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

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Healthcare
• Deep Learning Architectures

Background:

• Segmentation accuracy is critical for safe medical treatments and reliable clinical decisions.
• Existing U-shaped networks and Transformer architectures have limitations in modeling, fusion, and interpretability.

Purpose of the Study:

• To introduce a novel deep learning framework, U-GRKAN, that enhances segmentation accuracy, reduces complexity, and improves interpretability.
• To address limitations in cross-regional modeling and adaptive cross-layer fusion in current segmentation networks.

Main Methods:

• Developed a multi-group rational KAN (MGR-KAN) by replacing B-splines with group-shared rational functions.
• Integrated MGR-KAN into a U-shaped network (U-GRKAN), reducing parameters by 48% and enabling function-level interpretability.
• Employed channel attention for adaptive cross-layer fusion.

Main Results:

• Achieved high IoU/F1 scores across four diverse datasets (BUSI, GlaS, CVC, COVID-19-CT-Seg), outperforming the second-ranked model.
• Demonstrated significant improvements in segmentation performance, with gains up to 2.63/2.01 (IoU/F1).
• U-GRKAN showed superior balance between accuracy, complexity, and interpretability.

Conclusions:

• U-GRKAN offers a more balanced and interpretable solution for medical image segmentation.
• The proposed method exhibits strong generalization potential across various imaging modalities.
• This framework advances the reliability and safety of AI-driven clinical decision-making.