Automatic Segmentation and Functional Assessment of the Left Ventricle using U-net Fully Convolutional Network

Abstract

A new method for the automatic segmentation and quantitative assessment of the left ventricle (LV) is proposed in this paper. The method is composed of two steps. First, a fully convolutional U-net is used for the segmentation of the epi- and endo-cardial boundaries of the LV from cine MR images. This step incorporates a novel loss function that accounts for the class imbalance problem caused by the binary cross entropy (BCE) loss function. Our novel loss function maximizes the segmentation accuracy and penalizes the effect of the class-imbalance caused by BCE. In the second step, the ventricular volume curves are constructed from which LV function parameter is estimated (i.e., ejection fraction). Our method demonstrated a statistical significance in the segmentation of the epi- and endo-cardial boundaries (Dice score of 0.94 and 0.96, respectively) compared with the BCE loss (Dice score of 0.89 and 0.86, respectively). Furthermore, a high positive correlation of 0.97 between the estimated ejection fraction and the gold standard was obtained.