Deep Learning-Based CAD System for ECG Image Classification

For the last decade, deep learning techniques have significantly enhanced the early identification of diseases in the healthcare domain, particularly in the detection of cardiovascular conditions, which are the leading cause of mortality in the United States. Myocardial infarction (MI), commonly known as a heart attack, accounts for approximately 85% of cardiovascular cases, which highlights the importance of the precise diagnosis of this condition. Several studies introduced deep learning models as a modern computer-aided diagnostics (CAD) technique for this problem by fine-tuning pre-trained networks through transfer learning to increase the accuracy of MI detection. This study focuses on reviewing and improving classification-based convolutional neural network (CNN) models trained on electrocardiogram (ECG) images to offer insights into key trends and challenges. To enhance classification accuracy, this paper proposes a novel ECG image-cleaning technique that effectively reduces ECG signal noise in the input image while preserving critical signal features. Moreover, MultiVision variants of standard deep learning CNN architectures are introduced to enhance the model's classification accuracy by enabling localized attention to each lead of the ECG image. Experimental results demonstrate that on one hand, the introduced ECG imagecleaning technique increases classification accuracy by 5.65%; on the other hand, the MultiVision model achieves a classification accuracy of 100%, outperforming the existing state-of-the-art CNN-based model on the ECG images dataset of cardiac patients. Results highlight the potential of the proposed CAD system to advance automated cardiovascular disease (CVD) diagnosis and improve patient outcomes.