Multi-scale random walk driven adaptive graph neural network with dual-head neighboring node attention for CT segmentation

Segmenting objects with indistinct boundaries and large variations from CT volumes is a challenging issue due to overlapping intensity distributions from neighboring tissues or long-distance semantic relations. We propose a multi-scale random walk (RW) driven graph neural network (GNN) to address this issue. A graph is first initialized to represent image regions and deep semantic features from the segmentation encoder by graph nodes and attributes. We then propose a multi-scale graph reasoning model where for each scale, graph node attribute embedding is obtained by an adaptive GNN with dual-head neighboring node attention, while graph topology is evolved by RW. The neighboring-node attention mechanism is designed to learn and incorporate the importance and influence of neighboring nodes on their connected nodes. Random walking to multi-order neighbors enhance the contextual information formulation and diffusion along graph edges. Finally, multi-scale knowledge learnt from graphs is adaptively fused by a new graph-wise attention fusion module before reshaping and feeding to the segmentation decoder. We evaluate the contributions of major innovations by ablation studies, comparison with other state-of-the-art models on public kidney and tumor segmentation dataset. The generalization ability of our model is validated by different segmentation backbones. Experimental results show that the novel multi-scale adaptive graph reasoning architecture and RW-enhanced GNN model improved the segmentation of objects from adjacent tissues.