Research map / Label-efficient and robust learning / Semi-supervised learning
Anatomical Knowledge Mining and Matching for Semi-supervised Medical Multi-structure Detection
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 Semi-akmm, a mean-teacher semi-supervised detector for multi-structure ultrasound detection that adds graph-based anatomical structure matching and adaptive proposal-proportion priors.
Problem
Semi-supervised detection of multiple anatomical structures in medical (mainly ultrasound) images with very few box labels; general SSOD methods from natural images transfer poorly.
Approach
Mines distributional priors (proposal area/proportion-based loss weights and dynamic per-structure thresholds) via APKT, and aligns topological anatomical graphs between teacher and student with cross-attention graph matching and a memory bank in ASM.
Key results
- Semi-akmm reaches 60.6 mAP on 4C-A at 10% labels, versus 57.4 for Label Matching and 56.5 for Unbiased Teacher (Table 1).
- On 3VT-A at 1% labels, Semi-akmm improves over the best prior SSOD method by 1.3% mAP and over supervised by 21.4% mAP.
- Ablation at 10% labels: ASM plus APKT yields +7.5 mAP on 4C-A, +6.1 on 3VT-A, +5.1 on EPV over the baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ASM to the baseline, omitting APKT (ablation) | Semi-akmm · 4C-A, 10% labeled | mAP: 53.1 -> 57.3 (+4.2) | Helped |
| Add APKT to the baseline, omitting ASM (ablation) | Semi-akmm · 4C-A, 10% labeled | mAP: 53.1 -> 56.2 (+3.1) | Helped |
| Add ASM to the baseline, omitting APKT (ablation) | Semi-akmm · 4C-B, 10% labeled | mAP: 52.8 -> 56.4 (+3.6) | Helped |
| Add APKT to the baseline, omitting ASM (ablation) | Semi-akmm · 4C-B, 10% labeled | mAP: 52.8 -> 57.2 (+4.4) | Helped |
| Add ASM to the baseline, omitting APKT (ablation) | Semi-akmm · 3VT-A, 10% labeled | mAP: 55.7 -> 58.2 (+2.5) | Helped |
| Add APKT to the baseline, omitting ASM (ablation) | Semi-akmm · 3VT-A, 10% labeled | mAP: 55.7 -> 59.3 (+3.6) | Helped |
| Add ASM to the baseline, omitting APKT (ablation) | Semi-akmm · 3VT-B, 10% labeled | mAP: 45.6 -> 47.5 (+1.9) | Helped |
| Add APKT to the baseline, omitting ASM (ablation) | Semi-akmm · 3VT-B, 10% labeled | mAP: 45.6 -> 47.9 (+2.3) | Helped |
| Add ASM to the baseline, omitting APKT (ablation) | Semi-akmm · EPV, 10% labeled | mAP: 43.5 -> 46.4 (+2.9) | Helped |
| Add APKT to the baseline, omitting ASM (ablation) | Semi-akmm · EPV, 10% labeled | mAP: 43.5 -> 45.5 (+2.0) | Helped |
Limitations
- Evaluated only on ultrasound datasets (fetal cardiac and early pregnancy views)
- Gains at 1% labels are modest on some structures and some configurations are not best
- Relies on pseudo-label quality and a fixed reliable-ratio hyperparameter C
Experimental setup
Datasets: Fetal Cardiac Structure (FCS) 4C-A FCS 4C-B FCS 3VT-A FCS 3VT-B Early Pregnancy View (EPV)
Benchmarks: FCS 4C-A FCS 4C-B FCS 3VT-A FCS 3VT-B EPV
Metrics: mAP@0.5
Baselines: Supervised Unbiased Teacher Soft Teacher STAC Label Matching Consistent Teacher Semi-DETR Sparse Semi-DETR
Builds on: Mean Teacher Faster R-CNN with FPN Unbiased Teacher Sinkhorn iteration
Closest papers by meaning
- Leveraging Anatomical Consistency for Multi-Object Detection in Ultrasound Images via Source-free Unsupervised Domain Adaptation (AAAI 2025 · similarity 0.89)
- Unsupervised Domain Adaptation for Anatomical Structure Detection in Ultrasound Images (ICML 2024 · similarity 0.89)
- U2-BENCH: Benchmarking Large Vision-Language Models on Ultrasound Understanding (ICLR 2026 · similarity 0.85)
- M3-UDA: A New Benchmark for Unsupervised Domain Adaptive Fetal Cardiac Structure Detection (CVPR 2024 · similarity 0.85)
- Representing Part-Whole Hierarchies in Foundation Models by Learning Localizability Composability and Decomposability from Anatomy via Self Supervision (CVPR 2024 · similarity 0.85)
- MATCH: Multi-faceted Adaptive Topo-Consistency for Semi-Supervised Histopathology Segmentation (NeurIPS 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.