Deep learning in Myocarditis: A novel approach to severity assessment

Area of Science:

• Cardiovascular Pathology
• Artificial Intelligence in Medicine
• Digital Pathology

Background:

• Myocarditis poses significant acute risks, but objective quantification of cardiomyocyte damage from biopsies remains challenging.
• Current diagnostic methods for myocarditis lack standardized, objective measures for assessing disease severity at the tissue level.

Purpose of the Study:

• To develop and validate deep learning models for deriving a pathology-based severity index from whole-slide biopsy images of myocarditis.
• To establish an objective, reproducible method for quantifying cardiomyocyte damage and predicting severe myocarditis outcomes.

Main Methods:

• Retrospective analysis of 305 endomyocardial biopsies (1,056 slides), with 145 meeting Dallas criteria for myocarditis.
• Development of two deep learning models: a multiple instance learning (MIL) classifier and a Transformer-based model utilizing MIL-ranked patches.
• Severity prediction based on severe myocarditis outcomes (cardiogenic shock, mechanical support, or death).

Main Results:

• Model 1 (logistic regression with object detection) showed strong association between inflammation and severe outcomes (AUROC 0.809).
• Model 2 (Transformer) achieved superior discrimination (AUROC 0.993), identifying inflammatory infiltrates, myocyte injury, and architectural disruption.
• A continuous pathology-based severity score was generated, independent of clinical variables.

Conclusions:

• Combining MIL and Transformer models enables comprehensive histologic feature extraction for assessing clinically severe myocarditis.
• The developed score provides an objective and reproducible tissue-injury index for myocarditis.
• Prospective validation is needed to confirm the clinical utility of this novel severity score.