ALIEN: Attention-guided cross-resolution collaborative network for 3D gastric cancer segmentation in CT images

Automatic segmentation of gastric cancer (GC) in CT imaging holds significant clinical diagnostic and research implications. Despite the advancement of deep learning in medical image segmentation, there is a scarcity of segmentation algorithms specifically designed for GC due to its low contrast, blurry boundaries, and substantial inter-individual variations. In this paper, we propose a 3D Attention-guided cross-resoLutIon collaborativE Network (ALIEN) that integrates a multi-attention fusion module, a cross-resolution fusion module, and a scale-aware activation module to address the above challenges. The three modules collaboratively exploit, integrate, and refine features, enhancing the semantic awareness ability of the model to achieve more accurate GC segmentation results. We conduct a comprehensive ablation study to demonstrate the complementarity among these modules. Experimental results reveal that the proposed method achieved superior results to current mainstream 3D segmentation networks, effectively mitigating the issues of over-segmentation and semantic edge blurring in GC. The code is publicly available at https://github.com/ZHChen-294/ALIEN.