<p>Automated interpretation of electrocardiograms (ECGs) is essential for early detection of cardiovascular disease; however, in routine clinical workflows, ECGs are often archived as rasterized image reports rather than structured digital signals, limiting the applicability of signal-based deep learning models. This study proposes an explainable morphology-aware convolutional neural network (CNN) for multi-class cardiovascular disease classification directly from ECG images. The framework differentiates four clinically relevant categories: normal rhythm, abnormal heartbeat, myocardial infarction, and history of myocardial infarction. A curated dataset of 928 standard 12-lead ECG images was expanded to 5,562 samples using medically consistent, geometrically constrained augmentation techniques designed to preserve biological validity. Stratified five-fold cross-validation was conducted to ensure statistical reliability. The proposed architecture achieved a mean classification accuracy of 98.3% ± 0.6%, significantly outperforming VGG16 and InceptionV3 under identical training conditions (p &lt; 0.05). Comprehensive computational benchmarking and ablation analysis confirmed the contribution of key architectural components such as specific kernel sizes and pooling strategies—to stable generalization and superior inference efficiency. Grad-CAM–based visualization demonstrated that model attention aligns with clinically meaningful waveform regions, and misclassification profiles were found to mirror true pathophysiological overlaps. These findings indicate that task-specific morphology-aware CNN architectures can improve diagnostic reliability and computational efficiency, supporting scalable ECG image–based decision support systems in clinical practice.</p>