Research map / Label-efficient and robust learning / Semi-supervised learning
Adaptive Bidirectional Displacement for Semi-Supervised 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 Adaptive Bidirectional Displacement (ABD), a plug-in input-perturbation module that swaps low- and high-confidence image patches between weak and strong augmented views (ABD-R for unlabeled data, ABD-I with inverse confidence for labeled data) to stabilize consistency learning for semi-supervised medical image segmentation.
Problem
Consistency learning for semi-supervised medical image segmentation becomes unstable when multiple perturbations (input, network) are combined, so single perturbations are used and cover only limited cases.
Approach
Bidirectionally displace the lowest-confidence patch of one augmented view with the most output-similar high-confidence patch from the other view (ABD-R on unlabeled data), and for labeled data displace the highest-confidence patch with the lowest-confidence patch (ABD-I, with labels displaced identically), so the mixed-perturbation consistency is controlled and uncertain regions are still learned
Key results
- On ACDC with 10% labels, Ours-ABD (BCP) reaches 89.81% DSC versus 88.84% for BCP, and Ours-ABD (Cross Teaching) reaches 88.52% vs 86.45% for the reproduced baseline.
- On ACDC with 5% labels, Ours-ABD (BCP) achieves 88.96% DSC, a 1.37% gain over BCP (87.59%).
- On PROMISE12 with 20% labels, Ours-ABD reaches 82.06% DSC, about 5% above SCP-Net (77.06%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ABD to BCP | BCP · ACDC test, 5% labeled | DSC: 87.59 -> 88.96 | Helped |
| Add ABD to BCP | BCP · ACDC test, 10% labeled | DSC: 88.84 -> 89.81 | Helped |
| Apply ABD | Semi-supervised segmentation · PROMISE12 test, 20% labeled | DSC: 77.06 -> 82.06 | Helped |
| Add input perturbation (ablation) | Cross Teaching · ACDC test, 10% labeled | DSC: 86.45 -> 86.25 | Hurt |
| Add ABD-R to input-perturbed Cross Teaching (ablation) | Cross Teaching · ACDC test, 10% labeled | DSC: 86.25 -> 87.42 | Helped |
| Add ABD-I to input-perturbed Cross Teaching (ablation) | Cross Teaching · ACDC test, 10% labeled | DSC: 86.25 -> 87.20 | Helped |
| Add both ABD-R and ABD-I (ablation) | Input-perturbed Cross Teaching · ACDC test, 10% labeled | DSC: 86.25 -> 88.52 | Helped |
| Use random patch selection instead of reliable-confidence displacement (ablation) | ABD-R · ACDC test, 10% labeled | DSC: 87.42 -> 86.55 | Hurt |
| Displace corresponding highest-confidence regions (Same) (ablation) | ABD-R · ACDC test, 10% labeled | DSC: 87.42 -> 87.22 | Hurt |
| Randomly alternate Same and Reliable displacement (ablation) | ABD-R · ACDC test, 10% labeled | DSC: 87.42 -> 87.38 | Hurt |
| Set patch number K to 16 instead of K=4 or K=64 (ablation) | ABD · ACDC test, 10% labeled | DSC: 87.39; 88.52; 87.54 | Mixed |
| Combine Colorjitter and Cutout (ablation) | ABD strong augmentation · ACDC test, 10% labeled | DSC: 88.52 | Helped |
Surprising findings
- Adding input perturbation alone to Cross Teaching lowers DSC (86.25% vs 86.45%), showing mixed perturbations destabilize consistency learning.
Limitations
- Displacement operates on fixed square patches (K=16 grid), and the authors note a patch-adaptive strategy is future work.
- Ablations are performed only on ACDC with Cross Teaching as the baseline.
Experimental setup
Datasets: ACDC PROMISE12
Benchmarks: ACDC PROMISE12
Metrics: DSC Jaccard 95HD ASD
Baselines: Cross Teaching BCP U-Net DTC URPC MC-Net SS-Net SCP-Net PLGCL CCT SLC-Net
Builds on: Cross Teaching BCP Cross Pseudo Supervision CutMix-style mixing
Closest papers by meaning
- GapMatch: Bridging Instance and Model Perturbations for Enhanced Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.87)
- Steady Progress Beats Stagnation: Mutual Aid of Foundation and Conventional Models in Mixed Domain Semi-Supervised Medical Image Segmentation (CVPR 2025 · similarity 0.86)
- DyCON: Dynamic Uncertainty-aware Consistency and Contrastive Learning for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.86)
- Gradient Alignment Improves Test-Time Adaptation for Medical Image Segmentation (AAAI 2025 · similarity 0.85)
- Combinatorial CNN-Transformer Learning with Manifold Constraints for Semi-supervised Medical Image Segmentation (AAAI 2024 · similarity 0.84)
- Semi-supervised TEE Segmentation via Interacting with SAM Equipped with Noise-Resilient Prompting (AAAI 2024 · similarity 0.84)
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.