Answers are grounded in this paper and cite their sources.
OBJECTIVE: We aimed to develop and evaluate artificial intelligence (AI)-based methods for automated grading of calcium pyrophosphate deposition (CPPD) in ultrasound (US) images according to the Outcome Measures in Rheumatology (OMERACT) US scoring system. METHODS: Within the OMERACT CrystaL Artificial Intelligence Monitoring (CLAIM) project, two deep learning strategies were explored using knee meniscus US images graded from 0 to 3. Image-level classification was evaluated using convolutional neural networks (CNN) trained directly on cropped meniscus images (n = 446). A semantic segmentation approach was developed using a subset of 279 images with manually annotated masks identifying meniscus and deposits. Models were evaluated using an 80/20 training-validation split and stratified 5-fold cross-validation. Segmentation was assessed using Intersection over Union (IoU). For automated grading, an area ratio (AR), representing CPPD pixels relative to meniscus and CPPD pixels, was applied to predicted masks. RESULTS: Image-level classification achieved an accuracy of 0.69 in the single-split analysis and a mean accuracy of 0.59 (SD 0.10) in 5-fold cross-validation. End-to-end AR grading achieved accuracies of 0.64 and 0.61 (SD 0.10), with a grade 3 recall of 0.25 in the single split. AR grading on manually annotated reference masks achieved an accuracy of 0.82. The best-performing segmentation model achieved mean IoUs of 0.86 for background, 0.60 for meniscus, and 0.37 for CPPD. CONCLUSIONS: These findings provide a methodological basis for automated CPPD US grading, but performance remains far from clinical applicability, highlighting the need for larger datasets and improved segmentation.
No takes yet. Share an insight, caveat, or question.
Fabbro et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: