Machine learning accurately quantifies epicardial adipose tissue from non-contrast CT images in coronary artery disease. (25th January 2023)

• Record Type: Journal Article
• Authors: Cheng, N; Tan, E W P; Leng, S; Baskaran, L; Teo, L; Yew, M S; Singh, M; Huang, W M; Chan, M Y Y; Ngiam, K Y; Vaughan, R; Chua, T; Tan, S Y; Lee, H K; Zhong, L
• Is Part Of: European heart journal. Volume 44(2023)Supplement 1
• Publication Date: 2023-01-25
• Subjects: Cardiology -- Periodicals; Heart -- Diseases -- Periodicals; 616.12005
• DOI: 10.1093/eurheartj/ehac779.124
• Languages: English
• ISSNs: 0195-668X

Abstract

Abstract: Funding Acknowledgements: Type of funding sources: Other. Main funding source(s): Industry Alignment Fund – Pre-positioning Programme Background: Epicardial adipose tissue (EAT) is the visceral fat deposit within the pericardium that surrounds the heart and the coronary arteries. EAT volume measured from non-contrast CT (NCCT) has been demonstrated to be significantly associated with adverse cardiovascular risk, 1 particularly in patients with coronary artery disease.2 However, routine measurement of EAT volume is still challenging in clinical practice, as it is a tedious manual process and prone to human error. Purpose: We aimed to develop a fully automated AI toolkit (i.e., AI EAT) for the quantification of EAT from routine NCCT scans and assess its performance in reference to clinical ground truth. Methods: This is a multicenter study which performs CT scans in 5000 Asian Admixture patients (APOLLO study NCT05509010). In the current stage of this study, NCCT data analysis were conducted in 551 patients with 26, 037 images. AI EAT was developed via a novel deep learning framework using an ensemble region-based UNet. The region-based UNet uses 2 component UNet models to perform segmentation of pericardium at the apex region and non-apex region (middle and basal). EAT volume was obtained by automated thresholding of the voxels (-190 to -30 Hounsfield Unit) within the pericardium (Figure 1). The network was trained in 501 patients with 23, 712 NCCT images and tested in 50 patients with 2, 325 NCCT images. The performance of AI EAT was evaluated with respect to clinical ground truth using Dice similarity coefficient (DSC), Pearson correlation, and Bland-Altman analysis. Results: The AI EAT quantification process took less than 10 seconds per subject, compared with 20-30 minutes for expert readers. Compared to clinical ground truth, our AI EAT achieved a DSC of 0.96±0.01 and 0.91±0.02 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) with minimal error of 7±5%. Conclusion: End-to-end deep learning system accurately quantifies epicardial adipose tissue in standard NCCT images without manual segmentation. Fig 1. Comparison of manual and AI EAT.