Multi-disease, Multi-view and Multi-center Right Ventricular Segmentation in Cardiac MRI Using Efficient Late-Ensemble Deep Learning Approach

Abstract

In many computer vision areas, deep learning-based models achieved state-of-the-art performances and started catching the attention in the context of medical imaging. The emergence of deep learning is a cutting-edge for the state-of-the-art methods of cardiac magnetic resonance (CMR) segmentation. For generalization and better optimization of current deep learning models for CMR segmentation problems, the M&Ms-2 (Multi-Disease, Multi-View & Multi-Center Right Ventricular Segmentation in Cardiac MRI) challenge proposed data that are acquired from three clinical centers of Spain and Germany using three different magnetic resonance scanner vendors (Siemens, General Electric and Philips). To cater to the generalization issue on a multi-Disease dataset, the proposed model used lightweight convolutional layers’ blocks before the proposed residual block which have been used as a skip connection with a series of several layers for boundary and structural information preservation. The residual blocks that are used after every encoder block help bridge the semantic gap between the encoder and decoder. The efficient expansion, depth-wise, and projection block (EDP) is used at each decoder block for efficiently improving the segmentation maps. The proposed 2D-based model is used to segment the right ventricle (RV) in short-axis (SA) images, as well as in long-axis (LA) cardiac MR images without any additional dataset and pretrained weights. The proposed model produced optimal dice coefficients (DC) and Hausdorff distance (HD) scores in validation and testing images and could be useful for the segmentation of the RV and LA in cardiac MRI.