Real-time Automatic Guidance During Shoulder Ultrasound Scanning with Artificial Intelligence: Reducing Operator Dependency in Rotator Cuff Assessment.
Authors
Affiliations (5)
Affiliations (5)
- Department of Ultrasound, Institute of Ultrasound in Musculoskeletal Sports Medicine, The Affiliated Guangdong Second Provincial General Hospital of Jinan University, Guangzhou, PR China (Y.H., H.W., W.Y., S.L., M.Z., T.B., J.Y., T.W., P.W., H.L.); The Second School of Clinical Medicine, Southern Medical University, Guangzhou, PR China (Y.H., H.L.).
- Department of Ultrasound, The Second People's Hospital of Shenzhen, The First Affiliated Hospital of Shenzhen University, Shenzhen, PR China (Z.L.).
- Department of Ultrasound, Institute of Ultrasound in Musculoskeletal Sports Medicine, The Affiliated Guangdong Second Provincial General Hospital of Jinan University, Guangzhou, PR China (Y.H., H.W., W.Y., S.L., M.Z., T.B., J.Y., T.W., P.W., H.L.).
- Shenzhen Mindray Bio-Medical Electronics Co., Ltd, Shenzhen, PR China (M.Z., S.H., Z.X., K.X., D.Z., S.L.).
- Department of Ultrasound, Institute of Ultrasound in Musculoskeletal Sports Medicine, The Affiliated Guangdong Second Provincial General Hospital of Jinan University, Guangzhou, PR China (Y.H., H.W., W.Y., S.L., M.Z., T.B., J.Y., T.W., P.W., H.L.); The Second School of Clinical Medicine, Southern Medical University, Guangzhou, PR China (Y.H., H.L.). Electronic address: [email protected].
Abstract
This study aimed to develop an artificial intelligence (AI)-guided system for real-time automatic classification and structural recognition of shoulder US planes, in order to standardize the US scanning process. This was a prospective multicenter study. 852 standard plane images and 74,909 frame images from 13,312 shoulder US videos from Center 1 were used for model training/internal testing. 8458 frame images from 480 videos from Center 2 were collected as external test set. Convolutional neural networks were developed. Performance was evaluated using classification metrics (AUC) and structure detection metrics (average precision, AP). The clinical utility was assessed by comparing the examination duration of primary residents with and without AI guidance. AI performance in real-time shoulder US was evaluated by musculoskeletal (MSK) US expert with classification accuracy and target structure detection accuracy. The EfficientNetB2-based AI system, capable of concurrently guiding the acquisition of 15 standard planes and localizing 27 key structures, demonstrated robust performance on an independent external validation set (AUC: 0.99; mAP: 0.89). For junior residents, AI guidance significantly reduced shoulder US examination time by 34% compared to unassisted scans ([10.06 ± 2.74] min vs. [15.26 ± 5.07] min; p = 0.014), achieving efficiency comparable to expert supervision. Independent expert evaluation confirmed the high real-time guidance accuracy of the system. Our AI system standardizes shoulder US acquisition by providing accurate, real-time guidance. It reduces operator dependency for novices and establishes a reproducible foundation for subsequent diagnostic interpretation, representing a critical advancement towards standardized MSK US imaging acquisition.