Computer-assisted segmentation and tracking of brain lesions in magnetic resonance images based on probabilistic reasoning in space and time

There is an urgent need to develop computer-based automated techniques to detect abnormal tissue (i.e. lesions) in medical images and track progression in size, shape and intensity. The best accepted measure of brain tumor viability is interval change in tumor size and decisions on efficacy of clinical treatments in a given patient and in clinical trials are most commonly based on this measure.; Medical imaging modalities include magnetic resonance imaging (MRI), computed tomography (CT), and positron emission tomography (PET). MRI provides the best structural imaging of soft tissue, such as the brain, and is used to provide the data for analysis in this dissertation. Even with the best medical imaging techniques, medical image segmentation is a difficult problem as the pixel intensities of various tissues overlap and borders between tissues are not always sharp. Automating techniques to detect and track progression of brain tumors in MR images has been an ongoing research goal, but routine clinical applications for fully automated segmentation do not exist. Since there is some relation between images of a patient acquired at different times, temporal information has the potential to improve segmentation of lesions in medical images. It is hypothesized that using both temporal and spatial properties of the 4D image set will improve the automatic segmentation of lesions (e.g. tumors) compared with techniques that independently detect lesions from one scan to the next or focus only on areas of change in the dynamic series. A proposed method for lesion segmentation that uses probabilistic reasoning over space and time is the basis of this dissertation.; In this dissertation, the hidden Markov model (HMM) is explored for the first time in the context of medical image segmentation and novel transition matrices are developed. By incorporating both spatial and temporal information, we show an improvement in the accuracy of segmentation over previous methods that use spatial or temporal information alone. The framework of the 4D segmentation and tracking method developed in this dissertation is general enough to be applied to other applications not related to medical imaging such as tracking of objects in video sequences.