Research map / Detection and segmentation / Open-set and few-shot detection
M3-UDA: A New Benchmark for Unsupervised Domain Adaptive Fetal Cardiac 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.
Introduces the FCS fetal cardiac structure dataset from two hospitals and M3-UDA, an unsupervised domain adaptive detector combining histogram, sub-structure (angle) and global-structure (graph) matching on an FCOS/ResNet-101 backbone.
Problem
Unsupervised domain adaptive detection of fetal cardiac anatomical structures across hospitals, where domain shifts in ultrasound quality exist and expert annotations in target domains are costly or unavailable.
Approach
Combines pixel-level histogram matching with two topology-aware regularizers: angular adjacency-matrix alignment between sub-structures and graph-based (GNN + cross-attention + Wasserstein) alignment of organ morphology, exploiting fixed anatomical layouts in fetal cardiac views.
Key results
- On FCS hospital 1→2 4CC, M3-UDA reaches 76.81% mAP, 8.26 points above CMT (68.55%).
- On FCS hospital 2→1 4CC, M3-UDA reaches 74.55% mAP, 7.08 points above SIGMA++ (67.47%).
- On CardiacUDA site G→R, M3-UDA reaches 94.95% mAP, 1.97 points above SIGMA (92.98%), but AcroFOD (few-shot) is higher at 96.21%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add HM (ablation) | M3-UDA baseline · FCS 4CC, Hospital 1→2 | mAP: 57.20 -> 69.16 | Helped |
| Add GM after HM (ablation) | M3-UDA · FCS 4CC, Hospital 1→2 | mAP: 69.16 -> 73.17 | Helped |
| Add SM after HM and GM (ablation) | M3-UDA · FCS 4CC, Hospital 1→2 | mAP: 73.17 -> 76.81 | Helped |
| Add HM (ablation) | M3-UDA baseline · FCS 3VT, Hospital 1→2 | mAP: 36.78 -> 51.49 | Helped |
| Add GM after HM (ablation) | M3-UDA · FCS 3VT, Hospital 1→2 | mAP: 51.49 -> 55.32 | Helped |
| Add SM after HM and GM (ablation) | M3-UDA · FCS 3VT, Hospital 1→2 | mAP: 55.32 -> 60.39 | Helped |
| Add HM (ablation) | M3-UDA baseline · FCS 4CC, Hospital 2→1 | mAP: 64.32 -> 67.41 | Helped |
| Add GM after HM (ablation) | M3-UDA · FCS 4CC, Hospital 2→1 | mAP: 67.41 -> 73.23 | Helped |
| Add SM after HM and GM (ablation) | M3-UDA · FCS 4CC, Hospital 2→1 | mAP: 73.23 -> 74.55 | Helped |
| Add HM (ablation) | M3-UDA baseline · FCS 3VT, Hospital 2→1 | mAP: 41.61 -> 44.44 | Helped |
| Add GM after HM (ablation) | M3-UDA · FCS 3VT, Hospital 2→1 | mAP: 44.44 -> 47.70 | Helped |
| Add SM after HM and GM (ablation) | M3-UDA · FCS 3VT, Hospital 2→1 | mAP: 47.70 -> 49.36 | Helped |
Surprising findings
- On 3VT hospital 2→1, the few-shot method AsyFOD outperforms M3-UDA by 8.74 mAP, showing that a few target labels can beat the unsupervised approach.
- On CardiacUDA, the few-shot AcroFOD beats M3-UDA, suggesting the domain gap there is small.
Limitations
- Evaluated on only two hospitals and two fetal cardiac views, with a single backbone and detector.
- Few-shot methods with target labels outperform M3-UDA in some settings (AsyFOD on 3VT 2→1, AcroFOD on CardiacUDA).
- Sub-structure and global matching rely on pseudo-labels that are unreliable early in training, requiring a 30-epoch warm-up for the SM module.
Experimental setup
Datasets: FCS CardiacUDA
Benchmarks: FCS CardiacUDA
Metrics: mAP@0.5
Baselines: AcroFOD AsyFOD ConfMix SIGMA LRA CMT SIGMA++ Without DA
Builds on: FCOS SIGMA SIGMA++ Mean Teacher
Closest papers by meaning
- Unsupervised Domain Adaptation for Anatomical Structure Detection in Ultrasound Images (ICML 2024 · similarity 0.92)
- Leveraging Anatomical Consistency for Multi-Object Detection in Ultrasound Images via Source-free Unsupervised Domain Adaptation (AAAI 2025 · similarity 0.89)
- Anatomical Knowledge Mining and Matching for Semi-supervised Medical Multi-structure Detection (AAAI 2025 · similarity 0.85)
- MAPSeg: Unified Unsupervised Domain Adaptation for Heterogeneous Medical Image Segmentation Based on 3D Masked Autoencoding and Pseudo-Labeling (CVPR 2024 · similarity 0.84)
- FETAL-GAUGE: A BENCHMARK FOR ASSESSING VISION-LANGUAGE MODELS IN FETAL ULTRASOUND (ICLR 2026 · similarity 0.84)
- MCAT: Visual Query-Based Localization of Standard Anatomical Clips in Fetal Ultrasound Videos Using Multi-Tier Class-Aware Token Transformer (AAAI 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.