Research map / Architectures and efficiency / CNN design
Rolling-Unet: Revitalizing MLP’s Ability to Efficiently Extract Long-Distance Dependencies for Medical Image Segmentation
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.
Proposes Rolling-Unet, a U-Net-style CNN-MLP hybrid whose Rolling-MLP (R-MLP) modules capture long-range dependencies along multiple directions for medical image segmentation at matched CNN cost.
Problem
CNNs struggle to capture long-distance dependencies, while Transformers are computationally expensive and weak at local features; existing MLP vision models have only local receptive fields in medical segmentation.
Approach
A Rolling operation (cyclic shift plus cropping per channel, then weight-shared channel projection) gives each channel a shifted spatial view, yielding linear global receptive fields in one direction; stacking orthogonal and complementary R-MLPs gives four-direction dependencies, which are fused with depthwise separable convolution in a Lo2 block costing about one 3x3 convolution.
Key results
- On BUSI, Rolling-Unet (L) reaches 67.81 IoU and 80.17 F1 versus 67.27 IoU for UCTransNet, with Rolling-Unet (S) at 1.78M parameters.
- On ISIC 2018 at 256x256, Rolling-Unet (S) reaches 84.15 IoU and 91.13 F1, the best among compared methods.
- Ablation on ISIC 2018 shows R-MLP alone scores 84.14 IoU versus 81.10 IoU for a regular MLP replacement.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Rolling-Unet(L) instead of DconnNet | Medical image segmentation models · BUSI | IoU: 67.16±0.61 -> 67.81±1.80 | Helped |
| Use Rolling-Unet(L) instead of UNet++ | Medical image segmentation models · GlaS | IoU: 87.99±0.52 -> 88.02±0.28 | No clear effect |
| Use Rolling-Unet(S) instead of UCTransNet | Medical image segmentation models · ISIC 2018 (256) | IoU: 83.96 -> 84.15 | Helped |
| Use Rolling-Unet(S) instead of DconnNet | Medical image segmentation models · ISIC 2018 (512) | IoU: 83.60 -> 84.14 | Helped |
| Use Rolling-Unet(L) instead of DconnNet | Medical image segmentation models · CHASEDB1 | IoU: 69.90±0.44 -> 70.40±0.43 | Helped |
| Add DSC to neither DOR-MLP nor another spatial module (ablation) | Rolling-Unet · ISIC 2018 (512) | IoU: 79.48 -> 80.62 | Helped |
| Add R-MLP without DSC (ablation) | Rolling-Unet · ISIC 2018 (512) | IoU: 79.48 -> 81.62 | Helped |
| Add both DSC and R-MLP (ablation) | Rolling-Unet · ISIC 2018 (512) | IoU: 79.48 -> 82.11 | Helped |
| Add OR-MLP without DSC (ablation) | Rolling-Unet · ISIC 2018 (512) | IoU: 79.48 -> 83.39 | Helped |
| Add both DSC and OR-MLP (ablation) | Rolling-Unet · ISIC 2018 (512) | IoU: 79.48 -> 83.84 | Helped |
| Add DOR-MLP without DSC (ablation) | Rolling-Unet · ISIC 2018 (512) | IoU: 79.48 -> 83.46 | Helped |
| Add DSC to DOR-MLP in parallel (ablation) | Rolling-Unet · ISIC 2018 (512) | IoU: 83.46 -> 84.14 | Helped |
Surprising findings
- Smaller Rolling-Unet (S) sometimes outperforms the larger (L) variant on ISIC 2018, which the authors attribute to overfitting or training fluctuation.
- Rolling-Unet (L) does not beat all baselines uniformly; on CHASEDB1 DconnNet and Att-UNet tie closely with it, and on BUSI S is slightly below UCTransNet on IoU.
Limitations
- Remote dependencies are only along four directions (horizontal, vertical, two diagonals), not truly global receptive fields.
- No pretrained weights used and results on small datasets have sizable standard deviations; performance gains on GlaS and CHASEDB1 are not significant.
- Evaluated only on 2D segmentation; 3D extension left to future work.
Experimental setup
Datasets: ISIC 2018 BUSI GlaS CHASEDB1
Benchmarks: ISIC 2018 BUSI GlaS CHASEDB1
Metrics: IoU F1 HD95
Baselines: U-Net UNet++ Att-UNet MedT UCTransNet UNeXt DconnNet
Builds on: U-Net MLP-Mixer Sparse MLP AS-MLP UNeXt
Closest papers by meaning
- E2ENet: Dynamic Sparse Feature Fusion for Accurate and Efficient 3D Medical Image Segmentation (NeurIPS 2024 · similarity 0.84)
- FuseUNet: A Multi-Scale Feature Fusion Method for U-like Networks (ICML 2025 · similarity 0.84)
- nnWNet: Rethinking the Use of Transformers in Biomedical Image Segmentation and Calling for a Unified Evaluation Benchmark (CVPR 2025 · similarity 0.84)
- EMCAD: Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation (CVPR 2024 · similarity 0.82)
- DDViT: Double-Level Fusion Domain Adapter Vision Transformer (Student Abstract) (AAAI 2024 · similarity 0.82)
- S2S2: Semantic Stacking for Robust Semantic Segmentation in Medical Imaging (AAAI 2025 · similarity 0.82)
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.