Research map / Label-efficient and robust learning / Domain adaptation and generalization

Unsupervised Domain Adaptation for Anatomical Structure Detection in Ultrasound Images

ICML 2024 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Geometric alignment and synthetic-to-real adaptation

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 ToMo-UDA, an unsupervised domain adaptation detector for fetal heart and head anatomical structures using topology and morphology knowledge transfer, and releases the multi-center FUSH2 ultrasound benchmark.

Problem

Cross-institution domain shift degrades anatomical structure detection in fetal ultrasound, where structures are dense and overlapping and target annotations are costly.

Approach

Exploits consistent anatomical topology (relative positions of substructures) and morphology (per-substructure appearance) across domains: a GNN over substructure centroids aligned by optimal transport (TKT), plus morph-aware GNN node alignment with a numerical-equilibrium step on feature statistics (MKT).

Key results

• On FUSH2 heart, ToMo-UDA reaches 71.33% mAP on center 1→2 and 88.71% on center 2→1, outperforming all listed UDA baselines (e.g., CMT 60.47%, SIGMA++ 57.83% on 1→2).
• On FUSH2 head, ToMo-UDA achieves 81.30% mAP on center 1→2 versus 61.58% Source Only.
• On MMWHS, ToMo-UDA outperforms the second-best method by 5.83% mAP (CT→MRI) and 7.32% mAP (MRI→CT).

Measured effects reported in the paper

Surprising findings

• On heart center 2→1, ToMo-UDA's adapted detector surpasses the target-only supervised model (88.71% vs 84.78% mAP).
• Several few-shot UDA methods (AcroFOD, AsyFOD) perform below Source Only on some heart settings, which the authors attribute to unrepresentative sample selection.

Limitations

• Performance on head center 2→1 remains well below target-only, attributed to lower resolution of center 2 images.
• Few-shot UDA baselines can underperform Source Only due to domain bias from sample selection.

Experimental setup

Datasets: FUSH2 MMWHS CardiacUDA

Benchmarks: FUSH2 MMWHS CardiacUDA

Metrics: mAP@0.5

Baselines: Source Only AcroFOD AsyFOD ConfMix SIGMA LRA CMT SIGMA++ AT

Builds on: FCOS Feature Pyramid Network Sinkhorn optimal transport GCN

Closest papers by meaning

• Leveraging Anatomical Consistency for Multi-Object Detection in Ultrasound Images via Source-free Unsupervised Domain Adaptation (AAAI 2025 · similarity 0.93)
• M3-UDA: A New Benchmark for Unsupervised Domain Adaptive Fetal Cardiac Structure Detection (CVPR 2024 · similarity 0.92)
• Anatomical Knowledge Mining and Matching for Semi-supervised Medical Multi-structure Detection (AAAI 2025 · similarity 0.89)
• U2-BENCH: Benchmarking Large Vision-Language Models on Ultrasound Understanding (ICLR 2026 · similarity 0.85)
• 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)
• One-to-Multiple: A Progressive Style Transfer Unsupervised Domain-Adaptive Framework for Kidney Tumor Segmentation (NeurIPS 2024 · similarity 0.84)