A Custom Dilated-Separable CNN for Automated Cardiovascular Disease Detection Using Electrocardiogram Images

Abstract

Cardiovascular diseases are a leading cause of millions of deaths worldwide, placing a huge burden on healthcare systems. Therefore, early and accurate diagnosis of cardiovascular diseases is essential in determining efficient treatment and preventing life-threatening complications. Traditional diagnostic approaches that rely on manual interpretation of electrocardiograms (ECGs) are often subject to inter-observer variability and are time-consuming, highlighting the urgent requirement for effective, accurate, and automated diagnostic support systems. In this article, a customized Convolutional Neural Network (CNN) based diagnostic system is proposed for automated detection of cardiovascular diseases using a publicly available 12-lead ECG images dataset. Extensive pre-processing and data augmentation were applied to the dataset to improve signal diversity and minimize overfitting. This proposed CNN incorporates several building blocks, involving dilated temporal and separable wide convolutional layers, aimed to possess fine-grained morphological patterns and wider spatial dependencies in ECG images. Additionally, several layers of batch normalization and dropout were utilized for stabilizing training and improving generalization. Experimental evaluations revealed a superior classification accuracy of 99% for the proposed system, outperforming state-of-the-art pre-trained CNNs and existing related systems. Moreover, a diagnostic support tool has evolved to facilitate real-time implementation in clinical environments, providing an effective and easy-to-interpret framework for detecting cardiovascular diseases.