Research map / Label-efficient and robust learning / Semi-supervised learning
GraphCL: Graph-based Clustering 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 GraphCL, a semi-supervised medical image segmentation method that adds a structure-aware GCN and a correlation-clustering loss on top of bidirectional copy-paste (BCP) with a V-Net/U-Net teacher-student setup.
Problem
Semi-supervised medical image segmentation with very few labeled scans, where existing methods ignore graph/data-structure information and suffer from distribution mismatch between labeled and unlabeled data.
Approach
Builds a dense instance graph over CNN features (structure-aware alignment) and voxel-level affinity graphs, processed by a GCN and optimized with a correlation clustering loss (k-less clustering), on top of the BCP copy-paste teacher-student framework.
Key results
- On LA with 10% labels, GraphCL reaches 90.24% Dice, 82.31% Jaccard, 6.42 95HD, versus 89.39% Dice for BCP.
- On ACDC with 5% labels, GraphCL reaches 88.68% Dice versus 86.83% for BCP.
- On Pancreas-NIH with 20% labels, GraphCL reaches 83.15% Dice versus 81.12% for BCP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add structure-aware alignment (SA) (ablation) | GraphCL · LA (5% labeled) | Dice: 88.13 -> 88.80 | Helped |
| add graph clustering loss (L_CC) (ablation) | GraphCL · LA (5% labeled) | Dice: 88.34 -> 88.80 | Helped |
| add structure-aware alignment (SA) (ablation) | GraphCL · LA (10% labeled) | Dice: 88.79 -> 90.24 | Helped |
| add graph clustering loss (L_CC) (ablation) | GraphCL · LA (10% labeled) | Dice: 90.00 -> 90.24 | Helped |
| add structure-aware alignment (SA) (ablation) | GraphCL · ACDC (5% labeled) | Dice: 87.80 -> 88.68 | Helped |
| add graph clustering loss (L_CC) (ablation) | GraphCL · ACDC (5% labeled) | Dice: 88.13 -> 88.68 | Helped |
| add structure-aware alignment (SA) (ablation) | GraphCL · ACDC (10% labeled) | Dice: 89.52 -> 89.31 | No clear effect |
| add graph clustering loss (L_CC) (ablation) | GraphCL · ACDC (10% labeled) | Dice: 89.53 -> 89.31 | Hurt |
| add structure-aware alignment (SA) (ablation) | GraphCL · Pancreas-NIH (20% labeled) | Dice: 82.32 -> 83.15 | Helped |
| add graph clustering loss (L_CC) (ablation) | GraphCL · Pancreas-NIH (20% labeled) | Dice: 82.47 -> 83.15 | Helped |
| place GCN at encoder layer 5 rather than layer 1 (ablation) | GraphCL · ACDC (5% labeled) | Dice: 0.030 -> 88.68 | Helped |
| place GCN at encoder layer 5 rather than layer 1 (ablation) | GraphCL · ACDC (10% labeled) | Dice: 0.008 -> 89.31 | Helped |
Limitations
- Gains over BCP are modest on some metrics (e.g., LA 10% ASD and ACDC 10% 95HD show ties or regressions relative to the best baseline in ablations)
- Graph construction may be noisy; authors note need for more reliable labels and graph accuracy
- Method description of graph construction and loss is informal and partly inconsistent, limiting reproducibility
Experimental setup
Datasets: ACDC LA (Atrial Segmentation Challenge) Pancreas-NIH
Benchmarks: LA ACDC Pancreas-NIH
Metrics: Dice Jaccard 95HD ASD
Baselines: V-Net U-Net UA-MT SASSNet DTC URPC MC-Net SS-Net BCP DAN ADVENT CoraNet
Builds on: Bidirectional Copy-Paste (BCP) graph convolutional network correlation clustering
Closest papers by meaning
- Neighbor Does Matter: Density-Aware Contrastive Learning for Medical Semi-supervised Segmentation (AAAI 2025 · similarity 0.86)
- Combinatorial CNN-Transformer Learning with Manifold Constraints for Semi-supervised Medical Image Segmentation (AAAI 2024 · similarity 0.86)
- SGTC: Semantic-Guided Triplet Co-training for Sparsely Annotated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.85)
- SCCS: Deep Neural Spectral Clustering for Self-Supervised Subcellular Structure Segmentation (AAAI 2025 · similarity 0.84)
- Semi-supervised TEE Segmentation via Interacting with SAM Equipped with Noise-Resilient Prompting (AAAI 2024 · similarity 0.84)
- Annotation Ambiguity Aware Semi-Supervised Medical Image Segmentation (CVPR 2025 · similarity 0.84)
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.