Statistical Classification and level Set Methods in Medical Image Analysis Quantitative Evaluation of Articular Cartilage in Knee MRI

This thesis presents research and development of methods for automatic segmentation and quantitative evaluation of anatomical structures from medical images. An automatic segmentation algorithm is the foundation on which this framework stands, therefore a substantial part of the thesis concerns this subject. The segmentation methods focus on statistical voxel classification, using feature selection to find representative features for the classes, and incorporation of level set methods into the classification is also investigated. The segmentation methods are evaluated on magnetic resonance imaging (MRI) data of the knee and heart. The thesis also contains a quantitative evaluation of articular cartilage in relation to osteoarthritis (OA), with the aim of developing and evaluating imagingbased OA biomarkers suited for clinical trials. Factors such as joint incongruity, cartilage lesions, and fibrillation can all be linked to OA, and they can in turn be related to cartilage surface curvature at different scales. Cartilage curvature is analyzed using mean curvature flow and by a comparison of surface normals. Besides curvature, other biomarkers founded on the automatic segmentation algorithm such as volume, thickness, and homogeneity are also evaluated as potential OA biomarkers. The automatic segmentation methods described in this work perform well compared to gold standard segmentations, and compare favorably to state-of-the-art segmentation methods. The OA biomarkers obtained founded on the automatic cartilage segmentation show potential as disease markers as they are clearly related to radiographic signs of OA.