1. Conferences Overview
    2. Deutscher Kongress für Orthopädie und Unfallchirurgie 2026 (DKOU 2026)
    3. Detecting glenoid loosening after total shoulder arthroplasty from plain radiographs using a convolutional neural network
Logo

German Congress of Orthopaedics and Traumatology 2026 (DKOU 2026)

Deutsche Gesellschaft für Orthopädie und Unfallchirurgie (DGOU), Deutsche Gesellschaft für Orthopädie und Orthopädische Chirurgie (DGOOC), Deutsche Gesellschaft für Unfallchirurgie (DGU), Berufsverband für Orthopädie und Unfallchirurgie (BVOU)
20.-23.10.2026
Berlin

Meeting Abstract

Detecting glenoid loosening after total shoulder arthroplasty from plain radiographs using a convolutional neural network

Tim Schneller - Schulthess Klinik, Zürich, Schweiz
Lorenzo Menarbin - Schulthess Klinik, Zürich, Schweiz
Florian Freislederer - Schulthess Klinik, Zürich, Schweiz
Asimina Lazaridou - Schulthess Klinik, Zürich, Schweiz; Department of Anesthesiology, Brigham & Women's Hospital and Harvard Medical School, Boston, USA
Markus Scheibel - Schulthess Klinik, Zürich, Schweiz; Center for Musculoskeletal Surgery, Charité-Universitätsmedizin, Berlin, Deutschland

Text

Objectives and questions: Implant loosening is a common complication after total shoulder arthroplasty. Glenoid loosening specifically can lead to significant pain and function loss and frequently necessitates revision surgery. Machine Learning is a novel technology with high potential that shows large potential and increasing adoption rates in the field of orthopedic surgery. The aim of this study was to develop and validate a machine learning model to detect implant loosening on plain radiographs.

Material and methods: This retrospective study included A/P radiographs following total shoulder arthroplasty. Each patient was revised due to intraoperatively confirmed glenoid implant loosening. The last radiograph before the revision was labeled as the positive ground truth (“loose”) and the 6-week postoperative radiograph was marked as the negative ground truth (“stable”). A train/validation/test split of 60/20/20 was employed. After splitting, several data augmentation techniques were employed, such as rotating, flipping, cropping or changing brightness and contrast. The radiographs were then used to retrain a previously trained model (ResNet50). Performance metrics included accuracy, sensitivity, specificity and the area under the curve.

Results: 84 Patients were included in the study, which were 44% male, with an average age at surgery of 61.8 (11.1). Indications for arthroplasty primarily included primary osteoarthritis (64%) and revision arthroplasty (22%). Implants used included mainly Promos (14%) and Eclipse (62%). Mean time to revision was 6.4 (SD: 3.5) years. Our model reached an accuracy of 91.2%, a sensitivity of 87.5%, a specificity of 94.4% and an AUC of 0.91. Saliency maps indicate that our model consistently focuses on the glenoid area for correctly identified cases.

Discussion and conclusions: Our model demonstrated excellent performance in detecting glenoid implant loosening from plain radiographs. However, future work should focus less on detecting and more on predicting the occurrence of loosening before symptoms occur, allowing timely intervention by healthcare providers. Limitations include the lack of external validation and data sparsity.