Artificial intelligence-assisted quality assessment of mid-trimester ultrasound examinations using large vision-language models.

objectivesTo develop and evaluate a deep learning and vision-language model (VLM) system for automated quality assessment of mid-trimester fetal ultrasound examinations.

methodsThis multicenter retrospective study included 22,544 ultrasound planes from 273 internal and 29 external examinations, combined with public data. A SonoNet-based convolutional neural network (CNN) was developed to identify standard planes and key anatomical structures, while a fine-tuned VLM synthesized CNN-derived image descriptions to evaluate overall examination completeness. Six obstetric sonographers with different experience levels participated in a two-phase reader study to assess the effect of AI assistance on diagnostic performance.

resultsThe fine-tuned VLMs demonstrated consistent accuracy in overall examination assessment, achieving F1-scores up to 0.944 on the internal and 0.922 on the external test sets, consistently outperforming both the CNN baseline and zero-shot models. In individual plane analysis, the best-performing VLM reached an overall F1-score of 0.820 and performed robustly across all error categories. In the reader study, AI assistance improved overall diagnostic consistency and recall across all experience levels, while also showing shorter reading times during case review. The system also provided interpretable feedback by identifying missing or low-quality planes.

conclusionsA combined CNN-VLM system enables accurate and interpretable automated quality assessment of fetal ultrasound examinations and enhances reader performance in detecting incomplete or poor-quality scans.