Robust brain tumor segmentation for overall survival prediction

Abstract

A Glioma brain tumor is the deadliest brain tumor with high mortality. Treatment planning by human experts depends on a proper diagnosis based on physical symp toms along with a medical image analysis. Highly variability of a brain tumor in terms of size, shape, location, and a high volume of Magnetic Resonance Images makes the analysis time-consuming. Such manual diagnosis is non-reproducible and depends highly on human expertise. Automatic/semi-automatic segmentation methods achieve a reduction in time with excellent reproducible results. Further, proper analysis follows the prediction of the overall survival of patients. The pre diction helps in planning post-operative treatments which improve the prognosis. This dissertation develops fully convolutional neural networks for tumor segmen tation followed by a random forest regressor to predict the overall survival of the patients. The first network to address the segmentation problem is the 2D encoder decoder network that segments the substructures of a tumor separately. A net work parameter initialization technique addresses the class imbalance problem and scarcity of the label data. The network is extended to incorporate dense connec tions between successive layers of the encoder part, which helps the network learn diversified features from the input. In addition, it helps the error gradient to reach the initial layers in case the deep layers are not able to learn from the input. The network is further extended to incorporate a three-dimensional voxel relationship for consistent performance on unseen data. Tumor segmentation follows classification of patients based on survival days. The classical approaches- random forest classifier and random forest regressor ad dresses the problem to calculate survival days based on an optimal set of input features, which are extracted from the segmentation network.