5963Automated quantification of epicardial adipose tissue from non-contrast CT on multi-center and multi-vendor data using deep learning

Abstract

Abstract Background Epicardial adipose tissue (EAT), a metabolically active visceral fat depot surrounding the coronary arteries, has been shown to promote the development of atherosclerosis in underlying coronary vasculature. Purpose We evaluate the performance of deep learning (DL), a sub-group of machine learning algorithms, for robust and fully automated quantification of EAT on multi-center cardiac CT data. Methods In this study, 850 non-contrast calcium scoring CT scans, from multiple cohorts, scanners and protocols, with manual measurements of EAT from 3 different readers were considered. The DL method was based on a convolutional neural network trained to reproduce the expert measurement. DL global performance was first assessed using all the scans, and then compared to inter-observer variability on a subset of 141 scans. Finally, automated EAT progression was compared to manual measurement using baseline and follow-up serial scans available for 70 subjects. The proposed model was validated using 10-fold cross validation. Results Automated quantification was performed in 1.57±0.49 seconds compared to 15 minutes for manual measurement. DL provided high agreement with expert manual quantification for all scans (R=0.974, p<0.001) with no significant bias (0.53 cm3, p=0.13). EAT volume was higher in patients with hypertension (+18.02 cm3, p<0.001, N=442), with diabetes (+18.33 cm3, p<0.001, N=75) and with hypercholesterolemia (+7.33 cm3, p=0.039, N=508). Manual EAT volumes measured by two experienced readers on 141 scans were highly correlated (R=0.984, p<0.001) but presented a significant difference of 4.35 cm3 (p<0.001). On these 141 scans, DL quantifications were highly correlated to both experts' measurements (R=0.973, p<0.001; R=0.979, p<0.001) with significant and non-significant bias for readers 1 and 2 (5.19 cm3, p<0.001; 0.84 cm3, p=0.26), respectively. In 70 subjects, EAT progression quantified by DL correlated strongly with EAT progression measured by the expert reader (R=0.905, p<0.001) with no significant bias (0.64 cm3, p=0.43), and was related to increased non-calcified plaque burden quantified from coronary CT angiography (5.7% vs 1.8%, p=0.026). Automated vs. manual EAT volume Conclusion Deep learning allows rapid, robust and fully automated quantification of EAT from calcium scoring CT. It performs as an expert reader and can be implemented for routine cardiovascular risk assessment. Acknowledgement/Funding 1R01HL133616/01EX1012B/Adelson Medical Research Foundation