Deep learning in Myocarditis: A novel approach to severity assessment.

BACKGROUND: Myocarditis is life-threatening in the acute phase, yet biopsy-the diagnostic gold standard-lacks an objective method to quantify cardiomyocyte damage. We developed deep learning models to derive a pathology-based severity index for myocarditis from whole-slide biopsy images. METHODS AND RESULTS: We retrospectively analyzed 305 consecutive patients (1,056 digitized hematoxylin-eosin slides) who underwent endomyocardial biopsy between 2002 and 2021 at the National Cerebral and Cardiovascular Center; 145 met Dallas criteria for myocarditis and were used for severity modeling. Severe myocarditis was defined by short-term in-hospital outcomes (SCAI-aligned cardiogenic shock, initiation of mechanical circulatory support, or death). A multiple instance learning (MIL) classifier was first trained on slide-level myocarditis labels. We then built two severity models: (1) logistic regression using lymphocyte density derived from a YOLOv8-based object detector (Model 1), and (2) a Transformer that processed the top MIL-ranked patches to predict severe versus non-severe myocarditis (Model 2). Model 1 confirmed a strong association between inflammatory burden and severe outcomes (AUROC 0.809). Model 2 achieved superior discrimination (AUROC 0.993) with higher accuracy and precision. Attention maps indicated that Model 2 focused not only on inflammatory infiltrates but also on myocyte injury and architectural disruption, suggesting broader histologic signal capture. The final output was a continuous pathology-based severity score; clinical variables were not input to the models. CONCLUSIONS: Combining MIL with a Transformer enables comprehensive extraction of histologic features associated with clinically severe myocarditis and yields an objective, reproducible tissue-injury index. This score is intended to standardize histologic severity assessment and complement, rather than replace, clinical evaluation; incremental clinical utility requires prospective, multi-center validation.