Development of a Deep Learning Model for Automated Detection of Calcium Pyrophosphate Deposition on Hand Radiographs
Elisabeth Rosoux (1), Thomas Hügle (2), Tobias Manigold (3), Guillaume Fahrni (1), Deborah Markham (4), Fabio Becce (1)
Affiliation(s):
1. Lausanne University Hospital (CHUV) and University of Lausanne, Radiology, Lausanne, Switzerland,
2. Lausanne University Hospital (CHUV) and University of Lausanne, Rheumatology, Lausanne, Switzerland,
3. Inselspital, Bern University Hospital and University of Bern, Rheumatology, Bern, Switzerland,
4. On behalf of Lausanne University Hospital (CHUV) and University of Lausanne, Rheumatology, Lausanne, Switzerland
Background: Identifying the presence of calcium pyrophosphate deposition (CPPD) on large radiograph datasets for epidemiological studies is labor intensive and limited by moderate inter- and intra-observer reliability. Here, we aimed to develop a deep learning approach for automatically and reliably detecting CPPD on hand radiographs in patients with various rheumatic diseases including gout and rheumatoid arthritis, focusing on the triangular fibrocartilage complex (TFCC) and 2nd and 3rd metacarpophalangeal joints (MCP-2, MCP-3) according to the 2023 ACR/EULAR classification criteria.
Methods: Two radiologists independently labeled a dataset of 926 hand radiographs, yielding 319 CPPD positive and 607 CPPD negative cases across the three sites of interest. CPPD presence was then predicted using a convolutional neural network. The model performance was assessed using the area under the receiver operating characteristic (AUROC), sensitivity, specificity, and balanced accuracy, with heatmaps (Grad-CAM) aiding in case discrimination.
Results: The algorithm for combined TFCC, MCP-2, and MCP-3 classification showed robust performance with an AUROC of 0.89 and a balanced accuracy of 0.79 (sensitivity of 0.94 and specificity of 0.65). The TFCC-alone, MCP-2-alone, and MCP-3-alone models also showed a good performance with an AUROC of 0.85, 0.81, and 0.86 and a slightly lower balanced accuracy of 0.76, 0.67, and 0.70, respectively. Heatmap analysis revealed activation in the regions of interest for positive cases (true and false positives), but unexpected highlights were encountered possibly due to correlated features in different hand regions.
Conclusion: This study highlights the potential of an automated model for CPPD detection on hand radiographs with high performance in the combined and TFCC-alone models. MCP-2-alone and MCP-3-alone models showed lower accuracy due to class imbalance. This algorithm could, for example, be used to screen large clinical databases or electronic medical records for CPPD cases. Future work includes dataset expansion, threshold optimization, preprocessing refinement, and validation with external datasets.
Disclosure of interest: Elisabeth Rosoux: None declared; Thomas Hügle: Roche, Novartis, BMS, GSK, Janssen, Galapagos, Atreon, Vtuls; Tobias Manigold: None declared; Guillaume Fahrni: None declared; Deborah Markham: None declared; Fabio Becce: Horizon Therapeutics.
