KACNet: Enhancing CNN feature representation with Kolmogorov-Arnold networks for medical image segmentation and classification

Abstract

Convolutional neural networks (CNNs) often face challenges in generalizing across datasets due to edge-blurring limitations of fixed activation functions. To address this, this paper introduces a novel approach by integrating KANs into CNNs, forming a module called KANConv, to build an innovative encoder architecture. The encoder is designed to augment the CNN’s capability to effectively distinguish targets in medical image segmentation and classification tasks. Specifically, the CNN component leverages its nonlinear structure to extract hidden local features from the input image, while the KAN component captures global feature information by processing the local features via a sophisticated structure composed of B-spline functions. The proposed multi-stage architecture enables a comprehensive information fusion of local and global features, significantly improving feature representation quality. To validate our encoder’s effectiveness, we perform extensive experiments on: (1) five benchmark 2D medical image segmentation datasets, where our model achieves state-of-the-art performance; and (2) six prominent 3D medical classification datasets, consistently demonstrating superior results compared to existing methods.