A GNN-based Network for Tissue Semantic Segmentation in Histopathology Image
PengHui He, Aiping Qu, Shuomin Xiao, Meidan Ding · Journal of Physics Conference Series · 2023
Abstract Segmentation of different tissue regions in pathological images hold an significant position diagnosis and prognosis of cancer. Although the convolutional neural network(CNN) and transformer which treat the image as a grid or sequence structure have been widely employed in this task, which is difficult to capture irregular and complex targets flexibly. So it is still a challenging task. At this paper, we employ a GNN-based segmentation method for pathological images which adopts the encoding-decoding structure. We first represent the input image as a graph structure which consists of a number of patches viewed as nodes and connect the nearest neighbors to build a graph. We also introduce ViG block to build a hierarchical pyramid architecture which consists of grapher module with graph convolution and FFN module with two linear layers. In addition, we utilize a pyramid CNN architecture decoder to aggregate graph information with multi-scales. The proposed method reaches 75.68% and 87.72% mean Dice on BCSS and LAUD-HistoSeg datasets respectively demonstrate the effectiveness.