Research map / Detection and segmentation / CNN features
nnWNet: Rethinking the Use of Transformers in Biomedical Image Segmentation and Calling for a Unified Evaluation Benchmark
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 WNet, a UNet variant that pairs convolutional local-scope blocks with transformer global-scope bridges to keep local and global features flowing continuously, and benchmarks it inside the nnUNet framework on eight 2D/3D biomedical segmentation datasets.
Problem
Hybrid convolution-transformer encoder-decoder models for biomedical image segmentation alternate global and local features, causing feature confusion and training instability; prior works also use inconsistent experimental setups, making comparisons unreliable.
Approach
Modify UNet into two cascaded encoder-decoders where convolutional blocks extract local details, transformer (pooling-token-mixer) bridges capture global context at each scale, and both streams exchange information via channel concatenation, so global and local features are transmitted continuously; evaluate all models under the same nnUNet pipeline.
Key results
- On ISIC-2017 nnWNet reaches 82.62 Jaccard / 89.44 Dice vs. nnUNet 82.30 / 89.18, with only 7M parameters versus nnUNet's 92M (2D table).
- On BTCV nnWNet reports 74.76 Jaccard vs. nnUNet 73.10 and 3D UX-Net 71.43; on Parse2022 it reports 73.71 Jaccard vs. nnUNet 72.67.
- Statistical t-tests on Jaccard show nnWNet significantly better in 28 of 32 comparisons (87.50%, p<0.05).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use nnWNet instead of nnUNet | nnUNet · DRIVE | Jaccard: 68.70±1.03 -> 69.86±1.26 | No clear effect |
| Use nnWNet instead of nnUNet | nnUNet · ISIC-2017 | Jaccard: 82.30±0.76 -> 82.62±0.71 | Helped |
| Use nnWNet instead of nnUNet | nnUNet · Kvasir-SEG | Jaccard: 83.92±2.30 -> 84.58±1.89 | No clear effect |
| Use nnWNet instead of nnUNet | nnUNet · CREMI | Jaccard: 77.57±0.25 -> 79.86±0.25 | Helped |
| Use nnWNet instead of nnUNet | nnUNet · Parse2022 | Jaccard: 72.67±1.55 -> 73.71±0.64 | Helped |
| Use nnWNet instead of nnUNet | nnUNet · AMOS22 | Jaccard: 78.12±0.71 -> 78.93±0.58 | Helped |
| Use nnWNet instead of nnUNet | nnUNet · BTCV | Jaccard: 73.10±2.07 -> 74.76±2.46 | Helped |
| Use nnWNet instead of nnUNet | nnUNet · ImageCAS | Jaccard: 70.77±0.27 -> 71.58±0.25 | Helped |
| Replace pooling token-mixer with DW Conv 7 × 7 (ablation) | WNet · ISIC-2017 | Jaccard: 82.62±0.71 -> 82.55±0.57 | No clear effect |
| Replace pooling token-mixer with DW Conv 11 × 11 (ablation) | WNet · Kvasir-SEG | Jaccard: 84.58±1.89 -> 84.49±2.08 | No clear effect |
| Replace pooling token-mixer with E-MHSA (ablation) | WNet · CREMI | Jaccard: 79.86±0.25 -> 78.37±0.20 | Hurt |
| Replace channel concatenation with element-wise multiplication (ablation) | WNet · CREMI | Jaccard: 79.86±0.25 -> 77.49±0.36 | Hurt |
Surprising findings
- Several pure transformer models (e.g., SwinUNet, nnFormer, D-Former) failed to converge or were not reported on some datasets under the unified benchmark.
- Under the unified nnUNet benchmark, some published hybrid models perform worse than a well-designed UNet baseline.
Limitations
- Gains over nnUNet on several datasets are small (e.g., ISIC-2017, BTCV Dice differences within std), and some comparisons are not significant (e.g., DRIVE vs TransAttUNet p=0.83).
- Several baselines have missing results (e.g., SwinUNet, nnFormer, D-Former, ConvFormer), limiting comprehensive comparison.
- Design claims about feature transmission are supported mainly by ablations and ERF visualizations rather than formal analysis.
Experimental setup
Datasets: DRIVE ISIC-2017 Kvasir-SEG CREMI Parse2022 AMOS22 BTCV ImageCAS
Benchmarks: DRIVE ISIC-2017 Kvasir-SEG CREMI Parse2022 AMOS22 BTCV ImageCAS
Metrics: Jaccard Dice 95HD ASD
Baselines: nnUNet BCUNet CMUNet GFANet U2Net SwinUNet H2Former ConvFormer DuAT TransAttUNet VNet nnFormer
Builds on: UNet nnUNet PoolFormer Residual blocks
Closest papers by meaning
- DDViT: Double-Level Fusion Domain Adapter Vision Transformer (Student Abstract) (AAAI 2024 · similarity 0.87)
- Rethinking U-Net: Task-Adaptive Mixture of Skip Connections for Enhanced Medical Image Segmentation (AAAI 2025 · similarity 0.86)
- Combinatorial CNN-Transformer Learning with Manifold Constraints for Semi-supervised Medical Image Segmentation (AAAI 2024 · similarity 0.86)
- EffiDec3D: An Optimized Decoder for High-Performance and Efficient 3D Medical Image Segmentation (CVPR 2025 · similarity 0.86)
- DTMFormer: Dynamic Token Merging for Boosting Transformer-Based Medical Image Segmentation (AAAI 2024 · similarity 0.86)
- Upping the Game: How 2D U-Net Skip Connections Flip 3D Segmentation (NeurIPS 2024 · similarity 0.86)
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.