Deep Learning Modeling to Differentiate Multiple Sclerosis From MOG Antibody–Associated Disease

Comprehensive Summary

In this study, Cortese et al. examined whether deep learning (DL) could improve clinicians’ ability to differentiate between multiple sclerosis (MS) and myelin oligodendrocyte glycoprotein antibody–associated disease (MOGAD). These two conditions can appear similar on imaging, but they require vastly different treatment approaches. The researchers collected 406 MRI scans from adults with MS and MOGAD across 19 centers and compared the accuracy of diagnoses between a previously developed MRI algorithm and a new algorithm with a DL-based classifier. The original MRI algorithm achieved 75% accuracy in the validation cohort, with very high sensitivity (96%) but relatively low specificity (56%). The DL classifier-based algorithm, in contrast, showed high accuracy (77%) with relatively high specificity (83%) and sensitivity (67%) as well. When the two models were combined, diagnostic performance improved substantially, with accuracy of 86%, sensitivity of 84%, and specificity of 89%. Probability attention maps (PAMs) identified thresholds for the volumes of key brain regions, such as the corpus callosum, precentral gyrus, thalamus, and cingulate cortex in RRMS, and the brainstem, hippocampus, and parahippocampal gyrus in MOGAD. The findings of this study suggest that using the two models in tandem may improve diagnostic accuracy and provides evidence of differing physical pathologies between the two diseases.

Outcomes and Implications

The ability to reliably distinguish MS from MOGAD is clinically important because the two conditions differ in both prognosis and treatment response. Misdiagnosis and misclassification can delay necessary care or lead to unnecessary exposure to therapies of the incorrect disease. By integrating a conventional clinical/MRI algorithm with a DL classifier, this work demonstrates a practical path toward more accurate diagnosis. In this case, the DL approach improves specificity while the clinical/MRI model contributes sensitivity, making the combined approach both specific and sensitive. Integrating this DL-based approach into the diagnostic process will allow clinicians to provide more accurate diagnoses more frequently and thus allow patients to access the appropriate treatment in a timely manner.