U-GRKAN: An Efficient and Interpretable Architecture for Medical Image Segmentation

Segmentation accuracy and consistency directly affect the safety of treatment and the reliability of decision-making in tumor delineation, organ-at-risk protection, preoperative planning, and follow-up evaluation. The U-shaped convolutional network still has limitations in cross-regional modeling and adaptive cross-layer fusion, while the Transformer/hybrid architecture, although capable of encoding global context, is computationally expensive and has limited interpretability. To address these issues, we propose a multi-group rational KAN (MGR-KAN) and embed it into the U-shaped framework to form U-GRKAN: replacing each edge B-spline with group-shared and inter-group diversified grouped rational functions, reducing the number of parameters by 48% compared to U-KAN, and achieving function-level interpretability through group-level response curves; further using channel attention for adaptive cross-layer fusion. We assessed four datasets (BUSI, GlaS, CVC, and COVID-19-CT-Seg) and achieved scores of 67.85/80.58, 88.25/93.75, 86.63/92.74, and 78.57/87.97 (IoU/F1), respectively, which represent improvements of 2.63/2.01, 0.75/0.42, 1.58/0.86, and 2.51/1.61 over the second-ranked model. Overall, U-GRKAN strikes a more balanced compromise between accuracy, complexity, and interpretability, and shows good generalization potential across various modalities.