Encoder Modified U-Net and Feature Pyramid Network for Multi-class Segmentation of Cardiac Magnetic Resonance Images

Abstract

ABSTRACT Cardiovascular diseases are leading cause of death worldwide. Timely and accurate detection of disease is required to reduce load on healthcare system and number of deaths. For this, accurate and fast segmentation of Cardiac Magnetic Resonance Images is required. In this study, we propose to develop a transfer learning-based end-to-end trainable method to segment left ventricle, myocardium, and right ventricle of heart. In the presented work, Feature Pyramid Network and U-Net architecture are used where encoder is modified with networks like DenseNet, ResNet, and VGG. Performance evaluation is done using dice score, Jaccard index, and Hausdorff distance according to which U-Net with VGG encoder gives best results. The mean dice score obtained is 0.958, 0.914, and 93.4 for LV, MYO, and RV respectively. Also, Hausdorff distance for the proposed methods is 1.69, 2.28, and 1.90 for LV, MYO, and RV respectively. The p-value for the obtained results is less than 0.05 (=0.0313) which shows the statistical significance of the proposed method. This automatic end-to-end trainable computer-based method requires less time and resources while giving better results than state-of-art methods. It can save the time of medical practitioners in analyzing cardiac diseases.