Three-dimensional Deep Convolutional Neural Networks for Automated Myocardial Scar Quantification in Hypertrophic Cardiomyopathy: A Multicenter Multivendor Study

Background Cardiac MRI late gadolinium enhancement (LGE) scar volume is an important marker for outcome prediction in patients with hypertrophic cardiomyopathy (HCM); however, its clinical application is hindered by a lack of measurement standardization. Purpose To develop and evaluate a three-dimensional (3D) convolutional neural network (CNN)-based method for automated LGE scar quantification in patients with HCM. Materials and Methods We retrospectively identified LGE MRI data in a multicenter ( = 7) and multivendor ( = 3) HCM study obtained between November 2001 and November 2011. A deep 3D CNN based on U-Net architecture was used for LGE scar quantification. Independent CNN training and testing data sets were maintained with a 4:1 ratio. Stacks of short-axis MRI slices were split into overlapping substacks that were segmented and then merged into one volume. The 3D CNN per-site and per-vendor performances were evaluated with respect to manual scar quantification performed in a core laboratory setting using Dice similarity coefficient (DSC), Pearson correlation, and Bland-Altman analyses. Furthermore, the performance of 3D CNN was compared with that of two-dimensional (2D) CNN. Results This study included 1073 patients with HCM (733 men; mean age, 49 years ± 17 [standard deviation]). The 3D CNN-based quantification was fast (0.15 second per image) and demonstrated excellent correlation with manual scar volume quantification ( = 0.88, < .001) and ratio of scar volume to total left ventricle myocardial volume (%LGE) ( = 0.91, < .001). The 3D CNN-based quantification strongly correlated with manual quantification of scar volume ( = 0.82-0.99, < .001) and %LGE ( = 0.90-0.97, < .001) for all sites and vendors. The 3D CNN identified patients with a large scar burden (>15%) with 98% accuracy (202 of 207) (95% confidence interval [CI]: 95%, 99%). When compared with 3D CNN, 2D CNN underestimated scar volume ( = 0.85, < .001) and %LGE ( = 0.83, < .001). The DSC of 3D CNN segmentation was comparable among different vendors ( = .07) and higher than that of 2D CNN (DSC, 0.54 ± 0.26 vs 0.48 ± 0.29; = .02). Conclusion In the hypertrophic cardiomyopathy population, a three-dimensional convolutional neural network enables fast and accurate quantification of myocardial scar volume, outperforms a two-dimensional convolutional neural network, and demonstrates comparable performance across different vendors. © RSNA, 2019