Transfer-Learned Convolutional Networks for Multiclass Recognition of Myocardial Infarction and Cardiac Abnormalities in ECG Images

Accurate interpretation of electrocardiograms (ECGs) is central to the early recognition of myocardial infarction and other cardiac abnormalities, yet visual assessment may be affected by subtle inter-class similarities, recording artifacts, and reader-dependent variability. This study presents a controlled comparison of three transfer-learned convolutional neural networks, ResNet-50, ConvNeXt-Tiny, and DenseNet-201, for four-class classification of ECG images. The dataset comprised 3,951 images assigned to abnormal heartbeat, history of myocardial infarction, myocardial infarction, and normal ECG categories. A stratified split allocated 3,159 images to training, 393 to validation, and 399 to independent testing. All models were initialized with ImageNet weights, fine-tuned at an input resolution of 224 × 224 pixels, and assessed using accuracy, macro-averaged precision, recall, F1-score, parameter count, floating-point operations, and inference latency. DenseNet-201 achieved the strongest diagnostic performance, with 99.50% accuracy, 99.48% precision, 99.29% recall, and a 99.37% F1-score, while using 18.10 million trainable parameters. ResNet-50 produced closely comparable accuracy (99.25%) and the lowest latency (0.6949 ms per image), indicating a favorable option for time-sensitive deployment. ConvNeXt-Tiny reached 66.67% accuracy and showed substantially weaker generalization under the common training protocol. The findings indicate that dense feature reuse is particularly effective for preserving subtle ECG morphology, whereas residual learning offers an advantageous accuracy–latency balance. External, patient-level validation remains necessary before clinical use, especially because highly accurate image classifiers may exploit dataset-specific acquisition cues in addition to physiologically meaningful patterns.