Machine Learning vs Deep Learning for Fully Automatic Image Segmentation in Neurosurgery

Segmentation of brain tumor is a delicate procedure which requires a very high degree of accuracy. This thesis aims at studying the intricacies of automatic segmentation of brain Magnetic Resonance Images with the ultimate goal of building a fully automatic segmentation software whose accuracy and reliability will be acceptable in clinical practice. Artificial intelligence has shown successes in general image analysis. In order to facilitate the studies of how Artificial Intelligence can be applied in automatic segmentation of brain tumor MR images, two automatic segmentation software, Bratumia and DeepNeuro's SegmentGBM are employed. BraTumia adopts a Machine Learning approach (with conditional random field for spatial regularization) while DeepNeuro's SegmentGBM adopts Deep Learning. During the thesis, experiments have been conducted to investigate how each component of the pre-processing and segmentation pipeline affects the accuracy of the overall segmentation.