Reproduce “ASM-UNet: Adaptive Scan Mamba Integrating Group Commonalities and Individual Variations for Fine-Grained Segmentation” on OpenResearch

About this paper

Bo Wang, Mengyuan Xu, Yue Yan, Yuqun Yang, Kechen Shu, Wei Ping, Xu Tang, Wei Jiang and 1 more

ASM-UNet augments a 3D U-Net with Mamba modules whose scan order adapts to both shared anatomical patterns and patient-specific structure. It improves Dice performance over listed baselines on ACDC and Synapse, and reaches 42.57% fine-grained Dice on the eight-class BTMS biliary tract dataset, although very small ducts remain difficult to distinguish.

Key insights

• Population-level anatomical regularities and patient-specific structural information provide complementary guidance for selecting Mamba scan orders.
• Ablations show progressive gains from adding Mamba, individual scoring, group scoring, and finally their combination, with the full model increasing Dice from 40.23% for the U-Net baseline to 42.57% on the most difficult BTMS metric.
• Coarse identification of the biliary system is considerably easier than assigning correct labels to individual ducts, exposing the limitations of current fine-grained segmentation methods.

Summary by alphaXiv.