GST: A Brain-Inspired Graph Signal Transmitter for Biomedical Image Segmentation

Automatic segmentation of biomedical image is of great significance for computer-aided diagnosis. Different from the existing segmentation models based on convolutional neural network or attention mechanism with pixelwise prediction mode, this paper proposes a new Graph Signal Transmitter (GST) which regresses patch probability directly from the evolved semantic node and edge features. The images pass through an encoder to derive the semantic features, each voxel of which is taken as a node, and the corresponding feature vector along the channel dimension is taken as the node feature of the initial graph layer, and the relative positions between different nodes are considered as the edge features. With the feedforward process, the node and edge features of the following layers are evolved by fusing the features of source nodes, destination nodes and edges. The node features of the last graph layers are directly mapped back to the patch in semantic space with fold module and generate segmentation map. In the semantic and instance segmentation tasks for gland dataset, comparative experiments show that the proposed model outperforms the state-of-the-art (SOTA) methods, and ablation experiments demonstrate the effectiveness of GST.