Left ventricle segmentation for cine MR using deep learning

Abstract

Functional analysis of the heart can be performed using cardiac magnetic resonance imaging. Cine imaging is the most widely used cardiac modality. After careful segmentation of the left ventricle (LV), physiological heart indexes can be calculated for heart function evaluation. However, the physician spends significant time and effort to segment the LV. Therefore, physicians need an automated tool for LV segmentation to save time and effort. This chapter proposes a new approach for the automatic segmentation of the epicardial and endocardial boundaries of the LV. The segmentation is performed on the original cine cardiac images using a fully convolutional neural network known as the U-net. In cardiac images, there is a severe problem called class imbalance. This problem happens because the LV region comprises a tiny proportion of the image compared with the background. Because of the fact that the background represents the majority class, the network became biased toward the background during the learning process of the network. To avoid the class imbalance problem, we present a new loss function into our network. We did not use the traditional binary cross-entropy loss alone, because it encourages learning bias in the framework. We modified the loss function to maximize the value of accuracy. On the other hand, it works on reducing the learning bias that happens due to the binary cross-entropy. Our approach results in good segmentation performances for the epi- and endocardial boundaries (Dice 0.941 and 0.962, respectively) compared with the conventional loss function (Dice 0.893 and 0.872, respectively).