Automatic determination of the resection plane for shoulder arthroplasty in arthritic humeri: a deep learning model.

Abstract

The humeral head resection plane relies on the anatomic neck as a reference for shoulder arthroplasty planning and accurate intraoperative placement of the humeral component. However, osteophytes and deformity associated with osteoarthritis often obscures this landmark increasing preoperative planning time and complicating placement of the humeral component. To date, no automated method for identifying the resection plane in arthritic humeri exists. Two orthopedic surgeons determined the resection plane of 62 3D-models of the humerus derived from computed tomography scans by digitizing along the anatomic neck and then fitting the resection plane to those points. A deep learning model was then trained on 80% of the humeri and their corresponding resections. Testing was performed on the remaining 20% by inputting 3D humerus models into the trained deep learning model and assessing the predicted resection plane's deviation from the surgeon-defined resection plane. Evaluation metrics included the mean Euclidean distance between the centroids of both planes and the angular error between their normal vectors to quantify model accuracy. For arthritic humeri, the predicted resection plane when compared to the surgeon selected plane had mean absolute errors of 1.4 ± 0.7 mm for the centroid and 3.9 ± 1.6° for the normal vector. In nonarthritic humeri the mean absolute errors were 0.3 ± 0.2 mm and 3.1 ± 1.6, respectively. Our deep learning model achieves state-of-the-art accuracy for predicting the location of the resection plane. The model maintains robust performance in arthritic shoulders, enabling further automation of shoulder arthroplasty planning in more complex arthritic cases.