Left Ventricle Segmentation in Cardiac MR Images Using Deep Neural Network Models

The established method for evaluating the functioning state of the heart is magnetic resonance imaging. In order to accurately measure the heart's volumetric functional properties, left ventricular (LV) segmentation is crucial. Traditional manual analysis takes time and is dependent on the observer. In order to enhance the clinical workflow of heart functional measurement, automated segmentation algorithms are required. In this study, the left ventricle was segmented using deep learning architectures, U-Net and Multires U-Net, using 1201 short axis cardiac MR images from the MICCAI (Medical Image Computing and Computer-Assisted Intervention) 2017 challenge database. Performance indicators like Dice coefficient, accuracy, and the number of trainable parameters were used to compare the findings. U-Net is created using Fully Convolutional Networks (FCN), while Multires U-Net is built using convolutional layers in an encoder-decoder fashion. The decoder up samples the image using techniques like transpose convolution to produce the cardiac ventricle segmentation, while the encoder down samples the image using strided convolution to create a compressed feature representation of the image. Convolutional blocks are replaced by Multires blocks in the Multires U-Net design, which is a variation of the U-Net architecture. Dice coefficients of 0.9372 and 0.9575, respectively, were reached in the U-Net and Multires U-Net designs. Multires U-Net based designs could be used to segment the left ventricle and offer better analysis.