Xiaotian Hu, Junwei Huang, Mingxuan Liu, Kasidit Anmahapong, Yifei Chen, Yitong Luo, Yiming Huang, Xuguang Bai, Zihan Li, Yi Liao, Haibo Qu, Qiyuan Tian
6 min
Abstract
Fetal ultrasound (US) is the primary imaging modality for prenatal screening, yet its interpretation relies heavily on the expertise of the clinician. Despite advances in deep learning and foundation models, existing automated tools for fetal US analysis struggle to balance task-specific accuracy with the whole-process versatility required to support end-to-end clinical workflows. To address these limitations, we propose FetalAgents, the first multi-agent system for comprehensive fetal US analysis. Through a lightweight, agentic coordination framework, FetalAgents dynamically orchestrates specialized vision experts to maximize performance across diagnosis, measurement, and segmentation. Furthermore, FetalAgents advances beyond static image analysis by supporting end-to-end video stream summarization, where keyframes are automatically identified across multiple anatomical planes, analyzed by coordinated experts, and synthesized with patient metadata into a structured clinical report. Extensive multi-center external evaluations across eight clinical tasks demonstrate that FetalAgents consistently delivers the most robust and accurate performance when compared against specialized models and multimodal large language models (MLLMs), ultimately providing an auditable, workflow-aligned solution for fetal ultrasound analysis and reporting.
Alex: So it's not just better at individual bits, but the coordination makes the whole chain stronger. What kinds of specialist tools are they pulling from?
Sam: The experts wrap specific vision models—AI systems tuned to "see" ultrasound patterns. For plane classification, they combine models like one that understands ultrasound language from vast training and others for fine details. Segmentation experts outline structures like the abdomen. Biometry pulls measurements, like abdominal circumference via ellipse fits on those outlines. All outputs stay in structured formats to keep things auditable and hallucination-free—just numbers and facts.
Alex: Got it—each expert handles a narrow skill, like outlining the stomach or estimating gestational age, which is how far along the pregnancy is from measurements. But real scans are videos, not still photos—how does it deal with that continuous stream?
Sam: Videos are sequences of frames from the probe moving over the belly, so sonographers pick key moments manually. FetalAgents automates that: first, a tuned model spots diagnostic keyframes across planes like femur or thorax. Then the coordinator analyzes each, runs experts frame-by-frame, and the summarizer aggregates into a report with patient details, like checking measurements against growth charts. This mirrors the full workflow, from raw video to typed notes with consistency flags.
Alex: Oh—so that's why it's workflow-aligned; it handles the messiness of real exams, not just perfect lab images. And the paper tests this on actual external data, right?
Sam: Yes, they trained on public datasets then validated on independent multi-center ones, like African clinics, for out-of-distribution robustness. Across eight tasks—plane ID, segmentations, biometry—FetalAgents topped specialized models and large multimodal language models. For instance, on abdomen segmentation, it reached a notably higher Dice score—a measure of outline match—than the next best.
Alex: That's a clear improvement, but does the ensembling always win, or are there catches?
Sam: Ensembling boosts robustness by fusing strengths, like correcting one model's edge slips with another's, but it relies on those pre-trained vision models working well together. The paper notes performance holds on tested tasks, yet may need tweaks for rare anomalies or new protocols—prospective trials in clinics are next.
Alex: Makes sense; it's a meaningful step for standardizing scans where experts are scarce. So zooming out, what could this mean for prenatal care down the line?
Sam: In low-resource areas, it could cut delays by automating reports from any ultrasound video, easing sonographer burdens and spotting issues earlier. The structured outputs aid audits too, aligning with global standards like WHO charts. Still, as the authors stress, real-world deployment needs those clinical trials to confirm safety across diverse cases.
Alex: Well put, Sam. This shows how agent coordination can tackle complex medical imaging without overpromising. Thanks for breaking it down—listeners, check the shownotes for the paper. Thanks for listening to ResearchPod.