Understanding Graph Convolutional Networks to detect Brain Lesions from Stroke

Abstract

Brain lesions derived from stroke episodes can result in disabilities for a patient. Therefore, the segmentation of brain lesions is an important task in neurology. Recently this task has been mainly tackled by machine learning approaches that demonstrated to be very successful. One of these approaches is Graph Convolutional Networks (GCN), where the input image is interpreted as a graph structure. As usual for neural networks, the interpretability is hard due to their black-box nature. We provide an interactive visualization of the activation inherent in the GCN, which is map from the original dataset. We visualize the activation values of the underlying graph network on top of the input image. We show the usability of our approach by applying it to a GCN that was trained on a real-world dataset.