A fully automated deep learning (DL) pipeline was developed to detect 5 MRI lesion types in sacroiliac joints (SIJs) in axial spondyloarthritis (axSpA) — bone marrow oedema (BMO), erosions, fat lesions, sclerosis, and ankylosis — trained on the MEASURE 1 trial (n=132) and validated on two independent phase III datasets (PREVENT, n=555; SURPASS, n=414) using paired T1-weighted and STIR sequences scored by the Berlin SIJ method.
The DL model matched expert interreader agreement across all datasets; ankylosis detection was strongest (MEASURE 1 AUC 0.97, SURPASS AUC 0.99; balanced accuracy 0.95–0.97), while BMO showed high AUCs (0.85–0.93) but lower balanced accuracy (0.72–0.74). Performance generalised to independent datasets without additional retraining.
BMO balanced accuracy was modest (0.72–0.74), suggesting reduced reliability for active lesion detection compared to structural lesions. The model was trained and validated exclusively on phase III clinical trial populations, which may limit generalisability to routine clinical or early axSpA cohorts. Consensus-based labels used for training may not fully reflect real-world single-reader variability.
This automated DL tool reliably detects structural SIJ lesions in axSpA and generalises across trial datasets without retraining, supporting its use to improve consistency and scalability of MRI scoring in clinical trials. Clinicians should note that active lesion (BMO) detection is less precise and may still require expert review.
Explore related topics
