A Deep Learning Framework For Myocardial Infarction Detection From Apical-4-Chamberechocardiograms: Combining Segmentation And Machine Learning Classification

Abstract

Purpose: Myocardial infarction is a highly fatal cardiac disease caused by reduced blood flow to parts of or the whole cardiac muscle. Early detection and immediate intervention can greatly reduce the severity of damage to the heart muscle. This work presents a framework to detect MI in echocardiogram videos to increase the swiftness of MI diagnosis.

Methods: The framework functions in three key stages. In the first stage, a LadderNet CNN segments the left ventricle wall out of the echocardiogram’s frames. In the next Stage, the segmented left ventricle in each frame is divided into six sections whose displacements are tracked across the consecutive frames. In the final stage, a machine learning model classifies the presence of MI by analyzing the displacements of the wall’s sections. Five classic algorithms were assessed and compared for the classification task. The HMC-QU benchmark dataset was used for the training and testing of the framework.

Results: The segmentation model demonstrated excellent performance with an Intersection over Union of 97.32%, an F1 score of 97.40%, an accuracy of 99.78%, a precision of 97.50%, a sensitivity of 97.31%, and a specificity of 99.89%. For MI classification, the optimal model (Random Forest) achieved 85.71% accuracy, 86.67% precision, 92.86% sensitivity, 71.43% specificity, and 89.66% F1 score.

Conclusion: The promising results of this work suggest that the proposed framework has the potential to contribute to the detection of myocardial infarction.