Myocarditis Diagnosis Using Semi-Supervised Generative Adversarial Network and Differential Evolution

Myocarditis, a significant public health issue, can lead to heart failure and death. While endomyocardial biopsy (EMB) provides accurate diagnoses, its invasiveness limits its use in severe cases. Alternatively, cardiac magnetic resonance (CMR) offers a non-invasive method to detect and monitor myocarditis, displaying myocardial involvement with high signal contrast. Traditional studies on myocarditis often rely on supervised learning, which faces challenges such as the need for large, well-labeled datasets. A diagnostic model for myocarditis is introduced in this paper, utilizing CMR images and employing a semi-supervised generative adversarial network (SS-GAN) to enhance classifier performance, complemented by differential evolution (DE) for hyperparameter optimization. The semi-supervised nature of SS-GAN is pivotal in scenarios where labeled data is scarce or imbalanced. A regularization strategy is included to address missing modes and stabilize SS-GAN training. Feature vectors extracted by the discriminator are utilized to guide the generator in producing accurate outputs. Additionally, a reconstruction loss is incorporated into the loss function of SS-GAN, compelling the generator to reconstruct outputs based on the discriminator's features, thereby aligning outputs closer to actual data configurations. This enhancement not only improves SS-GAN performance but also ensures high-quality outputs. To minimize reliance on hyperparameters, the Random Key method is employed, optimized using the DE algorithm. The efficacy of the model is demonstrated on the Z-Alizadeh Sani myocarditis dataset, with further validation achieved through experiments on the EMIDEC dataset, assessing transfer learning (TL) effects. F-measures of 91.124% and 89.511% are achieved on the Z-Alizadeh Sani myocarditis and EMIDEC datasets, respectively, indicating superior performance over existing techniques.