Classification of CMR images between healthy and diseased myocardium: A deep learning approach

Abstract

Heart failure is a major cause of morbidity and mortality in the world. Late gadolinium enhanced (LGE) cardiac magnetic resonance (CMR) imaging can be used to directly visualise the presence of myocardial damage, a predictor of heart failure. Deep Learning (DL) models, specifically convolutional neural networks (CNNs), can help in assisting medical personnel in the analysis and classification of various diseases, including the identification of myocardial damage from CMR images. However, due to the lack of labeled medical images, training DL models is a big challenge. Transfer learning, a technique used for adapting models that have previously been trained on a larger dataset, has proven effective in overcoming this problem. This study compares three state-of-the-art CNNs (VGG16, VGG19 and Inception V3) in the classification of healthy and diseased myocardium using transfer learning on an imbalanced dataset. To compensate for the imbalance, three subsets of the dataset were created using undersampling and class weighting techniques. These were also used to train the models and subsequently compared. The aim was to see which model could be used in a clinical setting. Results showed that all models classified a significantly higher amount of times images as healthy, as opposed to diseased and that the balanced dataset provided slightly better outcomes. It was determined that although none of the models were appropriate to assist physicians at the moment, VGG19 could be further tuned to perform better.