Myocarditis Detection Using Proximal Policy Optimization and Mutual Learning

Myocarditis poses a significant health threat, often leading to severe outcomes such as heart failure or death. While endomyocardial biopsy (EMB) is a reliable diagnostic method, its invasive nature limits its use to severe cases. Cardiac magnetic resonance (CMR) offers a less invasive alternative by utilizing enhanced signal contrast to detect and monitor myocarditis, effectively highlighting affected myocardial areas. This study introduces an advanced model designed to overcome challenges like initial weight sensitivity and class imbalance, which are often overlooked in existing research. The model employs multiple convolutional neural networks (CNNs) to extract feature vectors from images for classification. To address class imbalance, a proximal policy optimization (PPO)-based algorithm is utilized, significantly improving the training process by preventing abrupt policy shifts and stabilizing them. This mechanism rewards precise classification of underrepresented classes more heavily than overrepresented ones, promoting balanced learning. Additionally, the model incorporates a novel pre-training method, the mutual learning-artificial Bee Colony (ML-ABC) technique, which dynamically adjusts initial weight settings through mutual learning components. This approach enhances the model's initial ability to recognize complex patterns in myocardial images, improving classification accuracy. The model's efficacy was validated using the Z-Alizadeh Sani myocarditis dataset, achieving an impressive F-measure of 83%. These results highlight the model's superiority over existing methods and confirm its potential for clinical use, marking a significant advancement in the early detection and treatment of myocarditis.