MedSegViG: Medical Image Segmentation with a Vision Graph Neural Network

Medical image segmentation is a crucial step toward automatic clinical diagnosis, which has received growing interest. Although some existing methods based on convolutional neural networks or transformers have achieved remarkable success in this task, they still show limitations in effectively modeling the relationships among different objects in images. In this paper, we propose a novel deep learning based model to address this issue by leveraging a vision graph neural network (ViG). Our model, MedSegViG, mainly consists of a hierarchical ViG encoder and a lightweight convolutional decoder. The hierarchical encoder extracts multi-level features from the image and captures the object relationships with graph neural networks. The lightweight decoder then fuses these features and generates the corresponding segmentation map. Extensive experiments are conducted on seven datasets for three typical medical image segmentation tasks: polyp segmentation, skin lesion segmentation, and retinal vessel segmentation. The results demonstrate the superiority of our MedSegViG over state-of-the-art models across various tasks and datasets. The code is released on https://github.com/Xinhong-Li/MedSegViG.