final ultra-sound images shown on the display. A typical ultrasound processing flowchart isshown in Fig 1.1. The front-end processing steps include channel data processing(analog to digital conversion (ADC), low noise amplification (LNA), time-gain-compensation (TGC), beamforming and RF data processing (demodulation, lowpass filtering and decimation). Channel data processing and beamforming enhance4Figure 1.1: Ultrasound front-end and back-end processing flow chart. Beam-forming is one of the most important processing steps to increase theultrasound signal to noise ratio.the SNR, while RF data processing removes the carrier frequency. The back-endprocessing steps include echo signal processing (envelope detection, log compres-sion, gamma correction), image enhancement, scan conversion and special imagingmodes such as color Doppler [39]. The echo signal processing aligns the range ofthe ultrasound data; the image enhancement suppresses the speckle noise caused byconstructive and destructive interference between scatters, while scan conversioncorrects the image geometry. The front-end processing is generally handled by afield programmable gate array (FPGA) and the back-end processing is handled insoftware (CPU/GPU).1.2 Ultrasound-guided lumbar spine needle injectionsFacet joint injections and epidural injections are commonly used lumbar spine nee-dle injections [40, 41]. They are widely used for pain management, labor deliveryand surgical anesthesia. Both applications require correct needle insertion pathplanning and real-time feedback on needle locations corresponding to surround-5ing anatomy structures to avoid tissue damage. The current gold standard to guidethe facet joint injection process is repeated fluoroscopic imaging. This results inradiation exposure on patients and operators [42]. For epidural injections, physi-cians often locate the needle insertion path by manual palpation and determine thereach of the epidural space by feeling relief of pressure to saline injection. Pre-vious studies show that this manual insertion procedure may result in undesiredneedle re-insertion and tissue damage especially by novices. Alternatively, ultra-sound imaging is increasingly being used to provide non-invasive guidance forspinal needle injections [10, 43–46]. Figure 1.2(a) shows an example image fromthe planning stage of an epidural injection. An anesthesiologist uses ultrasoundimaging to mark the location of each lumbar spine (typically the spinous process)and then decides the needle insertion location and path. Figure 1.2(b) shows anepidural needle insertion plot, where the needle passes through the skin and musclelayers and then penetrates along the gap between neighbouring spinous processesuntil the tip of the needle reaches the epidural space (shown as the red circle).Figure 1.3(a) shows common spine ultrasound imaging planes including thepara-sagittal view (shown as the purple slice) and the transverse view (shown asthe yellow slice). An example of the ultrasound para-sagittal view image is shownin Figure 1.3(b) and an example of the ultrasound transverse view image is shownin Figure 1.3(c) [47]. Specifically in Figure 1.3(b), the bone surface (laminae) isthe saw-tooth shape hyper-echoic wave. The region underneath the bone surfaceis darker since most acoustic energy is reflected back from the soft-tissue bonesurface. The spine surface is typically blurred and thick due to reasons such asnon-perpendicular reflections, reverberations and off-axis interference. Depend-ing on the relative angle between bone surface and the transmit ultrasound beam,some parts of the bone surfaces are hyper-echoic while others are weak and coulddisappear. The contrast between the bone and surrounding soft tissue is not high.As a result, there is general belief that ultrasound images of the spine are hardto interpret, which limits the application of ultrasound in needle guidance in thespine [48].One interesting line of research is generating a statistical vertebrae anatomicalshape model from CT images [49]. Then, based on detected bone surface andmodel fitting, the vertebrae shape is superimposed on the ultrasound image [50,6Figure 1.2: (a) shows an example image from the planning stage of an epidu-ral injection. The physician uses ultrasound to locate and mark eachlumbar spine level and then decide the needle insertion path. (b) showsan actual needle insertion plot where the needle has to penetrate throughthe gap between vertebrae and reach the epidural space.51]. In this case, bone structures including layers underneath the top bone surfacecould be revealed by the model, which increases interpretability. One pre-requisitefor correct model fitting is clear extraction of bone surfaces [52]. Apart from bonevisualization improvement, other research also investigated the general problem ofneedle detection and real-time needle position feedback [53–56].This thesis focuses on bone visualization improvement. I aim to achieve this bydesigning beamforming methods dedicated for bones. My goal is to improve thevisualization of the blurred bone surface and provide clear bone surface with highcontrast. By combing my methods with other techniques such as needle trackingand model fitting, the usability of ultrasound-guided spinal needle injections canbe improved. It should be noted that there is likely a trade-off of bone visualiza-tion with soft tissue visualization, but this trade-off is acceptable in many of theapplications listed above.7Figure 1.3: (a) shows common spine ultrasound imaging planes include thepara-sagittal view (shown as the purple slice) and the transverse view(shown as the yellow slice). An example of ultrasound para-sagittalview image is shown in (b) and an example ultrasound transverse viewimage is shown in (c). According to the para-sagittal view image, thebone surface (laminae) are the saw-tooth shape hyper-echoic wave. Thetissue underneath the bone surface is dark since most acoustic energy isreflected back from the bone surface. The bone surface is blurry due toreasons mentioned in the previous section.1.3 Specular reflectionFigure 1.4 shows three common interactions between the ultrasound wave and thevertebrae surface. The transducer array is displayed on the top of the figure and thethick solid lines shown in the bottom represent the vertebrae surface. The arrowsshow the center of the ultrasound beam direction and the curves around the arrowdisplay the actual beam shape. Due to the size and smoothness of vertebrae struc-8tures, the soft-tissue vertebrae interface can be considered as a specular reflectioninterface [12, 14, 57]. Figure 1.4(a) shows a scenario that the ultrasound beam isperpendicular to the bone surface. After time delay correction, the channel datawill be aligned horizontally and the peak of the energy is in the middle of the re-ceive aperture. Example channel data can be seen in Figure 1.5(a). Figure 1.4(b) isa scenario that the ultrasound beam interacts with the bone surface in an angle of90 - θi degrees (the incision angle, θi, is not zero). The peak of the transmit energyis therefore reflected to one side of the receive aperture. According to Snell’s law,the reflection angle, θr, is equal to the incision angle, θi. Figure 1.5(b) shows an ex-ample of the channel data received with the scenario two (θi is around 10 degrees),where the peak of the receive energy is shifted to the left. Although the time delaycalculation is still accurate for the center of the beam, signals from the side of thebeam including side lobes could be misaligned with centre signals due to the travelroute changes. Therefore, the receive channel data are undesirably tilted to the leftafter delay correction. Figure 1.4(c) shows the third scenario where the center ofthe beam does not interact with the bone but the side lobes do. This case is de-fined as off-axis interference since the signal is not intended. Because time delaycalculation is based on the center of the beam, the off-axis interference is tilted inthe delay compensated channel data. This can be seen from Figure 1.5(c). For softtissue surrounding the bone surface, diffusive scattering occurs and the character-istic of the transmit pulse is not preserved. This results in larger phase fluctuationscompared with three scenarios above. My goal is to highlight the actual signalsfrom the vertebrae surface (scenarios in Figures 1.4(a) and (b)) while suppressingoff-axis interference (Figure 1.4(c) and soft tissue).1.4 BeamformingAs mentioned, ultrasound beamforming is an important front-end processing step,which first aligns the phase of the received signal from each individual element, andthen performs summation. Since the signal is correlated and the noise is random,the signal to noise ratio (SNR) is improved after summation. A typical ultrasoundsystem performs beamforming on 32 to 256 different elements / channels. Moreelements generally mean higher SNR. To align the phase correctly, the time delay9Figure 1.4: Three common interaction scenarios between the ultrasoundwave and the vertebrae surface. The transducer head is displayed onthe top of the figure and the thick solid lines shown in the bottom repre-sent the vertebrae surface. The arrows show the center of the ultrasoundbeam direction and the curves around the arrow display the actual beamshape. (a) The ultrasound beam is perpendicular to the bone surface. (b)The ultrasound beam interacts with the bone surface in an angle of 90 -θi degrees. (c) The center of the beam does not interact with the bonebut the side lobes do (the off-axis interference).differences between any imaging points along the depth to all the elements are pre-computed based on their relative distance. For the geometry listed in Figure 1.4,the corresponding time delay, τs, of an element, Es, is calculated by:τs =(√(Z+R)2 +R2−2(Z+R)Rcos(α)+Z)cFs, (1.1)where Z is the distance between the transducer surface to the receive focusing pointalong the center of the receive aperture. R is the transducer radius and α is the an-gle between the center of the receive aperture and that specific transducer element.c is the speed of sound. Fs is the receive sampling frequency. Then, the aligneddata from each channel are summed to generate the beamforming output. Theconventional implementation is called delay-and-sum (DAS) beamforming. Thedetailed mechanism of the DAS procedure is shown in Figure 1.6. The rectanglesare transducer elements. The green pulses are the data received from each trans-ducer element (channel data). The number of active receive elements/channels iscalled the receive aperture (in this example: 128, from channel 0 to channel 127).10Figure 1.5: The normalized delay-compensated channel data from three ul-trasound bone interaction scenarios (in vivo data sets). (a) The ultra-sound beam is perpendicular to the bone surface. (b) The ultrasoundbeam is non-perpendicular to the bone surface (θi is around 10 degrees).(c) The center of the beam does not interact with the bone but the sidelobes do interact with the bones (off axis interference).The red arrow lines show the pulse travel time difference between each individualchannel to the center channel (channel 63). This difference reduces with the in-crease of depth and eventually becomes close to zero (the curves become more flatas the depth increases). After summation, the resulting pulse shown on the right11Figure 1.6: Delay-and-Sum (DAS) Beamforming mechanism for 128 receivechannels. The number of active receive elements/channels is called thereceive aperture (in this example: 128, from channel 0 to channel 127).The red arrow lines show the pulse travel time difference between eachindividual channel to the center channel (channel 63). This differencereduces with the increase of depth and eventually becomes close to zero(the curves become flatter as the depth increases). After summation, theresulting pulse shown on the right has a higher SNR since the signal iscorrelated while the noise is not.has a higher SNR since the signal is correlated while the noise is not.1.5 Adaptive beamforming methodsAdaptive beamforming methods have been widely studied to improve the imageresolution and suppress the side lobe/grating lobe artifacts. Instead of using pre-designed apodization weight for each channel within the receive aperture, adaptivebeamformers dynamically apply the weight based on the data statistics to improvethe resolution [22, 23]. Examples include minimum variance based beamform-ing [24, 25] and coherence based weighting [26, 27]. Several promising adaptive12beamforming methods are of special interest due to their computation simplic-ity. Typical adaptive beamforming methods include coherent-factor beamforming,generalized coherence factor beamforming, Wiener beamforming and phase coher-ence factor beamforming.In coherence factor (CF) based beamforming, the apodization weight is esti-mated by the ratio between the coherent summation (with sign) and in-coherentsummation (without sign) of the channel data [31]. In frequency domain, this rep-resents the portion of energy in the DC component over the energy across the spec-trum. If the channel data are strongly coherent, the weight is close to 1. Otherwiseit is close to zero. The equation for calculating CF is:CF =(∑N−1n=0 X(n))2N∑N−1n=0 X(n)2, (1.2)where X(n) is the data from channel n. N is the total number of active channels.Similar as CF, the generalized coherence factor (GCF) method increases the signalband from DC only to adjacent low frequency band which offers a more robustsignal estimation compared with CF when the noise level is high [32]. The equationto calculate GCF is:GCF =EsE, (1.3)where Es is the energy in the signal band of the Fourier transformed channel data(close to DC). E is the total energy of the Fourier transformed channel data. Fordelay-corrected channel data, energy close to DC is considered as the signal.Wiener beamforming is proposed for ultrasound imaging to adaptively weightthe ultrasound beamformed result based on the input variance [28]. Previous stud-ies show that the Wiener post filter can improve the image contrast and cancel outoff-axis/reverb noise in transcranial imaging [29]. Same as CF and GCF beam-forming, Wiener beamforming applies suppression directly on the beamformeddata:yout put =|S|2|S|2 +α−→w HPN−→wyinput , (1.4)where yinput is the regular beamformed result after applying proper time delay (thedelay-and-sum output), −→w is the weight vector across elements (typically a vec-13tor of ones for the conventional DAS beamforming), PN is the power of the noisecomponent, S is the signal component and α is a constant controlling the degree ofsuppression. The power of the noise is estimated through measuring the varianceof the channel data before beamforming. The signal component is estimated as theDAS beamforming output. Increasing α will result in increase in suppression lev-els. When the scale factor is set to the channel number (N = 192), the performancefor the Wiener post filter beamforming is the same as CF beamforming (This hasbeen reported in [28].).Similar as the Wiener beamforming, PCF beamforming analyzes the varianceamong the channel data but in the phase domain [30]. It sets the beamformingweight based on the standard deviation of phase values across the aperture. Theequation to estimate the PCF is:PCF = (1−√std(cos(φ))2 + std(sin(φ))2)p, (1.5)where p is a side lobe control parameter (set to 1). φ is the phase values across thechannel data after delay correction. Notice that the phase of each individual chan-nel data is estimated along the depth direction (not the aperture direction). Usingphase information instead of energy level could be beneficial since it is more sen-sitive to actual variations in the data regardless of amplitude. This is desirable tosuppress weak soft tissue noises since they generally have low amplitude and largevariation in phases. While for bone signals, the phase variation across the receiveaperture is typically small due to the wide-stretch in channel data. Therefore, utiliz-ing the phase change across the aperture direction can differentiate between bonesand soft tissue.1.6 ObjectivesMy goal is to highlight the actual signal from the vertebrae surface and suppressthe signal from the off-axis interference and soft tissue. Based on the characteris-tics of channel data from bone targets (specular reflectors, hyper-echoic), a usefulmechanism is therefore tracking the phase changes in the delay compensated chan-nel data across the aperture direction (channel direction). Another way to increase14the vertebrae signal is aligning the undesirably tilted bone signals in channel data(due to specular reflection) back to the beamforming direction. Signal cancella-tion due to the summation of positive and negative peaks can then be avoided inthe beamforming process. One straightforward solution is applying directionalfiltering on the channel data to regulate the tilted bone signals towards the beam-forming direction. To achieve a better alignment, a more sophisticate way wouldbe performing cross-correlation on data from neighbouring channels provided thelocation of bone is known. All these methods are performed in or prior to the beam-forming stage. Based on the analysis above, the following objectives are set in thisthesis: (1) appropriate simulation strategy and phantom experiments to mimic thebone-ultrasound interaction should be designed. Reasonable performance metricsfor evaluating the bone visualization performance should also be adopted. (2) Anadaptive beamforming method that tracks the phase change across the aperture di-rection will be designed and evaluated for bone visualization improvements. (3)Directional filtering and its effect on bone signal enhancement will be studied. (4)A beamforming method which aligns the tilted bone channel data through cross-correlation based on bone locations as the input will be designed and studied.15Chapter 2Simulation and experimentalsetupTo evaluate the performance of proposed beamforming methods and existing meth-ods on bone structures, simulations, phantom studies and in vivo volunteer studiesneed to be designed. Appropriate performance evaluation metrics also need to beselected. All these will be discussed in this chapter.2.1 Field II simulationThere is value in first simulating the specular reflection effects using a widely ac-cepted simulator such as Field II [58]. Since the current Field II software doesnot support direct simulation on specular reflectors, a two-step simulation strat-egy is developed to mimic specular reflection. The effects of non-perpendicularreflection are then analyzed in this new simulation environment. In a separate sim-ulation, channel data from soft tissue is simulated so the contrast between ‘bone’and surrounding soft tissue can be evaluated. Details are provided below.162.1.1 Specular reflection Field II simulation methodsField II has been widely used in simulating transducer impulse response, trans-mit pulses, pressure field and convolution with point scatterers. Previous researchdemonstrates its successful simulation on point targets, soft tissue structures andflow signals. However, current Field II software does not support direct simulationon specular reflections [58]. The reflected ultrasound pulse is not deflected to adifferent angle when the reflector is not perpendicular to the transmit ultrasoundbeam. Therefore, a two-step simulation strategy is proposed to mimic specular re-flection in Field II. Firstly, a set of ‘bone’ point targets are linearly placed in thefield. This can be seen from the blue horizontal line shown in Figure 2.1. Ultra-sound pulses are transmitted into the field from the transducer above (shown asblack array). This is defined as the first transmit event. Then, the ‘bone’ points aretreated as virtual transducer elements sending the specularly reflected ultrasoundwave back to the transducer surface. This is defined as the second transmit event.The transducer element wave forms in the second transmit event are signals re-ceived at the ‘bone’ locations from the first transmit event. The impulse responseof bones is set to one because there is no actual transducer involved in the sec-ond transmit event. Since signals arriving from the first transmit event have a delayacross the ‘bone’ points due to the beam shape and bone geometry, pulses transmit-ted in the second event can be tilted. In this case, the effect from specular reflectionis simulated. The detailed Field II simulation parameters are shown in Table 2.1.My simulation is designed based on simplicity but represents common ultrasoundimaging settings. It contains a linear array transducer with a transmit center fre-quency of 10 MHz, pulse length of 1.5 cycles and a pitch size around 0.7 of thewavelength (0.105 mm). The number of elements for the first transmit event is 127while the number of ‘bone’ points (the virtual transducer for the second transmitevent) is 1024. Although 127 does not match the actual number of elements thatwill be used in ultrasound experiments discussed in the following sections (192),it is chosen because of similar pitch size/wavelength ratio (which generates similarlevel of grating lobes) and less simulation time.The position and angles of the ‘bone’ points are changeable to mimic the tilting17and translation of the bones. According to Figure 2.1, the ‘bone’ can be located in adepth that is between 2 cm to 5 cm away from the transducer surface. The transmitfocus is fixed at 4 cm. To evaluate the performance on off-axis interference sup-pression, the ‘bone’ points are shifted right from 0 to 4 cm away from the center ofthe transmit beam. To mimic the effect of tilting generated by specular reflection,the ‘bone’ points are rotated counter clock-wisely from 0 to 30 degrees. All re-ceived channel data will go through the same delay correction process to generatean ultrasound scan line that is located along the center of the transducer geometry(regardless of the translation and tilting of ‘bone’ points).In soft tissue simulation, the number of point scatters used for simulation de-pends on the size of point spread function [58]. For a PSF size of 0.2 mm by 0.5mm by 7 mm, at least 10 scatters should be used. Higher density of scatters canbring more accurate simulation results but also requires higher computation time.12 is chosen in our simulation.2.1.2 Specular reflection Field II simulation resultsFigure 2.2 shows example channel data received from my simulation design. (a)and (b) show the received channel data before and after time delay correction whenthe ‘bone’ points are located in the center of the transmit aperture with a depth of4 cm. The time delay correction aligned the curved channel data from (a) intostraight lines in (b). (c) and (d) show the received channel data before and aftertime delay correction when the ‘bone’ points are shifted to one cm right of thetransmit aperture. In this case, the channel data become tilted and cannot be cor-rected by time delay adjustment.Figure 2.3 shows example channel data when the ‘bone’ targets are tilted. (a)and (b) show the received channel data before and after time delay correction whenthe ‘bone’ points are rotated with an angle of 0.1 radians counter clock-wisely (5.7degrees). (c) and (d) show the received channel data before and after time delaycorrection when the ‘bone’ points are rotated with an angle of 0.2 radians counter18Figure 2.1: The simulation experiment setup. The transducer is located onthe top and the mimicked ‘bone’ points are shown as the blue lines andlocated between 2 to 5 cm in depth.clock-wisely (11.5 degrees). Same as off-axis signals, delay correction would notbe able to align the channel data. Since the transmit beam has a beam width, oncethe bone is tilted, signals arrived from one side of the bone will arrive faster thanthe other side. This causes tilting in the received channel data. The amount oftilting in the channel data increases as the beam width increases.19Figure 2.2: The tilting in receive channel data caused by translation. (a) and(b) show the received channel data before and after time delay correctionwhen the ‘bone’ points are located in the center of the transmit aperture.(c) and (d) show the received channel data before and after time delaycorrection when the ‘bone’ points are shifted 1 cm to the right of thetransmit aperture.2.2 Ultrasound experiments: vertebrae phantom studiesIn addition to Field II simulation, vertebrae phantom studies are performed. A BKultrasound system with a curved linear array transducer is used for collecting pre-beamformed channel data (Analogic Corp., BK3500, Peabody, MA). A focusedtransmit with a focal depth of 5 cm and a center frequency of 3.5 MHz is used fortransmit and a 14 MHz sampling frequency is used for receive. The maximum re-ceive aperture for each scan line used in dynamic receive beamforming is 192. Forthe phantom experiment, a plastic artificial vertebrae (Xincheng Scientific Indus-20Figure 2.3: The tilting in receive channel data caused by rotation (non-specular reflection). (a) and (b) show the received channel data beforeand after time delay correction when the ‘bone’ points are rotated withan angle of 0.1 radians counter clockwise (5.7 degrees). (c) and (d)show the received channel data before and after time delay correctionwhen the ‘bone’ points are rotated with an angle of 0.2 radians counterclockwise (11.5 degrees).tries Co., Shanghai, China) made from polyvinyl chloride (PVC) is chosen. Theplastic vertebrae is merged in a water tank when channel data are collected. Theacoustic impedance of PVC is 3.27×106 kg/(m2s), while the acoustic impedanceof bones is 7.8× 106 kg/(m2s). This acoustic impedance difference means moreacoustic energy will penetrate through the water-bone surface than soft tissue-boneinterface. However, due to the size and smoothness of the phantom surface, specu-lar reflection can be generated in the water-plastic surface, which makes the phan-tom study anatomically realistic.21Table 2.1: The parameters for the Field II simulation.Transducer pitch 0.105 mmTransducer number of elements 128Transmit F number 3Transmit center frequency 10 MHzTransmit pulse length 1.5 cyclesTransmit pulse 6 dB bandwidth 0.9 of center frequencyTransmit focus 4 cmNumber of bone points 1024Length of bone points 4 cmDepth of bone points 3 - 5 cmTilting angles 0 - 30 degreesTranslation distance 0 - 4 cmThe first goal of the phantom study is to validate the Field II simulation re-sults. Figure 2.4(a) and (b) show the channel data when the vertebrae surface isperpendicular to the ultrasound beam without and with delay correction. Accord-ing to the figure, the channel data are parallel to the horizontal direction after delaycorrection. Figure 2.4(c) and (d) show the channel data when the ultrasound beaminteracts the vertebrae surface with an angle of 70 degrees (rotate 20 degree coun-terclockwise). The delay-corrected channel data tilt towards left. Figure 2.4(e) and(f) show the channel data when the ultrasound beam is 0.3 cm left of a vertebraesurface. As a result, the channel data after delay correction also tilt towards left.This tilting follows the same trend as the simulation results shown in Figure 2.2and Figure 2.3. As the rotation and translation for the bone surface increase, theamount of tilting in the channel data also increase.The next step of the phantom study is to evaluate the performance of differentbeamforming algorithms. The goal for using a phantom is not only evaluating theimprovements in contrast for a proposed method but also confirming there is no22Figure 2.4: (a) and (b) show the channel data when the vertebrae surface isperpendicular to the ultrasound beam without and with delay correction.According to the figure, the channel data are parallel to the horizontaldirection after delay correction. (c) and (d) show the channel data whenthe ultrasound beam interacts the vertebrae surface with an angle of70 degrees (rotate 20 degrees counterclockwise). The delay-correctedchannel data tilt towards left. (e) and (f) show the channel data whenthe ultrasound beam is 0.3 cm left of a vertebrae surface. As a result,the channel data after delay correction also tilt towards left. This tiltingfollows the same trend as the Field II simulation results.large distortion generated in this method. A fixed F-number of 1 is used to dy-namically assign the receive aperture size along the depth. The output scan lineis generated by summing the delay corrected channel data horizontally. Hanningapodization is employed for DAS beamforming. Each transmit and receive eventtypically results in one beamformed scan line and this transmit and receive pro-cess is repeated until all scan lines have been generated to form an output image.Detail system parameters used for the phantom study are shown in Table 2.2. Thebeamformed data pass through demodulation, envelope detection, log compressionand scan conversion. To have a fair comparison, the output images from different23beamforming methods are displayed in the same dynamic range (e.g., -80 to 0 dB).No post processing methods are performed after beamforming to avoid additionalfactors that may affect the results.2.3 Ultrasound experiments: in vivo volunteer studiesApart from phantom studies, in vivo experiments are also conducted. 60 datasets from 5 volunteers were collected including several common views for verte-brae ultrasound imaging (sagittal and transverse). Of these 60 data sets, 30 are fromsagittal view and 30 are from transverse view. All volunteers (4 male and 1 female)were healthy and had no history of spinal disease. Following written consent, vol-unteers were seated on a stool while the probe was placed over the lumbar region(L1 - L5). The ultrasound system parameters are the same as phantom studies. Thecollected channel data were stored offline and later used for evaluating differentbeamforming algorithms. To have a fair comparison, the output images from dif-ferent beamforming methods are displayed in the same dynamic range (e.g., -60 to0 dB). In our experiments, spatial compounding is not employed since the currentBK system does not support channel data collection when the spatial compoundingis on.2.3.1 Quantitative analysis for proposed beamforming methodsTo quantitatively evaluate the effect of the proposed method, performance evalua-tion metrics need to be chosen. Based on literature search, the contrast ratio (CR)metric proposed in [59] is a good match since it is a widely accepted contrast evalu-ation metric and has been used to quantify the bone visualization performance [59].In my CR measurement, signal ROIs are chosen corresponding to the vertebraestructure and noise ROIs are chosen corresponding to adjacent background. Thesurface of the vertebrae structure was manually outlined by independent ultrasoundresearchers with guidance from professional sonographers. Examples are shown inFigure 2.5. The vertebrae surface is shown in white and the background is shownin black. Notice that the ROI is generated based on DAS beamforming result forunbiased comparison.24Table 2.2: The system parameters for the vertebrae phantom and volunteerstudies.Transducer type. Curved linearCenter Freq. 3.5 MHzSampling Freq. 14 MHzTransducer Pitch 0.327 mmFocus depth 13 cmBeamforming Freq. 14 MHzNum. of Receive Channels 192Transmit F number 3Receive F number 1Apodization in DAS HanningInput I,Q range 16-bitCR=|I1− I2|√I21 + I22, (2.1)where I1 and I2 are the mean intensities of selected ROIs representing signal andnoise, respectively. Therefore, the range of CR is between 0 to 1.Other than the CR analysis, variance and sensitivity/specificity analyses arealso performed. These analyses are based on normalized beamforming outputs andmanually outlined bone ROIs. The mean variance inside and outside the bone ROIsfor different adaptive beamforming weights are compared. The sensitivity is de-fined as the percentage of pixels classified as bone pixels correctly inside the boneROI. The specificity is defined as the percentage of pixels classified as backgroundpixels (not bone pixels) correctly outside the bone ROI. The ROC curve can thenbe created. The area under the ROC curve can be analyzed to estimate sensitivityand specificity.Other useful metrics to evaluate the bone surface quality include the ability to25Figure 2.5: The manually outlined bone surface region and background. Theregion shown in white is considered as the bone surface signal and thebackground region surrounding the bones is shown in black. The CR iscompared based on the mean intensities between these regions follow-ing Equation (2.1).enhance weak vertebrae surface, level of false enhancement on soft tissue ligamentlayers above the bones, sharpness of the bone surface etc. All these factors will beevaluated in this thesis.2.3.2 SummaryIn this chapter, the Field II simulation strategy, phantom and in vivo experimentsand reasonable performance metrics for evaluating the performance in bone visu-alization are introduced. In next chapters, beamforming methods that can improvethe visualization of bones will be proposed. For each proposed algorithm, simula-tion, phantom and in vivo experiments discussed in this chapter will be employed.26Chapter 3Accumulated angle factor basedbeamformingBased on the channel data plot in Figure 1.5 and the Field II simulation results,it seems possible to distinguish between signals from bones and off-axis interfer-ence by tracking the phase changes in the delay compensated channel data acrossthe aperture direction (channel direction). Therefore, an accumulated angle fac-tor based beamforming method (AAF) is proposed for bone surface enhancement.This approach applies a Hilbert transform on delay compensated channel dataacross the receive aperture. The accumulated phase change across the receive aper-ture is then calculated and utilized as the weight in the beamforming output.3.1 Sine wave demonstrationThe description of AAF is started by offering a simplified demonstration of reflec-tion off a bone surface (after time delay correction) using the following 2D sinewave:y= sin(2piff s(t+ kd)), (3.1)where f is the center frequency of the ultrasound beam. f s is the sampling fre-quency. t is the number of samples along the depth direction (y direction). d isthe number of channels along the transducer aperture direction (x direction). k27is the slope of the phase offset and controls the tilt in the receive channel data.Since channel data from non-perpendicular reflections and off-axis interferencehave larger tilt levels than perpendicular specular reflections (after delay correc-tion), the data received with perpendicular reflections, non-perpendicular reflec-tions and off-axis interference can be mimicked with different k values. In casesof perpendicular reflection, there is no phase offset between neighbouring receivechannels and k is set to 0. In cases of non-perpendicular reflection and off-axisinterference, there is a tilt in the receive channel data and k would be a non-zerovalue. Assuming k is a constant, for a certain depth tc, the sine wave across the xdirection has the following format:y′ = sin(2pif kfs(d+tck)). (3.2)The accumulated phase change across the aperture direction can be calculatedby:A= 2Npif kfs, (3.3)where N is the total number of active channels. As the tilting increases, the accu-mulated phase change across the aperture direction, A, also increases. Figure 3.1shows simulation examples with k = 0, k = 0.05 and k = 0.1. The simplifieddemonstration sine wave echo has 4 full wavelengths along the depth direction(y direction) and 192 samples along the channel direction (x direction). As k in-creases from 0 to 0.1, the accumulated phase change across the aperture directionincreases from 0 to 12.Therefore, 1A is set as the beamforming weight so the off-axis interference canbe suppressed and the bone signal can be maintained. This method is definedas the accumulated angle factor based beamforming. A is defined as AAF, theaccumulated angle factor, and the beamforming weight, w, is defined as the inverseof AAF :w=1AAF. (3.4)The AAF method is compared with adaptive beamforming methods such as28CF, GCF, Wiener and PCF beamforming. In coherence factor (CF) based beam-forming, the beamforming weight is estimated based on the energy between the DCcomponent and the overall energy across the spectrum. The generalized coherencefactor (GCF) method increases the signal band to offer a robust signal estimationwhen the noise level is high. Wiener beamforming suppresses the output based onthe variance among the channel data. Similarly, PCF beamforming analyzes thevariance among the channel data but in the phase domain. The detail introductionon these beamforming methods are listed in chapter 1.Figure 3.2(a) plots the relations between the weight value and the slope value,k, for each adaptive beamforming method. Notice that the weight is the meanweight across the depth direction. The ratio between fs and f is set to 12. In manyultrasound systems the sampling frequency is about 4 times of the ultrasound fre-quency. The sampling frequency is set high to generate a smooth demonstrationresult that will therefore emphasize the concepts with less influence from sam-pling. The range of k is set between 0.001 to 0.4. Low k values mimic the re-ceived data from perpendicular specular reflections. Medium k values representnon-perpendicular specular reflections and off-axis interference. Diffuse scatteringis mimicked by much higher k values.As shown from the result, most beamforming methods show continuous reduc-tion trend in weight values as the slope value increases. This means high level ofsuppression for tilted signals, off-axis interference and soft tissue. Wiener, CF andPCF have similar weight response but Wiener shows high weight values includ-ing large ripples, indicating less suppression levels compared with other methods(more noise). GCF shifts the response to higher k values compared with CF. Theproposed accumulated phase method follows a 1k function in weight reduction sincethe accumulated phase increases linearly with k. It offers a continuous reduction inweight values as the tilting in sine wave increases.293.2 Equation and flow chart for AAFIn the demonstration with the simple sine wave, the accumulated phase change,A, increases linearly with k. In actual beamforming, the value of k varies withinthe receive aperture based on spatial relation between the ultrasound beam, bonesurface locations/angles and transducer geometry. The accumulated phase changecan be estimated from the difference between the maximum and the minimumof the unwrapped phase values across the receive aperture. The phase values areestimated from the Hilbert transform of the aperture data after delay correction.The equation to calculate AAF is:AAF = max(ϖH (X))−min(ϖH (X)), (3.5)where H represents the Hilbert transform. H (X) means the angle operation onthe Hilbert transformed result. ϖ denotes the unwrap operation. X(n) is the sam-pled delay compensated channel data at a given depth. n is the channel number andN is the receive aperture size. To maintain the dynamic range and avoid specialcases like AAF equals 0, AAF is set to 1 if it is less than one. The channel dataX(n) are Hilbert transformed across the aperture direction to allow for the angleestimation. The accumulated angle change is therefore the difference between themaximum to the minimum of the unwrapped phase changes across the aperture.The actual receive aperture increases linearly with depth (from 1 to 192) based onthe receive F number. The inverse of accumulated angle change across the effectiveaperture is then used as the weight to the summation of the channel data to generatethe beamforming output. The effective aperture, N, varies with depth.Out put =1AAF∑Nn=1X(n). (3.6)Figure 3.2(b) shows an example on the accumulated angle factor, AAF . Boneswith perpendicular reflection like in Figure 1.5(a) have AAF values around 8. Sig-nals from non-perpendicular reflection and off-axis interference have AAF valuesaround 50 to 100. The mean AAF value for overall vertebrae is around 28. The softtissue has a mean AAF value larger than 120. Therefore, the proposed accumulateangle factor method can successfully suppress signals from soft tissue and off-axis30interference.The flow chart of the proposed algorithm is shown in Figure 3.3. Notice oneimportant thing is that the angle information is estimated from the Hilbert trans-form at a single depth across A-lines, not from phases at each A-line (like in thephase coherence factor method introduced in chapter 1). The accumulated angleis therefore an indicator on how much the signal is fluctuated across the aperture.Low fluctuation generally means the signal is likely from hyper-echoic targets.3.3 Results3.3.1 Field II simulation resultsFigure 3.4 shows the weight changes for each beamforming method when the bonepoints are shifted right from 0 to 4 cm away from the center of the transmit beam(lateral translation experiment). (a), (b) and (c) correspond to different simulationswhere the ‘bone’ points are located in different depth (3 cm, 4 cm and 5 cm). 4cm is the transmit focus. All weights are normalized to the maximum weight atthe same depth and showed in dB scale. The proposed AAF method is shown inblack. Compared with other beamforming methods, the AAF method has a gradualreduction in weight as the ‘bone’ moves away from the center of the transmit beam.This is similar as the sine wave demonstration result. Lower weight means oncethe bone moves away from the ultrasound beam, these off-axis interference can besuppressed well in the AAF method. On the other hand, beamforming methodssuch as GCF, CF and Wiener have a quick drop in the weight values as the trans-lation distance increases. This is because these adaptive beamforming methods arebased on the amplitude variance among the channel data. A little tilting in the highamplitude ‘bone’ signal creates large change on the result. The amount of energyaway from DC increases quickly. As the tilting gets larger, the portion of signalsthat are from the bones reduces and its impact on the gain measurement also re-duces. This results in a rise of the gain values for larger translations. This trend canbe seen in all depths ((a), (b) and (c)). For phase based methods such as AAF andPCF, the weight is based on phase information instead of amplitude. AAF is theaccumulated phase change and PCF is variance of the phase. These methods offer31a gradual reduction in the weight values as the tilting in the channel data increases.Compared with PCF, the AAF method has a higher suppression ratio when the tilt-ing is large. This is also because the portion of the bone signal reduces as the tiltingin channel data gets very large.Figure 3.5 shows the weight changes for each beamforming method when thebone points are rotated counter clock-wisely from 0 to 30 degrees (rotation exper-iment). (a), (b) and (c) correspond to different simulations where the ‘bone’ pointsare located in different depth (3 cm, 4 cm and 5 cm). 4 cm is the transmit focus.All weights are normalized to the maximum weight at the same depth and showedin dB scale. The proposed AAF method is shown in black. According to the fig-ures, most beamforming methods show similar behaviour when the tilting angle ofthe ‘bone’ points increases. Wiener shows high weight values (low suppression)which matches the sine wave simulation results. According to Figure 3.5, the AAFmethod shows slightly higher level of suppression for most depths. Even thoughthe bone is tilted, it is still considered as signals. Therefore, over suppression ontilted bone signals could be one potential limitation for the proposed AAF method.On the other hand, AAF can suppress the artifacts generated by off-axis bones.The specular reflection Field II simulation shows continuous reduction on weightswhen the tilting and translation increase in specular reflection. In a separate dif-fusive scattering Field II simulation, channel data from soft tissue are simulatedso the improvement in contrast between ‘bone’ and surrounding soft tissue causedby AAF can be evaluated. The normalized mean weights for Wiener, CF, GCF,PCF and AAF in soft tissue region over ‘bone’ are: 0.405, 0.090, 0.241, 0.082 and0.047, respectively. AAF shows the highest contrast between ‘bone’ and soft tissuein Field II simulation.3.3.2 Ultrasound experiment: vertebrae phantom resultsThe vertebrae phantom results from six different beamforming methods, (a) DAS,(b) Wiener, (c) PCF, (d) CF, (e) GCF and (f) AAF method, are shown in Figure 3.6.The dynamic range for all beamforming methods is set to 80 dB. The proposed32AAF method shows the highest contrast with no large distortion in the vertebraesurface compared with other beamforming methods. Notice that due to the smallacoustic impedance difference between the water-plastic interface, the amount ofenergy under the plastic bone surface is higher compared with real in vivo scenar-ios. This results in high intensity values under the bone surface in DAS and manyadaptive beamforming methods except for AAF.3.3.3 Ultrasound experiment: vertebrae in vivo resultsAs introduced in the Chapter 2, 60 data sets from 5 volunteers were collected in-cluding several common views for vertebrae ultrasound imaging (30 sagittal and30 transverse). Figure 3.7 shows a sagittal view of vertebrae from a healthy vol-unteer for six different beamforming methods, (a) DAS, (b) Wiener, (c) PCF, (d)CF, (e) GCF and (f) the proposed AAF method. The laminae are the hyper-echoicwave shapes in the middle of the image. The skin layers are the lines in the nearfield. Between them is the muscle/fat tissue layer. The region under the laminaeshows dark intensities since most of the ultrasound signals are reflected back in thevertebrae surface. Compared with the vertebrae phantom, the in vivo laminae sur-face is weaker and less continuous due to the soft tissue layers above the vertebraesurface. The speckle noise is also stronger above the laminae. The dynamic rangefor all in vivo data sets is set to 60 dB due to lower signal to noise ratio comparedwith the phantom. The CR between the laminae and the surrounding soft tissuefor DAS, Wiener, PCF, CF, GCF and AAF is 0.40, 0.56, 0.86, 0.73, 0.71 and 0.93,respectively.Figure 3.8 shows a transverse view of vertebrae from healthy volunteers for sixdifferent beamforming methods, (a) DAS, (b) Wiener, (c) PCF, (d) CF, (e) GCF and(f) the proposed AAF method. In this angle, the vertebrae surface shows a series of‘discontinued’ surfaces. Most of bright lines correspond to lamina and transverseprocesses. The spinous process is almost disappeared due to the large incisionangle between the ultrasound beam and the spinous process surface which makesthe energy reflecting out of the receive aperture. In this example, the existence ofthe spinous process can only be verified by the shadow (dark region) underneaththe vertebrae surface. The CR between the laminae and the surrounding soft tissue33for DAS, Wiener, PCF, CF, GCF and AAF is 0.49, 0.62, 0.82, 0.75, 0.74 and 0.90,respectively. Although the spinous process still cannot be visualized, the proposedAAF method shows high contrast for the lamina and transverse process with lownoise in the soft tissue region.In in vivo results, the CR between bones and surrounding tissues are typicallylower compared with the phantom results. This is due to the existence of softtissue in in vivo cases. Although the background structure is different, the trend ofobtaining high CR for AAF remains the same, which indicates the effectiveness ofAAF method.Figure 3.9 shows results from several other in vivo examples. Since PCF showsbetter results compared with CF, GCF and Wiener beamforming, only PCF isshown with DAS and the proposed AAF results. According to the results, theAAF shows the highest CR. Among the 60 in vivo studies data sets, the mean CRfor the DAS is 0.49, Wiener is 0.64, PCF is 0.82, CF is 0.77, GCF is 0.76 and AAFis 0.91. The mean and variance of CR for each beamforming method is shown inFigure 3.10. The t test results prove the significant difference between the proposedAAF method with other beamforming methods (p < 0.001).Other than the CR analysis, variance and sensitivity/specificity analyses arealso performed. The mean variance inside and outside the manually outlined boneROIs for different adaptive beamforming weights are: Wiener (0.2793 and 0.1975),PCF (0.0317 and 0.0153), CF (0.0799 and 0.0451), GCF (0.1569 and 0.0969), andAAF (0.0237 and 0.0075), respectively. The proposed AAF method has the lowestweight variance both inside and outside the bone ROIs. The sensitivity and speci-ficity analysis is then performed based on the same ROIs. The background ROIis defined as the red box shown in Figure 3.8(a). The signal ROI is the manuallyoutlined bone surface within the background ROI. The sensitivity is defined as thepercentage of pixels classified as bone pixels correctly inside the singal ROI. Thespecificity is defined as the percentage of pixels classified as background pixels(not bone pixels) correctly outside the signal ROI but within the background ROI.The ROC curve for the in vivo data set in Figure 3.8 is shown in Figure 3.11. Thearea under the ROC curve is 0.755 for Wiener, 0.863 for PCF, 0.833 for CF, 0.8765for GCF and 0.938 for AAF. The proposed AAF method has higher sensitivity andspecificity than other methods.343.4 DiscussionWith recent advancements in multi-modalities registration and more usage in ul-trasound for real-time spinal surgery/anesthesia guidance, there is a need in im-proving the quality for ultrasound spinal imaging. My goal is to clearly identifythe vertebrae surface in ultrasound imaging. This is challenging due to complexinteractions between the ultrasound beam and hyper-echoic vertebral surface. TheAAF beamforming is proposed to achieve this goal. Sine wave demonstration,Field II simulation, phantom and in vivo studies are performed so the improvementfrom the proposed AAF method can be compared with other adaptive beamform-ing methods. Among the selected adaptive beamforming methods to compare,Wiener beamforming applies a weight based on the variance among the channeldata. PCF analyzes variance on the phase values instead of the amplitude. CF es-timates the weight based on the energy ratio from the transmit direction to the restof beam directions. GCF increases the signal band to offer a more robust estima-tion. The proposed AAF method estimates the weight based on accumulated phasechanges across the aperture. The sine wave demonstration confirms the reductionof weights for all selected beamforming methods as the tilting angle in channeldata increases. Wiener beamforming shows less suppression capability comparedwith other methods. Changing the ratio between the sampling frequency, fs, orthe center frequency f could have an impact on the accumulated angle change.For example, reducing the center frequency to half would reduce the AAF by halfas well. However, both the signals from the bone and soft tissue will have thesame change. This means the curve shown in Figure 3.2(a) will just be shifted /scaled left. The relative contrast improvement is therefore unchanged for the pro-posed method. Field II specular reflection simulation mimics two scenarios: (1)the off-axis interference caused by bone targets moving away from the center ofthe transmit ultrasound beam. (2) non-perpendicular reflection caused by tilting inbone surfaces. The first scenario proves the benefit of the proposed AAF methodsince the weight reduces continuously as the bone moves away from the centerof the transmit aperture. All other selected methods do not have this property.The second scenario shows the AAF method behaves similarly to the other beam-forming methods in suppressing non-perpendicular reflections. Field II diffusive35scattering simulation shows the contrast improvement between bone and soft tis-sue in AAF. Phantom and in vivo studies confirm the high contrast ratio for bonesobtained using the proposed method.Admittedly the AAF method is designed mainly for the CR improvement onbone structures. Although normal adaptive beamforming methods did not showa high contrast in the vertebrae surface, it does not mean they are not superior inother applications. In cases with soft-tissue ultrasound imaging with non-specularreflections and high lateral resolution requirements, GCF, CF, PCF and Wienercould be more appropriate.Practically, a small F number is appreciated (F number ≤ 1) in the proposedmethod since a large integration window (the actual aperture) is important for ac-curate overall phase estimation. One potential limitation of the proposed methodis not being able to enhance weak bones or surfaces with a large incision anglewith the transmit beam (like the spinous process in Figure 3.8). In those cases,the overall phase change is large. On the other hand, the reverberation artifact canbe suppressed since it spans fewer channels compared with the vertebrae surfaceitself, which means the overall phase fluctuation is higher.The proposed AAF method still uses the same delay correction mechanism asDAS beamforming. It only needs additional processing for the Hilbert transform,angle estimation, unwrapping and maximum/minimum estimation. The increasein computation requirement is therefore not large. The overall processing timeincrease for AAF compared with DAS is less than 10 percent in Matlab processing.In addition, this processing change is only a software/firmware modification, whichmeans it can be implemented into ultrasound systems without hardware changes.Although the computation requirement is low for the AAF method, it does notaddress the fundamental issue of signal cancellation in tilted bone signals caused byspecular reflections. Bones surfaces with large angles with the transmit ultrasoundbeam, such as spinous process, are hard to visualize. As a result, the contrast ofbones over strong soft tissue echoes are not very high. In the following chapters,these issues are going to be addressed.363.5 ConclusionAn AAF beamforming method is proposed to enhance the strong surfaces in spinalultrasound imaging. Phantom and in vivo studies show significant improvementsin contrast on hyper-echoic targets. Although AAF does not address the signalcancellation issue in tilted bones, it offers a simple solution to visualize bones withhigh contrast.37Figure 3.1: Demonstration of simplified bone-reflected channel data (afterdelay correction) using a 2D sine wave with different tilting levels. Thesine wave has 4 full wavelengths along the depth direction (y direction)and 192 samples along the channel direction (x direction). Signals fromperpendicular specular reflections have small or no tilting (a) while sig-nals from non-perpendicular specular reflections and off-axis interfer-ence have large tilting angles (b) and (c). The left column shows the2D sine wave with k = 0, 0.05 and 0.1. The middle column shows thephase change for the black horizontal line shown in the first column.Notice that the phase change is wrapped around. The unwrapped phasechange is shown in the right column. As the tilting angle increases, theunwrapped phase change gets larger.38Figure 3.2: (a) The relation between the weight and the slope value, k, foreach adaptive beamforming method in a demonstration 2D sine wavestudy. The 2D sine wave represents the signal coming from the spec-ular reflection targets (bones). The tilt of the sine wave is controlledby changing the slope value, k. Low k values mimic the received datafrom perpendicular specular reflections. Medium k values representnon-perpendicular specular reflections and off-axis interference. Dif-fuse scattering is mimicked by much higher k values. The proposedmethod brings a continuous weight reduction as k increases. (b) The es-timated AAF values from an in vivo data set. The mean AAF value foroverall vertebrae is around 28. Signals from non-perpendicular reflec-tion and off-axis interference have AAF values around 50 to 100. Thesoft tissue has a mean AAF value larger than 120. Therefore, the pro-posed accumulate angle factor method can successfully suppress signalsfrom soft tissue and off-axis interference.39Figure 3.3: For AAF beamforming, the time delay corrected and dynamicallyapodized channel data at a given sample depth are Hilbert transformedto estimate the angle variation across the aperture. These angle valuesare unwrapped. The accumulated angle change is then calculated bythe difference between maximum and minimum of the unwrapped an-gles across the aperture. The weight is one over this accumulated anglechange.40Figure 3.4: Field II specular reflection simulation results. This figure showsthe weight changes for each beamforming method when the bone pointsare shifted right from 0 to 4 cm away from the center of the transmitbeam (lateral translation experiment). (a), (b) and (c) correspond todifferent simulations where the ‘bone’ points are located in differentdepth (3 cm, 4 cm and 5 cm). 4 cm is the transmit focus. All weightsare normalized to the maximum weight at the same depth and showed indB scale. The proposed AAF method is shown in black. Compared withother beamforming methods, the AAF method has a gradual reductionin weight as the ‘bone’ moves away from the center of the transmitbeam.41Figure 3.5: Field II specular reflection simulation results. This figure showsthe weight changes for each beamforming method when the bone pointsare rotated counter clock-wisely from 0 to 30 degrees (rotation exper-iment). (a), (b) and (c) correspond to different simulations where the‘bone’ points are located in different depth (3 cm, 4 cm and 5 cm). 4cm is the transmit focus. All weights are normalized to the maximumweight at the same depth and showed in dB scale. The proposed AAFmethod is shown in black. According to the figures, most beamformingmethods show similar behaviour when the tilting angle of the ‘bone’points increases. Wiener shows high weight values (low suppression)which matches the sine wave simulation results.42Figure 3.6: The beamforming results for the vertebrae phantom study: (a)DAS, (b) Wiener, (c) PCF, (d) CF, (e) GCF and (f) AAF method. Theproposed AAF method shows the highest contrast with no large distor-tion in the vertebrae surface compared with other methods.43Figure 3.7: Vertebrae in vivo results. The sagittal view of vertebrae for sixdifferent beamforming methods: (a) DAS (b) Wiener, (c) PCF, (d) CF,(e) GCF and (f) the proposed AAF method. The laminae are the hyper-echoic wave shapes in the middle of the image. The skin layers are thelines in the near field. Between them are the muscle / fat tissue layers.The region under the laminae show dark image since most of ultrasoundsignals are reflected back in the vertebrae surface. The dynamic rangefor all beamforming methods was set to 60 dB due to lower signal tonoise ratio compared with the phantom. The CR between the laminaeand the surrounding soft tissue for DAS, Wiener, PCF, CF, GCF andAAF is 0.40, 0.56, 0.86, 0.73, 0.71 and 0.93, respectively. The AAFmethod generates a high contrast laminae surface against soft tissue.44Figure 3.8: Vertebrae in vivo results. The transverse view of vertebrae for sixdifferent beamforming methods: (a) DAS (b) Wiener, (c) PCF, (d) CF,(e) GCF and (f) the proposed AAF method. The skin layers are the linesin the near field. The hyper-echoic lines in the image represent differentvertebrae structures such as the spinous process, laminae and transverseprocess. Between them are the muscle / fat tissue layer. Same as thesagittal view, the dynamic range for all beamforming methods was setto 60 dB. The CR between the laminae and the surrounding soft tissuefor DAS, Wiener, PCF, CF, GCF and AAF is 0.49, 0.62, 0.82, 0.75,0.74 and 0.90, respectively. The AAF method generates a high contrastlaminae surface against soft tissue. The ROI is defined as the red boxshown in (f).45Figure 3.9: The results from several other in vivo studies. Since PCF showsbetter results compared with CF, GCF and Wiener beamforming, onlyPCF is shown with DAS and the proposed AAF results. According tothe results, the AAF shows the highest CR.46Figure 3.10: The mean and variance of CR in 12 in vivo data sets. The meanCR for the DAS is 0.49, Wiener is 0.64, PCF is 0.82, CF is 0.77, GCF is0.76 and AAF is 0.91. The t test results prove the significant differencebetween the proposed AAF method with other beamforming methods(p < 0.001).47Figure 3.11: The ROC curve for the in vivo data set in Figure 3.8. The sen-sitivity is defined as the percentage of pixels classified as bone pixelscorrectly inside the bone ROI. The specificity is defined as the per-centage of pixels classified as background pixels (not bone pixels) cor-rectly outside the bone ROI. The area under the ROC curve is 0.755for Wiener, 0.863 for PCF, 0.833 for CF, 0.8765 for GCF and 0.938 forAAF.48Chapter 4Directional filtering and structuretensor-based maskingIn the previous chapter, the AAF method is introduced to improve the contrast ofbone surfaces in the spine. Although it has the benefit of low computation require-ment, it does not address the signal cancellation issue due to tilting in the channeldata caused by specular reflections. It also cannot enhance weak bones or surfaceswith a large angle with the transmit beam (like the spinous process in Figure 3.8).Therefore, in this chapter, directional filtering on the channel data is proposed toimprove the visualization of the bones. The challenges are: (1) The tilted bone sig-nals caused by specular reflection can be regulated back to the horizontal directionso they are not cancelled out during the beamforming. (2) Large tilting off-axisinterference, reverberations and soft tissue should be suppressed so the contrast ofthe bone surface can be enhanced.Quadrature directional filters have been a popular tool in the image process-ing/computer vision field [60] for estimating local image information, such as ra-dial/angular frequency and directional energy. The log-Gabor filter is one of theexamples of quadrature filters. The ability to control the filter pass band and angleof interest is appealing in my ultrasound beamforming application. This means thepass angle can be controlled so only signals with no or little tilt angle along thehorizontal direction can pass. One potential drawback of the directional filtering isthat soft tissue ligaments and reverberations can also be enhanced as long as they49are aligned close to the horizontal direction after delay correction. To suppressthese, a tensor-based energy masking method is employed which only keeps sig-nals with large gradient and edges [61]. Only bone signals are maintained after themasking process.The proposed method is defined as the log-Gabor directional filtering withtensor-based masking method (DFTM). The pre-beamformed channel data are firstdelay compensated. Then, directional log-Gabor filtering is applied on the alignedchannel data to enhance signals along or close to horizontal directions. A tensor-based energy masking is then employed to clean out the noise generated from di-rectional filtering and avoid false enhancement. To validate my method, both simu-lation and phantom/in vivo volunteer studies (with written consents) are performed.4.1 Theoretical background and algorithm overviewDue to the size and smoothness of vertebrae structures, the soft-tissue vertebraeinterface can be considered as a nearly specular reflection interface. After timedelay correction, the channel data corresponding to hyper-echoic targets are shownas lines aligned horizontally or with some tilting angles. Reverberations from thebone interface will sit in the same direction but with less stretch and amplitude.Off-axis interference typically sits with an angle to the horizontal direction. Thisis because the channel delay is calculated based on the center of the beam not forthe side lobes. Signals coming from diffusive targets are weaker and show highfluctuations along the horizontal directions. In some cases, the tilting caused byspecular reflection can be large (e.g., reflections from spinous processes). Thelarge tilting angle is caused by the surface geometry and shallow depth.All signals coming from vertebrae surfaces directly are considered as the sig-nal. The off-axis interference, reverberations and reflections/scattering from softtissue structures are considered as the noise. Instead of direct summation of thechannel data horizontally (the delay-and-sum beamforming (DAS)), directionalfiltering is employed to align the signals with little tilting towards the horizontaldirection and suppress signals with large tilting. Then, weak signals from softtissue and reverberations are masked out based on the energy of gradient values.Log-Gabor filtering is desirable for the filtering process since the pass angle and50frequency band can be custom designed. A tensor-based method for energy mask-ing is chosen since representing local structure using a tensor is an effective way toidentify structures with large gradients [61]. It also offers robust estimation sinceno thresholding is involved in the masking process.Figure 4.1 shows the flow chart of the proposed algorithm. The pre-beamformedchannel data from each transmit event are first delay compensated. Then, a regionof interest (ROI) along the depth direction is selected. To avoid the near field strongreflections, the ROI is chosen starting with 1 cm below the transducer surface andall the way to 15 cm which covers most of the spinal structures. The ROI is thenfiltered using log-Gabor directional filtering. The noise generated from the direc-tional filter along with reverberation/soft tissue artifacts are then masked out usingthe tensor-based method. Finally, the filtered and masked channel data within theROI are summed along the channel direction to generate a single scan line just asregular DAS beamforming. This process is repeated until the whole ultrasoundimage is generated.4.1.1 Log-Gabor directional filteringThe log-Gabor filter has been widely used in image and signal processing field toisolate desired frequency components in a certain band or direction [62]. In thetime delay corrected channel data, the signal of interest is represented as a series ofhorizontally oriented lines (or with a small amount of tilt). The off-axis interferenceartifacts are represented as energies along other directions. The log-Gabor filter hasthe following formulation:LG(ω) = exp(− log2( ωω0 )σ f) cosA(θ −θ0), (4.1)where LG() is the frequency response of the filter, ω0 is the normalized filter centerfrequency, σ f is a parameter controlling the frequency bandwidth of the filter, θ0is the center angle of the filter. The spread of the pass band in the angle domaincan be controlled by A, which is inversely proportional to the bandwidth. In myexample, θ0 is set to 90, i.e., the ultrasound transmit direction, since lines closeto the horizontal directions are of special interest. In addition, ω0, σ f and A canbe changed to control the angle and frequency pass-band of the filter. Figure 4.251Figure 4.1: The processing chain of the proposed DFTM method. The pre-beamformed channel data from each transmit event are first delay com-pensated. Then, a region of interest (ROI) along the depth direction isselected. To avoid the near field strong reflections, the ROI is chosenstarting with 1 cm below the transducer surface and all the way to 15cm which covers most of the spinal structures. The ROI is then fil-tered using log-Gabor directional filtering. The noise generated fromthe directional filter along with reverberation/soft tissue artifacts arethen masked out using the tensor-based method. Finally, the filteredand masked channel data within the ROI are summed along the channeldirection to generate a single scan line just as regular DAS beamform-ing. Notice that the procedure is repeated for each transmit event togenerate the whole image.shows two examples of the frequency response for a wide pass band and a narrowpass band both in the angle and frequency direction.After constructing the filter response in radial and directional components, thespatial domain filter kernel can be generated by applying the inverse Fourier trans-form (F ) on LG. Since the frequency response is not symmetrical, the spatial52Figure 4.2: Two example frequency response of the proposed log-Gabor di-rectional filter. (a) has a wide pass band both in the angle and frequencydirection and (b) has a narrow pass band. In (a), the ω0 is set to 0.6, σ fis set to 1.6, A is set to 1. In (b), the ω0 is set to 0.6, σ f is set to 0.18,A is set to 8. ω0 is normalized spatial frequency. Selecting it as 0.6 isto ensure the pass band can cover most spatial frequency range. σ f andA are chosen based on the angle and frequency pass band. The 2D filterkernel size is set as 15 so two full wavelengths of the input channel datacan be covered.domain of the filter kernel is complex. The real part of the filter can be convolvedwith delay compensated channel data, C:LGF =C ∗ real(inv(FLG)). (4.2)4.1.2 Tensor-based maskingRepresenting local structure using structure tensor (ST) is an effective way to iden-tify strong line structures and suppress noise [61, 62]. ST is generally estimatedby convolution of a set of directional filters with the data. The number of filtersdepends on total angles of interest. To cover the 2D space, a series of three anglesgenerally means 0, 60 and 120. A series of four angles means 0, 45, 90 and 135.More angles give a better local tensor estimation but also have more computationrequirements. These directional filters have positive frequency response over thedesired filter direction and zero over other frequency. The frequency response of53the directional filter is defined as:F(µ) = R(||µ||)(µ ·nk)2, when µ ·nk > 0, (4.3)where µ is the frequency domain coordinates and R(||µ||) is the radial part of thefrequency response, typically a high-pass log-nominal function. (µ · nk)2 is thedirectional part of the frequency response and it varies with the region of interest.The convolution output of the directional filter with the aligned pre-beamformeddata, Yk, is therefore proportional to:Yk ∝ (na ·nk)2, (4.4)where na is the direction that has the largest gradient difference. Finally, the STcan be estimated by:T =∑k||Yk||(nk ·nTk ). (4.5)Since (λ I−T )x= 0, for a given tensor matrix, T = a bc d , the eigenvalues, λ1, λ2 and eigenvectors, x1, x2 can be calculated by:λ 2−λ (a+d)+ad−bc= 0, (4.6)x= bλ −a,λ −dc. (4.7)The construction of 2D structure tensor is important since eigenvalues and thecorresponding eigenvectors summarize the distribution of the gradient surroundinga given data point. If λ1 is much larger than λ2 and λ2 is relatively small, thismeans the highest gradient direction is aligned with eigenvector x1. This typicallycorresponds to a sharp edge (hyper-echoic targets). If λ1 is close to λ2 and theyare both much larger than zero, it corresponds to a point target (which is rare inthe channel data). If λ1 is similar to λ2 and they are both close to zero. In thiscase the gradient has no predominant direction, which typically corresponds to anoise region (soft tissue or off-axis interference). Based on this characteristic, theeigenvalues can be remapped so large eigenvalues can be further increased while54small eigenvalues are reduced. This remapping process can maintain or enhancestrong edges while suppressing noises. The remapping process is shown in thefollowing equations:λ ′1 =λ1αλ1α +(1−λ1)α , (4.8)λ ′2 =λ2λ1βλ2λ1β+(1− λ2λ1 )β, (4.9)where α and β are control constants. Then a reconstructed control tensor, T ′ iscalculated based on the updated eigenvalues λ ′1, λ ′2 and old eigenvectors x.T ′ = λ ′ · x · xT , (4.10)The control tensor is then used to generate the gradient mask along the ultra-sound transmit beam direction (perpendicular to the horizontal direction):Mask = T ′ · (n3 ·nT3 ), (4.11)where n3 ·nT3 is -0.25 00 0.75 .The tensor-masked output is calculated by:Out put =Mask ·LGF. (4.12)Applying the mask on the log-Gabor filtered channel data will suppress thehorizontal noise. Notice that no hard thresholding is used in my methods. Allmasking is purely based on the tensor analysis results.554.2 Results4.2.1 Field II simulation resultsFigure 4.3 shows Field II specular reflector simulation results. Figure 4.3(a) showsthe received channel data collected when the specular reflector is perpendicular tothe transmit ultrasound pulse (along the horizontal direction). After delay correc-tion, the received channel data are aligned horizontally except for weak gratinglobes. Figure 4.3(b) shows the same channel data as Figure 4.3(a) but with theproposed DFTM method using parameters listed in Figure 4.2(b). The horizontalcomponent is maintained in the DFTM result while grating lobes along other direc-tions are clearly suppressed. Figure 4.3(c) shows the received channel data whenthe specular reflector is tilted 15 counterclockwise. After the same time delay cor-rection, the peak of the received energy shifts to the left side since the center ofthe transmit beam tilts after specular reflection. As the channel number increases,the data tilt upwards because the travel distance reduces. Figure 4.3(d) shows thesame channel data as Figure 4.3(c) but with the proposed DFTM method. TheDFTM method successfully enhances the signal close to the horizontal directionon the left side of the aperture. On the other hand, it suppresses the energy fromother directions. Tilted signals including grating lobes are removed. Since the fil-tered signal is aligned horizontally, it avoids signal cancellation from summationbetween positive and negative peaks during the beamforming process. As a result,the bone signals become stronger as the tilting angle increases.Figures 4.3(e) and (f) show the peak beamformed channel data changing withtilting angle with and without the DFTM method. The DAS beamforming resultis shown in the dashed line, the DFTM method with wide spectrum is shown inthe solid line in figure 4.3(e). The DFTM method with narrow spectrum is shownin the solid line in figure 4.3(f). The filtered result maintains the same peak sig-nal level as DAS when the tilting angle is less than 10. As the angle increases,the bone signal starts to drop due to cancellation and the filtered result becomesstronger compared with DAS. When using the DFTM method with wide spectrum,the signal from the bones is 2 to 6 dB higher than DAS beamforming. If using thenarrow-band directional filter, the bone signal is 3 to 10 dB stronger than the DAS56Figure 4.3: Field II specular reflector simulation results. (a) shows the re-ceived channel data collected when the specular reflector is perpendic-ular to the transmit ultrasound pulse. (b) shows the same channel dataas (a) but with the proposed DFTM method. (c) shows the channel datacollected when the specular reflector is tilted 15 degrees counterclock-wise. (d) shows the same channel data as (c) but with the proposedDFTM method. (e) and (f) show the peak beamformed channel datachanging with tilting angle with and without the DFTM method. Thesignal from the bones is 2 to 6 dB higher than DAS beamforming if us-ing the wide-band directional filter and 3 to 10 dB stronger if using thenarrow-band filter.57Figure 4.4: The phantom results comparison between the DAS and DFTMmethod on a vertebrae phantom. (a), (c) and (e) are DAS results. (b),(d) and (f) are the corresponding results using DFTM. The quantitativeresult shows that the CR for vertebrae structures is 0.86 in the DASresult and 0.98 in the proposed DFTM result.beamforming. The narrow pass-band directional filter regulates signal towards thehorizontal direction and therefore resulting in the boosted SNR.Field II simulation results show the possible benefits of filtering the channeldata along the horizontal direction. In real ultrasound data, the bone signals areoften not aligned as well as the simulation cases due to presence of soft tissue,off-axis interference, reverberations and other artifacts. The benefit from DFTMis expected to be larger in real cases. Apart from bone signals improvement, it isalso important to verify that DFTM will not create significant distortion in the bonedata (false enhancement). Phantom experiments in the next sub-section will aim toaddress these issues.4.2.2 The phantom experiments resultsFigure 4.4 shows the results of phantom experiments for DAS and DFTM methodson the vertebrae phantom. Figures 4.4(a), (c) and (e) are DAS results from thetransverse view, the para-sagittal view and an in-between view. Figures 4.4(b),(d) and (f) are the corresponding results using DFTM. All results are shown in5860 dB dynamic range. Since there is no tissue structure around the bone surface,most of the top surface of the vertebrae structures can be clearly visualized. InFigures 4.4(a), (c) and (e), there are regions of gray pixels under the bone surface.These are the reverberation artifacts. The boundary of the bone structure is alsoblurry due to the effect from off-axis interference. The spinous process (in thetop middle of Figures 4.4(a) and (b)) is weaker compared with other bone surfacesbecause of the specular reflection and the relatively large angle between the spinousprocess and ultrasound beam. Since there is no soft tissue above the phantom, thevertebrae surface can be visualized in both cases.DFTM results (Figures 4.4(b), (d) and (f)) show clear enhancement of bonestructures. Meanwhile, artifacts from reverberation and off-axis interference areboth suppressed. Vertebrae structures can be visualized with sharp boundaries andhigh contrast. The overall contrast between the vertebrae surface and surroundingartifacts is greatly improved. The quantitative result shows that the contrast ratio(CR) for vertebrae structures is 0.86 in the DAS result and 0.98 in the proposedDFTM result. In addition, the vertebrae surfaces generated by DFTM do matchwith the corresponding bone surfaces in the DAS results. No visible distortion inbone structures is created.4.2.3 The in vivo resultsSame as the Chapter 3, 60 data sets collected from 5 volunteers were used in thein vivo study. Figure 4.5 shows an in vivo vertebrae transverse view example. TheDAS beamformed image is shown in Figure 4.5(a). The dynamic range was set to50 dB. The channel data used to beamform the red line in Figure 4.5(a) are shownin Figure 4.5(c). Delay correction has been applied on the channel data. The topbright region in the channel data corresponds to the spinous process. Log-Gabor di-rectional filtered channel data are shown in Figure 4.5(d). According to the figures,signals along or close to the horizontal direction are enhanced while other signalsare suppressed. In addition, signals with little tilting are now regulated towards thehorizontal direction. Not only the spinous process is enhanced in Figure 4.5(d),noise along the horizontal direction is also enhanced (top region in Figure 4.5(d)).Since the noise generally has lower amplitude compared with bones, a tensor-based59energy masking is proposed to suppress them. Figure 4.5(e) shows the tensor-basedmasking result where weak horizontal responses are cleared out while strong echosfrom spinous process are maintained. The same DFTM method is repeated onchannel data for every scan line. The resulting transverse view image is shownin Figure 4.5(b). Compared with Figure 4.5(a), the DFTM method suppresses thesoft tissue around the vertebrae surface. The reverberation artifact underneath thevertebrae surface is also reduced. Bone surfaces including the spinous process canbe clearly visualized in the DFTM beamformed data.Figure 4.6 shows more in vivo transverse view examples from different volun-teers. Figures 4.6(a), (c) and (e) are DAS results. The dynamic range is set to 50dB in all cases. Compared with the phantom data sets, the soft tissue creates lots ofgray scale noise surrounding vertebrae structures. A big portion of the ultrasoundenergy is reflected back in the soft tissue-bone interface, resulting a dark shadowunderneath the bone. In all DAS cases the spinous process cannot be clearly visual-ized due to the angle and location of the surface. Its location can only be estimatedby locating the shadow region underneath the bone surface. Figure 4.6(b), (d) and(f) show the corresponding results using DFTM. Same as the phantom studies, thecontrast of the vertebrae surface is greatly enhanced in the DFTM results. Thespinous process can be clearly visualized. Soft tissue, reverberation and off-axisartifacts are suppressed.Figure 4.7 shows an in vivo para-sagittal view example. The DAS beamformedimage is shown in Figure 4.7(a). The dynamic range is set to 50 dB. The channeldata used to beamform the red line in Figure 4.7(a) are shown in Figure 4.7(c).Delay correction has been applied on the channel data. The middle bright region inthe channel data corresponds to the lamina surface. Log-Gabor directional filteredchannel data are shown in Figure 4.7(d). Same as the transverse view examples,directional filtering aligns the channel data to the horizontal direction and removesdata with large tilting angles in the channel data. Tensor-based masking furthercleans out the signal, leaving a bone-only output after beamforming (shown inFigure 4.7(e)). The same DFTM method is repeated on channel data for every scanline. The resulting para-sagittal view image is shown in Figure 4.7(b). The contrastof vertebrae surface is greatly enhanced compared with the DAS result.Figure 4.8 shows more in vivo para-sagittal view examples. Similar to the60Figure 4.5: An example of in vivo transverse view image. The DAS beam-formed image is shown in (a). The dynamic range was set to 50 dB. Thechannel data used to beamform the red line in (a) are shown in (c). De-lay correction has been applied on the channel data. The middle brightregion in the channel data corresponds to the lamina surface. Log-Gabordirectional filtered channel data are shown in (d). The tensor-basedmasked channel data are shown in (e).transverse view results, directional log-Gabor filtered results (Figures 4.8(b), (d)and (f)) show clear suppression of soft tissue, off-axis interference and reverber-61Figure 4.6: Examples of in vivo transverse view images. (a), (c) and (e) areDAS results. (b), (d) and (f) are DFTM outputs. The directional log-Gabor filtered results show clear suppression on the soft tissue, off-axisinterference and reverberation artifacts. On the other hand, strong sig-nals along the horizontal direction are maintained.ation artifacts. In 60 in vivo data sets, the overall contrast of vertebrae structuresis improved from 0.49 in DAS to 0.96 in DFTM. Other than bone signals, strongligaments above the bone surface can also be enhanced by DFTM. This effect isclear in para-sagittal view due to the wave-like laminae surface, which creates non-perpendicular reflections. The bone signals are therefore weakened and show sim-ilar intensity levels as ligaments. Thus, directional filtering enhances both boneand ligament signals and tensor-based masking would not be able to suppress liga-ments. Example of enhanced ligaments can be seen in the lines above the s-shapevertebrae surface in Figures 4.8(d) and (f) (shown with the red arrows).Changing the directional pass-band for the log-Gabor filter can have a largeimpact on the beamforming results. Figure 4.9 shows an example of a transverseview image filtered using two different pass-band as shown in Figure 4.2. TheDFTM method using wide-band directional filtering is shown in Figure 4.9(b). TheDFTM method using narrow-band directional filtering is shown in Figure 4.9(c).62Figure 4.7: An example of in vivo para-sagittal view image. The DAS beam-formed image is shown in (a). The dynamic range was set to 50 dB. Thechannel data used to beamform the red line in (a) are shown in (c). De-lay correction has been applied on the channel data. The middle brightregion in the channel data corresponds to the lamina surface. Log-Gabordirectional filtered channel data are shown in (d). The tensor-basedmasked channel data are shown in (e).The original channel data for constructing the red line in Figure 4.9(a) are shown inFigure 4.9(d). The corresponding narrow-band and wide-band directional filtered63Figure 4.8: Examples of in vivo para-sagittal view images. (a), (c) and (e) areDAS results. (b), (d) and (f) are DFTM outputs. The directional log-Gabor filtered results show clear suppression on the soft tissue, off-axisinterference and reverberation artifacts. On the other hand, strong sig-nals along the horizontal direction are maintained, resulting high con-trast in bone surfaces. (c) and (d) show the anterior complex disappearsin the DFTM result (red boxes). (d) and (e) show the soft tissue seg-ments get falsely enhanced in the DFTM result (red arrows).channel data are shown in Figures 4.9(e) and (f), respectively. The narrow-band di-rectional filter regulates the tilted bone signals (caused by non-perpendicular spec-ular reflection) back to the horizontal direction, resulting stronger signals. Thewide-band filter slightly enhances the signal along the horizontal direction but didless angle correction. The tensor-masked results for both cases are shown in Fig-ures 4.9(g) and (h). Strong bone signals are maintained in the masked output whichleads to a sharper bone surface for the narrow-band result. Compared with thewide-band filter output in Figure 4.9(b), the narrow pass band results provide asharper and cleaner bone surface in Figure 4.9(c).Figure 4.10 shows an example of para-sagittal view image filtered using twodifferent pass-bands as shown in Figure 4.2. Same as the transverse view re-64sults, narrow-band filtering clearly regulates the signal back to the horizontal direc-tion. In addition, data with large tilting angle is better suppressed using a narrow-passband filter. The resulting laminae is therefore cleaner and sharper. Overall, theCR for the narrow-band filter is 0.96 while it is 0.89 for the wide-band filter.4.3 DiscussionThere has been increasing interest in using ultrasound to image bone and ligamenttissue. Example applications include scoliosis detection and ultrasound-guidedspinal needle injections [3]. Although such ultrasound applications are promis-ing, it is still challenging to clearly image bone structures and locate the bonelocations accurately [12] due to off-axis interference, reverberations, strong soft-tissue structures and non-perpendicular specular reflections. The bone surface istypically thickened and blurred and in some cases disappeared, therefore affectingthe accuracy in extracting the bone locations [13].The proposed DFTM method is effective in removing the off-axis interferencedue to its directional suppression nature. It also regulates hyper-echoic signals withsmall tilting angle back to the horizontal direction. This adjustment avoids signalcancellation in the beamforming process and increases the visualization of struc-tures such as spinous process. Employing narrow-band directional filtering canachieve a better enhancement on signals with little tilting angles and better sup-pression on signals with large tilting angles. Tensor-based energy masking acts asa post-processing step which keeps high intensity structures and suppresses weaksoft tissue and reverberation artifacts. As a result, the contrast for hyper-echoicstructures, such as bones, is significantly improved. Even in cases where non-perpendicular specular reflections are prominent, the proposed method success-fully suppresses the noise and highlights hyper-echoic signals. Structures such asthe spinous process can be recovered in the DFTM method. Compared with theAAF beamforming, the CR for bones in the in vivo data sets is improved from 0.91to 0.96.One of the advantages of the DFTM method is the computation efficiency.Compared with DAS and AAF methods, the increase in computation requirementfor DFTM is not large. This is because in DFTM, the output is generated scan65line by scan line instead of pixel by pixel. For generating each output scan line,one directional high-pass filtering plus a structure tensor analysis are needed. Thedirectional high-pass filter is a 2D convolution filter with kernel sizes between 7x7and 15x15. The structure tensor estimation requires a series of 2D convolutionwith tap sizes of 7x7 for four directions. The eigen value estimation is based on 2by 2 matrix and the eigen value remapping is achieved using a pre-computed LUT.Although the bone surface enhancement is promising, there are several limita-tions for the proposed DFTM method. First, due to the normalization procedure inthe tensor-based masking process, only relatively strong signals within the channeldata ROI can be enhanced. As a result, weak bone structures in deep depth cannotbe clearly visualized. For example, surface-of-interest such as the back-side of thevertebrae body (inside the red boxes of Figure 4.8(c) could be suppressed sincethey are weaker compared with the vertebrae surface above. Currently, a singlechannel data ROI covering ranges between 1.5 cm and 12 cm depth is selected forthe DFTM processing. To visualize these weak vertebrae surfaces, separate ROIsfocusing on different depths should be used. For example, one ROI can be selectedfrom 5 to 12 cm depth. In this case, the back-side of the vertebrae body can bemaintained since they are strong signals in that depth range.Another limitation for the DFTM method is falsely recognizing strong ligamenttissue above the laminae surface as the bone surfaces. Examples can be seen fromhorizontal line segments above the curved lamina surface (pointed by arrows inFigures 4.8(d) and (f)). The channel data from these ligament lines are horizontallines which have similar intensities as the bone surface underneath them (due tonon-perpendicular reflections). Thus, they are hard to remove from directionalfiltering and/or tensor masking.Finally, the proposed DFTM method is effective on enhancing signals withstrong echos and little tilting away from the horizontal direction. Bone structureslocated deep (results in weak amplitudes) with an orientation that is far from theperpendicular direction of the ultrasound pulse are not enhanced. The narrow pass-band filter suppresses the largely tilted channel data and the tensor-based maskingremoves weak amplitudes. These can be seen from Figures 4.8(d) and (f) wherethe vertical part of a lamina in the center is missing.664.4 ConclusionA DFTM method which applies directional filtering and tensor-based masking onchannel data is proposed to enhance the visualization of vertebrae surface in ul-trasound imaging. It achieves this by regulating the bone ultrasound signals backto the horizontal direction prior to beamforming. Simulation, phantom studies andin vivo volunteer studies show clear improvements in contrast of vertebrae struc-tures without introducing large distortions. Compared with the AAF method, bonesignals show higher contrast. Structures that are normally disappeared in AAF orDAS, such as spinous process, can now be visualized.67Figure 4.9: An example of transverse view image filtered using two differ-ent pass-band as shown in Figure 4.2. The DAS beamformed output isshown in (a). The DFTM method using wide-band directional filteringis shown in (b). The DFTM method using narrow-band directional fil-tering is shown in (c). The original channel data for constructing the redline in (a) are shown in (d). The corresponding narrow-band and wide-band directional filtered output for input (d) are shown in (e) and (f),respectively. The tensor-masked narrow-band directional filtered chan-nel data are shown in (g). The tensor-masked wide-band directionalfiltered channel data are shown in (h).68Figure 4.10: An example of para-sagittal view image filtered using two dif-ferent pass-bands as shown in Figure 4.2. The DAS beamformed out-put is shown in (a). The DFTM method using wide-band directionalfiltering is shown in (b). The DFTM method using narrow-band direc-tional filtering is shown in (c). The original channel data for construct-ing the red line in (a) are shown in (d). The corresponding narrow-bandand wide-band directional filtered output for input (d) are shown in (e)and (f), respectively. The tensor-masked narrow-band directional fil-tered channel data are shown in (g). The tensor-masked wide-banddirectional filtered channel data are shown in (h).69Chapter 5Region of interest basedclosed-loop beamforming forspinal ultrasound imagingChapter 4 introduces the DFTM method which regulates the tilted bone signalsback to the horizontal direction through directional filtering. Tensor-based mask-ing clears out the soft tissue noise, offering high contrast for the bones. On theother hand, there are several limitations associated with the DFTM method. Sincedirectional filtering regulates signals back to the horizontal direction, the resultbone surface is often thickened and lack of sharpness. In addition, strong softtissue ligaments can also be falsely enhanced along with the bone surface.Ideally, if the location of the bones is known, only the region of interest needsto be processed. Instead of using directional filtering, the channel data can bealigned through cross-correlation, the sharpness of the bone surface can be main-tained. Bone signals with large tilting can also be corrected using this approach. Toperform the cross-correlation correctly, tensor-based filtering is employed whichclears out the soft tissue noise and offers clean and easy-to-track bone channeldata.In this chapter, I build on the directional-filtering in Chapter 4 and proposea two-step closed-loop beamforming method to alleviate the above issues. Thismethod is composed of two major processing steps: (1) Generate an initial beam-70forming output and estimate bone locations. (2) Feed the bone location infor-mation back to the beamforming process so the tilted bone channel data can bealigned back to the horizontal direction through cross-correlation. The resultingbone surface is therefore sharp with high contrast. Since the cross-correlation isperformed based on data from bones (with closed-loop feedback), strong echoesfrom ligaments above the bones can be suppressed while echoes from bones canbe enhanced. To detect the bone location with high sensitivity, the directional fil-tered method discussed in Chapter 4 is used to generate the initial beamformingoutput. The bone locations are identified automatically by selecting the first inten-sity peak from deep to shallow direction (bottom to top). To obtain clean channeldata for cross-correlation, a structure tensor (ST)-based noise suppression methodis employed to suppress the soft-tissue/reverberation clutters.5.1 Algorithm overviewThe proposed method is defined as tensor-based region-of-interest (ROI) closed-loop beamforming. The detail processing steps are shown in Figure 5.1. The pro-cessing sequence of each processing block is marked with numbers. First, thechannel data are delay-compensated. Directional filtering is then applied to alignthe channel data followed by beamforming. These steps are repeated to generate allthe scan lines in an ultrasound image. Then, a rectangular ROI containing bones isselected in the beamformed image. The bone locations are automatically detectedby locating the first intensity peak from deep to shallow direction along each beam-formed line within the ROI (bottom to top). For each scan line, only one bone pointis identified since multiple bone layers are unlikely to be seen in ultrasound imagesalong the same scan line.Meanwhile, a tensor-based image filtering is applied on the delay correctedchannel data. This filter suppresses soft tissue clutter and only maintains strongsignals in the channel data. Cross-correlation is then performed on filtered chan-nel data based on identified bone locations. This acts like a closed-loop correctionstep. As a result, only tilted bone signals are aligned back to the beamformingdirection. Signals from strong ligaments above the bone surface could be sup-pressed since cross-correlation is referenced based on bone targets instead. After71Figure 5.1: The flow chart of the proposed closed-loop method. The process-ing sequence of each processing block is marked with numbers. First,the channel data are delay-compensated. Directional filtering is then ap-plied to align the channel data followed by beamforming. These stepsare repeated to generate all the scan lines in an ultrasound image. Then,a rectangular ROI containing bones is selected in the beamformed im-age. The bone locations are automatically detected by locating the firstintensity peak from deep to shallow direction along each beamformedline within the ROI (bottom to top). Meanwhile, a tensor-based im-age filtering is applied on the delay corrected channel data. This filtersuppresses soft tissue clutter and only maintains strong signals in thechannel data. Cross-correlation is then performed on filtered channeldata based on identified bone locations. After cross-correlation align-ment, another beamforming is performed. This process is repeated untilall scan lines of an ultrasound image is generated. To further improvethe bone visualization, the closed-loop process (steps 4, 6 and 7) can berepeated until the improvement of bone is minimal.cross-correlation alignment, another beamforming is performed. This process is re-peated until all scan lines of an ultrasound image is generated. To further improvethe bone visualization, the closed-loop process (steps 4, 6 and 7) can be repeateduntil the improvement of bone is minimal.5.1.1 Bone location detectionIn this study, the bone location is automatically determined by the first intensitypeak from bottom to top. This is to avoid mislabeling strong ligaments above the72bone surface as bone structures. Bones may not be the strongest echo in ultrasoundimages due to reasons such as specular reflection angle and depth. In addition,many bone targets such as the spinous process can only be identified in ultrasoundimages by locating the shadow region underneath the bones. Selecting the first peakbottom to top can therefore detecting these relatively weak echoes from bones. An-other potential challenge in detection of bone pixels is the reverberation/soft tissuenoise underneath the bone surface. Directly working on DAS beamforming resultscould be challenging for accurate bone detection due to low contrast of bones. InChapter 4, I discussed a log-Gabor based directional filtering method to align thetilted bone signals back to horizontal direction and suppress weak reverberation/-soft tissue noise. It shows significant improvement on the contrast of bone surfacecompared with surrounding soft tissue. In addition, weak bones including highly-tilted spinous process can be enhanced. Therefore, the DFTM method is used as thebeamforming method in my first beamforming step to improve the bone detectionperformance.5.1.2 Tensor-based filteringIn the closed-loop beamforming, cross-correlation is performed on the channeldata to align the tilted bone signals back to the beamforming direction. Due to theexistence of soft tissue clutter and reverberation, correct channel data alignmentcould be challenging. Therefore, filtering on the channel data is needed prior to thecross correlation process.In the previous chapter, structure tensor-based masking is employed to clearout the soft tissue noise, offering high contrast for the bones. Due to the promisingresult, we extend this masking process to a filtering process. Instead of only usingthe control tensor to mask out the horizontally-filtered output (DFTM), the tensor isused to guide all four directional filtered results. This tensor-based filter is appliedon the pre-beamforming channel data. The goal is to provide a clean channel datacomposed of hyper-echoic targets only.The tensor-filtered output, I, can be calculated by:I =4∑1(T ′ ·M ·H), (5.1)73where H is the log-Gabor filtered results along the same direction along 0, 45, 90and 135, M is the dual tensor basis (DTB) for these angles:M = 0.7500−0.25,0.250.50.50.25,−0.25000.75,0.25−0.5−0.50.25. (5.2)Once the channel data pass through the tensor-based filtering, weak signalsfrom diffusive scattering and off-axis interference can both be suppressed. On theother hand, signals with strong edges (from hyper-echoic targets) can be main-tained/enhanced.This filtering process not only facilitates the channel data alignment processbut also suppresses weak soft tissue/reverberation clutters. By constructing thestructure tensor, strong line structures can be identified and enhanced while weakspeckle noise can be suppressed. No hard-coded thresholding is used in the noisesuppression process.5.1.3 Data alignmentAfter the tensor-based filtering, channel data are aligned through cross correla-tion between neighbouring channels. The channel that locates in the center of thecurrent transmit/receive aperture is first picked out. Then the data correspondingto the predicted bone locations (from the first beamforming step) are used as thecorrelation reference. The size of the data is decided by the ultrasound transmitfrequency and the sampling frequency (12 MHz in this study). It needs to cover afew cycles of the ultrasound pulses for correct data alignment (e.g., 5 cycles). Thesearch range is the same as the data size. A large search range is not needed sincethe shift is estimated between data in neighbouring channels. The amount of shiftis decided by the maximum cross-correlation coefficient, C:C(s) =L+N∑n=L−NA(n)×B(n+ s)),s=−P...P, (5.3)where P is search range, s is the channel number, ±N is the channel data windowsize, L is the bone location estimated from the first processing step. Since the shiftis estimated between neighbour channels, the final shift for each channel is there-74fore the accumulated shift from that specific channel to the picked center channelin this transmit/receive event.5.1.4 Simulation and experimentsAs discussed in Chapter 2, Field-II simulations, phantom and in vivo volunteerstudies are conducted to evaluate the performance of the closed-loop beamformingmethod. To mimic the specular reflection effect in Field II, I simulate two differentsimulation events where the point targets in the first step are used as the transducerelement in the second simulation event. The ‘bone’ targets are tilted counterclock-wise from 0 to 15 degrees to mimic the effect when the bone surface is not perpen-dicular to the transmit ultrasound wave. In phantom and in vivo volunteer studies,a BK ultrasound system with a curved linear array transducer is used for collect-ing pre-beamformed channel data (Analogic Corp., BK3500, Peabody, MA). Afocused transmit with a focal depth of 5 cm and a center frequency of 3.5 MHz areused for transmit. The maximum receive aperture is 192 and the receive F numberis 1. Dynamic receive beam delays are applied on the data based on relative loca-tions to the receive aperture. The output scan line is generated by summing the pre-beamformed data horizontally. Each transmit and receive event typically results inone beamformed scan line and this transmit and receive process is repeated until allscan lines have been generated to form an output image. All results are shown inthe same dynamic range, 45 dB. A plastic artificial vertebrae (Xincheng ScientificIndustries Co., Shanghai, China) is used for the phantom study. To demonstratethe ability of the tensor-based filtering on ultrasound data, a standard ultrasoundtissue phantom with hyper-echoic point and disk targets (Gammex 406 LE, Mid-dleton, Wisconsin, USA) is used. My goal is to prove that tensor-based filteringcan suppress weak speckles while maintaining hyper-echoic targets. In in vivo ex-periments, 60 data sets from 5 volunteers are collected including typical transverseand para-sagittal view following written consent. To quantitatively evaluate theeffect of the proposed Tensor-based beamforming method, the contrast ratio (CR)evaluation metric is employed. The bone surface is first manually segmented asthe ground truth. Then the contrast ratio for bones and surrounding structures areevaluated. Sensitivity and specificity studies are also performed where the sensi-75Figure 5.2: Point and disk phantom experiment results. (a) shows the pre-beamformed channel data received after a single transmit event. Dy-namic receive beam delay has been applied on the data based on relativelocation to the receive aperture. (b) shows the same pre-beamformeddata but after structure tensor based filtering. (c) and (d) are the normal-ized beamformed results for the point and disk phantom using DAS andtensor-based filtering, respectively. The CR of the point target and diskare both enhanced.tivity is defined as the percentage of pixels classified as bone pixels correctly andthe specificity is defined as the percentage of pixels classified as background pixelscorrectly.5.2 Results5.2.1 Tensor-based filtering resultsFigure 5.2(a) shows the pre-beamformed channel data collected from a point-and-disk phantom. Dynamic receive beam delay has been applied on the data based onrelative locations to the receive aperture. The number of receive channels is 128.76The output scan line is generated by summing the pre-beamformed data horizon-tally. Each transmit and receive event typically results in one beamformed scanline and this transmit and receive process is repeated until all scan lines have beengenerated to form an output image. The channel data shown in Figure 5.2(a) arebeamformed to generate the scan line 60 in Figure 5.2(c). The high intensity edgesin Figure 5.2(a) correspond to the hyper-echoic point target in Figure 5.2(c). Sincethe point target locates around scan line 56, the channel data is a little tilted towardsleft. The hyper-echoic data are surrounded by lower-amplitude speckles from thesoft-tissue phantom.Figure 5.2(b) shows the ST filtered channel data of Figure 5.2(a). After pass-ing the ST-based filtering, weak signals from the speckles can be clearly suppressedwhile strong signals from point targets are kept. Notice that the shape and directionof the hyper-echoic target are minimally changed. Figures 5.2(c) and (d) show the2D beamformed image without and with ST-based filtering. The DAS beamform-ing is used for Figure 5.2(c). The contrast of the point target and disk is enhanced inFigure 5.2(d). The contrast ratio for the point target is improved from 0.71 in DASto 0.85 in the ST-based result. For the disk target, the contrast ratio is increasedfrom 0.32 to 0.43. The CR improvement for the disk phantom is not as clear as thepoint target because parts of signals from the disk are relatively weak so they aresuppressed in the filtering process. Notice that all filtering is conducted on chan-nel data (not the beamformed image). The computation requirement is not largelyincreased since the ST-based filtering is handled scan line by scan line instead ofpixel by pixel.5.2.2 Simulation resultsFigure 5.3(a) shows an example of channel data received from simulated bones inField II. In this case, the ‘bone’ points are rotated counterclockwise 12 degreesfrom horizontal. As a result, the received channel data after time delay correctionare tilted. This can be seen from the signals around the 4 cm depth. Notice thesignal from a grating lobe is shown in the weak lines on the top left corner ofthe image. Figure 5.3(b) shows the channel data after tensor-based filtering. Themain-lobe signal from bones is maintained while other noises such as from grating77Figure 5.3: Simulation results. (a) shows an example channel data receivedfrom simulated bones in Field II. In this case, the ‘bone’ points are ro-tated counterclockwise for 10 degree from the horizontal position. Asa result, the received channel data after time delay correction are tilted.This can be seen from the signals around the 4 centimeter depth. Noticethe signal from a grating lobe is shown in the weak lines on the top leftcorner of the image. (b) shows the channel data after tensor-based filter-ing. The main-lobe signal from bones are maintained while other noisesuch as from grating lobe are suppressed since they are relatively weak.(c) shows the result after the cross-correlation alignment. The signalsare aligned back to the horizontal direction. As a result, the positive andnegative data will not be cancelled out during the beamforming process.The peak amplitude will not drop as the tilting increases.lobe are suppressed since they are relatively weak. Figure 5.3(c) shows the resultafter the cross-correlation alignment. The signals are aligned back to the horizontaldirection. As a result, the positive and negative data will not be cancelled out duringthe beamforming process. Figure 5.4 shows the beamformed output amplitudechanges as the tilting angle increases. In DAS beamforming results, the amplitudedrops quickly as the tilting angle increases. It becomes 30 dB less when the tiltingreaches 10 degrees. On the other hand, the amplitude maintains the same aftercross-correlation alignment when the tilting angle is less than 15 degrees. Oncethe shift is larger than the cross-correlation search range, the signal begins to dropquickly. This simulation study proves the effectiveness for the propose method inmaintaining the bone signal strength in cases of tilting.78Figure 5.4: The beamformed output amplitude change in dB as the tiltingincreases from 0 to 15 degrees. The signal drops quickly in the DASbeamforming while it maintains the same in the proposed closed-loopbeamforming method until it reaches the search boundary.5.2.3 Phantom resultsThe goal of the phantom study is to prove the proposed closed-loop beamformingmethod can increase the contrast of the bones without distortion. Figure 5.5(a) isthe DAS beamformed result of a vertebrae in its transverse view. The correspond-ing anatomy is shown as the small figure in the top left corner of (a). Since theplastic vertebrae does not have a large impedance mismatch on its surface, a lot ofsignals penetrate through the bone surface. Thus, multiple bone surfaces can be vi-sualized along the same scan line. In my proposed method, only one bone locationis picked in each scan line for closed-loop beamforming. Figure 5.5(c) shows thebeamformed result prior to scan conversion using the directional filtering method(DFTM) introduced in the previous chapter. DFTM clears out the noise under thebone surface which simplifies the bone detection procedure. The automatically de-79Figure 5.5: A transverse view of a vertebrae phantom. (a) shows the DASbeamformed result. (b) shows the result from the proposed closed-loopbeamforming method. (c) shows the beamformed result prior to scanconversion using the directional filtering method (DFTM) introduced inthe previous chapter. DFTM clears out the noise under the bone surfacewhich simplifies the bone detection procedure. The automatically de-tected bone surface points are shown as the red line. They are the firstintensity peaks from the bottom to the top of each scan line. The con-trast of the bones are greatly improved in (c) compared with the DASresult (a). Notice that the bone segment that have been fed back to thebeamforming process is especially sharpened. The shape of the boneitself matches the result from the DAS result. There is no detectabledistortion generated from the proposed closed-loop method.tected bone surface points are shown as the red line. They are the first intensitypeaks from the bottom to the top of each scan line. Figure 5.5(b) shows the finalresult from the proposed closed-loop beamforming method. The contrast of thebones are greatly improved compared with the DAS result. The bone surface isalso sharper and thinner compared with the DFTM method. The shape of the boneitself matches the result from DAS and there is no obvious distortion generated.Notice that only the bottom half of the bone surface is enhanced in Figure 5.5(b)(sharp and thin surfaces) while the top bone surface is not improved as much (thicksurfaces). This is because cross-correlation is only performed based on bottombone locations. However, this is not an issue for in vivo cases since typically thereis no more than one bone surface can be visualized in the same scan line due to thelow level of ultrasound signal penetration.805.2.4 In vivo resultsSame as Chapters 3 and 4, 60 data sets collected from 5 volunteers were usedin the in vivo study. Figure 5.6 shows a para-sagittal view of the vertebrae [11].The corresponding anatomy is shown as the small figure in the top left corner ofFigure 5.6(a). The laminae are shown as hyper-echoic wave shapes in the mid-dle of the image in Figures 5.6(a)-(f). The DAS beamforming results are shownin Figures 5.6(a) and (b). The red line in (b) marks the scan line-of-interest and isgenerated by the channel data shown in (g). According to (g), signals from the lam-ina surface are high amplitude echoes at the depth of 4 cm. They are tilted due tothe specular reflection and the relative angle between the lamina surface and ultra-sound pulses. Soft tissue clutters are shown as regions with low amplitudes abovethe bone surface. Reverberation artifacts are shown under the bone surface with thesame tilting. Conventional DAS beamforming sums the channel data horizontallyto generate the output scan line. Signals from the tilted bones are therefore sup-pressed due to the addition of positive and negative peaks. The spread of the bonesignals across the depth direction is also widened, resulting in a thick bone surface.On the other hand, soft tissue/ligament structures above the bone often have lesstilting and are not cancelled in the beamforming process (top right horizontal linesin (g)). Figure 5.6(h) shows the tensor-filtered channel data. Signals from softtissue/ligaments and reverberations are removed since their amplitudes are weakercompared with bones. On the other hand, strong bone echoes are maintained. Thisshows the benefit of applying filtering prior to beamforming. Once the bone sig-nals are diminished after summation, it is harder to distinguish them from strongligaments. After tensor-based filtering, cross-correlation is applied on channel dataso signals can be aligned to horizontal. Same as phantom studies, the region-of-interest used for cross-correlation is determined by the bone locations in the firstbeamforming step. Compared with (a) and (b), directional filtered results (shownin (d) and (e), before scan conversion) provide clear bone surfaces with high con-trast. This allows us to detect the bone surface automatically (the red line in (e)).The red rectangular box shown in (d) represents the ROI box used for contrast andROC analysis. Notice that the ROI box is displayed in the pre-scan converted data.The aligned channel data from (h) are shown in (i). After beamforming, the result81bone surface is sharp and clean with high contrast (f) (after scan conversion). Asa reference, the directional filtered result is shown in (c) (after scan conversion).Compared with (c) and (f), not only the contrast and thickness of the bone surfaceare improved, the soft-tissue clutter is also suppressed in the tensor-based filteringprocess.Figure 5.7 shows a transverse view of the vertebrae [11]. The DAS beamform-ing results are shown in (a) and (b). The corresponding anatomy is shown as thesmall figure in the top left corner of (a). In this particular view, the spinous pro-cess is the hyper-echoic region inside the blue rectangular box. The laminae aretwo horizontal line segments around 3 cm depth. The transverse processes are twotilted hyper-echoic regions near 4 cm in depth and -4 and 4 cm laterally. Sincethe laminae and transverse process in this viewing angle are close to perpendicularto the ultrasound incision angle, they can be visualized relatively well. The mainchallenge is the spinous process. (g) shows the channel data for the spinous processused to generate the scan line (shown as red) in (b). Large tilting in the channel datamakes the spinous process ‘disappeared’ in the DAS beamformed result. Instead itis shown as a hyper-echoic cloud.The tensor-based filtering suppresses weak data from surrounding soft tissueand keeps strong signals from the spinous process (as shown in (h)). The cross-correlation then aligned the filtered channel data back to the horizontal direction,which minimizes the signal cancellation during the summation (results in (i)). No-tice that since the spinous process does not provide continuous echoes in the chan-nel data, the effect of alignment is not as clear as other bones. (d) and (e) showdirectional-filtered results which have similar enhancement for spinous process(before scan conversion). The automatically detected bone surface is shown asthe red line in (e). The result bone surface from the closed-loop method is shownin (f). Although the improvement for the spinous process is small compared withdirectional filtered results in (d) and (e), both the laminae and transverse processshow high contrast and sharpness improvements.Figure 5.8 shows more in vivo results for DAS ((a), (d), (g)), directional filter-ing ((b), (e), (h)) and closed-loop beamforming ((c), (f), (i)). In all cases, both thedirectional filtering and closed-loop beamforming increase the contrast for bones.The soft tissue clutter is greatly reduced. Compared with the directional filtered82result, the closed-loop beamforming method offers a sharper and thinner bone sur-face. Another difference between the closed-loop method and the directional filter-ing method is the suppression of strong soft tissue ligaments above the bones. Thesuppression is achieved by data alignment based on the bone signal below (not theligament signal). As a result, the non-tilted ligament signals become tilted after thedata alignment. This creates a clear difference between (b) and (c) in the regionhighlighted by a blue box in Figure 5.8(b). In (b), the ligament blends togetherwith the s-shape laminae surface while it is suppressed in (c).From the results of 60 in vivo data sets, the overall CR of bones is improvedfrom 0.49 in DAS to 0.91 in AAF method, 0.96 in directional filtering (DFTM) to0.99 in the closed-loop beamforming with p < 0.0001. The ROC curve for the invivo data set in Figure 5.6 is shown in Figure 5.10. The area under the ROC curveis 0.93 for the AAF method, 0.97 for the directional filtering method and 0.99 forthe closed-loop beamforming method. It shows that the closed-loop method hasthe highest sensitivity and specificity of the bone surfaces.5.3 DiscussionIn this chapter, I propose a closed-loop beamforming method that is specificallydesigned to improve the sharpness and contrast of bones. The method is evaluatedon commonly-used standard views in spinal ultrasound imaging and the resultsshow clear contrast enhancement of the bones. The generated spine surfaces aresharp and thin. Structures with highly-tilted surfaces such as the curved spinousprocess can also be visualized. Strong ligaments above the bone surface could besuppressed provided that the bone location is correctly fed back to the beamformingprocess.Although there are benefits from the proposed method, there are several lim-itations associated with it. First of all, proper selection of the ROI is needed toachieve correct enhancement. The bone surface needs to be included in ROI. Onthe other hand, strong tissue layers in the near field should be excluded since theywill be dominant features in the tensor analysis, resulting suppression in weak bonetargets. Secondly, successful bone signal alignment depends on correct feedbackof bone locations. Wrong-labeled bone points could result in false suppression of83bones and thickened bone surface. Since the proposed method generates bone sur-faces with high contrast, sensitivity and specificity, bone locations can be updatedbased on the closed-loop result and fed back to the beamforming. This process canbe repeated until changes are negligible. Thirdly, challenges exist when the bonesurface and the ultrasound beam is close to parallel. In this case the received bonesignal is often very weak or completely disappeared. These weak bone signals aretreated as the noise and filtered out in the tensor-based filtering process. Examplebroken bone surfaces can be seen in red boxes from Figure 5.9(b). Currently, onlyone bone point is identified along each scan line. To recover these ‘broken’ lines,multiple bone points can be assigned in a single scan line to connect the brokenbone points. For each bone point, a cross-correlation alignment is performed torecover signals from this bone point. Specially tailored tensor-filtering parameterscan be used so weak signals are not suppressed in this case (or without tensor filter-ing). This addition is defined as vertical points enhancement. Figure 5.9 shows theresult from DAS (a), closed-loop beamforming with one iteration (b) and closed-loop beamforming with two iterations plus vertical enhancement (c). The bonesurface shown in red boxes are more continuous and smooth in (c). The improve-ment becomes minimal when the closed-loop beamforming is applied for the thirdtime.My proposed method is customized for visualizing bones. In cases whereboth bone and soft-tissue need to be visualized, the generated bone surface can befused with DAS ultrasound images, providing images with bone enhancement. Thecomputation requirement for the closed-loop beamforming is higher than DAS. Itinvolves directional filtering, first-stage beamforming, tensor-based filtering andcross-correlation. However, since most processing steps are only needed once perscan line, instead of pixel by pixel, the overall processing increase is not exponen-tial. Currently, the closed-loop beamforming requires 6.0 ms on a AMD RadeonR9 200 GPU for channel data of 900 by 192, 600 ms for a 100-scan line image(based on a single closed-loop iteration). With the advancement in GPU parallelbeamforming, it is feasible to implement the algorithm in real-time.845.4 ConclusionA closed-loop beamforming method is proposed to enhance the visualization ofhyper-echoic structures in ultrasound imaging. Phantom and in vivo studies showsignificant improvements in the contrast and sharpness of vertebrae structures. Sig-nal cancellation due to the specular reflection in tissue-bone interface is avoided.Compared with the DFTH method, sharper and clearer bone interface can be gen-erated. Weak bone surfaces can be enhanced due to bone locations feedback andcross-correlation alignments.85Figure 5.6: A para-sagittal view of the vertebrae. The laminae are shown ashyper-echoic wave shapes in the middle of the image in (a)-(f). TheDAS results are (a) and (b). The red line in (b) marks the position ofthe scan line which is generated by the channel data in (g). The resultfrom directional filtering is (c). The same result as (c) but without scanconversion is (d). The automatically-detected bone surface is shownas the red line in (e). In (g), signals from the lamina surface are highamplitude echos at the depth of 4 cm. (h) shows the tensor-filtered (g).The cross-correlation-aligned channel data are (i). The final closed-loop beamforming result is (f). Compared with (c) and (f), not onlythe contrast and sharpness of the bone surface are improved, the soft-tissue clutter is also suppressed. The red rectangular box shown in (d)represents the ROI box used for contrast and ROC analysis.86Figure 5.7: A transverse view of the vertebrae. The DAS results are (a) and(b). The spinous process is shown as the hyper-echoic region inside theblue rectangular box. The result from directional filtering is (c). Thesame result as (c) but without scan conversion is (d). The automatically-detected bone surface is shown as the red line in (e). (g) shows thechannel data for the spinous process along the scan line listed with thered line shown in (b). (h) shows the tensor-filtered (g). The cross-correlation aligned channel data are (i). The final closed-loop beam-forming result is (f). Compared with (c) and (f), not only the contrastand sharpness of the bone surface are improved, structures such as thespinous process can be clearly visualized.87Figure 5.8: More in vivo results for comparison for DAS, directional filteringand closed-loop beamforming. In all cases, both the directional filter-ing and closed-loop beamforming increase the contrast for bones. Thesoft tissue clutter is greatly reduced. Compared with the directional fil-tered result, the closed-loop beamforming method further suppressedsoft tissue noise through tensor-based filtering. The bone surface is alsosharper and thinner after the channel data alignment. One important im-provement is the suppression of strong soft tissue ligaments above thebones. This is very clear in the difference between (b) and (c).88Figure 5.9: Multiple iteration in closed-loop beamforming. The result fromDAS is shown in (a), closed-loop beamforming with one iteration isshown in (b) and closed-loop beamforming with two iterations plus ver-tical enhancement is shown in (c). Compared with (b), the bone surfaceshown in red boxes are more continuous and smooth in (c).Figure 5.10: The ROC curve for the in vivo data set in Figure 5.6. The areaunder the ROC curve is 0.93 for the AAF beamforming method, 0.97for the directional filtering method and 0.99 for the proposed closed-loop beamforming method.89Chapter 6Conclusions and future work6.1 ConclusionsIn this thesis, I first propose a simple AAF method which is designed based onthe characteristics of channel data from bone targets (specular reflectors, hyper-echoic). The AAF method improves the contrast of bones since the overall phasechange across the receive aperture is small for hyper-echoic targets like bones. Themain advantage of this method is it is easy to implement and does not introducemuch additional processing except for the overall phase estimation. The poten-tial drawback is lacking of enhancement since it does not address the tilted signalcancellation issue in the beamforming process.By reorganizing the undesirably tilted bone signals in channel data (due tospecular reflection) back to the beamforming direction, signal cancellation in thebeamforming can be avoided and bone signals can be enhanced. To achieve thisgoal, a DFTM method is first developed which selects the pass angle band and reg-ulates the tilted channel data back to the horizontal direction. Then the alignmentperformance is improved by feeding the position of bones back to the beamformingprocess and performing localized cross-correlation (closed-loop beamforming). Toalign the data correctly, a structure tensor-based filtering is applied to clean out softtissue noise and highlight the bones. Phantom and in vivo studies show further im-provements in contrast, sensitivity and specificity for the DFTM and closed-loopbeamforming method compared with the AAF method. The closed-loop beam-90forming shows the highest improvement. Bones which sit in orientations that arefar from perpendicular to the transmit ultrasound beam can now be enhanced withthe closed-loop method.To summarize, three beamforming methods which are dedicated to enhancethe visualization of bones are proposed. It starts with the simplest AAF beamform-ing, followed by the directional filter based method and the most sophisticatedclosed-loop beamforming method. The directional filtering method can be usedas a pre-processing step (bone surface detection) for the closed-loop beamformingmethod for better performance. Appropriate simulations, experiments and reason-able performance metrics are also developed for evaluating the performance of theproposed methods in bone visualization.The main contributions from my work are: (1) Design a strategy to simu-late specular reflection in Field II. (2) Propose an AAF beamforming method tohighlight the hyper-echoic targets. (3) Developed a DFTM method to regulate thetilted bone signals in channel data back to the horizontal direction. (4) Apply andevaluate the structure-tensor based filtering on ultrasound channel data. (5) Feedback the bone location information back to the beamforming process to perform aclosed-loop correction.In Chapter 3, an accumulated angle factor based beamforming method (AAF)customized for bone surface enhancement is proposed. This approach analyzesthe phase change across the aperture direction. It first applies a Hilbert trans-form on delay compensated channel data across the receive aperture. The accu-mulated phase change across the receive aperture is then calculated and utilizedas the weight in the beamforming output. Sine wave demonstration, Field II sim-ulation, phantom and in vivo studies show the contrast of bones can be improvedsince soft tissue/highly tilted off-axis interference often have large accumulatedphase changes across the receive aperture. The benefits from AAF are: (1) Easyto implement. AAF only introduces less than 8 percent additional processing thanthe DAS method. (2) Good contrast improvement. Based on in vivo study results,AAF improves the contrast ratio from 0.49 (DAS) to 0.91.However, AAF does not address the bone signal beamforming cancellationissue caused by non-perpendicular reflection. The accumulated phase change forhighly tilted bone signals are also large, resulting in signal suppression. Structures91like spinous process can not be visualized in AAF results. Weak bone structuresare not enhanced.In Chapter 4, a log-Gabor directional filtering with tensor-based masking methodis proposed to address the above issue. Log-Gabor directional filtering is used toalign the tilted bone signals back to the beamforming direction. This adjustmentavoids bone signal cancellation in the beamforming process and increases the vi-sualization of structures such as spinous process. Since directional filtering alignssignals to the beamforming direction, not only the bone signals are enhanced, re-verberations, off-axis interference and soft tissue are often enhanced as well. Tomaintain the contrast for bones, tensor-based masking is employed to mask thedirectional filtered channel data based on intensity levels. As a result, weak sig-nals from soft tissue and reverberations are masked out. This method is defined asthe directional filtering with tensor-based masking method (DFTM). The contrastfor bones is further improved. Compared with the AAF beamforming, the CR forbones in the in vivo data sets is improved from 0.91 to 0.96. Structures such as thespinous process can be recovered. The sensitivity of bone is also improved.Compared with DAS and AAF methods, the increase in computation require-ment for DFTM is not large. This is because in DFTM, the output is generatedscan line by scan line instead of pixel by pixel. The DFTM method is implementedin open CL and the overall time to perform DFTM on channel data of 900 by 192(one scan line) is only 1.1 ms for a AMD Radeon R9 200 series GPU. For an imageof 100 scan lines, the processing time is 110 ms.There are several limitations for the proposed DFTM method. First, due tothe normalization procedure in the tensor-based masking process, only relativelystrong signals within the channel data ROI can be enhanced. As a result, weakbone structures in deep depth cannot be clearly visualized. In addition, strongligament tissue above the bone surfaces can be falsely enhanced since bone signalsmay not be stronger than ligaments due to non-perpendicular reflection. Finally,DFTM often generates thick bone surface since the alignment process (directionalfiltering) is not ideal. Signals surrounding the bone surface such as reverberationscould also be enhanced.Ideally, if the location of the bones is known, the channel data corresponding tothe bones can be aligned through cross-correlation. The sharpness of the bone sur-92face can be maintained. Weak bone signals with large tilting can also be correctedusing this approach. Strong ligaments can be suppressed since the cross-correlationalignment is conducted based on the tilted bone signal (not the ligaments). There-fore, I propose a closed-loop beamforming method which utilizes the bone locationinformation to guide the beamforming process. The detail is described in Chapter5. Since successful correlation requires correct information on bone locations, theDFTM method is employed to increase the sensitivity and contrast of bones. Thebone locations are then identified automatically by selecting the first intensity peakfrom deep to shallow direction (bottom to top). Tensor-based filtering is applied onthe channel data prior the alignment to suppress soft tissue noise. Cross-correlationis then performed based on identified bone locations, which acts like a closed-loopcorrection step. This closed-loop process can be repeated until the improvement ofbone visualization is minimal.From the results of 60 in vivo data sets, the overall CR of bones is improvedfrom 0.49 in DAS to 0.91 in AAF method, 0.96 in directional filtering to 0.99 inthe closed-loop beamforming with p < 0.0001. The area under the ROC curve is0.93 for the AAF method, 0.97 for the directional filtering method and 0.99 for theproposed closed-loop beamforming method. It shows that the closed-loop methodhas the highest sensitivity and specificity of the bone surfaces.The biggest limitation of the closed-loop method is high computation require-ment. It requires 6.0 ms on a AMD Radeon R9 200 GPU for channel data of 900by 192, 600 ms for a 100-scan line image (a single closed-loop iteration). Properselection of the ROI and robust bone location detection could also be challengingin actual clinical applications.In practice, the appropriate beamforming method should be chosen based onapplication requirement. If sharp and accurate bone surfaces are needed and thereis no real-time requirement, the closed-loop beamforming method should be cho-sen. In cases where high sensitivity and contrast of bones are desirable, the DFTHmethod should be selected. For normal bone signal enhancement with high frame-rate requirements, the accumulated-angle-factor based method should be used.Notice that no spatial compounding is employed in our study due to systemlimitation. However, we believe our proposed methods are still effective in spatialcompounding cases since tilting in channel data caused by specular reflection al-93ways exist regardless of the transmit angle. We do expect the bone surface wouldbe more connected with the addition of information from multiple transmit angles.Applying our methods on spatial compounded data remains one of the importantfuture tasks.It is anticipated that the improved vertebrae surface can benefit bone surfaceextraction and therefore facilitating the deployment of ultrasound imaging in spineanesthesia procedures [11, 13]. In case of ultrasound guided spinal procedures,both the vertebrae surface and soft tissue are needed. In this case, the extracted ver-tebrae surface can be merged back to the DAS beamformed image. This requiresthe system running the proposed bone visualization method and DAS beamform-ing in parallel. The bone surface is first extracted from the generated bone image.It is then overlaid on top of the DAS beamformed images. As a result, both of thebenefits from DAS (clear soft tissue structures) and my method (the bone surface)can be combined and displayed to the user. This combined ultrasound image canbe very useful for identifying the path for needle injections.6.2 Future work6.2.1 Algorithm accelerationThe proposed method focuses on aligning the titled channel data (caused bythe specular reflection) back to the beamforming direction. The computation costis high especially when large 1D arrays or 2D arrays are applied. In those cases,simplifying the algorithm and reducing the computation cost become critical. Oneof the future work is therefore focus on speeding up the processing time for direc-tional filtering and tensor analysis. My goal is to implement the proposed beam-forming algorithms into the ultrasound system in real-time, i.e., at least 5 framesper second. To achieve this goal, the proposed algorithm needs to be optimized tofit the modern GPU structures. Fox example, two dimensional convolution (one ofthe most expensive processing units in my algorithm) can be separated into a se-ries of one dimensional convolution. Since directional filtering kernel is typicallynot Gaussian (not separable), an error analysis needs to be performed to quantify94the trade-off. Pre-loading the channel data into on-chip memory can avoid unnec-essary data transfer between external and on chip memory, which can also speedthe processing up. GPU profile tools such as codeXL can be helpful for furtherperformance improvement. By employing these optimization steps, along with theadvancement of modern GPU architecture, the proposed algorithms can be exe-cuted in real-time, which is critical for real-time guidance.6.2.2 Robustness verificationThe goal of the research is increasing the visualization of bone structures inultrasound imaging and therefore provide real-time guidance on spinal needle in-jections. To achieve this goal, the robustness of the algorithm needs to be evaluated.I tested the proposed methods in 60 in vivo data sets from 5 volunteers and was ableto use the same processing parameters (e.g., filter kernel size, angle and frequencypass-band, cross-correlation search region) for all data sets. However, there arelimitations for these data sets. First of all, the number of volunteers is limited andthey are all healthy thin patients. In real life the size of patient varies and signalsfrom hyper-echoic targets are different. In addition, different imaging depth andultrasound frequency are used for different patients. Therefore, further verifica-tion of the robustness and effectiveness of proposed algorithms is needed. I planto increase the number of in vivo study subjects to 50 and include half of acousti-cally ‘difficult’ patients. Changes in algorithm parameters are not expected sincethe relative ratio between sampling frequency and ultrasound transmit frequencyis unchanged (always four times). In addition, since structure tensor-based mask-ing and filtering use soft thresholding and analyze the relative strength of signals,changing the channel data intensities is unlikely to change the performance.6.2.3 Translation to clinical practiceOnce the real-time implementation requirement is fulfilled and robustness/ ef-fectiveness of the algorithms are verified, the ultimate goal is transferring the pro-posed methods to clinical practice. I plan to focus on ultrasound-guided spinal nee-dle injection procedures such as facet joint injections and epidural. The enhancedbones generated from the proposed methods will be merged on top of the regular B-mode ultrasound images. It can be color-coded such as using yellow or red. Since95the bone surface and B-mode ultrasound images use the same ultrasound data, theframe rate should be unchanged. To evaluate the benefits of proposed algorithms,the procedure imaging time and/or procedure preparing time will be recorded. Thenumber of needle injections will also be analyzed and compared with cases withoutbone enhancement. I expect the time should be significantly reduced.6.2.4 Combination with artificial intelligenceWith the resurgence of machine learning and artificial intelligence for the pastdecade, there have been increasing interest in applying artificial intelligence tech-niques on ultrasound imaging and guided procedures. Previous studies have showngreat success in tissue classification, image segmentation, object detection and im-age quality improvements using deep learning. Among those studies, many havebeen dedicated for the ultrasound-guided spine needle injections [63, 64]. For myresearch, one interesting future topic is therefore automatically identifying bonepixels in the beamformed image through deep learning and then feedback the loca-tion of bones to the beamforming process. This is more accurate than the currentfirst intensity peak method. It can also select the ROI for tensor-based filteringbased on bone location distribution to avoid suppression on weak bone pixels. Thesize of ROI for each filter can vary from scanline-to-scanline depending on thebone channel data strength. For example, weak bone pixels correspond to smallROI sizes for filtering and cross-correlation alignment and hyper-echoic bone pix-els correspond to large ROI sizes. In addition, the deep learning-based classifica-tion can be extended to the channel data stage where each sample in the channeldata can be classified as signals from bones or surrounding soft tissue/off axisinterference. Then, only signals from bones are beamformed to generate the out-put. The main drawback for this approach would be long processing time due tothe much larger size of pre-beamformed channel data compared with beamformedoutput images. However, previous research also proves the feasibility of trainingthe translation between fast-acquired, low-quality ultrasound images with slowly-acquired, sophisticated, high-quality ultrasound images [65]. Therefore, it couldbe possible to generate a lookup table (LUT) showing the relations between DASbeamformed images and closed-loop beamforming output. Then the closed-loopoutput can be generated in real-time based on the DAS images and LUT.96Although deep learning based algorithms bring more opportunities, it does notmean there is less use for traditional signal and image processing algorithms. Infact, good signal processing methods can improve the quality of information pass-ing into the learning process and help the classification sensitivity and specificity.For example, the phase-symmetry information is used as a feature when identify-ing the epidural space through deep learning [66]. In my case, it is expected to bebeneficial utilizing the closed-loop, DFTM or AAF beamformed bone surfaces asthe input for Epidural space detection.97Bibliography[1] T. Szabo. Diagnostic ultrasound imaging inside out. Elsevier AcademicPress, 2004. → page 1[2] H. Holm and B. Skjoldbye. Interventional ultrasound. Ultrasound inMedicine and Biology, 22:773–789, 1996. → page 1[3] J. Horter R. Conradi E. Martin T. Grau, R. Leipold and J. Motsch.Paramedian access to the epidural space: the optimum window forultrasound imaging. Journal of Clinical Anesthesia, 13:213–217, 2001. →pages 1, 3, 65[4] A. Kirby A. Moulton R. Burwell, R. Aujla and J. Webb. The early detectionof adolescent idiopathic scoliosis in three positions using the scoliometerand real-time ultrasound: Should the prone position also be used? Studies inHealth Technology and Informatics, 88:74–80, 2002. → page 1[5] A. Khamene M. Callstrom W. Wein, S. Brunke and N. Navab. Automaticct-ultrasound registration for diagnostic imaging and image-guidedintervention. Medical Image Analysis, 12:577–585, 2007. → page 1[6] R. Peters J. Thomas and P. Jeanty. Automatic segmentation of ultrasoundimages using morphological operators. IEEE Transactions on MedicalImaging, 10:180–186, 1991. → page 1[7] F. Langlotz H. Talib J. Kowal, C. Amstutz and M. Ballester. Automatedbone contour detection in ultrasound b-mode images for minimally invasiveregistration in computer-assisted surgery an in vitro evaluation. TheInternational Journal of Medical Robotics and Computer Assisted Surgery,3:341–348, 2007.[8] A. Digioia D. Amin, T. Kanade and B. Jaramaz. Ultrasound registration ofthe bone surface for surgical navigation. Computer Aided Surgery, 8:1–16,2003. → page 198[9] A. Rick M. Stockheim B. Brendel, S. Winter and H. Ermert. Registration of3d ct and ultrasound datasets of the spine using bone structures. ComputerAided Surgery, 7:146–155, 2005. → pages 1, 3[10] J. Pelletier S. Chen D. Tampieri C. Yan, B. Goulet and D. Collins. Towardsaccurate, robust and practical ultrasound-ct registration of vertebrae forimage-guided spine surgery. The International Journal of Medical Roboticsand Computer Assisted Surgery, 6:523–537, 2011. → page 6[11] R. Rohling A. Rasoulian and P. Abolmaesumi. Feature-based multibodyrigid registration of ct and ultrasound images of lumbar spine. MedicalPhysics, 39:3154–3166, 2012. → pages 1, 81, 82, 94[12] M. Tiouririne F. Mauldin, K. Owen and J. Hossack. The effects of transducergeometry on artifacts common to diagnostic bone imaging with conventionalmedical ultrasound. IEEE Transactions on Ultrasonics Ferroelectrics andFrequency Control, 59:1101–1114, 2012. → pages 1, 2, 9, 65[13] A. Jain and R. Taylor. Understanding bone responses in b-mode ultrasoundimages and automatic bone surface extraction using a bayesian probabilisticframework. Medical Imaging: Ultrasonic Imaging and Signal Processing,5373:131–142, 2004. → pages 1, 65, 94[14] D. Christensen. Ultrasonic bioinstrumentation. John Wiley and Sons, 1988.→ pages 1, 9[15] M. McKeag M. Anderson and G. Trahey. The impact of sound speed errorson medical ultrasound imaging. Journal of the Acoustical Society ofAmerica, 107:3540–3548, 2000. → page 1[16] T. Johansen S. Mehdizadeh, A. Austeng and S. Holm. Minimum variancebeamforming applied to ultrasound imaging with a partially shaded aperture.IEEE Transactions on Ultrasonics Ferroelectrics and Frequency Control,59:683–693, 2012. → page 2[17] J. Hermans E. Samset R. Bandaru, A. Sornes and J. D’hooge. Delay andstandard deviation beamforming to enhance specular reflections inultrasound imaging. IEEE Transactions on Ultrasonics Ferroelectrics andFrequency Control, 63:2057–2068, 2016. → page 2[18] A. Hodgson I. Hacihaliloglu, R. Abugharbieh and R. Rohling. Bone surfacelocalization in ultrasound using image phase-based features. Ultrasound inMedicine and biology, 35:1475–1487, 2009. → page 299[19] J. Li Z. Wang and R. Wu. Time-delay-and time-reversal-based robust caponbeamformers for ultrasound imaging. IEEE Transactions on MedicalImaging, 24:1308–1322, 2005. → page 3[20] J. Leendertz A. Preece M. Klemm, I. Craddock and R. Benjamin. Improveddelay-and-sum beamforming algorithm for breast cancer detection.International Journal of Antennas and Propagation, 2008:1–9, 2008. →page 3[21] Z. Thomenius. Evolution of ultrasound beamformers. IEEE InternationalUltrasonics Symposium, pages 1615–1622, 1996. → page 3[22] M. Sasso and C. Cohen-Bacrie. Medical ultrasound imaging using the fullyadaptive beamformer. IEEE Conference of Acoustics, Speech, and SignalProcessing, 2:489–492, 2005. → pages 3, 12[23] J. Mann and W. Walker. A constrained adaptive beamformer for medicalultrasound: Initial results. IEEE International Ultrasonics Symposium,pages 1807–1810, 2002. → pages 3, 12[24] S. Wang and P. Li. Mvdr-based coherence weighting for high-frame-rateadaptive imaging. IEEE Transactions on Ultrasonics Ferroelectrics andFrequency Control, 57:281–289, 2009. → pages 3, 12[25] B. Asl and A. Mahloojifar. Eigenspace-based minimum variancebeamforming applied to medical ultrasound imaging. IEEE Transactions onUltrasonics Ferroelectrics and Frequency Control, 57:2381–2390, 2010. →pages 3, 12[26] A. Austeng J. Synnevag and S. Holm. Adaptive beamforming applied tomedical ultrasound imaging. IEEE Transactions on UltrasonicsFerroelectrics and Frequency Control, 54:1606–1613, 2007. → pages 3, 12[27] C. Seo and J. Yen. Sidelobes suppression in ultrasound imaging using dualapodization with cross-correlation. IEEE Transactions on UltrasonicsFerroelectrics and Frequency Control, 55:2198–2210, 2008. → pages 3, 12[28] C. Nilsen and S. Holm. Wiener beamforming and the coherence factor inultrasound imaging. IEEE Transactions on Ultrasonics Ferroelectrics andFrequency Control, 57:1329–1346, 2010. → pages 3, 13, 14[29] V. Shamdasani P. Kalman D. Maxwell F. Vignon, W. Shi and J. Powers.Transcranial image quality improvement with a multi-step approach. IEEE100International Ultrasonics Symposium, pages 1284–1287, 2013. → pages3, 13[30] M. Parrilla J. Camacho and C. Fritsch. Grating-lobes reduction byapplication of phase coherence factors. IEEE International UltrasonicsSymposium, page 341344, 2009. → pages 3, 14[31] K. W. Rigby K. W. Hollmand and M. ODonnell. Coherence factor ofspeckle from a multi-row probe. IEEE International UltrasonicsSymposium, page 12571260, 1999. → pages 3, 13[32] P. Li and M. Li. Adaptive imaging using the generalized coherence factor.IEEE Transactions on Ultrasonics Ferroelectrics and Frequency Control,50:128–141, 2007. → pages 3, 13[33] J. Buskenes J. Asen and C. Nilsen. Implementing capon beamforming on agpu for real-time cardiac ultrasound imaging. IEEE Transactions onUltrasonics Ferroelectrics and Frequency Control, 61:76–85, 2014. → page3[34] B. Yiu and A. Yu. Gpu-based minimum variance beamformer for syntheticaperture imaging of the eye. Ultrasound Med. Biol., 41:871–883, 2015. →page 3[35] C. Chen X. Zeng and Y. Wang. Eigenspace-based minimum variancebeamformer combined with wiener postfilter for medical ultrasoundimaging. Ultrasonics, 52:996–1004, 2012. → page 3[36] F. Gran I. Holfort and J. Jensen. Broadband minimum variancebeamforming for ultrasound imaging. IEEE Transactions on UltrasonicsFerroelectrics and Frequency Control, 56:314–325, 2009. → page 3[37] V. Chan K. Chin, A. Macfarlane and R. Brull. The use of ultrasound tofacilitate spinal anesthesia in a patient with previous lumbar laminectomyand fusion: A case report. Journal of Clinical Ultrasound, 37:482–485,2009. → page 3[38] V. Chan and A. Perlas. Atlas of ultrasound-guided procedures ininterventional pain management. Basics of Ultrasound Imaging, 2011. →page 3[39] W. Hedrick. Image and signal processing in diagnostic ultrasound imaging.Journal of Diagnostic Medical Sonography, 5:231–239, 1989. → page 5101[40] D. Rubin. Epidemiology and risk factors for spine pain. Neurologic Clinics,25:353–371, 2007. → page 5[41] A. Parr. Lumbar interlaminar epidural injections in managing chronic lowback and lower extremity pain: A systematic review. Pain Physician, 12:163–188, 2009. → page 5[42] N. Sehgal E. Dunbar M. Boswell, J. Colson and R.Epter. A systematicreview of therapeutic facet joint interventions in chronic spinal pain. PainPhysician, 10:229–253, 2007. → page 6[43] S. Gill G. Fichtinger E. Chen, P. Mousavi and P. Abolmaesumi. Ultrasoundguided spine needle insertion. The International Society for OpticalEngineering, 7625:7625 – 7268, 2010. → page 6[44] D. Bainbridge C. Wedlake A. Wiles D. Pace T. Peters J. Moore, C. Clarke.Image guidance for spinal facet injections using tracked ultrasound. MedicalImage Computing and Computer-Assisted Intervention, pages 516–523,2009.[45] E. Al-Atta’s V. Lessoway S. Massey D. Tran, A. Kaman and R. Rolling.Single-operator real-time ultrasound-guidance to aim and insert a lumbarepidural needle. Canadian Journal of Anesthesia, 57:313–321, 2010.[46] R. Jalal-M. Welch I. Ayukawa S. Nagpal A. Lasso M. Jaeger D. BorschneckG. Fichtinger P. Mousavi T. Ungi, P. Abolmaesumi. Spinal needle navigationby tracked ultrasound snapshots. IEEE Transactions on BiomedicalEngineering, 59:2766–2772, 2012. → page 6[47] K. Kirkham and K. Chin. Ultrasound for central neuraxial blockade.Current Anesthesiology Reports, 3:242–249, 12 2013. → page 6[48] Tran M Yoon SH, OBrien SL. Ultrasound guided spine injections:Advancement over fluoroscopic guidance? Current Physical Medicine andRehabilitation Reports, 3:104–113, 1 2013. → page 6[49] A. Rasoulian, R. Rohling, and P. Abolmaesumi. Lumbar spine segmentationusing a statistical multi-vertebrae anatomical shape+pose model. IEEETransactions on Medical Imaging, 32:1890–1900, 2013. → page 6[50] J. Osborn-A. RasoulianS.Nouranian V. Lessoway R. Rohling P. Abolmaesumi A. Seitel, S. Samira.Ultrasound-guided spine anesthesia: Feasibility study of a guidance system.Ultrasound in Medicine and Biology, 42:3043–3049, 2016. → page 6102[51] P. Mousavi A. Rasoulian, P. Abolmaesumi. Featurebased multibody rigidregistration of ct and ultrasound images of lumbar spine. Medical Physics,39:3154–3166, 2012. → page 7[52] R. Rohling-P. Abolmaesumi I. Hacihaliloglu, A. Rasoulian. Local phasetensor features for 3-d ultrasound to statistical shape+ pose spine modelregistration. IEEE Transactions on Medical Imaging, 33:2167–2179, 2014.→ page 7[53] J.Osborn S. Sojoudi S. Nouranian V. Lessoway R. Rohling P. AbolmaesumiA. Rasoulian, A. Seitel. Ultrasound-guided spinal injections: a feasibilitystudy of a guidance system. International Journal of Computer AssistedRadiology and Surgery, 10:1417–1425, 2015. → page 7[54] T. Salcudean V. Lessoway G. Ng P. Beigi, R. Rohling. Needle trajectory andtip localization in real-time 3-d ultrasound using a moving stylus.Ultrasound in Medicine and Biology, 41:2057–2070, 2015.[55] K. Dickie C. Leung B. Zhuang, L. Pelissier. Freehand ultrasound imagingsystems and methods for guiding fine elongate instruments. US Patent8,556,815.[56] S. Salcudean G. Ng P. Beigi, R. Rohling. Spectral analysis of the tremormotion for needle detection in curvilinear ultrasound via spatiotemporallinear sampling. International Journal of Computer Assisted Radiology andSurgery, 11:1183–1192, 2016. → page 7[57] N. Vasilyev Y. Suematsu R. Cleveland J. Huang, J. Triedman and P. Dupont.Imaging artifacts of medical instruments in ultrasound-guided interventions.Journal of Ultrasound, 26:1303–1322, 2007. → page 9[58] J. A. Jensen. Simulation of advanced ultrasound systems using field ii.International Symposium on Biomedical Imaging from Nano to Macro, 1:636–639, 2004. → pages 16, 17, 18[59] T. Johansen S. Mehdizadeh, A. Austeng and S. Holm. Eigenspace-basedminimum variance beamforming applied to ultrasound imaging ofacoustically hard tissues. IEEE Transactions on Medical Imaging, 31(1912-1921), 2012. → page 24[60] J. Noble D. Boukerroui and M. Brady. On the choice of band-passquadrature filters. Journal of Mathematical Imaging and Vision, 21(5380),2004. → page 49103[61] H. Knutsson and C. Westin. Normalized and differential convolution. IEEEComputer Society Conference on Computer Vision and Pattern Recognition,21(5380), 1993. → pages 50, 51, 53[62] G. Granlund and H. Knutsson. Signal processing for computer vision.Kluwer, 1995. → pages 51, 53[63] S. Salcudean G. Ng P. Beigi, R. Rohling. Casper: Computer-aidedsegmentation of imperceptible motiona learning-based tracking of aninvisible needle in ultrasound. International Journal of Computer AssistedRadiology and Surgery, 12:1857–1866, 2017. → page 96[64] V. Gunka P. Abolmaesumi R. Rohling J. Hetherington, V. Lessoway. Slide:Automatic spine level identification system using a deep convolutionalneural network. International journal of computer assisted radiology andsurgery, 12:11891198, 2017. → page 96[65] E. Roux H. Liebgott D. Friboulet M. Gasse, F. Millioz. Accelerating planewave imaging through deep learning-based reconstruction: An experimentalstudy. IEEE International Ultrasonics Symposium, pages 1–4, 2017. →page 96[66] M. Pesteie, V. Lessoway, P. Abolmaesumi, and R. N. Rohling. Automaticlocalization of the needle target for ultrasound-guided epidural injections.IEEE Transactions on Medical Imaging, 37:81–92, 2018. → page 97104

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:
https://iiif.library.ubc.ca/presentation/dsp.24.1-0372823/manifest

Comment

Related Items