Automated Medical Image Processing Using Efficient Shape Descriptors: Principles and Applications

This thesis focuses on the development of fast and automated methods for detecting and segmenting anatomical structures in various modalities of medical imaging using efficient shape descriptors. The identification and delineation of these targeted objects are fundamental steps for further computer assisted applications, such as surgery planning, surgery interventional guidance, computer-aided detection and diagnosis or information fusion. First of all, a brief introduction of fundamental topics in medical image processing will be given, including segmentation, registration, detection and classification. Common techniques and methodologies used in these domains are generally reviewed. Then, a review of shape descriptors that are commonly used in medical image processing is conducted. The review will categorize these techniques with respect to different dimensions, such as complexity, efficiency, degree of user interactions and sensitivity to parameters, etc. Afterwards, to demonstrate how these shape descriptors are applied in each individual clinical task, several segmentation, registration, classification and detection tasks, where a variety of shape descriptors serve as the mainstay, will be described in detail. Specifically, the tasks consist of the segmentation of femur heads in fluoroscopic images using a Gabor-based Hough shape descriptor, the segmentation and registration of breasts in magnetic resonance images, the detection of nipples in 3D breast ultrasound images and the segmentation of liver vessels in multi-phase computer tomography images using a variety of Hessian-based shape descriptors. Meanwhile, a computer-assisted diagnostic tool dedicated to breast lesion classification is proposed, on the basis of a series of sphere packing shape descriptors. For each task, clinical background will be first explained, and the state-of-the-art techniques that attempted to resolve the problem will be reviewed.