Research map / 3D vision / Depth and stereo
Upping the Game: How 2D U-Net Skip Connections Flip 3D 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 a U-shaped Connection (uC) that replaces 3D U-Net skip connections with a simplified 2D U-Net plus a Dual Feature Integration module, yielding uC 3DU-Net that beats SOTA 3D segmentation models with fewer parameters on five CT/MRI/OCT datasets.
Problem
3D CNN segmentation underutilizes high-density axial-slice plane features because isotropic 3D convolutions are parameter-inefficient and anisotropic medical volumes have uneven information density across axes.
Approach
Replace standard 3D skip connections with a lightweight 2D U-Net (uC) applied slice-wise, and fuse its outputs with 3D decoder features via a DFi block (concatenation + addition-based attention), so the network captures in-slice detail cheaply while keeping 3D context.
Key results
- uC 3DU-Net reaches 94.13 Dice on OIMHS vs 92.53-93.53 for prior baselines, with 21.7M params and 286.4G FLOPs.
- On FLARE2021 uC 3DU-Net averages 94.1 Dice with 21.7M params, about 40.9% of 3D UX-Net's parameter count, and improves mean Dice by 0.7%.
- Adding uC to SegResNet, TransBTS, SwinUNETR and 3D UX-Net improves Dice by 0.65%-1.21% on FLARE2021.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace 3D U-Net skip connections with uC | 3D U-Net · FLARE2021 | Dice (Mean): 93.4 -> 94.1 | Helped |
| replace 3D U-Net skip connections with uC | 3D U-Net · FeTA2021 | Dice (Mean): 87.4 -> 87.8 | Helped |
| replace 3D U-Net skip connections with uC | 3D U-Net · OIMHS | Dice: 93.53 -> 94.13 | Helped |
| replace 3D U-Net skip connections with uC | 3D U-Net · AbdomenCT-1K | Dice: 92.36±3.20 -> 93.35±2.60 | Helped |
| replace 3D U-Net skip connections with uC | 3D U-Net · BTCV | Dice: 82.31±3.06 -> 82.74±3.18 | Helped |
| use 2D rather than 3D convolutions in uC (ablation) | 3D U-Net (8 channels) + uC · OIMHS | Dice: 91.84±2.35 -> 94.31±1.62 | Helped |
| add DFi to stage 1 uC (ablation) | 3D U-Net + uC stage 1 · OIMHS | Dice: 94.55±2.16 -> 94.79±1.56 | Helped |
| add DFi to stage 2 uC (ablation) | 3D U-Net + uC stage 2 · OIMHS | Dice: 94.37±1.92 -> 94.53±1.93 | Helped |
| add DFi to stage 3 uC (ablation) | 3D U-Net + uC stage 3 · OIMHS | Dice: 93.75±2.12 -> 94.13±2.07 | Helped |
| add DFi to uC stages 1–3 (ablation) | 3D U-Net + uC stages 1–3 · OIMHS | Dice: 94.84±1.68 -> 95.03±1.68 | Helped |
| add uC at stage 1 (ablation) | 3D U-Net · OIMHS | Dice: 92.79±2.09 -> 94.55±2.16 | Helped |
| add uC at stage 2 (ablation) | 3D U-Net · OIMHS | Dice: 92.79±2.09 -> 94.37±1.92 | Helped |
Surprising findings
- A 16-channel uC 3DU-Net outperforms the 32-channel 3D U-Net with similar parameters and FLOPs.
- Slicing along the axial-slice plane (D dimension) gives the best uC gains, while 2D uC outperforms 3D uC at equal channel depth.
Limitations
- Gains on the isotropic FeTA2021 dataset are smaller than on anisotropic datasets.
- Relies on fixed per-dataset preprocessing and small public datasets; results from five-fold cross-validation on some datasets and single splits elsewhere.
- Improvements over baselines are often within overlapping standard deviations on some metrics.
Experimental setup
Datasets: FLARE2021 OIMHS FeTA2021 AbdomenCT-1K BTCV
Benchmarks: FLARE2021 OIMHS FeTA2021 AbdomenCT-1K BTCV
Metrics: mIoU Dice ASSD HD95 AdjRand VOE
Baselines: 3D U-Net SegResNet RAP-Net nnU-Net TransBTS UNETR nnFormer SwinUNETR 3D UX-Net SASAN UNETR++ D-LKA Net
Builds on: 3D U-Net U-Net DenseNet nnU-Net
Closest papers by meaning
- Rethinking U-Net: Task-Adaptive Mixture of Skip Connections for Enhanced Medical Image Segmentation (AAAI 2025 · similarity 0.90)
- UniPro: Unified Multi-Mode Medical Image Segmentation from 2D Images to 3D Volumes via Propagation (NeurIPS 2026 · similarity 0.88)
- EffiDec3D: An Optimized Decoder for High-Performance and Efficient 3D Medical Image Segmentation (CVPR 2025 · similarity 0.88)
- FuseUNet: A Multi-Scale Feature Fusion Method for U-like Networks (ICML 2025 · similarity 0.87)
- VISTA3D: A Unified Segmentation Foundation Model For 3D Medical Imaging (CVPR 2025 · similarity 0.87)
- Integrating Deep Metric Learning with Coreset for Active Learning in 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.