Research map / Image generation / Autoregressive and latent generation
U-KAN Makes Strong Backbone for Medical Image Segmentation and Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces U-KAN, a U-Net variant with tokenized Kolmogorov-Arnold Network (KAN) blocks near the bottleneck for medical image segmentation, and extends it into a diffusion noise predictor (Diffusion U-KAN).
Problem
U-Net backbones for medical segmentation and diffusion models are limited to linear modeling of channel patterns and lack interpretability; the paper asks whether KANs can improve these vision backbones.
Approach
Insert tokenized KAN blocks (patch tokens projected linearly, passed through stacked KAN layers with depth-wise conv and residuals) at the high-level, near-bottleneck stages of a convolutional U-Net, and reuse the same backbone as a DDPM noise predictor with time embeddings injected into KAN blocks.
Key results
- Seg. U-KAN reaches 63.38 IoU / 76.40 F1 on BUSI, outperforming U-Mamba (61.81 / 75.55), averaged over three runs.
- Average IoU across the three datasets is 78.69 vs 77.87 for U-Mamba, with 14.02 Gflops and 6.35M params versus U-Mamba's 2087 Gflops and 86.3M params.
- Diffusion U-KAN achieves FID 101.93 on BUSI, 41.55 on GlaS and 46.34 on CVC, beating the U-Net diffusion variants in the reported table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace U-Mamba with Seg. U-KAN | medical image segmentation models · BUSI | IoU: 61.81±3.24 -> 63.38±2.83 | Helped |
| Replace U-Mamba with Seg. U-KAN | medical image segmentation models · GlaS | IoU: 87.01±0.39 -> 87.64±0.32 | Helped |
| Replace U-Mamba with Seg. U-KAN | medical image segmentation models · CVC | IoU: 84.79±0.58 -> 85.05±0.53 | Helped |
| Replace U-Net diffusion middle block with KANBlock | Diffusion U-Net · BUSI | FID: 104.95 -> 101.93 | Helped |
| Replace U-Net diffusion middle block with KANBlock | Diffusion U-Net · GlaS | FID: 42.63 -> 41.55 | Helped |
| Replace U-Net diffusion middle block with KANBlock | Diffusion U-Net · CVC | FID: 49.30 -> 46.34 | Helped |
| Use 3 KAN layers instead of 1 (ablation) | U-KAN · Segmentation ablation | IoU: 64.20 -> 65.26 | Helped |
| Use 3 KAN layers instead of 2 (ablation) | U-KAN · Segmentation ablation | IoU: 64.56 -> 65.26 | Helped |
| Use 3 KAN layers instead of 4 (ablation) | U-KAN · Segmentation ablation | IoU: 64.72 -> 65.26 | Helped |
| Use 3 KAN layers instead of 5 (ablation) | U-KAN · Segmentation ablation | IoU: 64.86 -> 65.26 | Helped |
| Use KAN layers rather than MLP layers (ablation) | U-KAN · Segmentation ablation | IoU: 63.49 -> 65.26 | Helped |
| Keep KAN instead of replacing the first layer with MLP (ablation) | U-KAN · Segmentation ablation | IoU: 64.12 -> 65.26 | Helped |
Surprising findings
- U-NeXt has far fewer parameters and FLOPs than U-KAN but notably lower segmentation accuracy, so U-KAN is not the most efficient model overall.
Limitations
- Evaluated only on three small 2D medical datasets and a 64x64 unconditional generation setting.
- Interpretability claims rest on a qualitative activation-map analysis with a single plausibility IoU metric.
- Diffusion evaluation uses FID and IS only on small medical image sets with no comparison to modern diffusion architectures.
Experimental setup
Datasets: BUSI GlaS CVC-ClinicDB
Benchmarks: BUSI GlaS CVC-ClinicDB
Metrics: IoU F1 Gflops Params FID IS
Baselines: U-Net U-Net++ Att-UNet U-NeXt Rolling-UNet U-Mamba
Builds on: U-Net Kolmogorov-Arnold Networks DDPM TransUNet
Closest papers by meaning
- U-Nets as Belief Propagation: Efficient Classification, Denoising, and Diffusion in Generative Hierarchical Models (ICLR 2025 · similarity 0.89)
- Rethinking U-Net: Task-Adaptive Mixture of Skip Connections for Enhanced Medical Image Segmentation (AAAI 2025 · similarity 0.87)
- KAN: Kolmogorov–Arnold Networks (ICLR 2025 · similarity 0.87)
- Upping the Game: How 2D U-Net Skip Connections Flip 3D Segmentation (NeurIPS 2024 · similarity 0.86)
- On the expressiveness and spectral bias of KANs (ICLR 2025 · similarity 0.86)
- Factorized Diffusion Architectures for Unsupervised Image Generation and Segmentation (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.