Heart Disease Classification Using EfficientNet B5 with Three-Dimensional Scaled Electrocardiogram Images

In this study, we applied transfer learning to improve the classification of ECG images into various heart disease categories.We utilized a pre-trained convolutional neural network, EfficientNetB5 which is a convolutional neural network architecture that employs a compound scaling method to uniformly scale depth, width, and resolution using fixed coefficients.Unlike traditional methods that independently adjust these dimensions, EfficientNet scales all three dimensions simultaneously with a consistent set of scaling factors.We modified the architecture of the EfficientNetB5 model to accommodate our specific classification task of eleven heart disease categories.This adaptation involved replacing the final classification layer of the pre-trained model with a new output layer tailored for eleven distinct classes.We then fine-tuned the modified model on our dataset of 9,185 12-lead generated ECG images from PTB-XL database, which included 4,005 distortion-free images, 2,050 images with text distortions, and 3,130 images with wrinkled distortions.This fine-tuning process involved training the remaining layers of the network while leveraging the pre-trained weights to enhance feature extraction and accelerate convergence.The model performed with an f measure of 0.227 and was ranked 12/16 on official scoreboard.