Deep learning meets clinical practice: A You Only Look Once-based framework for accurate and real-time detection of
1Huainan Xinhua Medical Group Xinhua Hospital (Xinhua Hospital Affiliated to Anhui University of Science and Technology), Department of Medical Ultrasound, China.
Abstract:
ObjectiveEarly and accurate detection of carotid vulnerable plaques is essential for preventing ischemic stroke. This study developed an automated deep learning framework using ultrasound images and compared the performance of various You Only Look Once models.MethodsA retrospective multicenter dataset of 1610 carotid ultrasound images from 368 patients was collected from 17 September 2024 to 17 March 2025. Plaques were classified as stable or vulnerable using standardized ultrasound criteria. The dataset was stratified and divided into training, validation, and test sets at a 6:2:2 ratio, with strict patient-level separation to prevent data leakage. Four You Only Look Once models (versions 7, 8, 9, and 10) were trained under identical conditions. Performance was evaluated using mean average precision at various intersection-over-union thresholds as well as precision, recall, and F1 score.ResultsYou Only Look Once version 9 showed the best overall performance, achieving the highest mean average precision at intersection-over-union thresholds of 0.5 and 0.95 in the validation set. Similar results were observed in the test set, with superior detection accuracy for both stable and vulnerable plaques. You Only Look Once version 9 also achieved the highest precision, recall, and F1 score.ConclusionThe You Only Look Once version 9-based framework enables accurate and efficient carotid plaque detection and classification, supporting real-time assessment of plaque vulnerability and the prevention of ischemic stroke.
