Performance of ensemble of UNets for epicardial adipose tissue quantification: a multi-center APOLLO study

Abstract

Abstract Background Epicardial adipose tissue (EAT) surrounding the heart and coronary arteries, has been proposed to be a potential therapeutic target for the prevention of cardiac remodeling and arrhythmias in cardiovascular disease. Specifically, EAT volume is positively correlated with these cardiovascular events and therapeutic options are being developed with the aim to reduce EAT volume. EAT volume can be measured from non-contrast CT scans, but requires manual segmentation of the pericardium to isolate the EAT, which is prone to inter-observer variability. Purpose We aimed to develop a new AI EAT toolkit for automated quantification of EAT volume from multi-center non-contrast CT data and assess its performance against clinical annotations from four trained human experts. Methods This is a multicenter study which performs CT scans in 5000 Asian Admixture patients (APOLLO study). In the current stage of this study, NCCT data analysis was conducted in 1003 patients (mean age of 58±11 years; 33% females, with ethnicity breakdown of 72% Chinese, 20% Indian/Malay, 8% Others). Of these patients, 52% had hypertension, 19% had diabetes, 69% had hyperlipidemia, and 31% had obesity. AI EAT tool was developed via a novel deep learning framework using an ensemble of UNets (See Figure 1 and 2). The performance of AI EAT tool was evaluated with respect to expert clinical segmentations of four independent raters for a test dataset of 50 patient scans using dice similarity coefficient (DSC), Pearson correlation, and Bland-Altman analysis. Results The AI EAT quantification process took less than 20 seconds per subject, compared with 20-30 minutes for expert readers. Compared to clinical ground truth, our AI EAT achieved a DSC of 96.0 ± 1.0% and 89.1 ± 3.0% for pericardium and EAT segmentations, respectively. There was strong agreement between the AI EAT and clinical ground truth in deriving the EAT volume (r=0.99, p<0.001) and minimal error of 5.3±2.6%. The Bland-Altman analysis revealed the mean bias between the predicted EAT volume and the ground truth volumes to be 0.95 cm³ with the limits of agreement as [-12.86 cm³, 14.77 cm³]. Conclusion Ensemble of UNets based AI EAT tool is a novel method with high performance for the quantification of epicardial adipose tissue volume, as a potential diagnostic and prognostic tool for cardiac remodeling and arrhythmias.Figure 1.Ensemble of UNets framework.Figure 2.UNet Architecture used.