UKAN+: Enhancing medical image segmentation via fusion attention with Kolmogorov-Arnold Networks

With the emergence of Kolmogorov–Arnold Networks (KANs), which are known for their strong nonlinear representation capabilities, integrating KANs with U-Net, referred to as UKAN, opens new avenues within the U-Net family. However, the relatively simple architectural design of UKAN constrains the effective utilization of KANs’ nonlinear representational power, particularly in medical imaging scenarios. Therefore, this work investigates the potential of KAN in medical image segmentation. We propose a novel architecture that integrates KAN to enhance the attention module and encoder–decoder architecture for vision tasks. Specifically, we design a fusion attention module, KS-CS, comprising two parallel branches: a KAN-enhanced Selective Kernel (KS) unit that replaces fixed-activation MLPs with learnable B-spline activations for precise dynamic weight generation, and a Convolution-Enhanced Spatial Attention (CS) unit. This enhancement yielded a new architecture, termed UKAN+. A comprehensive evaluation of UKAN+ was conducted across eight distinct medical imaging modalities: Microscopy, Colonoscopy, CT, Interferometry Imaging, X-ray, OCT, Ultrasound, and MRI. Experimental results demonstrate the consistent superiority of UKAN+ over existing models, achieving a 1.97% improvement in average Dice and a 6.32% improvement in average IoU compared to the contemporary UKAN. These gains are robust across diverse anatomical structures and imaging protocols, while maintaining a clinical real-time segmentation speed of 30 FPS. These efforts provide significant insights and highlight the potential of combining KAN with attention mechanisms and U-Net to establish a robust framework for medical image segmentation. The source code will be released upon publication.