A deep learning-based model for significantly enhancing diagnostic accuracy and efficiency in novice sonographers: development and validation of a YOLOv13-powered real-time abdominal ultrasound anomaly detection system

Abstract Objective To develop and validate a deep learning model for real-time identification and localization of 17 common abdominal anomalies on ultrasound images, providing reliable technical support for primary care sonography. Methods Retrospectively collected 39 530 B-mode ultrasound images from 18 947 patients with abdominal anomalies at the Second Affiliated Hospital of Army Medical University from January 2017 to September 2024). A stratified random sample of 1 891 images formed an independent testing set; the remaining 37 639 images were split 7:1:2 into training (26 258), validation (3 840), and internal test (7 541) sets. After standardized annotation, RetinaNet, Faster R-CNN, RT-DETR, and YOLOv13 were compared on the internal test set using IoU, precision, recall, F1-score, and mAP@50. The optimal model was further evaluated against 6 sonographers (3 novice, 3 intermediate) and novice sonographers assisted by the model on the independent test set. Results YOLOv13 achieved superior performance on the internal testing set: mAP@50 0.653 (95%CI:0.608 to 0.669) versus RetinaNet 0.546 (95%CI:0.511 to 0.563), Faster R-CNN 0.578 (95%CI:0.531 to 0.591), RT-DETR 0.627 (95%CI:0.585 to 0.643); precision 0.633 (95%CI:0.624 to 0.643), recall 0.715 (95%CI:0.704 to 0.725), F1-score 0.671 (95%CI:0.663 to 0.681). On the independent testing set (IoU=0.5), YOLOv13’s F1-score (0.624) exceeded all novice sonographers’ mean F1-score (0.484; P