Abstract: Total shoulder arthroplasty is a standard restorative procedure practiced by orthopedists to diagnose shoulder arthritis in which a prosthesis replaces the whole joint or a part of the joint. It is often challenging for doctors to identify the exact model and manufacturer of the prosthesis when it is unknown. This paper proposes a transfer learning-based class imbalance-aware prosthesis detection method to detect the implant’s manufacturer automatically from shoulder X-ray images. The framework of the method proposes a novel training approach and a new set of batch-normalization, dropout, and fully convolutional layers in the head network. It employs cyclical learning rates and weighting-based loss calculation mechanism. These modifcations aid in faster convergence, avoid local-minima stagnation, and remove the training bias caused by imbalanced dataset. The proposed method is evaluated using seven well-known pre-trained models of VGGNet, ResNet, and DenseNet families. Experimentation is performed on a shoulder implant benchmark dataset consisting of 597 shoulder X-ray images. The proposed method improves the classifcation performance of all pre-trained models by 10–12%. The DenseNet-201-based variant has achieved the highest classifcation accuracy of 89.5%, which is 10% higher than existing methods. Further, to validate and generalize the proposed method, the existing baseline dataset is supplemented to six classes, including samples of two more implant manufacturers. Experimental results have shown average accuracy of 86.7% for the extended dataset and show the preeminence of the proposed method.

Key words: Biomedical engineering , · Artifcial intelligence , · Total shoulder arthroplasty , · Prosthesis identifcation , · Biomedical image classifcation