Affinity Graph Based End-to-End Deep Convolutional Networks for CT Hemorrhage Segmentation

Brain hemorrhage segmentation in Computed Tomography (CT) scan images is challenging, due to low image contrast and large variations of hemorrhages in appearance. Unlike the previous approaches estimating the binary masks of hemorrhages directly, we newly introduce affinity graph, which is a graph representation of adjacent pixel connectivity to a U-Net segmentation network. The affinity graph can encode various regional features of the hemorrhages and backgrounds. Our segmentation network is trained in an end-to-end manner to learn the affinity graph as intermediate features and predict the hemorrhage boundaries from the graph. By learning the pixel connectivity using the affinity graph, we achieve better performance on the hemorrhage segmentation, compared to the conventional U-Net which just learns segmentation masks as targets directly. Experiments in this paper demonstrate that our model can provide higher Dice score and lower Hausdorff distance than the conventional U-Net training only segmentation map, and the model can also improve segmentation at hemorrhagic regions with blurry boundaries.