Research map / Detection and segmentation / Weakly supervised detection and segmentation
Integrating Deep Metric Learning with Coreset for Active Learning in 3D 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 Group-based Contrastive Learning (GCL), an NT-Xent extension using patient/volume/adjacent-slice groupings, to learn a metric for K-Center Coreset slice-based active learning in 3D medical segmentation, evaluated with weak and full annotations.
Problem
Active learning for 3D medical segmentation is costly; existing AL methods are volume-based, ignore domain data groupings, and are not studied with weak (scribble) annotations or slice-based selection.
Approach
Replace the Euclidean metric in Coreset with a learned metric from a SimCLR-style ResNet-18 encoder trained with a group contrastive loss that pulls together slices from the same patient, volume, or adjacent-slice group, plus a batch sampler that makes same-group slices co-occur in batches.
Key results
- On ACDC with weak annotation at 2%, the method reaches 52.3 DICE vs 45.2 for Coreset (best baseline), per the main table.
- Achieves best performance on 21 of 27 comparison points across tables (per the paper's count).
- With pre-trained weights on ACDC, mean DICE over all annotation points is 79.5 vs 78.6 for Coreset.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use metric-learning Coreset instead of Coreset | Slice-based active learning with weak supervision · ACDC, 2% annotation | DICE: 45.2 -> 52.3 | Helped |
| Use metric-learning Coreset instead of CoreGCN | Slice-based active learning with full supervision · ACDC, 2% annotation | DICE: 67.7 -> 70.9 | Helped |
| Use metric-learning Coreset instead of the strongest listed baseline | Slice-based active learning with full supervision · CHAOS, 2% annotation | DICE: 80.7 -> 80.5 | Hurt |
| Use metric-learning Coreset instead of the strongest listed baseline | Slice-based active learning with full supervision · DAVIS, 10% annotation | DICE: 43.6 -> 42.8 | Hurt |
| Add contrastive losses to vanilla Coreset (ablation) | Coreset on ACDC weak annotations · 2-5% annotation datapoint | mDICE: 58.5 -> 59.5; 58.5 -> 58.1 | Mixed |
| Add volume group loss only (ablation) | Coreset on ACDC weak annotations · 2-5% annotation datapoint | mDICE: 58.5 -> 60.7 | Helped |
| Add NT-Xent, patient, and volume group losses (ablation) | Coreset on ACDC weak annotations · 2-5% annotation datapoint | mDICE: 58.5 -> 65.4 | Helped |
| Add adjacent slice group loss to NT-Xent, patient, and volume group losses (ablation) | Coreset on ACDC weak annotations · 2-5% annotation datapoint | mDICE: 65.4 -> 61.1 | Hurt |
Surprising findings
- Adding the slice-group (adjacent slice) contrastive loss generally hurts performance in the ablation, and the volume-group loss gives the best additive gains.
- Random sampling is the strongest baseline on DAVIS and in several full-supervision settings, and the method does not beat all baselines on every dataset (e.g., 40% weak on ACDC is tied).
Limitations
- Limited compute restricted the number of experimental runs and models trained.
- Training set bias and fairness for underrepresented groups were not considered.
- Group-loss weights were tuned on ACDC with the best combination reported, and the paper's clinical utility claims for DICE 0.5-0.9 are general rather than measured.
Experimental setup
Datasets: ACDC CHAOS MS-CMR DAVIS
Benchmarks: ACDC CHAOS MS-CMR DAVIS
Metrics: 3D DICE score mean DICE
Baselines: Random sampling Coreset (K-Center Greedy) CoreGCN TypiClust Stochastic Batches VAAL Deep Ensembles (Variance Ratio) BALD
Builds on: Coreset (k-Center Greedy) NT-Xent / SimCLR DMPLS UNet CoreGCN
Closest papers by meaning
- SGTC: Semantic-Guided Triplet Co-training for Sparsely Annotated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.88)
- VISTA3D: A Unified Segmentation Foundation Model For 3D Medical Imaging (CVPR 2025 · similarity 0.86)
- Upping the Game: How 2D U-Net Skip Connections Flip 3D Segmentation (NeurIPS 2024 · similarity 0.86)
- E2ENet: Dynamic Sparse Feature Fusion for Accurate and Efficient 3D Medical Image Segmentation (NeurIPS 2024 · similarity 0.86)
- ToNNO: Tomographic Reconstruction of a Neural Network's Output for Weakly Supervised Segmentation of 3D Medical Images (CVPR 2024 · similarity 0.86)
- 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.