PDSA-DBUNet: dual-branch U-Net for abdominal medical image segmentation driven by polynomial dynamic sparse attention

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Abdominal medical image segmentation is essential for computer-aided diagnosis, lesion localization, quantitative organ analysis, and treatment planning. However, abdominal computed tomography (CT) images often contain large anatomical variations, complex tissue structures, low contrast, blurred boundaries, and small tumor regions, which make accurate and efficient segmentation challenging. Convolutional neural networks (CNNs) are effective in extracting local textures, edges, and spatial details, but their limited receptive fields restrict long-range dependency modeling. Transformers can capture global contextual relationships, yet dense token modeling introduces redundant background information and high computational cost, while fine-grained local details may be insufficiently preserved. To address these limitations, we propose a Polynomial Dynamic Sparse Attention-driven Dual-Branch U-Net, termed PDSA-DBUNet, for abdominal medical image segmentation. The network adopts a parallel CNN–Transformer architecture, where the CNN branch preserves local boundary and texture information, and the Transformer branch models global semantic dependencies. A dynamic sparse visual token generation module adaptively aggregates dense image tokens into a compact set of discriminative sparse tokens, emphasizing organ regions and blurred boundaries while suppressing irrelevant background responses. A PoM-based polynomial mixing mechanism is further incorporated into the dynamically generated sparse token set to replace conventional multi-head self-attention, enabling efficient global contextual interaction with reduced computational and memory costs. In addition, a Dual-Scale Parallel Attention Fusion (DSPAF) module integrates CNN-derived local details with Transformer-derived global semantics to enhance boundary localization and small-target representation. Experiments on the LiTS and KiTS19 datasets show that PDSA-DBUNet achieves Dice scores of 96.79% and 86.13% on LiTS-Liver and LiTS-Tumors, and 94.03% and 89.06% on Kidney and Kidney-Tumor segmentation tasks, respectively. The model requires 72.87 GFLOPs, contains 41.62M learnable parameters, and has a checkpoint storage size of 129.93 MB. These results demonstrate that PDSA-DBUNet achieves a favorable balance between segmentation accuracy and computational efficiency.