Abstract
Cardiovascular diseases (CVDs) are among the leading causes of mortality worldwide and are increasingly affecting not only elderly individuals but also younger populations. In this study, we investigate the application of computer vision techniques for the automatic classification of electrocardiogram (ECG) images to detect heart diseases. Computer vision is widely used in medical diagnostics due to its ability to accurately analyze complex visual data and assist in non-invasive early detection.We developed and evaluated two deep learning models using the publicly available ECG image dataset from the National Heart Foundation of Bangladesh. The first model employed a custom convolutional neural network (CNN) architecture, while the second utilized a transfer learning approach based on the ResNet50 architecture pretrained on ImageNet. Both models were trained to classify ECG images into four clinically relevant categories: normal condition, abnormal heartbeat, acute myocardial infarction, and history of myocardial infarction.The results demonstrate high classification performance from both approaches. The ResNet50-based model achieved a training accuracy of 83.4% and a validation accuracy of 81.2%, indicating strong generalizability and robust convergence. In contrast, the custom CNN model showed a training accuracy of 100% and a validation accuracy of 99.1%, reflecting excellent discriminative power within the dataset. These outcomes confirm the effectiveness of deep convolutional architectures in processing medical image data and highlight the potential for ECG-based automated diagnostics in supporting cardiology workflows.