Automated deep learning for real-time focal liver lesions detection in ultrasound videos a multicenter study

Early detection of focal liver lesions (FLLs) is crucial for clinical practice, but ultrasound performance heavily depends on operator experience. We developed Auto-DFLLs, an automated deep learning model based on ResNet and FPN architectures to detect FLLs in ultrasound videos. It was trained and validated on 5258 prospectively collected videos from three hospitals. On internal validation, Auto-DFLLs achieved an AP50 (average precision at IoU = 50%) of 0.7772, Pr70 (precision at 70% recall) of 0.7967, and FP70 (false positives at 70% recall) of 3.4688. Validation study showed that Auto-DFLLs significantly improved junior sonographers' detection (AFROC-AUC: 79.52 vs. 71.55, P = 0.021) and enhanced senior sonographers' performance (AFROC-AUC: 78.64 vs. 74.57, P = 0.0366), especially for small lesions (< 10 mm, P = 0.034). Auto-DFLLs maintained stable detection across different lesion size, echogenicity, location, and ultrasound equipment from different manufacturers. Auto-DFLLs reduces operator-dependent variability and offers a reliable assistive tool for real-time FLLs screening, particularly valuable in resource-limited areas.