Artificial intelligence to automatically identify reverse shoulder arthroplasty implant brands on postoperative radiographs including uncertainty quantification

Identifying the brand of reverse shoulder arthroplasty (rTSA) implanted is key in the postoperative evaluation of patients, a process that can be time-consuming and prone to error. This is particularly relevant when rTSA is considered for shoulders where the implant brand is not available or accessible in the patient record. Identification of the prosthesis model by subjective assessment of preoperative radiographs is possible, but challenging. To assist surgeons in accurately and efficiently identifying prosthetic brands on radiographs, this study aimed to develop a computer vision artificial intelligence (AI) algorithm capable of classifying 8 commonly used rTSA prosthesis brands. Using a previously developed deep learning (DL) shoulder classifier, 5,256 anteroposterior and axillary shoulder radiographs corresponding to 1,368 shoulders that had undergone rTSA surgery were identified. Four human observers reviewed and labeled the radiographs according to 1 of 8 rTSA prosthesis manufacturers and model brands. All labeled radiographs were divided into developmental and testing sets, with stratified splitting to maintain consistent ratios of each prosthesis model across both sets. A DL classification algorithm based on EfficientNet was then trained on the developmental set using a five-fold cross-validation approach to classify these 8 prosthesis models, with additional conformal prediction applied for uncertainty quantification. The trained DL algorithm and uncertainty quantification were evaluated on the testing set by calculating accuracy, F1 score, efficiency, and coverage. The trained AI algorithm achieved an average accuracy of 0.958 and an average F1 score of 0.946. After applying conformal prediction, average coverage was 0.978, with an average efficiency of 0.72. Integrated maps generated for the final model showed weighted pixels around the key features of the implant when making the prediction, including center screw, glenosphere, humeral tray, and the tip of the stem. The final model was able to classify images at a rate of approximately 45 images per second. Our results demonstrate the ability of DL-based models to accurately and efficiently identify commonly used implants used in rTSA. This automated identification pipeline can particularly improve the clinical workflow for revision rTSA and has broader implications for the curation of radiographic registries.