SquareNet: multi-scale progressive difference and scale-cross attention network for volumetric medical image segmentation

Abstract Accurate medical image segmentation is critical for computer-aided diagnosis and clinical treatment but is challenged by complex object region features, such as diverse sizes, locations, and shapes. We propose SquareNet, a 3D multiscale progressive difference and cross-scale attention network for robust volumetric medical image segmentation. SquareNet employs a dual encoder-decoder architecture with a multiscale progressive difference (MSPD) branch to extract discriminative features and resolve scale conflicts, and a group scale-cross attention (GSCA) branch to expand the receptive field and capture long-term voxel dependencies. A hierarchical group feature aggregation (HGFA) module fuses global and local features from both branches. Evaluations on the LiTS2017, 3Dircadb, and WORD datasets demonstrate that SquareNet achieves superior segmentation accuracy compared to state-of-the-art methods, as validated by qualitative and quantitative results.