CGCA-KAN: Correction-Guided Cluster-Aware Attention and KAN Enhanced Architecture for Medical Image Segmentation

Accurate medical image segmentation relies on collaborative modeling of local details and global semantics, especially in small-volume structures, blurred boundaries, and fine-grained anatomical regions under low signal-to-noise conditions. However, existing transformer-based methods typically suffer from feature redundancy problems caused by inaccurate attention mechanism focus and nonlinear modeling defects caused by insufficient expression ability of feedforward networks, which leads to attention bias and nonlinear modeling bias, and ultimately degrades segmentation performance. To address these challenges, we propose CGCA-KAN, a novel transformer-based framework that employs a collaborative correction mechanism (CCM) to jointly mitigate attention bias and nonlinear modeling bias. Specifically, we introduce a cluster-aware self-attention module (CASAM) to mitigate attention bias by refining semantic focus and suppressing redundancy through token group compression, thereby enhancing attention to small-volume structures. Additionally, we design a Kolmogorov-Arnold Network enhanced feedforward network (KAN-EFFN) to mitigate nonlinear modeling bias through adaptive nonlinear transformations, thereby improving the model's ability to delineate blurred boundaries. Extensive experiments on LiTS2017, Synapse, and BraTS2020 datasets demonstrate state-of-the-art performance in fine-grained segmentation of ambiguous lesions in the liver, complex structures of multiple organs, and brain tumors. Our results highlight CGCA-KAN as a promising solution for addressing modeling biases in medical image segmentation.