Dynamic ultrasound motion metrics combined with deep learning for clinical differentiation of subacromial impingement syndrome.

purposeThis prospective study evaluated the diagnostic performance of deep learning models in predicting subacromial impingement syndrome (SIS) during dynamic shoulder ultrasonography, comparing a faster region-based convolutional neural network (Faster R-CNN) with a self-transfer learning CNN (STL-CNN). The utility of integrating a one-dimensional convolutional neural network (1D-CNN) for SIS classification was also examined. MATERIALS AND

methodsParticipants underwent shoulder abduction and adduction during ultrasound imaging. Faster R-CNN and STL-CNN were trained to localize anatomical landmarks, and the better-performing model was paired with a 1D-CNN to differentiate SIS. Subacromial motion metrics-including acromiohumeral distance (AHD), horizontal AHD (hAHD), and vertical AHD (vAHD)-were used as classification features.

resultsAmong 59 SIS patients and 59 controls, Faster R-CNN demonstrated significantly lower mean distance errors than STL-CNN for the greater tuberosity (0.1302 cm vs. 0.4835 cm, p = 0.03) and lateral acromion (0.0585 cm vs. 0.2634 cm, p = 0.02). vAHD yielded superior discrimination compared with AHD and hAHD. Using Faster R-CNN-derived trajectories, the 1D-CNN achieved 94% accuracy for vAHD, surpassing results based on ground-truth annotations.

conclusionFaster R-CNN enabled more accurate landmark localization than STL-CNN, while vAHD enhanced SIS identification. Combining faster R-CNN with a 1D-CNN demonstrated high diagnostic accuracy, underscoring the potential of deep learning for automated SIS assessment during dynamic ultrasonography. However, the current workflow requires offline video analysis, and future advances should focus on real-time implementation and improved generalizability.