Research map / Label-efficient and robust learning / Semi-supervised learning
Steady Progress Beats Stagnation: Mutual Aid of Foundation and Conventional Models in Mixed Domain 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 SynFoC, a mean-teacher framework that synergistically trains MedSAM (LoRA-tuned) and U-Net with instance-wise self-mutual confidence pseudo-label ensembling and consensus-divergence consistency regularization for mixed-domain semi-supervised medical image segmentation.
Problem
Semi-supervised medical image segmentation under domain shift (mixed-domain unlabeled data, single-domain labels), where foundation models like MedSAM make overconfident wrong pseudo-labels and conventional U-Nets overfit to labeled data.
Approach
Instead of letting the foundation model dominate, train a foundation model and a conventional model jointly: pseudo-labels are ensembled per instance with a ratio set by self-confidence (teacher-student stability) and mutual confidence (agreement between the two models), and consensus regions get entropy minimization while divergent regions get MSE alignment.
Key results
- On Prostate with 20 labels, SynFoC MedSAM reaches 79.30% average DSC (best among compared methods), a 10.31% DSC gain over the best competitor per the paper.
- On M&Ms with 5 labels, SynFoC MedSAM achieves 80.84% average DSC versus 78.16% for SymGD (second best), roughly 2% improvement.
- On BUSI, SynFoC MedSAM reaches 71.61% (1/8 labels) and 74.75% (1/4 labels) average Dice on the cancer class, outperforming other methods by about 2.7% and 2.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Synergistically train U-Net and MedSAM with SMC and CDCR (ablation) | U-Net and MedSAM · Prostate | DSC: 81.14 -> 87.16 | Helped |
| Add SMC to Base training (ablation) | U-Net and MedSAM · Prostate | DSC: 81.14 -> 86.19 | Helped |
| Add CDCR to Base training (ablation) | U-Net and MedSAM · Prostate | DSC: 81.14 -> 82.88 | Helped |
| Use SMC and CDCR together (ablation) | U-Net and MedSAM · Prostate | DSC: 86.80 -> 87.16 | Helped |
| Set the pseudo-label ensemble ratio to 0.5 (ablation) | U-Net and MedSAM · Prostate | DSC: 86.80 -> 85.93 | Hurt |
| Use the CPS pseudo-label integration strategy (ablation) | U-Net and MedSAM · Prostate | DSC: 86.80 -> 84.30 | Hurt |
| Use a time-dependent linear ensemble ratio (ablation) | U-Net and MedSAM · Prostate | DSC: 86.80 -> 86.19 | Hurt |
| Use self-confidence only to set the ensemble ratio (ablation) | U-Net and MedSAM · Prostate | DSC: 86.80 -> 86.23 | Hurt |
| Use mutual confidence only to set the ensemble ratio (ablation) | U-Net and MedSAM · Prostate | DSC: 86.80 -> 86.29 | Hurt |
| Use SMC-based SynFoCMedSAM instead of CPC-SAM | MedSAM · Prostate | DSC: +10.31% | Helped |
| Use SMC-based SynFoCMedSAM instead of the strongest reported baseline | MedSAM · Fundus | DSC: 87.66 -> 88.60 | Helped |
| Use SMC-based SynFoCMedSAM instead of the strongest reported baseline | MedSAM · M&Ms | DSC: 77.81 -> 80.84 | Helped |
Surprising findings
- The conventional U-Net trained from scratch can correct high-confidence mispredictions of MedSAM, so the conventional model is kept as an active partner rather than discarded.
Limitations
- Training takes roughly 1 extra hour versus MedSAM alone and requires training two models jointly.
- Evaluated mainly on 2D slices from four multi-domain datasets with very few labels; generality to other domains not established beyond the supplementary SSMIS/UDA experiments.
- Pseudo-label ensembling relies on a fixed confidence threshold (0.95) and hand-designed self/mutual confidence formulas.
Experimental setup
Datasets: Prostate Fundus M&Ms BUSI ACDC MS-CMRSeg
Benchmarks: Prostate Fundus M&Ms BUSI ACDC MS-CMRSeg
Metrics: DSC Jaccard 95HD ASD
Baselines: UA-MT FixMatch SS-Net BCP CauSSL ABD SymGD SAMed SemiSAM H-SAM CPC-SAM
Builds on: MedSAM Mean Teacher SymGD SAMed BCP
Closest papers by meaning
- Constructing and Exploring Intermediate Domains in Mixed Domain Semi-supervised Medical Image Segmentation (CVPR 2024 · similarity 0.91)
- GapMatch: Bridging Instance and Model Perturbations for Enhanced Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.89)
- Adaptive Bidirectional Displacement for Semi-Supervised Medical Image Segmentation (CVPR 2024 · similarity 0.86)
- SemiDAViL: Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation (CVPR 2025 · similarity 0.86)
- beta-FFT: Nonlinear Interpolation and Differentiated Training Strategies for Semi-Supervised Medical Image Segmentation (CVPR 2025 · similarity 0.86)
- Versatile Medical Image Segmentation Learned from Multi-Source Datasets via Model Self-Disambiguation (CVPR 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.