Thesis (M.App.Sc.) - Carleton University, 2007. / Includes bibliographical references (p. 82-89). Also available in electronic format on the Internet.

This thesis addresses the problem of analysing clinical MRI using modern computer vision methods for a variety of clinical and research-related tasks. We use automated machine learning algorithms to develop a spinal MRI analysis framework for a number of tasks such as vertebrae detection, labelling; disc and vertebrae segmentation, and radiological grading, and we validate the framework on a large, heterogeneous dataset of 300 symptomatic back pain patients from multiple clinical sites and scanners. Our framework has a number of back pain research and other spine-related clinical applications and could hopefully find application in a clinical workflow in the future. Our framework has five steps -- detection, labelling, segmentation, support regions and features, and machine learning for radiological measurements. The framework works in full 3D and has currently been implemented on sagittal T2 slices. We use Deformable Part Models along with a chain model to detect and label vertebrae, and a powerful graph cuts based method for vertebrae and disc segmentation. The labelled detections and segmentations are used to place support regions for feature extraction, which are mapped into a number of radiological measurements -- namely Pfirrmann grade, disc space narrowing, and herniation/bulge. The radiological ground truth was provided by a clinical radiologist with 25 years experience. We demonstrate a high performance in the measurement in each. The measurements are performed using support vector machines and support vector regressors learned on training data. We next investigate the problem of what is the best method of obtaining support regions. We first used pixel intensity features to predict the Pfirrmann grade, narrowing and bulge/herniation, with vertebrae segmentation to localise their support regions. Since segmentation of spine images, especially intervertebral discs is an unsolved problem and algorithms are prone to failure, we then ask the question, to segment or not to segment. To answer the question, we compare results on Pfirrmann grade prediction with three different points on the no segmentation to full disc segmentation involving no segmentation, vertebrae segmentation, or disc segmentation and find that vertebrae segmentation suffices. We finally show preliminary results in distinguishing between different radiological conditions related to the posterior side of the disc more finely than before in literature, taking information from both sagittal and axial slices to attempt to distinguish between herniated and bulged discs.

Indiana University-Purdue University Indianapolis (IUPUI) / A novel algorithm was developed to convert raw 2-dimensional endoscope videos into 3-dimensional view. Minimally invasive surgeries aided with 3D view of the invivo site have shown to reduce errors and improve training time compared to those with 2D view. The novelty of this algorithm is that two cues in the images have been used to develop the 3D. Illumination is the rst cue used to nd the darkest regions in the endoscopy images in order to locate the vanishing point(s). The second cue is the presence of ridge-like structures in the in-vivo images of the endoscopy image sequence. Edge detection is used to map these ridge-like structures into concentric ellipses with their common center at the darkest spot. Then, these two observations are used to infer the depth of the endoscopy videos; which then serves to convert them from 2D to 3D. The processing time is between 21 seconds to 20 minutes for each frame, on a 2.27GHz CPU. The time depends on the number of edge pixels present in the edge-detection image. The accuracy of ellipse detection was measured to be 98.98% to 99.99%. The algorithm was tested on 3 truth images with known ellipse parameters and also on real bronchoscopy image sequences from two surgical procedures. Out of 1020 frames tested in total, 688 frames had single vanishing point while 332 frames had two vanishing points. Our algorithm detected the single vanishing point in 653 of the 688 frames and two vanishing points in 322 of the 332 frames.

Medical imagery is increasingly evolving towards higher resolution and throughput. The increasing volume of data and the usage of multiple and often novel imaging modalities necessitates the use of mathematical and computational techniques for quicker, more accurate and more robust analysis of medical imagery. The fields of computer vision and machine learning provide a rich set of techniques that are useful in medical image analysis, in tasks ranging from segmentation to classification and population analysis, notably by integrating the qualitative knowledge of experts in anatomy and the pathologies of various disorders and making it applicable to the analysis of medical imagery going forward. The object of the proposed research is exactly to explore various computer vision and machine learning methods with a view to the improved analysis of multiple modalities of brain and cardiac imagery, towards enabling the clinical goals of studying schizophrenia, brain tumors (meningiomas and gliomas in particular) and cardiovascular disorders. In the first project, a framework is proposed for the segmentation of tubular, branched anatomical structures. The framework uses the tubular surface model which yields computational advantages and further incorporates a novel automatic branch detection algorithm. It is successfully applied to the segmentation of neural fiber bundles and blood vessels. In the second project, a novel population analysis framework is built using the shape model proposed as part of the first project. This framework is applied to the analysis of neural fiber bundles towards the detection and understanding of schizophrenia. In the third and final project, the use of mass spectrometry imaging for the analysis of brain tumors is motivated on two fronts, towards the offline classification analysis of the data, as well as the end application of intraoperative detection of tumor boundaries. SVMs are applied for the classification of gliomas into one of four subtypes towards application in building appropriate treatment plans, and multiple statistical measures are studied with a view to feature extraction (or biomarker detection). The problem of intraoperative tumor boundary detection is formulated as a detection of local minima of the spatial map of tumor cell concentration which in turn is modeled as a function of the mass spectra, via regression techniques.

Maternal diet can have a great impact on the health and development of the fetus. Poor fetal nutrition has been linked to the development of a set of conditions in later life, such as coronary heart disease, type 2 diabetes and hypertension, while restricted growth can result in hypogylcemia, hypocalcemia, hypothermia, polycythemia, hyperbilirubinemia and cerebral palsy. High alcohol consumption during pregnancy can result in Fetal Alcohol Syndrome, a condition that can cause growth retardation, lowered intelligence and craniofacial defects. Current biometric assessment of the fetus involves size-based measures which may not accurately portray the state of fetal development, since they cannot differentiate cases of small-but-healthy or large-but-unhealthy fetuses. This thesis aims to outline a set of more appropriate measures of accurately capturing the state of fetal development. Specifically, soft tissue area and liver volume measurement are examined, followed by facial shape characterisation. A number of tools are presented which aim to allow clinicians to achieve accurate segmentations of these landmark regions. These are modifications on the Live Wire algorithm, an interactive segmentation method in which the user places a number of anchor points and a minimum cost path is calculated between the previous anchor point and the cursor. This focuses on giving the clinician intuitive control over the exact position of the segmented contour. These modifications are FA-S Live Wire, which utilises Feature Asymmetry and a weak shape constraint, ASP Live Wire, which is a 3D expansion of Live Wire, and FA-O Live Wire, which uses Feature Asymmtery and Local Orientation to guide the segmentation process. These have been designed with each of the specific biometric landmarks in mind. Finally, a method of characterising fetal face shape is proposed, using a combination of the segmentation methods described here and a simple shape model with a parameterised b-spline meshing approach to facial surface representation.

Indiana University-Purdue University Indianapolis (IUPUI) / Mammographic density is an important risk factor for breast cancer, detecting and screening at an early stage could help save lives. To analyze breast density distribution, a good segmentation algorithm is needed. In this thesis, we compared two popularly used segmentation algorithms, EM-MPM and K-means Clustering. We applied them on twenty cases of synthetic phantom ultrasound tomography (UST), and nine cases of clinical mammogram and UST images. From the synthetic phantom segmentation comparison we found that EM-MPM performs better than K-means Clustering on segmentation accuracy, because the segmentation result fits the ground truth data very well (with superior Tanimoto Coefficient and Parenchyma Percentage). The EM-MPM is able to use a Bayesian prior assumption, which takes advantage of the 3D structure and finds a better localized segmentation. EM-MPM performs significantly better for the highly dense tissue scattered within low density tissue and for volumes with low contrast between high and low density tissues. For the clinical mammogram, image segmentation comparison shows again that EM-MPM outperforms K-means Clustering since it identifies the dense tissue more clearly and accurately than K-means. The superior EM-MPM results shown in this study presents a promising future application to the density proportion and potential cancer risk evaluation.

Indiana University-Purdue University Indianapolis (IUPUI) / We present a novel, variational and statistical approach for model-based segmentation. Our model generalizes the Chan-Vese model, proposed for concurrent segmentation of multiple objects embedded in the same image domain. We also propose a novel shape descriptor, namely the Multi-Compartment Distance Functions or mcdf. Our proposed framework for segmentation is two-fold: first, several training samples distributed across various classes are registered onto a common frame of reference; then, we use a variational method similar to Active Shape Models (or ASMs) to generate an average shape model and hence use the latter to partition new images. The key advantages of such a framework is: (i) landmark-free automated shape training; (ii) strict shape constrained model to fit test data. Our model can naturally deal with shapes of arbitrary dimension and topology(closed/open curves). We term our model Active Geometric Model, since it focuses on segmentation of geometric shapes. We demonstrate the power of the proposed framework in two important medical applications: one for morphology estimation of 3D Motor Neuron compartments, another for thickness estimation of Henle's Fiber Layer in the retina. We also compare the qualitative and quantitative performance of our method with that of several other state-of-the-art segmentation methods.

Indiana University-Purdue University Indianapolis (IUPUI) / Object segmentation, quantification, and shape modeling are important areas inmedical image processing. By combining these techniques, researchers can find valuableways to extract and represent details on user-desired structures, which can functionas the base for subsequent analyses such as feature classification, regression, and prediction. This thesis presents a new framework for building a three-dimensional (3D) hippocampal atlas model with subfield information mapped onto its surface, with which hippocampal surface registration can be done, and the comparison and analysis can be facilitated and easily visualized. This framework combines three powerful tools for automatic subcortical segmentation and 3D surface modeling. Freesurfer and Functional magnetic resonance imaging of the brain's Integrated Registration and Segmentation Tool (FIRST) are employed for hippocampal segmentation and quantification, while SPherical HARMonics (SPHARM) is employed for parametric surface modeling. This pipeline is shown to be effective in creating a hippocampal surface atlas using the Alzheimer's Disease Neuroimaging Initiative Grand Opportunity and phase 2 (ADNI GO/2) dataset. Intra-class Correlation Coefficients (ICCs) are calculated for evaluating the reliability of the extracted hippocampal subfields. The complex folding anatomy of the hippocampus offers many analytical challenges, especially when informative hippocampal subfields are usually ignored in detailed morphometric studies. Thus, current research results are inadequate to accurately characterize hippocampal morphometry and effectively identify hippocampal structural changes related to different conditions. To address this challenge, one contribution of this study is to model the hippocampal surface using a parametric spherical harmonic model, which is a Fourier descriptor for general a 3D surface. The second contribution of this study is to extend hippocampal studies by incorporating valuable hippocampal subfield information. Based on the subfield distributions, a surface atlas is created for both left and right hippocampi. The third contribution is achieved by calculating Fourier coefficients in the parametric space. Based on the coefficient values and user-desired degrees, a pair of averaged hippocampal surface atlas models can be reconstructed. These contributions lay a solid foundation to facilitate a more accurate, subfield-guided morphometric analysis of the hippocampus and have the potential to reveal subtle hippocampal structural damage associated.