Research map / Label-efficient and robust learning / Semi-supervised learning
Contrastive learning of global and local features for medical image segmentation with limited annotations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use corresponding-volume slice partitions to define global contrastive pairs (G^D) rather than random pairs (G^R) (ablation) | Global contrastive pre-training · ACDC, Prostate, MMWHS | DSC: 0.631 -> 0.691; 0.521 -> 0.579; 0.500 -> 0.553 (|X_tr|=1) | Helped |
| Add positive pairs across corresponding partitions of different volumes (G^D instead of G^D-) (ablation) | Global contrastive pre-training · Across the 9 dataset and training-volume settings | DSC | Mixed |
| Add local contrastive decoder pre-training using L^R (ablation) | G^R global pre-training · ACDC, Prostate, MMWHS | DSC: 0.631 -> 0.668; 0.521 -> 0.557; 0.500 -> 0.528 (|X_tr|=1) | Helped |
| Use corresponding-volume local-region positives (L^D) instead of random local sampling (L^R) (ablation) | Local contrastive decoder pre-training · Across the 9 dataset and training-volume settings | DSC | Mixed |
| Increase global pre-training batch size from 40 to 450 (ablation) | G^D encoder pre-training · ACDC, |X_tr|=1 | DSC: 0.691 -> 0.668 | Hurt |
| Use 6 rather than 4 partitions per volume (ablation) | G^D encoder pre-training · ACDC | DSC: 0.691 -> 0.652; 0.789 -> 0.760; 0.870 -> 0.859 | Hurt |
| Increase local region size from 1×1 to 3×3 (ablation) | Local contrastive decoder pre-training · Prostate, MMWHS | DSC: Prostate: 0.554 -> 0.567 and 0.614 -> 0.607; MMWHS: 0.559 -> 0.574 and 0.674 -> 0.681 | Mixed |
| Pre-train all 5 decoder blocks rather than a shorter decoder (ablation) | Local contrastive decoder pre-training · ACDC | DSC: 0.703 -> 0.690 (3×3 regions, |X_tr|=1; d3 -> d5) | Hurt |
| Jointly optimize global and local losses rather than use stage-wise training (ablation) | Proposed pre-training · ACDC | DSC: 0.725 -> 0.634 (|X_tr|=1); 0.789 -> 0.741 (|X_tr|=2) | Hurt |
| Combine proposed pre-training with Mixup | Proposed initialization · ACDC, Prostate, MMWHS | DSC: 0.725 -> 0.757; 0.579 -> 0.588; 0.569 -> 0.617 (|X_tr|=1) | Helped |
| Combine proposed pre-training with self-training | Proposed initialization · Prostate, MMWHS | DSC: 0.579 -> 0.607; 0.569 -> 0.647 (|X_tr|=1) | Helped |
| Use proposed contrastive initialization instead of random initialization | UNet segmentation · ACDC, Prostate, MMWHS | DSC: 0.614 -> 0.725; 0.489 -> 0.579; 0.451 -> 0.569 (|X_tr|=1) | Helped |
Closest papers by meaning
- Self-Paced Contrastive Learning for Semi-supervised Medical Image Segmentation with Meta-labels (NeurIPS 2021 · similarity 0.92)
- Neighbor Does Matter: Density-Aware Contrastive Learning for Medical Semi-supervised Segmentation (AAAI 2025 · similarity 0.87)
- SGTC: Semantic-Guided Triplet Co-training for Sparsely Annotated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.87)
- Revisiting MAE Pre-training for 3D Medical Image Segmentation (CVPR 2025 · similarity 0.87)
- DyCON: Dynamic Uncertainty-aware Consistency and Contrastive Learning for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.87)
- Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation (CVPR 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.