MRI-based interpretable deep learning radiomics for predicting treatment response in axial spondyloarthritis.

BACKGROUND: The clinical efficacy of tumor necrosis factor inhibitors (TNFi) in axial spondyloarthritis (axSpA) is limited by high non-response rates, necessitating accurate pre-treatment stratification. This study aimed to develop and validate MRI-based interpretable deep learning radiomics (DLR) models to predict treatment response to TNFi in patients with axSpA. METHODS: In this prospective study, patients diagnosed with axSpA who underwent sacroiliac joint MRI before TNFi initiation were enrolled. Patients were allocated to training and test sets at an approximate 4:1 ratio. The endpoints were major (ASAS40) and moderate (ASAS20) improvement, as defined by the Assessment of SpondyloArthritis International Society criteria. Deep learning and radiomics features were extracted from MRI. A four-step selection process, including reproducibility analysis, univariable filtering, redundancy reduction, and LASSO regression, was used to distill robust predictors. These were then integrated with clinical data to build DLR-clinical (DLRC) models using L2-regularized logistic regression. RESULTS: A total of 183 patients (mean age, 26.5 ± 9.0 years; 121 [66.1%] males) were analyzed. The DLRC models were validated on the independent test set, with high area under the receiver operating characteristic curve (AUC) values for predicting ASAS40 (0.876, 95%CI: 0.738-0.973) and ASAS20 (0.886, 95%CI: 0.753-0.978) responses. Calibration curves indicated good model agreement, and decision curve analysis (DCA) confirmed their clinical utility. Compared to the clinical-only model, the DLRC models showed significant integrated discrimination improvement (IDI) for ASAS40 and ASAS20 (0.234 and 0.283, all p < 0.05) prediction, respectively. CONCLUSION: MRI-based interpretable DLR models accurately predicted TNFi treatment response in axSpA patients, offering a non-invasive tool to guide personalized therapy.