Research map / Label-efficient and robust learning / Domain adaptation and generalization
A Unified Degradation-Robust Approach to SSL and UDA for 3D Medical Images
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 unified 3D medical segmentation framework combining wavelet-based cross-component exchange, a re-weighted diffusion VNet, and cross-decoder pseudo training to handle semi-supervised learning, domain adaptation, and degraded inputs.
Problem
Medical image segmentation under limited annotations and domain shift, where test images may also be degraded by noise or blur; prior SSL and UDA methods are task-specific and degrade on degraded images.
Approach
Swap high-frequency Haar wavelet components between labeled and unlabeled volumes to enhance robustness, use a diffusion VNet whose backbone and skip features are re-weighted via FFT-based scaling, and train three decoders with cross-decoder pseudo-labels (two vanilla VNets plus the diffusion decoder).
Key results
- On MMWHS CT-to-MR, the method reaches 83.4% average Dice and 2.23 ASD versus 71.4 for GenericSSL.
- On the LA dataset with 10% labels, Dice 91.07 versus 90.31 for GenericSSL.
- On degraded MR-to-CT test images, the method retains about 90-91 Dice across noise and blur conditions, outperforming GenericSSL with and without preprocessing.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed unified framework instead of Genericssl | LA with 10% labeled data · LA | Dice: 90.31 -> 91.07 | Helped |
| use the proposed unified framework instead of MLRPL | Pancreas with 20% labeled data · Pancreas | Dice: 81.53 -> 82.81 | Helped |
| use the proposed unified framework instead of Genericssl | MMWHS MR-to-CT adaptation · MMWHS MR to CT | Avg Dice: 90.1 -> 91.1 | Helped |
| use the proposed unified framework instead of Genericssl | MMWHS CT-to-MR adaptation · MMWHS CT to MR | Avg Dice: 71.4 -> 83.42 | Helped |
| add a vanilla VNet decoder (ablation) | D-VNet self-training on MMWHS CT to MR · MMWHS CT to MR | Dice: 60.6 -> 65.4 | Helped |
| add a second vanilla decoder and Cross-Decoder Pseudo training (ablation) | D-VNet with an additional vanilla VNet decoder on MMWHS CT to MR · MMWHS CT to MR | Dice: 65.4 -> 76.1 | Helped |
| add the RW-DVNet reweighting strategy (ablation) | D-VNet with CDP on MMWHS CT to MR · MMWHS CT to MR | Dice: 76.1 -> 78.8 | Helped |
| add WCCE (ablation) | RW-DVNet with CDP on MMWHS CT to MR · MMWHS CT to MR | Dice: 78.8 -> 83.4 | Helped |
| add WCCE (ablation) | the proposed method on MMWHS CT-to-MR under Gaussian noise σ = 0.2 · MMWHS CT to MR, Gaussian noise σ = 0.2 | Dice: 81.88 -> 83.06 | Helped |
| add WCCE (ablation) | the proposed method on MMWHS CT-to-MR under motion blur k = 10 · MMWHS CT to MR, motion blur k = 10 | Dice: 79.02 -> 81.74 | Helped |
| add WCCE (ablation) | the proposed method on MMWHS CT-to-MR under Rician noise σ = 0.1 · MMWHS CT to MR, Rician noise σ = 0.1 | Dice: 78.65 -> 82.31 | Helped |
Surprising findings
- Coupled cross-pseudo supervision outperforms the decoupled flow used by GenericSSL, contrary to that prior work's design.
Limitations
- Poisson noise simulation on reconstructed images does not reproduce sinogram-level CT noise.
- Hyperparameters (b, s, beta) are set empirically per dataset.
- Degraded-image evaluation uses synthetic corruptions applied to test images only.
Experimental setup
Datasets: Left Atrium (LA) Pancreas MMWHS
Benchmarks: Left Atrium Pancreas MMWHS
Metrics: Dice Jaccard HD95 ASD
Baselines: VNet UA-MT URPC DTC SASSNet MC-Net LMISA-3D SS-Net SimCVD BCP MLRPL GenericSSL
Builds on: GenericSSL (Wang and Li 2024) Diffusion VNet cross pseudo supervision Denoising Diffusion Implicit Models FFT-based re-weighting (Si et al. 2024)
Closest papers by meaning
- MAPSeg: Unified Unsupervised Domain Adaptation for Heterogeneous Medical Image Segmentation Based on 3D Masked Autoencoding and Pseudo-Labeling (CVPR 2024 · similarity 0.89)
- S2S2: Semantic Stacking for Robust Semantic Segmentation in Medical Imaging (AAAI 2025 · similarity 0.87)
- Samba: Severity-aware Recurrent Modeling for Cross-domain Medical Image Grading (NeurIPS 2024 · similarity 0.87)
- Unleashing the Potential of SAM for Medical Adaptation via Hierarchical Decoding (CVPR 2024 · similarity 0.86)
- Learning Generalized Medical Image Segmentation from Decoupled Feature Queries (AAAI 2024 · similarity 0.86)
- One-to-Multiple: A Progressive Style Transfer Unsupervised Domain-Adaptive Framework for Kidney Tumor 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.