CSAIL Event Calendar: Previous Series
Brain Tumor Segmentation
Speaker: David Gering , Medical Vision Group - MIT AI Laboratory
A method is presented for automatic segmentation of brain tumors. The approach applies to structures that are too irregular, in both shape and texture, to permit construction of comprehensive training sets. By training on the surroundings instead of the target, the algorithm is able to recognize deviations from normalcy, and subsequently hone in on boundary delineation. The technique is instantiated within the framework of adaptive segmentation using the EM algorithm and Markov random fields. As in earlier work, the classifier simultaneously corrects for signal inhomogeneities while labeling healthy tissue according to its class. Additionally, the method computes a fitness map over the image to represent the probability of pathology. Such fitness is computed based on intensity and very elementary properties of shape and position.