Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2026 Aug 20.
Published in final edited form as: Circ Cardiovasc Imaging. 2026 Jul 2;19(7):e019726. doi: 10.1161/CIRCIMAGING.125.019726

Automated AI-Based Aortic Measurements from Attenuation Correction CT as an Adjunctive Cardiovascular Risk Biomarker: An International Multicenter Study

Anna M Marcinkiewicz a,b,*, Aakash Shanbhag a,c,*, Panithaya Chareonthaitawee d, Wenhao Zhang a, Hasanian Al-Jilaihawi a, Ryan Zaid a, Sebastien Cadet a, Giselle Ramirez a, Mark Lemley a, Jirong Yi a, Waseem Hijazi a, Valerie Builoff a, Joanna X Liang a, Vinicius F Calsavara e, Andrew J Einstein f, Edward Miller g, Attila Feher g, Terrence D Ruddy h, Viet T Le i, Steve Mason i, Stacey Knight i,j, Erick Alexanderson k, Isabel Carvajal-Juarez k, Leandro Slipczuk l, Mark I Travin l, Thomas L Rosamond m, Samuel Wopperer d, Daniel S Berman a, Damini Dey a, Marcelo Di Carli n, Robert JH Miller a,o, Piotr J Slomka a
PMCID: PMC13489667  NIHMSID: NIHMS2195366  PMID: 42389791

Abstract

Background:

Aortic enlargement is a powerful predictor of dissection and rupture, yet it is rarely evaluated during routine myocardial perfusion imaging, despite the widespread availability of CT attenuation correction scans. The aim was to determine whether fully automated, opportunistically derived artificial intelligence (AI)-based aortic measurements from MPI CTAC scans are associated with adverse outcomes in a large multicenter cohort.

Methods:

CTAC scans from patients undergoing PET/CT and SPECT/CT MPI across 10 centers were included. A deep learning model automatically segmented the thoracic aorta and postprocessing algorithm extracted maximum ascending and descending diameters. Aortic size index (1) was calculated by indexing diameter to body surface area.

Results:

A total of 29,339 patients (56% male, median age 66, interquartile range [IQR]:58–75 years) were included. Over a median 3.5 years follow-up (IQR:1.9–5.0), 5083 (17.3%) patients died. Median ascending and descending ASI values were 1.8 cm/m2 (IQR:1.6–2.0) and 1.5 cm/m2 (IQR:1.4–1.6), respectively, with an increase with age, and higher values in females. Elevated ASI thresholds (Ascending >2.2 cm/m2; descending >1.6 cm/m2) were significantly associated with increased all-cause mortality (ascending: adjusted HR 1.16 [95% CI 1.07–1.26], p<0.001; descending: adjusted HR 1.23 [95% CI 1.14–1.31], p<0.001). Notably, the prognostic value of abnormal ASI persisted independent of age, sex, and perfusion abnormalities.

Conclusions:

AI can unlock previously unused information within routine MPI CTAC scans by rapidly and automatically quantifying aortic size at scale. Opportunistic aortic measurements derived from CTAC may serve as an adjunctive risk biomarker and could support the prognostic value of standard MPI without additional imaging or radiation.

Keywords: artificial intelligence, myocardial perfusion, computed tomography attenuation correction, single-photon emission computed tomography, aortic diameter

Graphical Abstract

graphic file with name nihms-2195366-f0001.webp

Tweet:

Routine CT attenuation scans, acquired with cardiac PET or SPECT can be used to automatically measure aortic diameters with AI - an opportunity to extract an additional risk biomarker from the images we already have.

#HybridMyocardialImagingPerfusion #AI #AorticDiameter

Introduction

Myocardial perfusion imaging (MPI) is among the most commonly performed noninvasive tests for the diagnosis and assessment of coronary artery disease1,2. As part of standard practice, low resolution non-contrast computed tomography (CT) attenuation correction (CTAC) scans are routinely acquired for cardiac PET and are increasingly incorporated into cardiac SPECT protocols3,4. Although historically used only to correct photon attenuation, CTAC images offer far more information than traditionally appreciated. For example, opportunistic assessment of coronary artery calcification on CTAC as emerged as an important adjunct to MPI interpretation and is now firmly integrated into the evaluation of patients with suspected CAD5.

Although the thoracic aorta is clearly visualized on the CTAC scans, these images are not routinely assessment for aortic size during MPI and are usually performed on much higher resolution contrast CT scans. This represents a missed opportunity as aortic dissection and rupture carry substantial risk of cardiovascular death (CVD)6 and all-cause mortality7,8. Aortic aneurysms are often clinically silent until discovered7–9. Importantly, the risk of catastrophic rupture rises steeply with increasing aortic diameter9,10.

While CAD remains the leading cause of mortality2,11, atherosclerosis is a systemic process that affects multiple vascular beds, including the aorta, where it contributes to aneurysm formation and progressive dilation9. Current American College of Cardiology/American Heart Association guidelines10 recommend electrocardiographically (ECG) gated, contrast-enhanced CT scans for accurate aortic assessment; however, prior work has shown that non-contrast CT using an outer-wall-to-outer-wall approach can also yield reliable diameter measurements12.

Recent advances in artificial intelligence (AI) have accelerated to automatically quantify aortic dimensions and detect aneurysms, particularly using high-quality ECG-gated, contrast-enhanced CT data sets13–16. Although some automatic approaches have been applied to non-gated, non-contrast CT scans17–19, these methods have not been extended to the CTAC images routinely acquired during cardiac PET and SPECT. To our knowledge, no prior study has leveraged ultra-low dose CTAC scans (an abundant but underutilized imaging resource) for AI-based aortic measurements and related these measurements to clinical outcomes.

Given the ubiquity of CTAC imaging in MPI, this untapped dataset offers a unique opportunity to extract meaningful vascular information at no additional time, cost, or radiation. Therefore, we aimed to evaluate whether fully automated AI-based 3D aortic diameter measurements obtained from ultra-low-dose CTAC scans are associated with CVD and All-Cause Mortality7 in a large, multicenter MPI cohort.

Methods

Data availability

To the extent allowed by data sharing agreements and IRB protocols, the deidentified data and data analysis code from this manuscript will be shared upon written request.

Code availability

The TotalSegmentator code is publicly available20, cLSTM code is publicly available under a Creative Commons BY-NC license at https://doi.org/10.5281/zenodo.10632288

Study population

We analyzed consecutive patients who underwent hybrid PET/CT or SPECT/CT MPI between 2007 and 2024 for evaluation of suspected or known CAD. Patients were drawn from 10 participating centers in the REgistry of Flow and Perfusion Imaging for Artificial Intelligence with positron emission tomography PET (REFINE PET)21 or the Registry of Fast Myocardial Perfusion Imaging with Next-Generation SPECT (REFINE SPECT)22,23 registries. Figure 1 shows the study flow diagram, including the exclusion criteria.

Figure 1. Study flow chart.

Figure 1.

Open in a new tab

Patients who underwent stress PET/CT or SPECT/CT myocardial perfusion imaging were included in the study. “No clinical data” included ≥1 variable missing among those used for hazard ratio adjustment (sex, age, hypertension, dyslipidemia, diabetes mellitus, smoking, stress total perfusion deficit, left ventricular ejection fraction). Abbreviations: CT – computed tomography; CTAC – computed tomography attenuation correction; PET – positron emission tomography; SPECT – single-photon emission computed tomography

The study was approved by institutional review board at each participating site and conducted in accordance with the Declaration of Helsinki. Sites either obtained written informed consent or waiver of consent for the use of the de-identified data. Demographic and clinical characteristics, including self-reported sex (male or female), were obtained from the respective registries22,23 CTAC acquisition parameters for all centers are summarized in Table S1.

Two clinical endpoints were evaluated 1) CVD and 2) ACM. Mortality outcomes were determined via national death certificate (United States) or administrative health databases (Mexico, Canada)21 Data on CVD were available only for patients who underwent PET/CT MPI.

MPI Analysis

Standard clinical software within the Cedars-Sinai Cardiac Suite (Cedars-Sinai Medical Center, Los Angeles) was used by trained technologists at the core laboratory (Cedars-Sinai Medical Center, Los Angeles) to quantify total perfusion deficit (TPD)24 in all studies. Stress TPD <5% was defined as normal myocardial perfusion. Additional methodological details are available in the supplementary methods.

Automated Measurements of Aortic Diameter

CTACs from stress PET MPI and stress SPECT MPI were included. A deep learning network based on nnUNet and a pre-trained model (TotalSegmentator) that was trained on 1204 retrospectively collected CT scans from another institution not included in our registry, was used to segment the aorta20,25. To measure maximal diameters of the ascending and descending aorta, an aortic centerline was first automatically generated from the aortic mask, and the aorta was subsequently segmented into ascending and descending portions. Centerline extraction was performed using a K-means clustering–based algorithm. The centerline was resampled at 3-mm intervals, and orthogonal cross-sectional planes were generated at each point. The maximal diameter measured across all orthogonal planes was defined as the aortic diameter for analysis. Differentiation between the ascending and descending aorta was based on the orientation of vectors relative to the centroid of the aortic volume, with anterior displacement classified as the ascending aorta and posterior displacement classified as the descending aorta. No additional model training was performed using the study population and therefore the analysis is an external validation on the proposed approach.

Abnormal Thresholds

The aortic size index was defined as the ratio of aortic diameter [cm]/BSA [m2]26. ASI >2.2 cm/m2 and a descending ASI >1.6 cm/m2 were considered abnormal according to 2024 ESC Guidelines for the Management of Peripheral Arterial and Aortic Diseases27. Indexing to body size is essential for appropriate risk stratification across the wide range of patient anthropometrics27. As a secondary analysis, we also report outcomes using previously established absolute (non-indexed) diameter thresholds (details provided in the supplementary materials).

Agreement with standard manual measurements

To assess concordance with experienced readers interpretation a subset of 273 consecutive patients from one of the randomly selected participating site underwent manual aortic measurements. Two experienced readers with 10 years (ML) and 3 years (WH) of imaging expertise manually measured maximal short-axis diameters of the ascending and descending aorta at the level of the pulmonary artery bifurcation12. These values were compared with the automatically derived measurements. Moreover, we used a subset of 60 paired scans (CT angiography [CTA] and CTAC), acquired within 30–40 seconds at a separate site (Mexico City). Manual measurements of aortic diameters were performed by the readers, and both inter-scan (CTA vs CTAC) and inter-reader agreement were evaluated.

Automated Coronary Artery and Thoracic Aortic Calcium Scoring

Coronary artery calcium (CAC) segmentation and scoring were performed using our previously validated deep learning model28. This framework incorporates two convolutional long short-term memory (convLSTM) networks - one for heart mask generation and one for CAC segmentation - which were externally tested on an independent dataset of 3,000 ECG-gated CT scans)29. CAC scores were derived automatically using previously established deep learning segmentation methods29.

Thoracic aortic calcium was quantified using the automated aortic volume of interest20. Voxels exceeding 130 Hounsfield Units (HU) were identified, and a connected-components algorithm was applied to exclude isolated noisy single pixels.

Statistical Analysis

Normally distributed continuous variables were expressed as mean ± standard deviation (SD), while non-normally distributed continuous variables were expressed as medians with interquartile range (IQR; [Q1-Q3). Categorical variables were presented as count and relative frequencies (percentages).

Baseline characteristics were compared between the two groups. Non-normally distributed continuous variables were compared using the Wilcoxon rank sum test, and categorical variables were compared using Pearson’s χ2 test. Kaplan-Meier survival curves were plotted for descriptive visualization, log-rank tests were used to compare unadjusted survival curves, and adjusted hazard ratios with 95% confidence intervals were estimated using Cox proportional hazards regression models. Multivariable Cox proportional hazards (PH) models were fitted to evaluate the association between the primary exposure variable and each outcome, adjusting for sex, age, hypertension, dyslipidemia, diabetes, smoking, deep learning-derived CAC and thoracic aorta calcium, stress TPD, and stress left ventricular ejection fraction. Adjusted hazard ratios (HRs) with corresponding 95% confidence intervals30 were reported. Confidence intervals were calculated using Wald-type intervals based on model-based standard errors. The PH assumption was assessed using Schoenfeld residuals and formally tested with the Grambsch-Therneau test and was found to be valid in all analyses.31 CVD and ACM were analyzed as separate time-to-event outcomes because they address distinct clinical questions; therefore, no formal multiplicity adjustment was applied across these analyses.

To account for potential heterogeneity between the SPECT and PET populations, shared frailty Cox models were fitted with imaging modality specified as the clustering variable. The shared frailty term was assumed to follow a gamma distribution with mean 1 and variance θ, where θ quantifies the degree of between-modality heterogeneity.

In secondary analyses, ascending and descending ASI were modeled as continuous variables using natural cubic spline functions with 3 degrees of freedom to assess potential non-linear associations with ACM. To assess the incremental prognostic value of ASI beyond established clinical and imaging risk factors, nested Cox PH models were fitted. Extended models additionally incorporated ascending ASI, descending ASI, or both. Model discrimination was evaluated using Harrell’s concordance index (C-index), and improvement in model fit was assessed using likelihood ratio testing and continuous net reclassification index (NRI) between nested models. Cardiovascular mortality was only available in the PET MPI population. The agreement between inter-observer manual measurements of aortic diameter and the inter-method comparison (AI vs manual measurements) was assessed using Bland–Altman analysis. For each paired comparison, the mean difference (bias) and 95% limits of agreement (LOA) were computed as bias ± 1.96×SD of the paired differences. For each paired comparison (Reader 1 vs Reader 2; AI vs Reader 1; AI vs Reader 2), paired differences were calculated as the difference between measurements obtained by the two methods (measurementA − measurementB). The mean paired difference was defined as the systematic bias. Statistical significance of the bias was assessed using a one-sample t-test comparing the mean paired difference to zero. Differences in measurement variability between paired methods were assessed using the Morgan–Pitman test. Reliability was quantified using the intraclass correlation coefficient (ICC) with a two-way random-effects model for absolute agreement for each pairwise comparison, including inter-reader agreement (Reader 1 vs. Reader 2) and inter-method agreement (AI-derived vs. manual measurements for Reader 1 and Reader 2), with corresponding 95% confidence intervals. ICC values < 0.50 indicated poor reliability, 0.50–0.75 moderate reliability, 0.75–0.90 good reliability, and > 0.90 excellent reliability.31 In an additional analysis, we assessed whether absolute diameters of larger than 45 mm were associated with clinical outcomes in younger patients. All hypothesis tests were performed as two-sided tests at a 5% significance level. Statistical analysis was performed with Python 3.11.5 (Python Software Foundation, Wilmington, DE, USA), and R version 4.3.2 (R Foundation for Statistical Computing, Vienna, Austria).

Results

Patient Characteristics

Study population

A total of 29,963 scans from 10 sites were screened. Of these, 25 (0.08%) CTAC scans were missing, 212 (0.7%) were excluded due to AI-segmentation failure, and 387 (1.3%) were excluded due to incomplete clinical data (≥1 missing variable required for multivariable modeling or absent height/weight) (Figure 1). The final analytic cohort comprised 29,339 patients. The distribution of the CTAC coverage based on modality is shown in Figure S1.

Baseline characteristics for the final population stratified by sex are shown in Table 1. Overall, males represent 56.1% of all participants, and the median age was 66 (IQR 58, 75) years. During a median follow-up of 3.5 (IQR 1.9, 5.0) years, 5083 (17.3%) patients died. Baseline characteristics stratified by ACM and by imaging modality are provided in Tables S2 and S3, respectively.

Table 1.

Baseline characteristics for all participants stratified by sex

All Male Female p-value Effect size
N (%) 29,339 16,455 (56.1) 12,884 (43.9)
Age [years] 66.0 (58.0, 75.0) 66.0 (58.0, 74.0) 67.0 (58.0, 75.0) <0.001* −0.04
BMI [kg/m2] 29.8 (25.7, 35.2) 29.4 (25.8, 34.3) 30.4 (25.5, 36.3) <0.001* −0.13
BSA 2.0 (1.8, 2.2) 2.1 (1.9, 2.3) 1.8 (1.7, 2.0) <0.001* 0.25
Hypertension 21790 (74.6) 12346 (75.4) 9444 (73.7) <0.001* 0.04
Diabetes mellitus 9837 (33.6) 5675 (34.5) 4162 (32.3) <0.001* 0.05
Dyslipidemia 19032 (64.9) 11179 (68.3) 7853 (61.3) <0.001* 0.15
Smoking 5381 (18.4) 3421 (20.9) 1960 (15.3) <0.001* 0.15
Family history of CAD 8746 (30.2) 4660 (28.7) 4086 (32.1) <0.001* −0.07
Prior CAD
 Prior Myocardial Infarction 4404 (15.0) 3013 (18.3) 1391 (10.8) <0.001* 0.21
 Past PCI 5431 (18.5) 3900 (23.7) 1531 (11.9) <0.001* 0.31
 Past CABG 2497 (8.5) 1949 (11.9) 548 (4.3) <0.001* 0.28
 Post TAVR 183 (0.8) 116 (0.8) 67 (0.6) <0.001* 0.02
 Post other open heart surgery 714 (3.0) 537 (3.9) 177 (1.7) <0.001* 0.13
ACM 5083 (17.3) 3065 (18.6) 2018 (15.7) <0.001* 0.08
ACM, 5-year follow-up 3.5 (1.9, 5.0) 3.5 (1.9, 5.0) 3.6 (2.0, 5.0) <0.001* 0.14
CT Quantitative Image Analysis Parameters
DL maximal ascending aorta diameter [mm] 35.9 (32.4, 39.2) 37.4 (37.1, 40.6) 34.1 (30.8, 37.0) <0.001* 0.36
DL maximal descending aorta diameter [mm] 29.3 (27.3, 31.6) 30.5 (28.7, 32,6) 27.8 (26.2, 29.7) <0.001* 0.49
Ascending Aortic Size Index [cm/m2] 1.8 (1.6, 2.0) 1.8 (1.6, 2.0) 1.9 (1.6, 2.1) <0.001* −0.10
Descending Aortic Size Index [cm/m2] 1.5 (1.4, 1.6) 1.5 (1.3, 1.6) 1.5 (1.4, 1.7) <0.001* −0.14
DL CAC score (log-CAC) 2.0 (0.0, 2.9) 2.5 (1.0, 3.2) 1.2 (0.0, 2.5) <0.001* 0.33
DL Thoracic Aorta Calcium 735.1 (100.1, 2725.9) 887.3 (134.6, 3058.0) 570.6 (66.5, 2284.6) <0.001* 0.10
MPI Acquisition and Quantitative Image Analysis Parameters
Stress Total Perfusion Deficit 4.0 (1.5, 9.4) 4.8 (1.9, 11.8) 3.2 (1.2, 7.2) <0.001* 0.18
Left Ventricular Ejection Fraction 66.0 (55.4, 74.3) 61.5 (50.6, 69.4) 72.1 (63.2, 79.1) <0.001* −0.45
Open in a new tab
*

significant p-values

Values are presented as N (%) or median (IQ1,IQ3). Significant p-values are in bold. Continuous variables were compared using the Wilcoxon rank sum test whereas the categorical variables were compared using Pearson’s Chi-squared test.

Effect sizes were estimated using standardized mean differences for non-skewed measures, for skewed distributions for all MPI acquisition and CT based parameters – Cliff’s delta was calculated. The interpretation of Cliff’s delta was as follows: negligible < 0.15, small 0.15, medium 0.33, and large 0.47.32

ACM – all-cause mortality; BMI – body mass index; BSA – body surface area; CABG – coronary artery bypass graft; CAC – coronary artery calcium; CAD – coronary artery disease; CT – computed tomography; DL – deep learning; MPI – myocardial perfusion imaging; N – number of patients; PCI – percutaneous coronary intervention; TAVR – transcatheter aortic valve replacement

Among PET/CT patients, CVD were available for 15,488 subjects, of whom 1,511 (9.8%) died from cardiovascular causes. Table S4 summarizes the baseline characteristics of PET/CT cohort stratified by CVD status

Automatic Aortic Measures

Distributions of ASI for the ascending and descending aorta stratified by sex in all patients are shown in Figure 2, with corresponding distributions of maximal aortic diameter distributions provided in Figure S2.

Figure 2. Distributions of Aortic Size Index (ASI) from CTACs in all patients.

Figure 2.

Open in a new tab

ASI of the ascending and descending aorta stratified by sex. Abbreviations: CTAC – computed tomography attenuation correction

In the overall cohort, the median ASI for the ascending aorta was 1.8 (1.6, 2.0) cm/m2, and the median descending ASI was 1.5 (1.4, 1.6) cm/m2. The median maximal diameters were 35.9 (32.4, 39.2) mm for the ascending aorta and 29.3 (27.3, 31.6) mm for the descending aorta.

Age- and sex-specific ASI percentiles are shown in Table S5 and Figure 3, with corresponding absolute-diameter percentiles in Table S6 and Figure S3. Both ASI and absolute diameters increased progressively with age in males and females, with generally higher values in female patients. For example, among males, the mean ascending ASI Increased from 1.6 in those younger than 60 years to 2.0 in those older than 80 years.

Figure 3. Distribution of ascending and descending aortic size index.

Figure 3.

Open in a new tab

The distribution of ASI as a function of sex and age is shown for all patients. ASI increases with age in female (N=12884) and male (N=16455) patients, with overall higher values in female patients

Comparison with manual measurements

Inter-observer and inter-method (manual vs AI-driven) comparison of aortic diameter measurements is shown in Table S7, whereas Figure S4 presents Bland-Altman plots comparing manual (2D) measurements with AI-derived 3D aortic diameter measurements on CTAC scans. For both readers, the manual and AI-based automatic measurements showed good reliability for the ascending aorta. For the descending aorta, reliability between reader 1 and the automatic measurements was moderate, whereas reliability between manual and automatic measurements for reader 2 was good. For the descending aorta, the mean difference between manual and AI-based measurements for reader 1 was −1.91 mm (95% LOA −6.17, 2.34), while for reader 2, −2.37 mm (95% LOA −5.33, 0.59) (Table S7). The inter-reader agreement is strong with near-zero bias.

To assess inter-scan agreement of the aortic diameter measurements, we included 60 paired scans (CTA and CTAC) from a separate site (Mexico City) with a mean age of 62.5 ± 11.3 years; 65% were male. On CTAs, the median ascending and descending aortic diameters were 33.5 (29.3, 36.0) mm and 25.8 (23.3, 27.4) mm, respectively. On CTACs, the median ascending and descending aorta diameters were 35.2 (32.8, 37.0) mm, and 28.4 (26.4, 31.3) mm, respectively. An inter-scan (CTA vs CTAC) agreement is presented in Table S8 and Figure S5. Figure S6 shows an inter-reader agreement for manual aortic diameter measurements on CTA (Figure S6A) and CTAC (Figure S6B). For inter-scan agreement between CTA and CTAC, the mean difference for reader 1 was 2.27 mm (95% LOA: −1.58 to 6.05) for the ascending aorta and 1.88 mm (95% LOA: −1.58 to 5.34) for the descending aorta.

All-Cause Mortality Prediction

Figure 4 displays Kaplan-Meier curves stratified by ASI in males and females. Figure S7 shows Kaplan-Meier curves stratified by ASI and maximal absolute aortic diameter in the overall cohort. In males, abnormal ASI identified a significantly higher ACM than those with abnormal absolute diameters for both the ascending and descending aorta (p<0.001, for both).

Figure 4. Kaplan-Meier (KM) curves for mortality in males and female patients.

Figure 4.

Open in a new tab

Survival stratified by ascending and descending aortic size index in males and females. Abnormal ascending ASI was defined as >2.2 cm/m2 ascending aorta and >1.6 cm/m2 for the descending aorta. Adjusted hazard ratios were adjusted by sex, age, hypertension, dyslipidaemia, diabetes mellitus, smoking, coronary artery calcium, thoracic aortic calcium, stress perfusion, and left ventricular ejection fraction. Modality was specified as the clustering variable in the shared frailty Cox models. Adjusted hazard ratios and 95% confidence intervals were estimated using multivariable Cox proportional hazards regression models. Abbreviations: CI – confidence interval; HR – hazard ratio; HRadj – adjusted hazard ratio

Table 2 summarizes the associations between ASI and mortality for the entire cohort and stratified by sex, age (<65 and ≥65 years), and perfusion category (TPD <5% and ≥5%). Table S9 shows the association between ASI and all-cause mortality stratified by site. Among patients aged <65 years, aortic diameter ≥45 mm (ascending or descending) was present in 3.4% of the cohort and was not associated with increased mortality (7.8% vs 10.3%; HR 1.03, 95% CI 0.73–1.44, p=0.88). Similarly, ascending aortic diameter ≥45 mm (3.3% of the cohort) was not associated with mortality (7.4% vs 10.3%; HR 0.99, 95% CI 0.70–1.42, p=0.97). A stepwise adjustment of the association between ACM and both ascending and descending ASI is presented in Table S10. Both abnormal ascending and descending ASI were independently associated with a significantly higher risk of death after adjustment for clinical data; clinical and MPI metrics; and clinical and MPI metrics combined with CAC and thoracic calcium (Table 2 and Table S10). Associations between ACM and ascending and descending ASI in patients undergoing PET/CT and SPECT/CT are presented in Table S11 and Table S12, respectively. In the PET/CT patients, abnormal ascending and descending ASI were both associated with a higher risk of death (adjusted HR 1.22, 95% CI 1.11–1.34, p<0.001, and adjusted HR 1.28, 95% CI 1.19–1.37, p<0.001, respectively). Similarly, in the SPECT/CT subgroup, abnormal ascending and descending ASI were associated with increased mortality risk (adjusted HR 1.37, 95% CI 1.14–1.64, p<0.001, and adjusted HR 1.43, 95% CI 1.20–1.71, p<0.001, respectively). Table S13 shows the incremental prognostic value of ASI beyond established clinical and imaging risk factors, with nested Cox models were compared. Addition of ascending ASI resulted in a model fit improvement (likelihood-ratio chi-square [LR χ2]=24.48, p<0.001) while the addition of descending ASI resulted in a modest increase in discrimination, with similar findings when both ASI indicators were included (Δ=0.0051; LR χ2=70.58, p<0.001). Modeling ASI as continuous variables resulted in a slightly greater improvement in discrimination (C-index 0.728 vs 0.737; Δ=0.0088; LR χ2=111.28, p<0.001). NRI analyses demonstrated statistically significant but modest improvements in risk classification. Subgroup analyses showed no significant modification by sex, but there was a significant interaction between TPD and descending ASI (interaction p=0.003) as seen in Table S14. High ASI for the descending aorta was independently associated with all-cause mortality regardless of age, with high ASI for ascending aorta significantly associated with all-cause mortality patients under age 65.

Table 2.

Association with ACM for ascending ASI >2.2 cm/m2 and descending ASI >1.6 cm/m2 in all patients

Ascending Aortic Size Index [cm/m2] Descending Aortic Size Index [cm/m2]
Unadjusted HR, 95% CI p-value Adjusted HR, (95% CI) p-value Unadjusted HR, 95% CI p-value Adjusted HR, (95% CI) p-value
All Patients (N=29,339)
1.58 (1.46, 1.70) <0.001* 1.16 (1.07, 1.26) 0.001* 2.13 (2.02, 2.25) <0.001* 1.23 (1.14, 1.31) <0.001*
Subgroups
Male (N=16,455) 1.65 (1.48, 1.85) <0.001* 1.18 (1.04, 1.32) 0.007* 2.24 (2.09, 2.41) <0.001* 1.25 (1.15, 1.36) <0.001*
Female (N=12,884) 1.64 (1.48, 1.83) <0.001* 1.15 (1.03, 1.32) 0.013* 2.17 (1.98, 2.37) <0.001* 1.17 (1.05, 1.31) 0.005*
<65 years (N=12,858) 1.41 (1.12, 1.77) 0.267 1.54 (1.21, 1.97) 0.001* 1.83 (1.60, 2.09) <0.001* 1.41 (1.22, 1.65) <0.001*
≥65 years (N=16,481) 1.28 (1.18, 1.38) <0.001* 1.08 (0.99, 1.19) 0.082 1.69 (1.58, 1.80) <0.001* 1.17 (1.08, 1.26) <0.001*
TPD <5% (N=12,706) 1.72 (1.54, 1.93) <0.001* 1.19 (1.05, 1.35) 0.008* 2.34 (2.15, 2.55) <0.001* 1.34 (1.20, 1.49) <0.001*
TPD ≥5% (N=16,633) 1.58 (1.42, 1.75) <0.001* 1.14 (1.02, 1.28) 0.021* 2.05 (1.91, 2.20) <0.001* 1.17 (1.07, 1.28) <0.001*
Open in a new tab

Adjusted model included sex†, age, hypertension, dyslipidemia, diabetes mellitus, smoking, DL coronary artery calcium, DL thoracic aortic calcium, stress TPD‡, left ventricular ejection fraction. Modality was specified as the clustering variable in the shared frailty Cox models. Adjusted hazard ratios and 95% confidence intervals were estimated using multivariable Cox proportional hazards regression models.

*

significant p-values,

†

except for HR stratified by sex,

‡

except for HR stratified by TPD

Abbreviations: ACM – all-cause mortality; ASI – aortic size index; CI – confidence interval; DL – deep learning; HR – hazard ratio; TPD – total perfusion deficit

Figure 5 illustrates the continuous nonlinear relationships between unadjusted HR and a range of ascending and descending ASI threshold, stratified by sex. Figure 6 provides representative examples of patients with abnormal (Figure 6A–C) and normal (Figure 6D–F) aortic measurements.

Figure 5. Association with death for ascending and descending aortic size index for males and females.

Figure 5.

Open in a new tab

Unadjusted hazard ratio (solid red line) and 95% confidence intervals (dotted blue line), derived from a restricted quadratic spline curve, illustrate how ascending and descending ASI related to risk of death in male and female subgroups. Abnormal ascending ASI was defined as >2.2 cm/m2 ascending aorta and >1.6 cm/m2 for the descending aorta. Vertical dashed orange line - threshold from 2024 ESC guidelines (27) Abbreviations: ASI – aortic size index

Figure 6. Examples of AI-based aortic evaluation from CT attenuation correction scan.

Figure 6.

Open in a new tab

Top: A 68-year-old male with an abnormal aortic size index (1) of the ascending and descending aorta, axial view (A) with a corresponding AI segmentation overlay (B). C. 3D reconstruction of the segmented aorta. Bottom: A 59-year-old female with normal ascending and descending ASI. Axial view (D) with a corresponding segmentation overlay (E), and a 3D reconstruction of the segmented aorta (F). Abbreviations: AI – artificial intelligence

Cardiovascular Death Prediction

Kaplan-Meier curves for CVD stratified by ASI in males and females who underwent PET/CT are shown in Figure S8. The association between ASI and CVD risk across sex, age, and perfusion categories are detailed in Table S15.

Among PET/CT patients, abnormal ASI was strongly associated with increased CVD risk. In unadjusted analyses, abnormal ascending ASI HR 2.07 [1.79, 2.38]) and abnormal descending ASI 2.30 [2.07, 2.54] were each linked to significantly higher CVD (p<0.001 for both). Elevated risk persisted after multivariable adjustment. In patients aged ≥65 years, abnormal ascending ASI remained independently associated with CVD (adjusted HR ascending ASI 1.51 [1.29, 1.77], as did abnormal descending ASI (adjusted 1.685 [1.49, 1.91], p<0.001 for both).

Figure S9 represents the association with CVD for ascending and descending ASI stratified by sex.

Discussion

In this large, multicenter study, we developed and validated an opportunistic, fully automated approach for quantifying thoracic aortic dimensions on ultra-low-dose, non-contrast CTAC scans obtained during routine hybrid MPI. Despite the ubiquity of CTAC imaging in clinical practice, these scans have not been leveraged for aortic diameter assessment to date. In the present study, we specifically investigated the impact of aortic diameter derived from CTAC scans on mortality prediction, a well understood biomarker which can be easily interpreted by physicians. Using a deep-learning framework that generates 3D aortic segmentations combined with postprocessing to derive centerline-based measurements, we demonstrated that automated extraction of both ascending and descending aortic diameters is feasible, rapid, and scalable across multiple sites and MPI modalities. Importantly, abnormal indexed aortic dimensions were associated with higher risk of CVD and ACM, independent of clinical risk factors and MPI findings. To our knowledge, this is the first study to apply automated aortic measurement to CTAC imaging and the first to demonstrate its adjunctive value as an additional risk biomarker in a broad, consecutive population undergoing MPI. With > 29,000 CTAC scans analyzed, this also represents the largest CTAC-based aortic evaluation performed to date. Although the incremental improvement in discrimination was significant, its magnitude was small after adjustments for sex, age, hypertension, dyslipidemia, diabetes mellitus, smoking, DL coronary artery calcium, DL thoracic aortic calcium, stress TPD, left ventricular ejection fraction, suggesting that CTAC-derived ASI should be considered a complementary risk marker rather than a standalone tool for clinical decision-making.

In addition to its prognostic significance, the AI-based framework demonstrated excellent technical performance. Automated measurements were successfully generated in more than 99% of CTAC scans, despite the ultra-low-dose nature, variable coverage, and heterogeneous acquisition protocols across 10 international sites and two imaging modalities. Agreement with manual 2D measurements was strong, with mean bias of ~3mm between, which is consistent with previous studies comparing 2D and 3D manual segmentation33. Importantly, the automated algorithm differentiated ascending from descending aortic segments and provided consistent results across age, sex, and body-size subgroups. These findings demonstrate the feasibility of AI-driven aortic evaluation and suggest that such approaches may facilitate more efficient and standardized assessment within MPI workflows.

Aortic diameter is a well-established determinant of aortic dilatation, aneurysm formation, and rupture risk9,10. Aortic dilatation is second only to atherosclerosis as the most common aortic pathology27 with prevalence exceeding 80% in selected populations such as patients with bicuspid aortic valve disease34. Given the wide availability of CTAC in routine MPI, our findings highlight a novel opportunity to incorporate automated aortic size measurement into standard MPI reporting without additional scanning time, contrast, or radiation. Identifying patients with unrecognized aortic enlargement could prompt more intensive management of modifiable risk factors, such as hypertension and smoking, and inform decisions regarding follow-up imaging. Nearly 75% of participants in our cohort had hypertension, underscoring a high-risk environment9 in which early detection of subclinical aortic enlargement could have meaningful clinical impact. This approach parallels the successful integration of coronary calcium scoring on CTAC5 and reflects the growing role of opportunistic imaging, in which existing scans are repurposed to enhance cardiovascular risk assessment.

Aortic dimensions are influenced by multiple factors, including age, sex, hypertension, smoking, and genetic predisposition10,35. Consistent with prior studies6,36 we observe larger absolute aortic diameters in men, whereas females more commonly reached higher index values later in life. These findings underscore the importance of indexed measurements - such as ASI - particularly in heterogeneous patient populations. The progressive increase in both absolute and indexed diameters with advancing age further supports routine, opportunistic evaluation of aortic size, especially among asymptomatic individuals undergoing MPI for other indications.

Although absolute aortic diameter has traditionally been used in clinical practice27 indexed measures - including ASI and height index - provide more accurate risk stratification and are recommended in several clinical contexts, including the 2024 ESC guidelines for conditions such as Turner syndrome26,27,37. Notably, in a restricted analysis to younger patients (<65 years), there was no association between the absolute aortic diameter thresholds (≥45 mm) and mortality, whereas ASI demonstrated a consistent association with outcomes. This finding aligns with previous studies reporting that non-indexed aortic diameters were not associated with all-cause mortality in contrast to BMI-indexed measures.6,38 The absence of an observed association may reflect differences in follow-up, competing risks, and most importantly the impact of clinical management including surgical intervention. Additionally abnormal ASI was associated with increased mortality in both sexes and across age strata, and these associations persisted after adjusting for clinical and imaging covariates, including perfusion abnormalities. Abnormal ASI measures were independently associated with all-cause mortality and provide additional prognostic information, the incremental improvement in risk prediction beyond established clinical and imaging variables was modest.

The aortic wall is typically 1–3 mm thick39. However, on non-contrast images, the wall cannot be distinguished, and therefore the reported values reflect the total aortic diameter, including the wall38. In addition, the aortic diameter varies by approximately 1–2 mm between systole and diastole38. Prior automated aortic measurements studies have focused primarily on ECG-gated, contrast enhanced CT angiography or diagnostic chest CT12–14,40, with fewer efforts applying automation to non-contrast, non-gated CT17–19. To our knowledge, no previous study has leveraged the substantially lower dose CTAC images used for attenuation correction in hybrid MPI. The mean inter-scan differences in manual measurements of ascending and descending aortic diameters ranged from 1.88 to 3.15 mm. These differences likely reflect the aortic wall thickness and variation in cardiac cycle phase in non-ECG-gated CTACs. The present findings demonstrate that clinically meaningful aortic measurements can be reliably extracted from the scans, expanding the potential reach of automated aortic assessment to the large global population undergoing MPI.

Study Limitations

This study has several limitations. First, we used ultra-low-dose, non-ECG gated CTAC scans, which may affect accuracy and could lead to overestimation of aortic dimensions in some cases. However, the objective was to evaluate the prognostic significance of aortic measurements that can be obtained opportunistically and quickly, without additional scan time, contrast, or radiation exposure. Nonetheless, further investigation of AI-derived aortic measurements is needed to minimize the risk of unnecessary downstream testing. Second, ASI was calculated using BSA, which may vary substantially throughout life and could influence risk assessment, particularly in obese individuals. Third, because CTAC often provides limited anatomic coverage, the aortic root, sinotubular junction, and, in some cases, part of the aortic arch may have been included with the ascending or descending aortic segments, potentially affecting anatomic position. Fourth, both the REFINE PET and REFINE SPECT registries lacked cause-specific mortality data. CVD cause-of-death information was collected in the REFINE PET registry, but this information was not collected in the REFINE SPECT registry; therefore, these patients could not be included in CVD-related analyses. We also did not have data on genetic or congenital contributors to aortopathy, such as aortic valve morphology, connective tissue disorders, or bicuspid valve status, nor information regarding subsequent management of aortopathy or aortic complications. Moreover, the heterogeneity between the PET and SPECT cohorts, including differences in baseline risk profiles, imaging characteristics, and follow-up duration, may influence the observed associations. Shared frailty modeling was used in an attempt to account for this heterogeneity. However, our finding that ASI was independently associated with cardiovascular outcomes in two separate multicenter registries suggest that the results are broadly generalizable. What is more, the multivariable models incorporated key clinical and imaging parameters. However, the observational, registry-based design of the study may still be susceptible to residual confounding. Unmeasured factors related to treatment selection, referral patterns, and comorbidity burden may have influenced both imaging findings and outcomes. Therefore, the observed associations should be interpreted as reflecting prognostic associations rather than definitive evidence of causal independence. Future prospective or randomized studies will be required to more clearly establish causal effects. Additionally, automated measurements were derived from 3D volumetric assessment with orthogonal diameter extraction, rather than the conventional 2D axial measurements performed on non-contrast CT. This methodological difference is likely the primary contributor to the observed discrepancies between AI- and expert-derived measurements.

Conclusions

Opportunistic AI-based aortic diameter measurements can be rapidly obtained from low resolution non-contrast ungated CTAC scans in patients undergoing MPI. These measurements represent an adjunctive imaging biomarker of cardiovascular risk. Incorporating routine aortic size assessment into MPI interpretation could help leverage the clinical utility of CTAC and potentially contribute to more comprehensive cardiovascular risk stratification. However, the clinical relevance of its incremental prognostic value warrants further validation.

Supplementary Material

Supplemental Table of Contents
Supplemental Methods
Supplemental Tables
Supplemental Figure S1
Supplemental Figure S2
Supplemental Figure S3
Supplemental Figure S4
Supplemental Figure S5
Supplemental Figure S6
Supplemental Figure S8
Supplemental Figure S9
Supplemental Figure S7

Supplemental Methods

Tables S1–S15

Figures S1–S9

References 41

Clinical perspective.

Hybrid imaging enhances the diagnostic accuracy of myocardial perfusion imaging (MPI) through the use of computed tomography attenuation correction (CTAC) scans. Beyond attenuation correction, CTACs offer additional insights into other structures and conditions, including aortic dilation, that may not be the primary focus of the examination. However, this information is not routinely utilized in clinical practice. We show that aortic size can be automatically quantified from CTACs using AI, enabling systematic evaluation in all patients undergoing hybrid MPI. Integration of AI-derived aortic measurements from the CTACs may improve cardiovascular risk assessment and identify patients who could benefit from further evaluation and surveillance.

Funding

This research was supported in part by grants R01HL089765 and R35HL161195 from the National Heart, Lung, and Blood Institute at the National Institutes of Health (PI: Piotr Slomka). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. The funders played no role in study design, collection, analysis, and interpretation of data, as well as in the writing of the report. The decision of the paper submission for publication was not influenced by any of the funders.

Disclosures

Dr. Marcinkiewicz received consulting fees from APQ Health. Prof. Chareonthaitawee received consulting fees from GE Healthcare, Cardiovascular Clinical Sciences, and IBA, and royalties from UpToDate. Prof. Al-Jilaihawi is a consultant to and has received institutional grants for clinical research in Pi-Cardia and holds stock options; holds equity and serves on the scientific advisory board for DASI Simulations; and has served as a consultant to Abbott, Edwards Lifesciences and Medtronic. Prof. Ruddy received research grant support from Siemens Medical Systems and Pfizer Global. Prof. Einstein has received speaker fees from Ionetix, consulting fees from Artrya and W. L. Gore & Associates, and authorship fees from Wolters Kluwer Healthcare. Prof. Einstein has also served on scientific advisory boards for Canon Medical Systems and Synektik S.A. and received grants to Columbia University from Alexion, Attralus, BridgeBio, Canon Medical Systems, Eidos Therapeutics, Intellia Therapeutics, International Atomic Energy Agency, Ionis Pharmaceuticals, National Institutes of Health, Neovasc, Pfizer, Roche Medical Systems, Shockwave Medical, and W. L. Gore & Associates. Dr. E. Miller received grant support from Pfizer, ARGO SPECT, Alnylam, Siemens Medical Systems and the National Institutes of Health. Dr. E. Miller received consulting fees from Pfizer, Alnylam, Synektik, and Eidos/BioBridge. Dr. Slipczuk has received institutional grants from Amgen and Philips. Prof. Di Carli received consulting fees from MedTrace, Valo Health and IBA and institutional grant support from Sun Pharma, Xylocor and Intellia. Prof. Dey, Prof. Slomka, and Dr. Slipczuk, and Prof. Berman declare equity interest in APQ Health. Prof. Slomka, and Prof. Berman participate in software royalties for QPS software at Cedars-Sinai Medical Center. Prof. Berman received research grant support from the Dr. Miriam and Sheldon G. Adelson Medical Research Foundation, served as a consultant for GE Healthcare. Dr. Robert Miller received consulting fees from Alnylam and Bayer and research support from Alberta Innovates. Prof. Slomka received research grant support from Siemens Medical Systems and consulting fees from Synektik S.A. Viet T Le has received research grant support from J&J/Janssen, has received honorarium from the American College of Cardiology for Editor-in-Chief role at Cardiosmart, and has served on advisory boards for Amgen, Amarin, Bayer, Boehringer Ingelheim, Esperion, Idorsia, iRhythm, Merck, Novartis, Novonordisk, and Pfizer. Other authors declare no competing interests.

List of abbreviations

AI

artificial intelligence

BSA

body surface area

CTAC

computed tomography attenuation correction

CVD

cardiovascular death

MPI

myocardial perfusion imaging

PET

positron emission tomography

SPECT

single-photon emission computed tomography

ASI

aortic size index

References

  • 1.Al-Mallah MH, Ahmed AI, Nabi F, Chang SM, Kleiman NS, Chamsi-Pasha MA, Shah A, Han Y, Zoghbi WA, Mahmarian JJ. Outcomes of patients with moderate-to-severe Ischemia excluded from the ischemia trial. J Nucl Cardiol. 2022;29:1100–1105. [DOI] [PubMed] [Google Scholar]
  • 2.Gulati M, Levy PD, Mukherjee D, Amsterdam E, Bhatt DL, Birtcher KK, Blankstein R, Boyd J, Bullock-Palmer RP, Conejo T, et al. 2021 AHA/ACC/ASE/CHEST/SAEM/SCCT/SCMR Guideline for the Evaluation and Diagnosis of Chest Pain: Executive Summary: A Report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. Circulation. 2021;144:e368–e454. [DOI] [PubMed] [Google Scholar]
  • 3.Heller GV, Links J, Bateman TM, Ziffer JA, Ficaro E, Cohen MC, Hendel RC. American Society of Nuclear Cardiology and Society of Nuclear Medicine joint position statement: attenuation correction of myocardial perfusion SPECT scintigraphy. J Nucl Cardiol. 2004;11:229–230. [DOI] [PubMed] [Google Scholar]
  • 4.Pazhenkottil AP, Ghadri JR, Nkoulou RN, Wolfrum M, Buechel RR, Kuest SM, Husmann L, Herzog BA, Gaemperli O, Kaufmann PA, et al. Improved outcome prediction by SPECT myocardial perfusion imaging after CT attenuation correction. J Nucl Med. 2011;52:196–200. [DOI] [PubMed] [Google Scholar]
  • 5.Vrints C, Andreotti F, Koskinas KC, Rossello X, Adamo M, Ainslie J, Banning AP, Budaj A, Buechel RR, Chiariello GA, et al. 2024 ESC Guidelines for the management of chronic coronary syndromes. Eur Heart J. 2024;45:3415–3537. [DOI] [PubMed] [Google Scholar]
  • 6.Rueda-Ochoa OL, Bons LR, Zhu F, Rohde S, El Ghoul K, Budde RP, Ikram MK, Deckers JW, Vernooij MW, Franco OH, et al. Thoracic aortic diameter and cardiovascular events and mortality among women and men. Radiology. 2022;304:208–215. [DOI] [PubMed] [Google Scholar]
  • 7.Marcinkiewicz AM, Buchwald M, Shanbhag A, Bednarski BP, Killekar A, Miller RJ, Builoff V, Lemley M, Berman DS, Dey D, et al. AI for multistructure incidental findings and mortality prediction at chest CT in lung cancer screening. Radiology. 2024;312:e240541. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Isselbacher EM. Thoracic and abdominal aortic aneurysms. Circulation. 2005;111:816–828. [DOI] [PubMed] [Google Scholar]
  • 9.Munden RF, Carter BW, Chiles C, MacMahon H, Black WC, Ko JP, McAdams HP, Rossi SE, Leung AN, Boiselle PH, et al. Managing Incidental Findings on Thoracic CT: Mediastinal and Cardiovascular Findings. A White Paper of the ACR Incidental Findings Committee. J Am Coll Radiol. 2018;15:1087–1096. [DOI] [PubMed] [Google Scholar]
  • 10.Isselbacher EM, Preventza O, Hamilton Black J, Augoustides JG, Beck AW, Bolen MA, Braverman AC, Bray BE, Brown-Zimmerman MM, Chen EP, et al. 2022 ACC/AHA Guideline for the Diagnosis and Management of Aortic Disease: A Report of the American Heart Association/American College of Cardiology Joint Committee on Clinical Practice Guidelines. Circulation. 2022;146:e334–e482. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.Knuuti J, Wijns W, Saraste A, Capodanno D, Barbato E, Funck-Brentano C, Prescott E, Storey RF, Deaton C, Cuisset T, et al. 2019 ESC Guidelines for the diagnosis and management of chronic coronary syndromes: The Task Force for the diagnosis and management of chronic coronary syndromes of the European Society of Cardiology (ESC). Eur Heart J. 2020;41:407–477. [DOI] [PubMed] [Google Scholar]
  • 12.Wolak A, Gransar H, Thomson LE, Friedman JD, Hachamovitch R, Gutstein A, Shaw LJ, Polk D, Wong ND, Saouaf R, et al. Aortic size assessment by noncontrast cardiac computed tomography: normal limits by age, gender, and body surface area. JACC Cardiovasc Imaging. 2008;1:200–209. doi: 10.1016/j.jcmg.2007.11.005 [DOI] [PubMed] [Google Scholar]
  • 13.Lo EMK, Chen S, Ng KHL, Wong RHL. Artificial intelligence-powered solutions for automated aortic diameter measurement in computed tomography: a narrative review. Ann Transl Med. 2024;12:116. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Boninsegna E, Piffer S, Simonini E, Romano M, Lettieri C, Colopi S, Barai G. CT angiography prior to endovascular procedures: can artificial intelligence improve reporting? Phys Eng Sci Med. 2024;47:643–649. [DOI] [PubMed] [Google Scholar]
  • 15.Martinez-Mera JA, Tahoces PG, Carreira JM, Suarez-Cuenca JJ, Souto M. Automatic characterization of thoracic aortic aneurysms from CT images. Comput Biol Med. 2015;57:74–83. [DOI] [PubMed] [Google Scholar]
  • 16.Zanfardino M, Punzo B, Maffei E, Saba L, Bossone E, Nistri S, La Grutta L, Franzese M, Cavaliere C, Cademartiri F. Unsupervised machine learning for risk stratification and identification of relevant subgroups of ascending aorta dimensions using cardiac CT and clinical data. Comput Struct Biotechnol J. 2024;23:287–294. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Rau A, Michel L, Wilhelm B, Raghu VK, Reisert M, Jung M, Kellner E, Schlett CL, Aerts HJWL, Lu MT, et al. Deep learning to predict cardiovascular mortality from aortic disease in heavy smokers. npj Cardiovasc Health. 2024;1:28. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Hamelink II, de Heide EEJ, Pelgrim G, Kwee TCT, van Ooijen P, de Bock GHT, Vliegenthart RR. Validation of an AI-based algorithm for measurement of the thoracic aortic diameter in low-dose chest CT. Eur J Radiol. 2023;167:111067. [DOI] [PubMed] [Google Scholar]
  • 19.Sedghi Gamechi Z, Bons LR, Giordano M, Bos D, Budde RPJ, Kofoed KF, Pedersen JH, Roos-Hesselink JW, de Bruijne M, et al. Automated 3D segmentation and diameter measurement of the thoracic aorta on non-contrast enhanced CT. Eur Radiol. 2019;29:4613–4623. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Wasserthal J, Breit H-C, Meyer MT, Pradella M, Hinck D, Sauter AW, Heye T, Boll DT, Cyriac J, Yang S, et al. TotalSegmentator: robust segmentation of 104 anatomic structures in CT images. Radiology: Artificial Intelligence. 2023;5:e230024. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21.Ramirez G, Lemley M, Shanbhag A, Kwiecinski J, Miller RJH, Kavanagh PB, Liang JX, Dey D, Slipczuk L, Travin MI, et al. The REgistry of Flow and Perfusion Imaging for Artificial INtelligEnce with PET (REFINE PET): Rationale and Design. J Nucl Cardiol 2025. 2025;52:102449. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Slomka PJ, Betancur J, Liang JX, Otaki Y, Hu LH, Sharir T, Dorbala S, Di Carli M, Fish MB, Ruddy TD, et al. Rationale and design of the REgistry of Fast Myocardial Perfusion Imaging with NExt generation SPECT (REFINE SPECT). J Nucl Cardiol. 2020;27:1010–1021. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23.Miller RJH, Lemley M, Shanbhag A, Ramirez G, Liang JX, Builoff V, Kavanagh P, Sharir T, Hauser MT, Ruddy TD, et al. The Updated Registry of Fast Myocardial Perfusion Imaging with Next-Generation SPECT (REFINE SPECT 2.0). J Nucl Med. 2024;65:1795–1801. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24.Slomka PJ, Nishina H, Berman DS, Akincioglu C, Abidov A, Friedman JD, Hayes SW, Germano G, et al. Automated quantification of myocardial perfusion SPECT using simplified normal limits. J Nucl Cardiol. 2005;12:66–77. [DOI] [PubMed] [Google Scholar]
  • 25.Isensee F, Jaeger PF, Kohl SA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat. Methods 2021;18:203–211. [DOI] [PubMed] [Google Scholar]
  • 26.Davies RR, Gallo A, Coady MA, Tellides G, Botta DM, Burke B, Coe MP, Kopf GS, Elefteriades JA. Novel measurement of relative aortic size predicts rupture of thoracic aortic aneurysms. Ann Thorac Surg. 2006;81:169–177. [DOI] [PubMed] [Google Scholar]
  • 27.Mazzolai L, Teixido-Tura G, Lanzi S, Boc V, Bossone E, Brodmann M, Bura-Rivière A, De Backer J, Deglise S, Della Corte A, et al. 2024 ESC Guidelines for the management of peripheral arterial and aortic diseases. Eur Heart J. 2024;45:3538–3700. [DOI] [PubMed] [Google Scholar]
  • 28.Miller RJH, Pieszko K, Shanbhag A, Feher A, Lemley M, Killekar A, Kavanagh PB, Van Kriekinge SD, Liang JX, Huang C, et al. Deep Learning Coronary Artery Calcium Scores from SPECT/CT Attenuation Maps Improve Prediction of Major Adverse Cardiac Events. J Nucl Med. 2023;64:652–658. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29.Williams MC, Shanbhag AD, Zhou J, Michalowska AM, Lemley M, Miller RJH, Killekar A, Waechter P, Gransar H, Van Kriekinge SD, et al. Automated vessel-specific coronary artery calcification quantification with deep learning in a large multi-centre registry. Eur Heart J Cardiovasc Imaging. 2024;25:976–985. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30.Grambsch PM, Therneau TM. Proportional hazards tests and diagnostics based on weighted residuals. Biometrika. 1994;81:515–526. [Google Scholar]
  • 31.Koo TK, Li MY. A guideline of selecting and reporting intraclass correlation coefficients for reliability research. J Chiropr Med. 2016;15:155–163. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Meissel K, Yao ES. Using Cliff’s Delta as a Non-Parametric Effect Size Measure: An Accessible Web App and R Tutorial. Practical Assessment, Research, and Evaluation. 2024;29:2. [Google Scholar]
  • 33.Dillavou ED, Buck DG, Muluk SC, Makaroun MS. Two-dimensional versus three-dimensional CT scan for aortic measurement. J Endovasc Ther. 2003;10:531–538. [DOI] [PubMed] [Google Scholar]
  • 34.Della Corte A, Bancone C, Quarto C, Dialetto G, Covino FE, Scardone M, Caianiello G, Cotrufo M. Predictors of ascending aortic dilatation with bicuspid aortic valve: a wide spectrum of disease expression. Eur J Cardiothorac Surg. 2007;31:397–404; discussion 404-395. [DOI] [PubMed] [Google Scholar]
  • 35.Biaggi P, Matthews F, Braun J, Rousson V, Kaufmann P, Jenni R. Gender, age, and body surface area are the major determinants of ascending aorta dimensions in subjects with apparently normal echocardiograms. J Am Soc Echocardiogr 2009;22:720–725. [DOI] [PubMed] [Google Scholar]
  • 36.Agmon Y, Khandheria B, Meissner I, Schwartz G, Sicks J, Fought A. Is aortic dilatation an atherosclerosis-related process? Clinical, laboratory, and transesophageal echocardiographic correlates of thoracic aortic dimensions in the population with implications for thoracic aortic aneurysm formation. J Am Coll Cardiol. 2003;42:1076–1083. [DOI] [PubMed] [Google Scholar]
  • 37.Zafar MA, Li Y, Rizzo JA, Charilaou P, Saeyeldin A, Velasquez CA, Mansour AM, Mahmood SUB, Ma W-G, Brownstein AJ. Height alone, rather than body surface area, suffices for risk estimation in ascending aortic aneurysm. J Thorac Cardiovasc Surg. 2018;155:1938–1950. [DOI] [PubMed] [Google Scholar]
  • 38.Mensel B, Mahnken AH, Kaiser E, Völzke H, Dörr M, Felix SB, Ittermann T, Lieb W, Lorbeer R. Association of aortic diameters and mortality in the general population-an MRI-based study. European Radiology. 2025;35:427–437. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39.Elefteriades JA, Mukherjee SK, Mojibian H. Discrepancies in Measurement of the Thoracic Aorta: JACC Review Topic of the Week. J Am Coll Cardiol. 2020;76:201–217. [DOI] [PubMed] [Google Scholar]
  • 40.de Carvalho Macruz FB, Lu C, Strout J, Takigami A, Brooks R, Doyle S, Yun M, Buch V, Hedgire S, Ghoshhajra B. Quantification of the thoracic aorta and detection of aneurysm at CT: development and validation of a fully automatic methodology. Radiol Artif Intell. 2022;4:e210076. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 41.Litmanovich D, Bankier AA, Cantin L, Raptopoulos V, Boiselle PM. CT and MRI in diseases of the aorta. AJR Am J Roentgenol 2009; 193: 928–940. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplemental Table of Contents
NIHMS2195366-supplement-Supplemental_Table_of_Contents.pdf (113.9KB, pdf)
Supplemental Methods
NIHMS2195366-supplement-Supplemental_Methods.pdf (130.3KB, pdf)
Supplemental Tables
NIHMS2195366-supplement-Supplemental_Tables.pdf (294.5KB, pdf)
Supplemental Figure S1
NIHMS2195366-supplement-Supplemental_Figure_S1.pdf (245.6KB, pdf)
Supplemental Figure S2
NIHMS2195366-supplement-Supplemental_Figure_S2.pdf (220.6KB, pdf)
Supplemental Figure S3
NIHMS2195366-supplement-Supplemental_Figure_S3.pdf (234.2KB, pdf)
Supplemental Figure S4
NIHMS2195366-supplement-Supplemental_Figure_S4.pdf (234.8KB, pdf)
Supplemental Figure S5
NIHMS2195366-supplement-Supplemental_Figure_S5.pdf (330.9KB, pdf)
Supplemental Figure S6
NIHMS2195366-supplement-Supplemental_Figure_S6.pdf (341.3KB, pdf)
Supplemental Figure S8
NIHMS2195366-supplement-Supplemental_Figure_S8.pdf (279.1KB, pdf)
Supplemental Figure S9
NIHMS2195366-supplement-Supplemental_Figure_S9.pdf (208.7KB, pdf)
Supplemental Figure S7
NIHMS2195366-supplement-Supplemental_Figure_S7.pdf (480KB, pdf)

Data Availability Statement

To the extent allowed by data sharing agreements and IRB protocols, the deidentified data and data analysis code from this manuscript will be shared upon written request.

RESOURCES