ADSAM-Net: A Dual-Stage Attention Enhanced Mamba Network for 3D Medical Image Segmentation
Abstract
Accurate 3D medical image segmentation is crucial for clinical use. CNNs and transformers have limitations: limited receptive fields or quadratic complexity. State space models like Mamba offer linear complexity with global perception, but existing methods (e.g., EM-Net) struggle with small organs due to poor spatial dependency encoding. We propose ADSAM-Net, integrating a Dual-stage Attention Module (DSAM) into Mamba layers and decoder stages. DSAM enables cross-dimensional feature calibration and detail enhancement. Experiments on Synapse and BTCV datasets show that ADSAM-Net achieves average Dice scores of $\mathbf{7 8. 5 0 \%}$ and $\mathbf{7 8. 5 8 \%}$, significantly improving small-organ segmentation (e.g., pancreas, gallbladder) while maintaining high performance on large organs. Hausdorff distance is also substantially reduced. ADSAM-Net offers an efficient and accurate solution for 3D medical image segmentation.