KANSeg: An efficient medical image segmentation model based on Kolmogorov-Arnold networks for multi-organ segmentation

Abstract

Currently, multi-organ segmentation methods based on convolution neural networks have achieved milestones in medical image analysis. However, there are some challenging issues such as the complex background, blurred boundaries between organs. These lead to poor boundary segmentation. To address this issue, we propose a multi-organ segmentation method based on Kolmogorov-Arnold Networks (KAN), called KANSeg. We develop a KAN-Activated Convolution module (KAN-ACM) to construct both the encoder and decoder, thereby enhancing the learning and interpretation of intricate patterns within multi-organ images. Moreover, to further augment the model's ability to represent nonlinear features, we design a KAN bottleneck module (KAN-BM) to extract more discriminative features. Finally, we conduct comprehensive experiments on two datasets. The proposed KANSeg can achieve Dice Score of 79.95%, 90.99% on the Synapse multi-organ dataset (Synapse) and the Automated cardiac diagnosis challenge (ACDC) datasets. The outcomes demonstrate that our method yields more accurate segmentation results compared with state-of-the-art methods.