Research map / Detection and segmentation / Open-set and few-shot detection

Leveraging Anatomical Consistency for Multi-Object Detection in Ultrasound Images via Source-free Unsupervised Domain Adaptation

AAAI 2025 · Detection and segmentation › Open-set and few-shot detection · approach: Domain-adaptive object detection · 7 citations (OpenAlex)

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 AATS, a source-free unsupervised domain adaptation mean-teacher framework for multi-organ detection in ultrasound, using clustering-based structure refinement and graph-based morphology alignment.

Problem

Source-free unsupervised domain adaptation for multi-object detection in medical ultrasound images, where source data is private and target annotations are unavailable; prior source-free work covered only single-object (polyp) detection or natural images.

Approach

A source-free mean-teacher detector that aligns teacher and student at the organ level via K-means-based soft-assignment (USR) and aligns hypergraph representations of organ features with optimal-transport-style matching (GMA), exploiting consistent anatomical topology across domains.

Key results

• On FH center 1→2, AATS reaches 77.71% mAP versus 73.02% for the best UDA method AT and 65.97% for the AT-SF source-free baseline.
• On FH center 2→1, AATS reaches 89.41% mAP, exceeding AT-SF by 7.46 points (81.95%).
• Ablation on AATS-EPV→EPV shows USR and GMA together add 13.57 mAP over the baseline (62.23 to 75.80).

Measured effects reported in the paper

Surprising findings

• Ablation on FH center 2→1 shows USR adds only +2.28 mAP, while GMA gives +6.96, suggesting the graph alignment contributes most in that setting.
• The method fails to train when the ROI confidence threshold is 0.9, since student proposals rarely meet the threshold.

Limitations

• Relies on a confidence threshold for proposals; at 0.9 training fails to converge.
• Evaluated on only two ultrasound datasets (FH and EPV-related) with a single VGG-16 Faster R-CNN backbone.
• Performance on the EPV-related tasks is lower, attributed to smaller and less distinct structures in early pregnancy.

Experimental setup

Datasets: Fetal Heart (FH) AATS-EPV EPV

Benchmarks: FH center 1→2 FH center 2→1 AATS-EPV→EPV EPV→AATS-EPV

Metrics: mAP@0.5

Baselines: Source Only AT SIGMA LRA CMT SIGMA++ M3-UDA ToMo-UDA AT-SF LODS SMPT++ A2SFOD

Builds on: mean teacher Faster R-CNN Hypergraph Sinkhorn iteration AT

Closest papers by meaning

• Unsupervised Domain Adaptation for Anatomical Structure Detection in Ultrasound Images (ICML 2024 · similarity 0.93)
• Anatomical Knowledge Mining and Matching for Semi-supervised Medical Multi-structure Detection (AAAI 2025 · similarity 0.89)
• M3-UDA: A New Benchmark for Unsupervised Domain Adaptive Fetal Cardiac Structure Detection (CVPR 2024 · similarity 0.89)
• U2-BENCH: Benchmarking Large Vision-Language Models on Ultrasound Understanding (ICLR 2026 · similarity 0.86)
• One-to-Multiple: A Progressive Style Transfer Unsupervised Domain-Adaptive Framework for Kidney Tumor Segmentation (NeurIPS 2024 · similarity 0.86)
• A Unified Degradation-Robust Approach to SSL and UDA for 3D Medical Images (AAAI 2025 · similarity 0.85)