Deep-Learning for Epicardial Adipose Tissue Assessment With Computed Tomography

Background Epicardial adipose tissue (EAT) volume is a marker of visceral obesity that can be measured in coronary computed tomography angiograms (CCTA). The clinical value of integrating this measurement in routine CCTA interpretation has not been documented. Objectives This study sought to develop a deep-learning network for automated quantification of EAT volume from CCTA, test it in patients who are technically challenging, and validate its prognostic value in routine clinical care. Methods The deep-learning network was trained and validated to autosegment EAT volume in 3,720 CCTA scans from the ORFAN (Oxford Risk Factors and Noninvasive Imaging Study) cohort. The model was tested in patients with challenging anatomy and scan artifacts and applied to a longitudinal cohort of 253 patients post-cardiac surgery and 1,558 patients from the SCOT-HEART (Scottish Computed Tomography of the Heart) Trial, to investigate its prognostic value. Results External validation of the deep-learning network yielded a concordance correlation coefficient of 0.970 for machine vs human. EAT volume was associated with coronary artery disease (odds ratio [OR] per SD increase in EAT volume: 1.13 [95% CI: 1.04-1.30]; P = 0.01), and atrial fibrillation (OR: 1.25 [95% CI: 1.08-1.40]; P = 0.03), after correction for risk factors (including body mass index). EAT volume predicted all-cause mortality (HR per SD: 1.28 [95% CI: 1.10-1.37]; P = 0.02), myocardial infarction (HR: 1.26 [95% CI:1.09-1.38]; P = 0.001), and stroke (HR: 1.20 [95% CI: 1.09-1.38]; P = 0.02) independently of risk factors in SCOT-HEART (5-year follow-up). It also predicted in-hospital (HR: 2.67 [95% CI: 1.26-3.73]; P ≤ 0.01) and long-term post–cardiac surgery atrial fibrillation (7-year follow-up; HR: 2.14 [95% CI: 1.19-2.97]; P ≤ 0.01). Conclusions Automated assessment of EAT volume is possible in CCTA, including in patients who are technically challenging; it forms a powerful marker of metabolically unhealthy visceral obesity, which could be used for cardiovascular risk stratification.

C oronary computed tomography angiography (CCTA) is used to evaluate coronary artery disease (CAD) risk, 1 with guidelines in Europe 2 and the United States 3 recommending CCTA for assessment of patients with chest pain.The use of CCTA for detecting CAD is increasing worldwide, and it is likely that further valuable information within CCTA scans that is currently not fully utilized in clinical practice could improve risk assessment and patient management for cardiometabolic diseases, with multiple such technologies being discovered through the uptake of artificial intelligence methods in research and practice. 4ipose tissue is recognized as a key regulator of cardiovascular health and disease, exerting both protective and deleterious effects on the cardiovascular system. 5Epicardial adipose tissue (EAT) is a metabolically active depot of visceral fat 5 and may be a feature of metabolically unhealthy obesity and metabolic syndrome. 6Indeed, EAT volume has been associated with multiple distinct cardiovascular diseases including CAD and atrial fibrillation (AF). 7The EAT volume is generally considered to be a marker of visceral obesity, as opposed to more sophisticated metrics such as the pericoronary Fat Attenuation Index, which specifically captures the degree of coronary inflammation, and has prognostic value over and above that of EAT volume. 8 Briefly, the ORFAN study (NCT05169333) is an international multicenter prospective cohort study that collects CCTA scans and patient clinical data from those who are undergoing or have undergone CCTA since 2005. 9For this analysis, data were utilized from across 4 National Health Service sites in England and 1 in the United States (Figure 1).The AdipoRedOx

Continued on the next page
West et al    West et al

CENTRAL ILLUSTRATION Continued
Automatic EAT Assessment for Risk Prediction    We selected the optimum unified cutoff for EAT volume prognostication for all-cause mortality, fatal/ nonfatal MI, and fatal/nonfatal stroke in SCOT-HEART by identifying the value that maximized the Youden J statistic (sum of sensitivity and specificity) on time-dependent receiver-operating characteristic curve analysis for all-cause mortality, fatal/nonfatal MI, and fatal/nonfatal stroke to ensure an optimum balance between sensitivity and specificity in our models.For consistency, the same approach was used to select the optimum cutoff for EAT volume prognostication for in-hospital postoperative and longterm AF risk within the severe CAD population of the AdipoRedOx study.Further discussion about the statistical analysis is in the Supplemental Methods.2. In all cohorts where EAT volumes were quantified, both manually and automatically, the values were normally distributed.

INTER-READER REPEATABILITY FOR EAT SEGMENTATION
AND WHOLE HEART SEGMENTATION.The interobserver variability between 2 expert analysts of the Oxford Academic Cardiovascular Computed Tomography Core Lab core lab was evaluated in 100 randomly selected patients from the UK sites of the ORFAN study.CCC for EAT volume was excellent between readers at 0.970, and the bias was nonsignificant at mean of 2.1 (95% agreement: À3.9 to 6.1) cm 3 (P ¼ 0.74) (Supplemental Figures 5A and 5B).For the whole heart segmentation volume, CCC was also excellent at 0.969 with nonsignificant bias mean of 15.2 (95% agreement: À7.5 to 23.1) cm 3 (P ¼ 0.08) (Supplemental Figures 5C and 5D).
At the time of the CCTA, application of the fully automated segmentation tool for quantification of EAT volume was found to be a significant independent predictor of the presence of AF at time of CCTA and obstructive CAD from CCTA (any 1 coronary vessel with $50% stenosis on CCTA), within 1,558 patients randomized to receive CCTA in the SCOT-HEART trial population.
When accounting for CVD risk factors the odds ratio of AF at time of CCTA per SD increase of EAT was 1.25 (95% CI: 1.08-1.40;P ¼ 0.03) (Figure 6A).
When accounting for CVD risk factors the odds ratio of obstructive CAD from the CCTA per SD increase of EAT was 1.13 (95% CI: 1.04-1.30;P ¼ 0.01) (Figure 6B).Results with statistically selected risk factor adjustment is shown in Supplemental Figure 4.

LONGITUDINAL EAT VOLUME CLINICAL CORRELATIONS.
Median follow-up for the 1,558 patients randomized to receive CCTA in the SCOT-HEART trial, which were analyzed was 4.8 years.There were 35 deaths of all causes (2.25%) of which 4 were related to coronary heart disease (0.25%).There were 8 fatal and nonfatal strokes (0.51%) and 39 fatal and nonfatal myocardial infarctions (2.5%).
The HR of all-cause mortality per SD increase of EAT was 1.28 (95% CI: 1.10-1.37);P ¼ 0.02, after accounting for CVD risk factors (Figure 6C).When other abbreviations as in Figures 1 and 2.
West et al adjusted for the same risk factors, the HR of noncardiac mortality per SD increase of EAT volume was 1.17 (95% CI: 1.07-1.33;P ¼ 0.04) (Figure 6D).This constitutes a DHR of À0.10, confirming that EAT is a measure of visceral adipose tissue related to multiple fatal pathologies, unrelated to CAD.When accounting for CVD risk factors, the HR of MI per SD increase of EAT was 1.26 (95% CI: 1.09-1.38;P ¼ 0.001) (Figure 6E).presented as Supplemental Figure 6, with EAT volume retaining significance in all models.There was no significant change in results for all postoperative AF risk analysis when BMI was replaced with waisthip ratio (Supplemental Figure 7).We developed a deep-learning model utilizing a single network for the fully automated and rapid quantification of EAT volume from CCTA.Previous automated models for EAT quantification have predominantly been in small cohorts. 12,13The most relevant model is by Commandeur et al 14  EAT volume has previously been found to be associated with CVD metrics and outcomes including the presence of atherosclerosis, 15 as well as coronary calcification progression. 16EAT is a source of numerous proinflammatory mediators that circulate well beyond the microcirculation of the heart to exert paracrine and endocrine effects on the cardiovascular and endocrine systems. 5We found EAT volume to be a significant predictor of all-cause mortality even with the exclusion of cardiac deaths.This suggests that the EAT may play a clinically significant role in broader metabolic diseases beyond atherosclerotic CAD.We propose that EAT volume should be treated as the gold standard for the detection of metabolically unhealthy visceral obesity and could form part of routine clinical interpretation of CCTA.This would shift the focus of CCTA examination as a purely structural assessment of the coronary arteries toward a more universal assessment of cardiovascular risk that considers a key visceral, metabolically sensitive, tissue depot.

DISCUSSION
EAT is in continuous bidirectional communication with the cardiovascular system. 5When there is discordance between adipose tissue and the cardiovascular system the former is thought to shift function and exerts detrimental effects on the vessels and the heart muscle, 17 which may predispose the patient to adverse outcomes such as those that we investigated.We found that EAT volume is predictive of nonfatal MI and stroke independent of BMI and following adjustment for other relevant disease risk factors.
We demonstrate that EAT volume is an indepen- CCTA provides the noninvasive gold standard measurement of EAT volume because of its excellent spatial resolution.However, manual quantification is laborious and currently falls outside the scope of routine CCTA interpretation.If clinical utility of automated EAT volume quantification could be demonstrated and found to be feasible in patients with technically challenging CCTA, it is possible that this measure could become part of standard of care.In this study we developed and validated a deep-learning network (DLN) for the automated quantification of EAT volume, which was then tested in real-world patients with CCTA with commonly encountered image quality issues to ensure validity.Then, we applied the fully automated EAT quantification tool to investigate the clinical association of EAT volume with relevant cross-sectional and longitudinal disease outcomes (Central Illustration).METHODS STUDY POPULATIONS.Each study (ORFAN [Oxford Risk Factors and Noninvasive Imaging Study], AdipoRedOx [Adipose tissue and cardiovascular RedOx regulation] study, and the SCOT-HEART [Scottish Computed Tomography of the Heart] trial) received ethical approval.The full ethics, population descriptions, variable definitions, laboratory techniques, and CCTA acquisition and retrieval protocols are outlined in the Supplemental Methods.
study involves patients who are undergoing cardiac surgery; as part of the study, patients undergo CCTA shortly after their operation and are prospectively followed up for clinical outcomes via National Health Service's NHS Digital (see the Supplemental Methods).The SCOT-HEART trial included clinical patients with suspected angina caused by coronary heart disease, who were followed up for 5 years post-CCTA for clinical outcomes (see the Supplemental Methods).OVERALL STUDY DESIGN.The overall approach to the development of the DLN, the internal and external validation, and the application of the DLN in external cohorts for ascertainment of clinical utility is outlined in Figure 2. In summary, 2,200 CCTA scans A B B R E V I A T I O N S A N D A C R O N Y M S AF = atrial fibrillation BMI = body mass index CAC = coronary artery calcium CAD = coronary artery disease CCC = Lin concordance correlation coefficient CCTA = coronary computed tomography angiography CVD = cardiovascular disease DLN = deep-learning network EAT = epicardial adipose tissue LA = left atrial MI = myocardial infarction Harefield National Health Service (NHS) Foundation Trust, London, United Kingdom; g Translational Cardiovascular Research Group, Department of Cardiology, Milton Keynes University Hospital, Milton Keynes, United Kingdom; h Faculty of Medicine and Health Sciences, University of Buckingham, Buckingham, United Kingdom; i Department of Cardiovascular Sciences and National Institute for Health Research Leicester Biomedical Research Centre, University of Leicester, Leicester, United Kingdom; j Royal CENTRAL ILLUSTRATION Development, Testing, and External Application of an Artificial Intelligence-Powered Epicardial Adipose Tissue Quantification Tool for Clinical Practice 2,200 UK scans manually segmented & 600 UK scans manually corrected for AI model feedback learning Residual-U-Net CNN for 3D volumetric segmentation of epicardial adipose tissue on clinical CCTA model for automated EAT quantification EAT improves detection of CAD & AF risk at time of scan, regardless of BMI and other risk factors EAT volume improves prognostic assessment for MI, stroke, postoperative AF, and mortality Testing of AI-Powered EAT Quantification Tool Anterior view of EAT Superior view of EAT Inferior view of EAT Excellent automated AI-powered EAT extraction vs human expert: CCC of 0.97 in external validation cohort of 720 U.S. CCTAs EAT quantification proven in challenging clinical populations: Elevated BMI, elevated CAC, recent heart surgery, & metallic artifact West HW, et al.J Am Coll Cardiol Img.2023;16(6):800-816.
United Kingdom = ORFAN sites who contributed data to this analysis = ORFAN sites that did not contribute data for this analysis The ORFAN (Oxford Risk Factors and Noninvasive Imaging Study) Arm 4 study is an international multicenter retrospective cohort study of patients undergoing clinically indicated CCTA.The initial cohort size is 75,000 patients within the United Kingdom and 25,000 internationally, with ethically approved expansion underway for 250,000 patients.Within the United Kingdom, the study includes 17 National Health Service (NHS) Trusts, 4 of which contributed data to the current study.Data collected for each participant includes the CCTA scan, data from the local hospital electronic patient record (EPR), and data from authorized third parties, including NHS Digital, all hospital event data from 2005 to now; NICOR (National Institute for Cardiovascular Outcomes Research), a national cardiac event registry; and SSNAP (Sentinel Stroke National Audit Programme), a national stroke event registry.CCTA ¼ coronary computed tomography angiography.

(
Top) A deep-learning model was trained to automatically extract the adipose tissue from CCTA.(Middle) The model performed excellently compared to human segmentation in internal and external testing, including in patient groups that are commonly occurring yet challenging for CCTA.(Bottom) The final automated artificial intelligence (AI) model for epicardial adipose tissue (EAT) quantification was applied to external clinical cohorts and revealed improved detection of prevalent disease risk for coronary artery disease (CAD) and atrial fibrillation (AF) and provided incremental prognostic benefit for key cardiovascular events such as myocardial infarction (MI), stroke, postoperative AF, and mortality in longitudinal cohorts.3D ¼ 3-dimensional; BMI ¼ body mass index; CAC ¼ coronary artery calcium; CCC ¼ Lin concordance correlation coefficient; CNN ¼ convolutional neural network; CCTA ¼ coronary computed tomography angiography; CVD ¼ cardiovascular disease; GPU ¼ graphics processing unit.from the ORFAN study were used for training the DLN for the detection of the whole heart within the bounds of the pericardium.Following training, an initial assessment of model performance was performed in 100 unseen ORFAN study scans (Supplemental Figures 1 and 2).Three separate groups of 200 unseen scans from the ORFAN study were used for fine-tuning the model through 3 iterations of feedback learning.The DLN was tested internally on 200 unseen ORFAN study scans from the UK sites of the study.External validation was performed on 720 unseen scans from the U.S. sites of the ORFAN study.The DLN was then applied to unseen scans from challenging clinical populations to test the model in patients with challenging anatomy and/or commonly occurring scan artifacts.Finally, the model was tested in unseen external scans of the SCOT-HEART trial for real-world evaluation of the prognostic value of EAT volume as a marker of metabolically unhealthy obesity.The model was also applied within the Adi-poRedOx study to test the prognostic value of EAT volume on the risk of in-patient post-cardiac surgery AF (>30 seconds of AF on monitoring) and long-term AF (paroxysmal, persistent, or chronic) following surgery were investigated.DEVELOPING THE DLN FOR AUTOMATED SEGMENTATION AND QUANTIFICATION OF EAT VOLUME.Manual segmentation of the 2,200 CCTA and the iterations of scans for feedback learning of the DLN, and the automated extraction of EAT volume from the heart segmentation were performed using CaRi-Heart (version 2.2.1, Caristo Diagnostics Ltd) (Supplemental Figure 3). 10A fully automated method for whole heart segmentation on CCTA scans was employed using a 3-dimensional Residual-U-Net neural network architecture for volumetric segmentation of CCTA (Supplemental Methods).The architecture of the DLN is demonstrated in Figure 3A.

FIGURE 2
FIGURE 2 Study Flowchart of Model Development, Testing, and External Application

FIGURE 3 ΔSigmoid
FIGURE 3 Schematic of the Deep-Learning Model for Automated Segmentation of the Whole Heart Within the Pericardium and Example Automated Segmentation

J 3
A C C : C A R D I O V A S C U L A R I M A G I N G , V O L . 1 6 , N O .6 , 2 0 2 West et al J U N E 2 0 2 3 : 8 0 0 -8 1 6 INTERNAL VALIDATION.A random sample of 200 sequestered CCTA from the UK sites in the ORFAN study were used for internal validation of the DLN.Human segmentation of these scans was undertaken blind to all other data.EXTERNAL VALIDATION.A sample of 720 unseen CCTA from the Cleveland Clinic site of the ORFAN study were used for external validation of the algorithm-as a broad external validation cohort. 11The manual segmentation of these scans was undertaken blind to all other data.STATISTICAL ANALYSIS.For inter-reader repeatability testing and human vs automated model assessment of testing data (preliminary testing, internal/external validation, and challenging clinical populations) agreement was assessed by using the Lin concordance correlation coefficients (CCC) with scatterplots and Bland-Altman analysis for significance of bias.When applied in a cohort-wide setting in the AdipoRedOx and SCOT-HEART studies, all EAT volumes were standardized by patient body surface area using the Du Bois formula.
geographic location, demographics, clinical risk factors, and CCTA scan technical parameters for all ORFAN study cohorts used in DLN training, validation, and external testing are shown in Table 1.The demographics and scan characteristics of the external clinical cohorts for which the DLN were applied following development are shown in Table 2.The relevant clinical outcomes for the prospective clinical cohorts are presented in Supplemental Table INTERNAL VALIDATION OF THE MODEL.Final internal validation occurred following 3 iterations of feedback learning to enhance the performance of the model.The median EAT volume in internal validation was 120.9 (IQR: 95.1-156) cm3 .When applied to 200 unseen scans from the UK sites of the ORFAN study, the CCC was 0.972 (Figure4A).The bias in Bland-Altman analysis (Figure4B) was also nonsignificant at 6.1 (IQR: À11.1 to 15.7) cm 3 (P ¼ 0.19).EXTERNAL VALIDATION OF THE MODEL.The final deep-learning model was applied to 720 unseen scans from the U.S. sites of the ORFAN study.The mean automated analysis time for the automated segmentation was 12.4 seconds compared with mean manual segmentation time of 18 minutes and 20 seconds.The median EAT volume in external validation was 169.3 (IQR: 111.6-241.7)cm 3 .The CCC for the automated deep-learning model vs human expert segmentation in the external validation cohort was excellent, at 0.970 (Figure 4C) and the bias in Bland-Altman analysis (Figure 4D) was nonsignificant at 3.2 (IQR: À13.6 to 17.2) cm 3 (P ¼ 0.20).VALIDATION OF THE AUTOMATED MODEL FOR EAT VOLUME QUANTIFICATION IN CHALLENGING CLINICAL POPULATIONS.Excellent CCC for automated EAT segmentation vs human expert segmentation was achieved in all challenging patient groups: Patient's with recent cardiac surgery (<6 weeks post operation) CCC ¼ 0.960 (Figure 5A, green); patients with body mass index (BMI) $40, CCC ¼ 0.962 (Figure 5A, red); patients with reported coronary artery calcium (CAC) score of $400 (Figure 5B, green), CCC ¼ 0.958; patients with significant metallic artifact within the pericardium, CCC ¼ 0.955 (Figure 5B, red), and a combined patient group of recent open-heart surgery, BMI $30 kg/m 2 and CAC $400, CCC ¼ 0.955 (Figure 5C).

FIGURE 4 D
FIGURE 4 Validation of the Deep-Learning Model

Finally
, when accounting for the same risk factors, HR of stroke per SD increase of EAT is 1.20 (95% CI: 1.08-1.32;P ¼ 0.02) (Figure 6F).Results with statistically selected risk factor adjustment are shown in Supplemental Figure 4. Adding EAT volume into a clinical model led to significant improvement in the ability to detect obstructive CAD on CCTA (any 1 coronary vessel

FIGURE 5
FIGURE 5 Validation of the Automated Deep-Learning Model in Challenging Clinical Populations

FIGURE 6 1 - 6 FIGURE 7 6 FIGURE 8 3 Model 1 (
FIGURE 6 Cross-Sectional and Longitudinal Associations Between EAT Volume and Clinical Outcomes in the SCOT-HEART Trial In this study we developed a deep-learning model for automated segmentation and quantification of EAT from CCTA images.The model was then validated in multiple cohorts, including commonly occurring challenging populations where manual segmentation is extremely difficult because of artifacts with good performance.Then we applied this automated model to the SCOT-HEART cohort, demonstrating a good prognostic value of EAT volume for all-cause mortality and cardiovascular events, as a possible measure of unhealthy visceral obesity relevant to cardiometabolic dysfunction, regardless of whether EAT volume was used as a continuous variable or when used with a cutoff.Contrary to pericoronary Fat Attenuation Index, which captures the degree of coronary artery inflammation and is predictive of cardiac (but not of noncardiac) mortality, we now demonstrate that EAT volume is predictive of noncardiac mortality, confirming its role as a broader biomarker of visceral obesity that affects survival in a broader sense.We also demonstrate that this measurement has important prognostic value for postoperative AF in patients undergoing cardiac surgery, beyond known postoperative risk models including LA volume and N-terminal pro-B-type natriuretic peptide.Fully automated measurement of EAT volume incorporated into routine interpretation of CCTA promises to significantly improve the risk stratification of patients across several important clinical outcomes.
who developed a convolutional neural network capable of automated EAT segmentation and tested in the EISNER (Early Identification of Subclinical Atherosclerosis by Noninvasive Imaging Research) cohort; however, no head-to-head comparison is made in this study.The model developed here achieves accurate EAT volume quantification in technically challenging yet commonly occurring clinical populations.The need for any artificial intelligence-based radiology approaches to be applicable for all-comers is fundamental to patient acceptance and the future uptake of such technology.The DLN reduced EAT quantification time from an average of 18 minutes when performed manually, to an average of 12 seconds, rendering this tool usable in the clinical environment without adding workload to clinical teams.
dent predictor providing incremental value for postoperative AF regardless of patient BMI, LA volume, N-terminal pro-B-type natriuretic peptide, and other AF risk factors, indicating an important role for EAT in driving postoperative arrhythmogenesis.It is proposed that proinflammatory cytokines diffuse locally from dysfunctional EAT into the myocardium and contribute toward atrial myopathy, 18 which drives the risk of AF.STUDY LIMITATIONS.We did not have detailed adiposity data (eg, waist-hip ratio) or mortality data available within the SCOT-HEART trial population to investigate the exact causes of noncardiac mortality that could be driving our finding of elevated risk of West et alJ A C C : C A R D I O V A S C U L A R I M A G I N G , V O L . 1 6 , N O .6 , 2 0 2 3Automatic EAT Assessment for Risk Prediction J U N E 2 0 2 3 : 8 0 0 -8 1 6

TABLE 1
Demographics and Scan Characteristics of ORFAN Study Cohorts Used in AI Model Development and Testing Values are n (%) or median (IQR).AI ¼ artificial intelligence; BMI ¼ body mass index; CAD ¼ coronary artery disease; EAT ¼ epicardial adipose tissue; GE ¼ General Electric; NA ¼ not applicable; ORFAN ¼ Oxford Risk Factors and Noninvasive Imaging Study.

TABLE 2
Demographics and Scan Characteristics of External Clinical Cohorts Values are n (%) or median (IQR).AdipoRedOx ¼ Adipose tissue and cardiovascular RedOx regulation; SCOT-HEART ¼ Scottish Computed Tomography of the Heart; other abbreviations as in Table1.