Medical image segmentation and diffusion weighted magnetic resonance image analysis.

摘要

Medical image segmentation plays an important role in diagnosis, surgical planning, navigation, and various medical evaluations. Medical images are frequently corrupted by high levels of noise, signal dropout and poor contrast along boundaries. Sometimes, their intensity might have multi-modal distribution. In this dissertation, I will present one method to segment images that are corrupted by noise and some dropout. In the model presented, prior points together with prior shape information are incorporated into a joint segmentation and registration model in both a variational framework and in level set formulation. This technique is applied to segment cardiac ultrasound images. A second model, which is based on applying non-parametric density approximation to simultaneously segment and smooth noisy medical images without adding extra smoothing terms, is presented. My goal is to develop a powerful and robust algorithm to locate objects with interiors having a complex multi-modal intensity distribution and/or high noise level. The model was applied to the problem of segmenting T1 weighted magnetic resonance images.; Diffusion weighted images render non-invasive in vivo information about how water diffuses into a 3D intricate representation of tissues. My work provides histological and anatomical information about tissue structure, composition, architecture, and organization. I have proposed several models to reconstruct human brain white matter fiber tracts, to recover intra-voxel structure, to classify intra-voxel diffusion, to estimate, smooth and characterize apparent diffusion coefficient profiles. A geometric flow is designed to segment the main core of white matter fiber tracts in diffusion tensor images.