FR-NET: Focal Loss Constrained Deep Residual Networks for Segmentation of Cardiac MRI

Abstract

Accurate segmentation of the left ventricle is an important step in evaluation of cardiac function. We have proposed a framework combining skip connection and focal loss together for left ventricle segmentation from cardiac MRI images. Residual neural networks (ResNet) have been used as the backbone of our method and have been shown to improve not only the segmentation accuracy of the left ventricle (LV) but also the network optimization process, thereby increasing the convergence rate of training due to improved gradient backpropagation. In addition, dice loss is trained with focal cross entropy loss by an alternative training strategy. Experiments show that our method achieves significant performance on the Sunnybrook public dataset.