Dynamic graph convolutional autoencoder with node-attribute-wise attention for kidney and tumor segmentation from CT volumes

Extraction and integration of semantic connections, spatial relations and dependencies are critical in volumetric image segmentation. This is a challenging issue, especially when there are long-distance objects with close semantic relations and neighboring objects with indistinct boundaries. We propose a novel dynamic graph convolution (DGC) autoencoder with node-attribute-wise attention (NodeAttri-Attention) for relation inference and reasoning, with applications on kidney and tumor segmentation from computerized tomography (CT) volumes. We first introduce a new graph construction strategy for 3D volumetric image data, where graph node attributes and connections represent topological relations and high-level correlations. Then NodeAttri-Attention mechanism is proposed to obtain attention-enhanced node attributes by discriminating adaptive contributions of various features. Finally, the DGC strategy is designed to learn and integrate the complex and underlying correlations across image regions. Our DGC dynamically updates graph topology and node attributes as the graph convolutional layer gradually deepens. Experimental results and ablation studies demonstrated the effectiveness of each of our major innovations in NodeAttri-Attention DGC, especially when objects are with weak boundaries, irregular shapes, and various sizes. The improved segmentation results of embedding NodeAttri-Attention DGC to different segmentation backbones show the generality of DGC autoencoder.