CKTNet : A Cross‐Attention KAN ‐Transformer Based Few‐Shot Learning for Pericardial Adipose Tissue Segmentation

ABSTRACT Accurate quantification of pericardial adipose tissue (PAT) through magnetic resonance (MR) images segmentation is crucial for the diagnosis of early cardiovascular diseases. The complex anatomical structures and locations of PAT exacerbate the blurring phenomenon at the boundaries of MR images. Although Unet has become the cornerstone of PAT segmentation, it still faces challenges such as weak long‐range dependency capture and insufficient nonlinear modeling. These limitations are even more prominent in the scenario of few‐shot learning where labeled MR dataset is scarce. Therefore, we propose a novel few‐shot learning‐based Cross‐attention KAN‐Transformer Network, named CKTNet. Specifically, we propose an adaptive multi‐scale feature fusion module for dynamically obtaining valuable low‐level and high‐level features, fusing multi‐scale contextual information to alleviate the challenges posed by complex anatomical structures in PAT. Secondly, we combine KAN with Depthwise Convolution to construct a KAN encoder and embed it into a Transformer architecture. The KAN encoder can effectively capture the nonlinear relationship of PAT structure complexity. Finally, we propose a Cross‐attention Transformer block, which addresses two major shortcomings: the difficulty of prototype based few‐shot learning models in capturing subtle differences between instances, and the Transformer's lack of attention to local context. We compared our model with the state‐of‐the‐art models on the MRPEAT dataset in the MR modality. Our model achieved an score of 0.799, precision of 0.762, recall of 0.842, and Hausdorff distance of 14.016. The experimental results show that our model significantly improves PAT MR images segmentation accuracy and interpretability, which is crucial for clinical applications.