Exploring Graph-Based Neural Networks for Automatic Brain Tumor Segmentation

Manual evaluation of medical images, such as MRI scans of brain tumors, requires years of training, is time-consuming, and is often subject to inter-annotator variation. The automatic segmentation of medical images is a long-standing challenge that seeks to alleviate these issues, with great potential benefits for physicians and patients. In the past few years, variations of Convolutional Neural Networks (CNNs) have established themselves as the state-of-the-art methodology for this task. Recently, Graph-based Neural Networks (GNNs) have gained considerable attention in the deep learning community. GNNs exploit the structural information present in graphical data by aggregating information over connected nodes, allowing them to effectively capture relation information between data elements. In this project, we propose a GNN-based approach to brain tumor segmentation. We represent 3D MRI scans of the brain as a graph, where different regions in the images are represented by nodes and edges connect adjacent regions. We apply several variations of GNNs for the automatic segmentation of brain tumors from MRI scans. Our results show GNNs give reasonable performance on the task and allow for realistic modeling of the data. Furthermore, they are far less computationally expensive and time-consuming to train than state-of-the-art segmentation models. Lastly, we assign Shapley value-based contribution scores to input MRI features to learn what features are relevant for a particular segmentation, generating interesting insights into explaining the predictions of the proposed model.