Key Takeaway
A coordinated team of specialized AI models can analyze fetal ultrasound videos end-to-end—extracting key frames, running plane-specific measurements, and producing structured clinical reports—while outperforming standalone vision models across multiple clinical tasks.
ON THIS PAGE
What They Found
FetalAgents organizes a central coordinator (a large language model) and multiple specialized vision experts to split responsibilities: identify the correct ultrasound plane, run classification and segmentation, compute biometric measurements, and synthesize results into a clinical report. It adds an automated keyframe extractor to summarize continuous video scans into diagnostically relevant frames, so the system handles video workflows rather than just single images. Across multi-center external datasets, the coordinated system achieved the best results compared with task-specific vision models and general multimodal language-vision models on a range of clinically relevant tasks. The architecture forces experts to return structured outputs (JSON), reducing risky free-form predictions and making downstream reporting more reliable. Orchestrator-Worker Pattern
Explore evaluation patternsSee how to apply these findings
Data Highlights
1Top performance on eight distinct clinical tasks (plane ID, segmentation, biometry, and gestational age among others) when evaluated against specialized vision models and multimodal language-vision baselines.
2Validated gestational-age and head circumference work on a 511-case subset from an established dataset used for external testing.
3Tested across multiple external, multi-center datasets (including African multi-center data) and a dedicated keyframe dataset to demonstrate end-to-end video summarization and multi-plane reporting.
Implications
AI engineers building clinical imaging pipelines can adopt the coordinator/expert/summarizer pattern to combine best-in-class vision models into one workflow without rewriting each model. Clinical AI teams, medical device vendors, and hospital IT leads can use this approach to automate reporting, reduce dependence on scarce expert sonographers, and improve consistency across sites. Research Agents
Key Figures
Fig 1: Figure 1: System diagram of FetalAgents. The framework comprises three distinct modules: (1) Agentic Clinical Workflow, where the Coordinator interprets queries and dispatches tasks; (2) Image Analysis, where specialized experts execute specific vision tasks (classification, segmentation, biometry); and (3) Video Summary, which aggregates multi-plane keyframes and temporal findings into a comprehensive report.
Fig 2: Figure 2: Qualitative evaluation of clinical reporting. (a) Image Caption Generation: FetalAgents correctly identifies the standard plane and performs biometry to generate a report concordant with human experts. (b) Video Summarization: The system autonomously extracts multi-plane keyframes from a continuous video stream and cross-references ultrasound findings with patient metadata (LMP) for clinical consistency.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreKeep in Mind
Results are based on retrospective datasets and multi-center external validation but not yet on prospective clinical trials, so real-world performance and clinical impact remain to be proven. The system depends on a high-quality central language model for orchestration; its behavior and safety need monitoring and governance to avoid edge-case failures. Integration into clinical workflows will require attention to data privacy, local imaging protocols, and human-in-the-loop checks before autonomous use. Model Risk Management
Deep Dive
FetalAgents groups specialized vision models and a large language model into a three-part workflow: a Coordinator that interprets the user query and identifies the ultrasound plane, modular Expert agents that wrap task-specific vision models and fuse multiple model outputs deterministically, and a Summarizer that composes structured findings into a human-readable clinical report. Experts return only structured JSON results, which reduces the risk of misleading free-form outputs. For video scans, an automated keyframe extraction module distills continuous recordings into diagnostically relevant frames that the experts analyze, allowing end-to-end video summarization and cross-plane reasoning. The system was trained on multiple public fetal ultrasound datasets for plane classification, segmentation, keyframe identification, and biometric tasks, then evaluated on several independent external datasets including multi-center African data and a 511-case subset for gestational-age testing. Across eight clinical tasks, FetalAgents produced the strongest results compared to single-task vision models, ultrasound-focused foundation models, and general multimodal models. Practical implications include standardized, automated reporting that mirrors how sonographers reason across planes and tasks, but prospective clinical evaluation, robust monitoring, and integration safeguards are needed before deployment in routine care. Model Context Protocol (MCP) Pattern Multi-Agent Video & Podcast Production
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
ArXiv preprint with no listed affiliations and low h-index values (mostly <5). Multiple authors but no institutional backing or top venues—fits 'emerging / limited info'.