Artificial intelligence-based decision support for axial spondyloarthritis integrating inflammatory and structural MRI information.
Authors
Affiliations (6)
Affiliations (6)
- Medical Artificial Intelligence (AI) Research Center, Research Institute for Future Medicine, Samsung Medical Center, Seoul, Republic of Korea.
- Department of Digital Health, Samsung Advanced Institute for Health Sciences & Technology (SAIHST), Sungkyunkwan University, Seoul, Republic of Korea.
- Department of Medicine, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Republic of Korea.
- Department of Radiology, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Republic of Korea.
- Department of Data Convergence and Future Medicine, Sungkyunkwan University School of Medicine, Seoul, Republic of Korea.
- Department of Biomedical Software Engineering, The Catholic University of Korea, Bucheon, Republic of Korea.
Abstract
Axial spondyloarthritis (axSpA) assessment on MRI is complicated by interpretation difficulty and by clinical variables that may be incomplete at the time of imaging. We aimed to develop an image-only deep learning framework to approximate the clinical assessment of axSpA from sacroiliac joint (SIJ) MRI, integrating fat-suppressed T1-weighted (FS T1) and short tau inversion recovery (STIR) sequences to capture structural and inflammatory information without clinical variables. This retrospective study included patients with chronic back pain who underwent SIJ MRI from January 2010 to December 2021. The treating rheumatologist's clinical diagnosis served as the reference standard. A two-stage framework integrated FS T1 and STIR features with Stage 1-derived binary bone marrow edema (BME) predictions to generate an axSpA probability. Performance in a test set was compared with the Assessment of SpondyloArthritis International Society (ASAS) definition of a "positive MRI" and the ASAS classification criteria using exact McNemar tests; discordant cases were analyzed separately. Among 291 patients (163 axSpA, 128 non-axSpA), the model achieved an AUROC of 0.9412, accuracy of 86.0%, sensitivity of 87.5%, and specificity of 84.0% in the test set (n=57). No significant paired differences in accuracy, sensitivity, or specificity were detected between the model and either ASAS comparator. In an exploratory analysis of 10 cases discordant between ASAS "positive MRI" assessment and clinical diagnosis, the model correctly classified 7 (70.0%). The model showed similar performance to the ASAS-based comparators in the test set without clinical variables, suggesting potential value as an image-based decision-support tool. External multicenter validation is needed to confirm generalizability.