Abdominal Tumor Segmentation Based on Lightweight-Attention-based and TransUNet

MRI is vital for abdominal tumor segmentation, but low contrast and organ overlap reduce accuracy. This paper proposes a model that combines a lightweight attention mechanism with Transformer architecture, using a scale attention mechanism for multi-scale fusion and depthwise separable convolutions to reduce parameters. Experiments on the Synapse dataset show the model outperforms UNet, DenseUNet, and TransUNet by 4.25%, 3.05%, and 2.73%, respectively, demonstrating its effectiveness in complex organ scenarios.