Background: Magnetic resonance imaging (MRI) is of value in patients suspected of RA. A disadvantage of MRI is that its visual assessment requires expertise and is time-consuming.

Objectives: To facilitate MR image interpretation, the A utomatic D eep learning-based M RI analysis of I nflammatory signs in RA ( ADMIRA ) system was developed using 1691 MRI scans of hands and feet, with the visual RA MRI Scoring method as reference. We here studied its validity.

Methods: Inflammation (total, synovitis, tenosynovitis, and osteitis) of MCP, wrist, and MTP joints on 1.5 Tesla contrast-enhanced MRIs of 180 early arthritis patients was quantified using ADMIRA. Content validity was evaluated by assessing agreement between ADMIRA scores and mean scores of two visual readers (reference). Construct validity was evaluated by assessing associations between ADMIRA scores and clinical parameters, testing previously observed associations with visual MRI scores and age, disability, hand function. To gain insight into ADMIRA assessment, outliers were visually inspected and heatmaps indicating the location of ADMIRA assessment reviewed.

Results: ADMIRA total inflammation, synovitis, tenosynovitis, and osteitis scores demonstrated good to excellent agreement with visual scores: Pearson correlations were 0.94, 0.94, 0.94, and 0.79 respectively, intraclass correlations 0.96, 0.97, 0.97, and 0.86 respectively. Older age associated with higher ADMIRA scores (β=1.04/year, 95% CI=1.03-1.05), similar to visual scores. Likewise, patients with higher ADMIRA scores had more physical disabilities (β=0.012/point, 95% CI=0.006-0.018). Patients with difficulties dressing had higher ADMIRA scores on hand MRIs (also higher visual scores; p=0.002). Outliers (n=10 scans) were mostly overestimations of osteitis. Review of heatmaps showed that ADMIRA-scored lesions corresponded to visually-scored lesions.

Conclusions: ADMIRA demonstrates good validity. This AI innovation shows promise as a time-efficient method for MRI interpretation in rheumatology.

REFERENCES: NIL.

Acknowledgments: NIL.

Disclosure of Interests: None declared.