A Fully Convolutional Neural Network Based on 2D-Unet in Cardiac MR Image Segmentation

Abstract

Cardiac MRI image segmentation is of great importance for evaluating cardiac function and diagnosing diseases. Manual segmentation is time-consuming and tedious, so automatic segmentation is very popular in practical applications. In this paper, we propose an improved full convolutional neural network based on 2D-Unet for automatic segmentation of the left ventricle, right ventricle and myocardium. Experiments were conducted on the ACDC 2017 Challenge Training dataset. The segmentation results were assessed by means of average vertical distance, Dice coefficient and Hausdorff distance. Our model reduces the amount of parameters, improves the training speed, uses the fusion loss function, and maintains a satisfactory segmentation accuracy of left ventricle, right ventricle and myocardium.