Deep learning-based quantification of knee effusion-synovitis volume on MRI – Technique development and validation

OBJECTIVE To develop and validate a deep learning (DL) model for automatic quantification of knee effusion-synovitis volume (ESV) on MRI, assess correlations of ESV with semiquantitative effusion-synovitis (sqES) scores, and compare associations of ESV and sqES with MRI features and symptoms of knee osteoarthritis. DESIGN A DL model was developed to quantify ESV on baseline right-knee MRI from the Osteoarthritis Initiative (n=4,698), which was trained and tested using manual segmentations from 101 randomly selected knees. Dice coefficients were used to quantify segmentation performance. Spearman correlations were computed between ESV and sqES scores from Whole-Organ Magnetic Resonance Imaging Score (WORMS ES ) and MRI Osteoarthritis Knee Score (MOAKS ES ). Paired differences in standardized β coefficients (Δβ std ) from linear models were used to compare associations of ESV with tissue-specific WORMS scores and the total score of the Western Ontario McMaster Universities Arthritis Index (WOMAC). RESULTS The DL model achieved a mean Dice coefficient of 0.79 (95% Confidence Interval [CI] 0.71-0.86) on the test set. ESV showed moderate correlations with WORMS ES (ρ=0.50, 95% CI 0.47-0.53) and MOAKS ES (ρ=0.65, 95% CI 0.63-0.67) and demonstrated larger effect sizes for associations with WORMS features (minimum Δβ std =0.03 [95% CI 0.00-0.07]) and knee OA symptoms (minimum Δβ std =0.03 [95% CI 0.00-0.07]) than sqES scores. CONCLUSIONS A DL model for automated quantification of knee ESV on MRI was developed and validated. While our results demonstrate the potential of ESV as a scalable imaging biomarker in osteoarthritis research, validation in independent cohorts is necessary to confirm its utility.