- Research Article
2
- 10.1088/2631-8695/ae1103
SquareNet: multi-scale progressive difference and scale-cross attention network for volumetric medical image segmentation
- Oct 16, 2025
- Engineering Research Express
- Huaxiang Liu + 5 more +5
Accurate segmentation of medical images can assist doctors in computer-aided diagnosis and clinical treatment. Due to the complexity of the object region features (e.g., size, location, and shape), it is challenging to fully extract semantic features in medical image segmentation. To address these issues, we propose a 3D multiscale progressive difference and cross-scale attention network for medical image segmentation. Specifically, we propose a dual encoder-decoder network architecture comprising a multi-scale progressive difference (MSPD) branch and a group scale-cross attention (GSCA) branch. In the MSPD branch, we introduce a progressive difference module as the basic skip connection layer to enrich more discriminative and detailed features across multiple scales and resolve scale conflicts. In the GSCA branch, a group scale-cross attention module is designed to enhance the receptive field and build long-term dependencies between voxels. By combining the features from the MSPD and GSCA branches, the hierarchical group feature aggregation (HGFA) module is designed to fuse the multi-scale global information and local spatial information. We conducted qualitative and quantitative evaluations on three publicly available datasets, including LiTS2017, 3Dircadb, and WORD. Experimental results show that our model can achieve better segmentation performance than the state-of-the-art models on these three datasets.
Read more