A fully automated deep learning MRI pipeline can reliably detect sacroiliac joint lesions in axial spondyloarthritis, matching expert readers and offering high accuracy for structural lesions like ankylosis, with potential to standardize imaging outcomes in trials and care.

The system operates in two stages: first delineating left and right SI joints, then classifying five lesion types (bone marrow edema is active; erosions, fat lesions, sclerosis, and ankylosis are structural) using paired T1-weighted and STIR sequences and the Berlin SIJ scoring method.

Training used consensus labels from MEASURE 1 to address interreader variability, with validation on two independent phase 3 datasets, PREVENT and SURPASS, to test robustness.

The study discloses funding from Novartis and notes some authors’ financial ties, and it states AI-assisted editorial tools were used with human review.

Led by Amir Jamaludin at the University of Oxford, the study was published online September 5, 2026 in the Annals of the Rheumatic Diseases.

Limitations include binary lesion assessment, absence of an erosion-sensitive sequence, reliance on sponsor data and consensus standards, and a disease-enriched axSpA cohort that may limit differentiation of axSpA-specific lesions from nonspecific findings.

Performance metrics are outstanding, with ankylosis AUCs of 0.97 (MEASURE 1) and 0.99 (SURPASS), and strong results for other lesions (erosions, fat lesions, BMO, sclerosis) across datasets.

The AI system’s workflow integrates SIJ segmentation with lesion typing and employs the Berlin scoring framework to quantify five lesion types across both SI joints.

Balanced accuracy for automated detection rivaled interreader agreement across all datasets, underscoring the system’s robustness in mirroring expert evaluation.