Research map / Label-efficient and robust learning / Semi-supervised learning
Neighbor Does Matter: Density-Aware Contrastive Learning for Medical Semi-supervised 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 DACL, a density-aware contrastive learning strategy that pushes low-density anchor features toward high-density positive-key cluster centers, combined with co-training, for semi-supervised medical image segmentation on ACDC and Synapse.
Problem
Semi-supervised multi-organ medical image segmentation suffers from scarce labels and low soft-tissue contrast; existing pseudo-labeling and consistency methods use single samples and ignore feature-space neighborhood structure and sparsity.
Approach
Estimate feature density via average cosine similarity to k-nearest neighbors (multi-scale, with a class memory bank), select low-density features as contrastive anchors and high-density features as positive keys approximating cluster centers, and weight each pair with a positiveness score in a soft contrastive loss added to co-training.
Key results
- On ACDC with 10% labels, DACL reaches 90.91% Dice vs 89.4% for ACTION++ (second best).
- On Synapse with 20% labels, DACL reaches 41.32% Dice vs 39.61% for DHC.
- Ablation on Synapse (10% labels): density-aware single-scale sampling adds 1.13% Dice over the CPS baseline, and the full model reaches 29.88% Dice.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DACL instead of ACTION++ | Semi-supervised segmentation · ACDC 5% labeled | Dice (%): 88.5$\pm$0.9 -> 90.12$\pm$0.4 | Helped |
| Use DACL instead of ACTION++ | Semi-supervised segmentation · ACDC 10% labeled | Dice (%): 89.4$\pm$1.1 -> 90.91$\pm$0.5 | Helped |
| Use DACL instead of ACTION++ | Semi-supervised segmentation · Synapse 10% labeled | Dice (%): 27.41$\pm$0.9 -> 29.88$\pm$0.7 | Helped |
| Use DACL instead of DHC | Semi-supervised segmentation · Synapse 20% labeled | Dice (%): 39.61$\pm$3.9 -> 41.32$\pm$1.2 | Helped |
| Add plain contrastive learning with random sampling (ablation) | CPS baseline · ACDC 10% labeled | Dice (%): 85.71$\pm$0.9 -> 87.50$\pm$0.8 | Helped |
| Add plain contrastive learning with random sampling (ablation) | CPS baseline · Synapse 20% labeled | Dice (%): 18.07$\pm$1.2 -> 24.32$\pm$0.6 | Helped |
| Add single-scale density-aware anchor and key sampling (ablation) | Plain contrastive learning · ACDC 10% labeled | Dice (%): 87.50$\pm$0.8 -> 88.60$\pm$0.7 | Helped |
| Add single-scale density-aware anchor and key sampling (ablation) | Plain contrastive learning · Synapse 20% labeled | Dice (%): 24.32$\pm$0.6 -> 25.45$\pm$0.6 | Helped |
| Add multi-scale density estimation (ablation) | Single-scale density-aware sampling · ACDC 10% labeled | Dice (%): 88.60$\pm$0.7 -> 89.30$\pm$0.6 | Helped |
| Add multi-scale density estimation (ablation) | Single-scale density-aware sampling · Synapse 20% labeled | Dice (%): 25.45$\pm$0.6 -> 26.30$\pm$0.5 | Helped |
| Add class memory for density estimation and extend positive keys with memory features (ablation) | Multi-scale density estimation · ACDC 10% labeled | Dice (%): 89.30$\pm$0.6 -> 89.80$\pm$0.5 | Helped |
| Add class memory for density estimation and extend positive keys with memory features (ablation) | Multi-scale density estimation · Synapse 20% labeled | Dice (%): 26.30$\pm$0.5 -> 27.12$\pm$0.5 | Helped |
Limitations
- Absolute Synapse Dice remains low (about 30-41%) in the semi-supervised settings.
- Sensitive to loss-weight choices; high contrastive weight is detrimental according to the authors' analysis.
Experimental setup
Datasets: ACDC Synapse
Benchmarks: ACDC Synapse
Metrics: Dice Jaccard ASD
Baselines: V-Net CPS MC-Net URPC SS-Net BCP CAML DHC Cross-ALD ACTION++
Builds on: CPS MC-Net SupCon Mean-Teacher masked average pooling
Closest papers by meaning
- Hunting Sparsity: Density-Guided Contrastive Learning for Semi-Supervised Semantic Segmentation (CVPR 2023 · similarity 0.91)
- SGTC: Semantic-Guided Triplet Co-training for Sparsely Annotated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.87)
- GraphCL: Graph-based Clustering for Semi-Supervised Medical Image Segmentation (ICML 2025 · similarity 0.86)
- DyCON: Dynamic Uncertainty-aware Consistency and Contrastive Learning for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.86)
- Semi-supervised TEE Segmentation via Interacting with SAM Equipped with Noise-Resilient Prompting (AAAI 2024 · similarity 0.85)
- Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.85)
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.