Hypertrophic Cardiomyopathy Diagnosis Using Deep Learning Techniques

Abstract

Hypertrophic cardiomyopathy (HCM), which can lead to serious cardiac problems, is often diagnosed using cardiovascular magnetic resonance (CMR) images obtained from patients. This research consisted in developing a deep learning technique to diagnose HCM, using a dataset consisting of 37,421 healthy and 21,846 HCM patients obtained over a period of two years. The proposed dataset is the largest publicly available CMR dataset for HCM diagnosis. Three experts inspected the images and determined whether each one showed a case of HCM or not. Novel data augmentation was used by employing color filtering. To classify the augmented images, a convolutional neural network (CNN) was designed and tuned. To the best of the authors' knowledge, none of the existing studies have tackled HCM diagnosis based on CMR images, and this paper is the first one in this regard. Comparing the designed algorithm output with the experts' opinions, the proposed method achieved accuracy of 98.53, recall of 98.70, and specificity of 95.21 on the augmented dataset. Experiments were also conducted with different optimizers and other methods of data augmentation to further evaluate the proposed method. Using the proposed data augmentation method, accuracy of 98.53 was achieved, which is higher than the best accuracy (95.83) obtained by the other evaluated methods of data augmentation. The paper presents the theoretical performance bound of the proposed method, and a comparison with existing papers which reveals the superiority of the proposed approach in terms of various performance metrics. The advantages of the proposed method include the elimination of the contrast agent and its complications, a lower CMR examination time, and lower costs for patients and cardiac imaging centers.