Research map / Label-efficient and robust learning / Semi-supervised learning
GapMatch: Bridging Instance and Model Perturbations for Enhanced 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 GapMatch, a semi-supervised medical segmentation framework that combines instance perturbation (FixMatch-style weak/strong consistency) with Gradient-based Adversarial Perturbation (GAP), a gradient-ascent parameter perturbation, and reports gains on ACDC and LA.
Problem
Semi-supervised medical image segmentation with very few labels, where consistency regularization typically uses only instance perturbations or only model perturbations, not both.
Approach
Uses the consistency loss gradient to form an adversarial parameter perturbation (GAP, θ* = θ + ε g/||g||) and enforces consistency under the perturbed model in addition to instance-level weak/strong consistency, iterating so the decision boundary moves to low-density regions with larger margins.
Key results
- On ACDC with 1% labels, GapMatch reaches 86.8 Dice / 0.81 ASD versus SS-Net's 63.4 / 2.94 at the same ratio.
- On LA with 10% labels, GapMatch reaches 91.0 Dice / 1.46 ASD, close to the fully supervised VNet (91.5 Dice).
- Adding GAP to ICT raises ACDC 1% Dice from 31.7 to 75.8; GAP also improves EM, Pseudo Label, MixUp, CutMix and VAT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add GAP to EM (ablation) | EM · ACDC, 1%, 5%, and 10% labeled | Dice (%): ↑13.9; ↑18.9; ↑3.2 | Helped |
| Add GAP to Pseudo Label (ablation) | Pseudo Label · ACDC, 1%, 5%, and 10% labeled | Dice (%): ↑25.4; ↑20.4; ↑2.35 | Helped |
| Add GAP to ICT (ablation) | ICT · ACDC, 1%, 5%, and 10% labeled | Dice (%): ↑44.1; ↑26.5; ↑3.4 | Helped |
| Add GAP to MixUp (ablation) | MixUp · ACDC, 1%, 5%, and 10% labeled | Dice (%): ↑43.0; ↑25.6; ↑3.6 | Helped |
| Add GAP to CutMix (ablation) | CutMix · ACDC, 1%, 5%, and 10% labeled | Dice (%): ↑16.1; ↑16.9; ↑2.9 | Helped |
| Add GAP to VAT (ablation) | VAT · ACDC, 1%, 5%, and 10% labeled | Dice (%): ↑7.6; ↑5.2; ↑1.5 | Helped |
| Add GAP to EM (ablation) | EM · ACDC, 1%, 5%, and 10% labeled | ASD (voxel): ↓18.7; ↓4.7; ↓1.67 | Helped |
| Add GAP to Pseudo Label (ablation) | Pseudo Label · ACDC, 1%, 5%, and 10% labeled | ASD (voxel): ↓5.76; ↓2.05; ↓1.04 | Helped |
| Add GAP to ICT (ablation) | ICT · ACDC, 1%, 5%, and 10% labeled | ASD (voxel): ↓26.8; ↓3.61; ↓1.15 | Helped |
| Add GAP to MixUp (ablation) | MixUp · ACDC, 1%, 5%, and 10% labeled | ASD (voxel): ↓24.3; ↓0.9; ↓1.84 | Helped |
| Add GAP to CutMix (ablation) | CutMix · ACDC, 1%, 5%, and 10% labeled | ASD (voxel): ↓9.46; ↓6.54; ↓0.93 | Helped |
| Add GAP to VAT (ablation) | VAT · ACDC 1% labeled | ASD (voxel): ↑0.11 | Hurt |
Surprising findings
- GapMatch trained from scratch outperforms pre-trained SOTA methods ACTION and ARCO on most metrics.
- GAP adds large gains to weak baselines at 1% labels, e.g. ICT Dice increases by 44.1 points.
Limitations
- Training time per iteration is increased relative to other model perturbation methods.
- Gains reported for only two medical benchmarks (ACDC and LA).
Experimental setup
Datasets: ACDC LA
Benchmarks: ACDC LA
Metrics: Dice ASD
Baselines: DAN DTC DCT UAMT SASSNet CPS GCL MC-Net SS-Net ACTION BCP ARCO
Builds on: FixMatch Virtual Adversarial Training Interpolation Consistency Training Mean Teacher BCP
Closest papers by meaning
- VQ-Seg: Vector-Quantized Token Perturbation for Semi-Supervised Medical Image Segmentation (NeurIPS 2025 · similarity 0.89)
- Steady Progress Beats Stagnation: Mutual Aid of Foundation and Conventional Models in Mixed Domain Semi-Supervised Medical Image Segmentation (CVPR 2025 · similarity 0.89)
- Adaptive Bidirectional Displacement for Semi-Supervised Medical Image Segmentation (CVPR 2024 · similarity 0.87)
- DyCON: Dynamic Uncertainty-aware Consistency and Contrastive Learning for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.86)
- Constructing and Exploring Intermediate Domains in Mixed Domain Semi-supervised Medical Image Segmentation (CVPR 2024 · similarity 0.86)
- S2S2: Semantic Stacking for Robust Semantic Segmentation in Medical Imaging (AAAI 2025 · 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.