Research map / Label-efficient and robust learning / Semi-supervised learning
DyCON: Dynamic Uncertainty-aware Consistency and Contrastive Learning 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 DyCON, a semi-supervised 3D medical segmentation framework that adds an entropy-weighted voxel consistency loss (UnCL) and a focal, entropy-aware patch contrastive loss (FeCL) to a mean-teacher setup.
Problem
Semi-supervised 3D medical image segmentation suffers from class imbalance and high uncertainty from pathology variation; existing uncertainty-based consistency methods discard high-uncertainty voxels and contrastive methods treat all pairs equally.
Approach
UnCL weights each voxel's consistency loss by teacher and student entropy (with a decaying beta and entropy regularization) so uncertain voxels are kept but down-weighted early; FeCL adds dual focal weights, entropy-based adjustment, and inter-network top-k hard negatives to patch-wise supervised contrastive learning.
Key results
- On ISLES-2022 with 10% labels, DyCON reaches 65.71% Dice versus 60.13% for GALoss and 58.84% for MagicNet.
- On BraTS2019 with 20% labels, DyCON reaches 88.75% Dice versus 86.63% for CML.
- On LA with 5% labels, DyCON reaches 90.96% Dice (3D UNet) and 91.18% (VNet) versus 89.12% for AC-MT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DyCON to the semi-supervised segmentation framework | 3D UNet on ISLES’22 · ISLES’22, 5% labeled | Dice (%): 53.53 -> 61.48 | Helped |
| add DyCON to the semi-supervised segmentation framework | 3D UNet on BraTS’19 · BraTS’19, 10% labeled | Dice (%): 85.26% and 87.05% | Helped |
| add UnCL without β to MT (ablation) | MT on ISLES’22 · ISLES’22, 10% labeled | Dice (%): 36.43 -> 58.78 | Helped |
| use adaptive β in UnCL (ablation) | MT + UnCL on ISLES’22 · ISLES’22, 10% labeled | Dice (%): 58.78 -> 64.52 | Helped |
| use adaptive β in UnCL (ablation) | MT + UnCL on BraTS’19 · BraTS’19, 10% labeled | Dice (%): 80.04 -> 85.97 | Helped |
| add dual focal weighting to SupCon with UnCL (ablation) | MT on ISLES’22 · ISLES’22, 10% labeled | Dice (%): 38.24 -> 63.78 | Helped |
| add inter-network top-k hard negatives (ablation) | MT + focal weights + UnCL on ISLES’22 · ISLES’22, 10% labeled | Dice (%): 63.78 -> 64.39 | Helped |
| add entropy-awareness to the focal weights (ablation) | MT + focal weights + UnCL on ISLES’22 · ISLES’22, 10% labeled | Dice (%): 63.78 -> 65.46 | Helped |
| combine FeCL with UnCL (ablation) | MT on ISLES’22 · ISLES’22, 10% labeled | Dice (%): 65.46 -> 66.07 | Helped |
| use FeCL instead of SupCon (ablation) | MT on small lesions · ISLES-2022 small lesions | Dice (%): 47.77% vs. 54.77% | Helped |
| use FeCL instead of SupCon (ablation) | MT on scattered lesions · ISLES-2022 scattered lesions | Dice (%): 65.36% vs. 70.71% | Helped |
Surprising findings
- Adding entropy weighting to MT without a scaling factor already gives large gains (ISLES Dice 36.43 to 58.78 at 10% labels), suggesting much of the gain comes from the loss weighting itself.
Limitations
- Evaluated only on four datasets, with ISLES'22 and BraTS restricted to single sequences (DWI, T2-FLAIR).
- Some hyperparameters (beta, gamma, tau, patch size) tuned per dataset, and the paper notes DyCON uses a bespoke gradient analysis in the supplement only.
Experimental setup
Datasets: ISLES-2022 BraTS2019 LA Pancreas-NIH
Benchmarks: ISLES-2022 BraTS2019 LA Pancreas-NIH
Metrics: Dice IoU HD95 ASD
Baselines: MT UA-MT DTC MCF AC-MT CML BCP MagicNet GALoss UPC URPC
Builds on: Mean Teacher SupCon Focal Loss UA-MT BCP
Closest papers by meaning
- ConDSeg: A General Medical Image Segmentation Framework via Contrast-Driven Feature Enhancement (AAAI 2025 · similarity 0.88)
- A Semantic Knowledge Complementarity based Decoupling Framework for Semi-supervised Class-imbalanced Medical Image Segmentation (CVPR 2025 · similarity 0.87)
- Minding Fuzzy Regions: A Data-driven Alternating Learning Paradigm for Stable Lesion Segmentation (CVPR 2025 · similarity 0.86)
- S2S2: Semantic Stacking for Robust Semantic Segmentation in Medical Imaging (AAAI 2025 · similarity 0.86)
- GapMatch: Bridging Instance and Model Perturbations for Enhanced Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.86)
- Neighbor Does Matter: Density-Aware Contrastive Learning for Medical Semi-supervised Segmentation (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.