Quantification of Left Ventricular Mass on Cardiac CT Using Deep Learning Approach

Left ventricular (LV) mass is a robust predictor of adverse events. Cardiac computed tomography angiography (CCTA) is a well-established modality used for the assessment of LV mass. In this study, we aimed to develop, validate, and test a fully automated deep learning (DL) system capable of quantifying LV mass on CCTA. A DL system using a convolutional neural network based on the U-Net architecture was first developed for the training purpose utilising 86 LV images. The system was able to segment the LV from axial series with no other reprojection required. The system was then validated on 13 more datasets, where its performance was further improved by tuning. Linear regression analysis with Pearson correlation, Sorensen–Dice coefficient, mean absolute percentage error and Bland-Altman analysis plot were generated to assess the performance of the system against the ground-truth LV mass data calculated by conventional manual-segmentation method. Eight-two randomly generated multi-vendor datasets were used for testing the performance (mean age 58±11 yrs, 75% male) of the DL system. Mean LV mass of the cohort was 126.8±39.4g. The average DL system analysis time was 21.34±6.49s. Excellent correlation was seen for the DL system vs ground-truth in detecting LV Mass (r=0.96, p<0.0001). The Sorensen-Dice coefficient for the test data was 0.92 with the mean absolute percentage error of 7.94%. Bland-Altman analysis reported bias of 6.2g for the DL method. Fully automated quantification of LV Mass on CCTA is feasible with high accuracy and excellent agreement.