U-KAN Makes Strong Backbone for Medical Image Segmentation and Generation

• DOI: 10.1609/aaai.v39i5.32491
• Publication Date: 2025-04-11T11:10:39Z
• Authors: Chenxin Li; Xinyu Liu; Wuyang Li; Cheng Wang; Hengyu Liu; Yifan Liu; Zhen Chen; Yixuan Yuan

U-Net has become a cornerstone in various visual applications such as image segmentation and diffusion probability models. While numerous innovative designs improvements have been introduced by incorporating transformers or MLPs, the networks are still limited to linearly modeling patterns well deficient interpretability. To address these challenges, our intuition is inspired impressive results of Kolmogorov-Arnold Networks (KANs) terms accuracy interpretability, which reshape neural network learning via stack non-linear learnable activation functions derived from Kolmogorov-Anold representation theorem. Specifically, this paper, we explore untapped potential KANs improving backbones for vision tasks. We investigate, modify re-design established pipeline integrating dedicated KAN layers on tokenized intermediate representation, termed U-KAN. Rigorous medical benchmarks verify superiority UKAN higher even with less computation cost. further delved into U-KAN an alternative noise predictor models, demonstrating its applicability generating task-oriented model architectures.