TOPLINE

A fully automated deep learning system detected active inflammatory and structural lesions in MRI scans of sacroiliac joints (SIJs) in patients with axial spondyloarthritis (axSpA), with performance comparable to expert readers and higher accuracy for structural lesions, especially ankylosis.

METHODOLOGY

• Researchers developed a two-stage automated pipeline using deep learning to delineate left and right SIJs and then detect five MRI-defined lesion types: one active lesion (bone marrow edema [BMO]) and four structural lesions (erosions, fat lesions, sclerosis, and ankylosis).
• The system was trained on data from the MEASURE 1 trial including 132 patients with axSpA, using consensus-based labels from multiple expert readers to address interreader variability.
• Validation was performed on two independent phase 3 clinical trial datasets: PREVENT (n = 555) and SURPASS (n = 414).
• Lesions were assessed at the quadrant or joint level using paired T1-weighted and Short Tau Inversion Recovery MRI sequences, and the Berlin SIJ scoring method.
• Performance metrics included area under the curve (AUC), balanced accuracy, sensitivity, specificity, and kappa statistics, with results compared against expert reader evaluations.

TAKEAWAY

• Performance of the AI system was strongest in structural lesion detection, particularly in ankylosis (in MEASURE 1 and SURPASS: AUC, 0.97 and 0.99, respectively; balanced accuracy, 0.95 and 0.97, respectively). 
• Detection of erosions (AUC: 0.87, 0.89, and 0.96 in MEASURE 1, PREVENT, and SURPASS, respectively), fat lesions (AUC: 0.89, 0.92, and 0.95 in MEASURE 1, PREVENT, and SURPASS, respectively), BMO (AUC: 0.85, 0.91, and 0.93 in MEASURE 1, PREVENT, and SURPASS, respectively), and sclerosis (AUC: 0.89, 0.88, and 0.98 in MEASURE 1, PREVENT, and SURPASS, respectively) was robust across all datasets. 
• Balanced accuracies for the automated detection of all lesions were comparable to interreader balanced accuracies across datasets.

IN PRACTICE

"[T]his study demonstrates that a fully automated MRI pipeline can reliably detect both active and structural SIJ lesions in axSpA in a cross-sectional setting with performance comparable with expert interreader agreement, generalisable across multiple phase III trial datasets. These findings support the potential of automated MRI analysis to enhance the consistency and scalability of imaging outcomes in clinical trials and observational studies, and provide a foundation for future evaluation in routine clinical care," the authors wrote.

SOURCE

The study was led by Amir Jamaludin, University of Oxford, Oxford, UK. It was published online on September 5, 2026, in Annals of the Rheumatic Diseases.

LIMITATIONS

The study was limited by binary lesion assessment, lack of an erosion-sensitive sequence, and the use of standardized, single-sponsor trial data and consensus-based reference standards. Additionally, the study included a disease-enriched axSpA cohort, which limited its ability to distinguish axSpA-specific lesions from nonspecific findings in populations assessed for potential axSpA. 

DISCLOSURES

The study was funded by Novartis. Several authors reported having financial or employment ties with various organizations including Novartis. Details are provided in the original article.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.