Download or read book A Deep-learning Convolutional Neural Network Framework for Mulitple Sclerosis Lesion Detection and Segmentation in Patient Brain Images written by Maor Zaltzhendler and published by . This book was released on 2016 with total page pages. Available in PDF, EPUB and Kindle. Book excerpt: "This thesis presents a convolutional neural network (CNN) based approach for detection and segmentation of Multiple Sclerosis lesions in brain magnetic resonance imaging (MRI). Automated pathology segmentation was presented in literature, starting from the early 1990s, and although reported to be a challenging task, could be highly beneficial for clinical trials labeling where large amounts of images are at hand. Robust detection of such pathology is still an open problem, and is prone to variabilities in: image non-uniformity, intensity distributions, acquisition artifacts, brain-structures, patients, scanners, configurations and sites. A CNN-based approach is proposed due to its recently reported high quality and generalization properties for computer vision tasks, providing a high degree of invariance and taking spatial correlation within the image structure into account. In order to address the task using both local and context-related information, a multi-scale approach is suggested, integrating the accuracy of several CNNs within a hierarchical framework for pathology segmentation. The presented model is general, and could be used for other pathology detection and segmentation contexts that require object delineation and classification in 3D magnetic resonance imaging. Several different architectures and experiments are presented throughout the document, while providing benchmarks and qualitative views over their results. Additional contributions of this thesis include: (a) learning CNN-based brain-features, evaluating their discriminative power, and observe appearance and constancy, (b) develop a general approach for MRI segmentation, while naturally incorporating the full 3D neighbourhood information rather than using 2D or augmented-2D with consecutive slices information. A comprehensive set of experiments is provided throughout this thesis, and performed over two different multi-site large scale proprietary clinical trials that were made available for this research. First, the method was configured and tested over the first clinical trial only. Once the hyper-parameters were set, no further tuning was allowed and the architecture was tested over the second clinical trial, which is much larger, and showed similar performance. The results of the method over this data yielded sensitivity values of up to 0.68, and Dice scores up to 0.59. The method achieved even higher metric scores of 0.86-1.00 true-positive rates when considering only larger lesions. The experiments performed show comparable performance to previously reported results from the literature over the same dataset. The data-driven features are presented, and shown to capture brain structures that lead to MS lesion discrimination both qualitatively and quantitatively." --