Enhancing Myocarditis Diagnosis Through Deep Learning and Data Augmentation: A Novel Framework Addressing Imbalance and Initialization Sensitivity

Abstract

Myocarditis poses a serious public health risk, with the potential to cause heart failure and sudden death. Traditionally, diagnosing myocarditis relies on non-invasive imaging, particularly cardiac magnetic resonance imaging (MRI), though MRI results can be vulnerable to operator bias. Our research addresses this by introducing an innovative deep-learning framework to tackle challenges frequently overlooked in past studies, including class imbalance, sensitivity to initial weight settings, and generalizability. Our model leverages convolutional neural networks (CNNs) to extract detailed feature vectors for highly precise classifying of myocarditis. Since the class imbalance problem is frequent in many training datasets, we will adopt a reinforcement learning (RL) strategy to shift more emphasis on the underrepresented classes for balanced learning. Additionally, our model involves a mutual learning-based artificial bee colony (ML-ABC) algorithm for efficient pretraining of weights. Improve training data diversity and volume further using online data augmentation with an improved version of the generative adversarial network (GAN). We further enhance the performance of the generator by considering the information provided by the features produced by the discriminator on which to base its output for making it realistic, hence increasing the accuracy of the generator. Our model, when applied to the Z-Alizadeh Sani myocarditis dataset, reaches an accuracy of 90.8%, outperforming previously reported techniques and reiterating its feasibility for clinical purposes. These results significantly advance early myocarditis detection and open new avenues for enhanced treatment strategies.