Deep learning of coronary calcium scores from PET-CT attenuation maps accurately predicts adverse cardiovascular events

Abstract

Background.

Assessment of coronary artery calcium (CAC) by computed tomography (CT) provides an accurate measure of atherosclerotic burden. CAC is also visible in CT attenuation correction (CTAC) scans, always acquired with cardiac positron emission tomography (PET/CT).

Objectives.

We aimed to develop a deep learning model capable of fully automated CAC definition from PET CTAC scans.

Methods.

The novel DL model, originally developed for video applications, was adapted to rapidly quantify CAC. The model was trained using 9,543 expert annotated CT scans and was tested in 4,331 patients from external cohort undergoing PET/CT with major adverse cardiac events (MACE) (follow-up 4.3 years), including same-day paired ECG-gated CAC scans available in 2,737 patients. We analyzed the MACE risk stratification in 4 CAC score categories (0; ​ 1–100; 101–400; >400​) and compared CAC scores derived from ECG-gated CT scans (standard scores) by expert observers with automatic DL scores from CTAC scans (DL-CTAC).

Results.

Automatic DL scoring required less than 6 seconds per scan. DL-CTAC scores provided stepwise increase in the risk of MACE across the CAC score categories (hazard ratio up to 3.2, p<0.001). Net reclassification improvement of standard CAC scores over DL-CTAC scores was non-significant (−0.02, 95% confidence interval: −0.11, 0.07). The negative predictive values for MACE of zero CAC with standard (85%) and DL-CTAC score (83%) were similar (p=0.19).

Conclusions.

DL-CTAC scores predict cardiovascular risk similarly to standard CAC scores quantified manually by experienced operators from dedicated ECG-gated CAC scans and can be obtained almost instantly with no changes to PET/CT scanning protocol.

Introduction

Positron emission tomography (PET) myocardial perfusion imaging (MPI) is well established in the assessment of patients with known and suspected coronary artery disease (CAD)(1); however, it does not measure coronary atherosclerosis directly. The extent of coronary artery calcium (CAC), a measure of atherosclerotic burden, provides powerful risk stratification and influences preventive therapies and lifestyle changes(2). Prior studies have shown that CAC provides incremental prognostic information to MPI(3) and can improve patient management(4,5).

Most current PET systems are offered with a hybrid configuration, including a computed tomography (CT) scanner. For PET/CT, a low-dose, ungated CT attenuation correction (CTAC) scan is always obtained(1). CAC information could be potentially extracted from CTAC scans to provide supplementary data and allow for more informed decision-making(4) after MPI. However, while subjective visual assessment of CAC is recommended, quantitative CAC scoring in CTAC scans is currently not performed. Therefore, the utility of these ubiquitous scans is not fully exploited clinically.

We aimed to develop and evaluate automated and rapid CAC quantification from CTAC scans using a novel deep learning (DL) approach that integrates the data from adjacent CT slices(6). We evaluated the prognostic value of CAC scores obtained from DL segmentations of CTAC scans (DL-CTAC scores) in the prediction of major adverse cardiac events (MACE) and compared DL-CTAC scores with clinical CAC scores obtained from standard CAC scans.

Methods

The general overview of the methods is shown in Figure 1. Information about the DL model architecture is given in the Supplement.

Figure 1. Overview of training and evaluation methods.

Top: the steps of automatic CAC scoring. Middle: training and testing characteristics. Bottom: evaluation techniques. Abbreviations: CAC – coronary artery calcium; conv-LSTM – convolutional long-short term memory neural network; CTAC – computed tomography attenuation correction; MACE – major adverse cardiac event.

Study Data

Training and internal testing cohort

For training, internal validation, and internal testing, we utilized data from three centers that included 9543 scans (1827 ECG-gated CAC scans and 7716 CTAC maps). More information on the dataset is given in the Supplement.

External held-out testing cohort

The separate 4^th center, held-out test cohort consisted of consecutive 4,761 patients who had their first-time PET-MPI scan in Cedars-Sinai between 2010 and 2018 with low-dose (tube voltage 100kV, current-time product 11–13mAs, gantry rotation speed 0.5s), shallow breathing CTAC scans. We excluded patients with missing follow-up, patients who had revascularization within 90 days from scan (n=417), and 13 patients with missing or corrupt CTAC scans (Figure 1). For a separate sub-analysis, we excluded all subjects with evidence of prior CAD (defined as history of revascularization or myocardial infarction). Analysis of agreement between DL-CTAC scores and standard CAC scores was performed in 2,737 patients from this cohort, who had a separate standard CAC scan performed on the same day, during the same imaging session.

Details on imaging protocols for training, internal testing, and external testing sets are given in Supplementary Table 1.

Calcium scoring by expert readers

Calcium scoring for training and internal testing

The training, internal validation, and internal testing sets included both CTAC maps and dedicated CAC scans (included in the training and internal testing cohorts to ensure better generalizability of the model across various scan protocols and qualities). All scans were manually annotated by two readers with at least 5 years of experience in CAC scoring using dedicated quantitative software (Cardiac Suite, Cedars Sinai). Manual annotation resulted in the creation of pixel-by-pixel masks for CAC lesions and for non-coronary calcifications for each CT slice, to allow the model to separate CAC from non-coronary opacities.

CAC scores were calculated based on the lesion annotations according to the standard clinical algorithm(7). All cases were categorized based on CAC score (category 1: CAC score = 0, category 2: CAC score 1–100, category 3: CAC score 101–400, category 4: CAC score >400)(8). In 20 randomly selected patients, readers' scoring time was recorded.

Calcium scoring of the External Testing Set

We utilized clinical CAC scores acquired from dedicated CAC scans at the time of reporting. The clinical scores were calculated using a standard dedicated workstation with quantitative software (ScImage Inc., Los Altos, CA, USA and Cardiac Suite, Cedars-Sinai, Los Angeles, CA, USA) by an experienced technologist and reviewed by the reporting cardiologist during clinical reporting.

Model evaluation

Internal testing

Internal testing was performed on a held-out data set with no overlap with the training or validation cases. In this set, we assessed the agreement in CAC score categories between scores coming from expert readers' annotations and DL segmentations.

External testing

We tested the trained model in the data from the separate center (Cedars-Sinai) cohort. All stress CTAC scans underwent automatic segmentation, and the DL-CTAC scores were calculated based on the segmentation masks. When stress CTAC scan was not available, rest CTAC scan was used instead. No data from this cohort was used to train the network.

To allow for comparison with standard calcium scoring, the DL-CTAC scores were separately evaluated on a subset of the external testing cohort with available clinical CAC scans acquired on the same date as PET/CT. We studied the relation of CAC score category based on DL-CTAC score as well as based on expert readers' annotations on dedicated CAC scans with the occurrence of MACE.

Clinical outcomes

The external testing cohort was followed for MACE (defined as all-cause mortality, late revascularization [percutaneous coronary intervention or coronary artery bypass grafting] that occurred > 90 days from scan date, admission for unstable angina, or myocardial infarction). Follow-up for all-cause mortality was obtained using internal hospital records and the Social Security Death Index, National Death Index, and California Non-comprehensive Death File until 12/2020. Information regarding myocardial infarction, unstable angina, and revascularization were collected from hospital records and verified by site physicians according to standard criteria(9).

Statistical analysis

Continuous variables were expressed by median and interquartile ranges (IQR). Two-sided median values were compared with Wilcoxon rank-sum test or Kruskal-Wallis test. Categorical variables were compared using Fisher's exact test. A p-value <0.05 was considered significant.

We used Kaplan-Meier curves to show the discrimination of MACE risk in 4 CAC categories in the whole test set and the subset with available same-day gated CAC scans.

Univariable proportional hazards Cox models were used to compare MACE-free survival in four CAC score categories. Additionally, to investigate the added value of standard CAC scores as well as DL-CTAC scores on top of PET MPI by creating multivariable Cox models adjusted for myocardial flow reserve as obtained by quantitative PET software (QPET, Cedars-Sinai)(10) and relative ischemia (summed difference score obtained from 17-segment stress and rest perfusion scores).

Categorical net reclassification index (NRI) was used to investigate if standard CAC scores yield better risk discrimination than DL-CTAC scores. Confidence intervals for NRI were assessed using bootstrapping. Negative predictive value of DL-CTAC score of zero in the prediction of no future MACE was assessed. In the subset with available CAC scans, negative predictive values were compared using generalized score statistics(11).

Agreement DL-CTAC with scores annotated by expert readers on the same scans was evaluated in the internal set in 4 CAC categories using linearly weighted Cohen's kappa and concordance index. Similar to other studies(12,13), we chose to use linear weighting to be conservative in our findings. Agreement between original CAC scores expert annotated on ECG-gated scans with DL-CTAC scores was analyzed in the subset of cases with available CAC scans.

Statistical analysis was performed using R and RStudio software. Concordance matrices were generated in Python. Linearly weighted Kappa coefficients with confidence intervals were calculated using VassarStats software (see Supplementary Table 2).

Compliance with machine learning reporting guidelines

This study was designed and conducted following the Proposed Requirements for Cardiovascular Imaging-Related Machine Learning Evaluation (PRIME)(14). To improve the transparency of reporting and the reproducibility of machine-learning algorithms, the PRIME checklist is included as Supplementary Table 5.

Ethical approval

This study complies with the Declaration of Helsinki. The institutional review board at Cedars-Sinai and the participating sites approved the collection of data for the registry. Informed consent has been obtained from the subjects (or their legally authorized representative).

Results

The development cohort included patients from 3 different centers with 6,944 low-dose CTAC scans and 1,643 ECG-gated CAC scans (Supplementary Figure 1). It took up to 8 minutes to score a CTAC scan by an experienced reader (median 105 seconds, IQR 80, 188). The end-to-end DL processing was <6 seconds per scan. A typical radiation dose associated with a CTAC scan was 0.2–0.3mSv, while a typical radiation dose for a dedicated CAC scan acquired at the time of PET MPI was 1–3 mSv.

Internal testing

The DL predictions were evaluated against annotations by expert readers in the internal held out set of 956 cases that included 212 CAC scans and 744 CTAC scans (details shown in Table 1 and Supplementary Table 3). The model was able to successfully score scans with various slice thicknesses (ranging from 2.5 to 5mm). Agreement in CAC score categories with those based on expert readers' annotations was very good across the variety of acquisition protocols (linearly weighted Cohens Kappa 0.83 [95% CI 0.8, 0.85], n=956). Concordance matrices are presented in Figure 2.

Table 1.

Training cohort characteristics

Characteristic	Overall, N = 9,543	Training, N = 7,631	Internal validation, N = 956	Internal testing, N = 956	p-value
Age	65 (56, 73)	64 (56, 73)	66 (57, 75)	65 (57, 73)	0.001
Sex male	5,023 (53%)	4,014 (53%)	504 (53%)	505 (53%)	>0.9
BMI	29 (26, 34)	29 (26, 34)	30 (26, 34)	29 (25, 34)	0.8
Hypercholesterolemia	5,253 (56%)	4,149 (55%)	557 (59%)	547 (58%)	0.013
Smoking	1,376 (15%)	148 (16%)	1,076 (14%)	152 (16%)	0.2
Hypertension	6,146 (65%)	4,878 (65%)	638 (68%)	630 (67%)	0.07
Diabetes	2,519 (27%)	1,991 (26%)	265 (28%)	263 (28%)	0.4
Post - PCI	972 (10%)	760 (10%)	97 (10%)	115 (12%)	0.13
Post - CABG	396 (4.4%)	316 (4.4%)	41 (4.6%)	39 (4.3%)	>0.9
History of CAD	1,240 (13%)	977 (13%)	119 (13%)	144 (15%)	0.13
CAC score	54 (0, 481)	43 (0, 503)	99 (1, 398)	100 (1, 398)	<0.001
CAC score category					<0.001
CAC = 0	3,214 (34%)	2,736 (36%)	239 (25%)	239 (25%)
CAC 1–100	2,112 (22%)	1,634 (21%)	239 (25%)	239 (25%)
CAC 101–400	1,588 (17%)	1,110 (15%)	239 (25%)	239 (25%)
CAC >400	2,629 (28%)	2,151 (28%)	239 (25%)	239 (25%)
ECG-gating	1,827 (19%)	1,431 (19%)	212 (22%)	184 (19%)	0.04

Statistics presented: median (inter quartile range), n (%); Abbreviations: BMI – body mass index; CABG – coronary artery bypass grafting, CAC – coronary artery calcification; CAD – coronary artery disease; PCI -percutaneous coronary intervention

Figure 2. Internal testing results.

Concordance matrices present agreement in 4 CAC categories between human expert readers on internal testing set (no overlap with training set). Top: the internal testing set comprised both low dose CTAC scans and ECG-gated CAC scans, results from both types of scans, and scans in the internal testing set combined. Bottom: results by site.

External testing

Patient characteristics

During the median follow-up of 4.3 years (IQR 2.5, 6.6) at least one event occurred in 1354 (31%) subjects. Death comprised 70% of observed first MACE (n=953) and occurred in 25% of patients (n=1070) within the follow-up period. MACE comprised late revascularization events (15%, n=210), myocardial infarction (7.8%, n=106), unstable angina (6.3%, n=85). Detailed characteristics of the external testing cohort are presented in Table 2 (summarized by the availability of the clinical CAC score) and in Supplementary Table 4 (summarized by follow-up status).

Table 2.

External cohort characteristics

Characteristic	Overall, N = 4,331	No CAC scan available, N = 1,594	With available CAC scan, N = 2,737	p-value
Age	71 (64, 79)	73 (65, 80)	70 (63, 78)	<0.001
Male	2,503 (58%)	1,121 (70%)	1,382 (50%)	<0.001
BMI	27 (24, 32)	27 (24, 31)	28 (24, 32)	0.001
Hypercholesterolemia	2,910 (67%)	1,239 (78%)	1,671 (61%)	<0.001
Smoking	320 (7.4%)	106 (6.6%)	214 (7.8%)	0.2
Hypertension	3,368 (78%)	1,334 (84%)	2,034 (74%)	<0.001
Diabetes	1,431 (33%)	582 (37%)	849 (31%)	<0.001
Post - PCI	978 (23%)	869 (55%)	109 (4.0%)	<0.001
Post - CABG	467 (11%)	455 (29%)	12 (0.4%)	<0.001
MFR	2.36 (1.81, 2.98)	2.14 (1.62, 2.80)	2.45 (1.93, 3.07)	<0.001
SDS	0 (0, 2)	1 (0, 4)	0 (0, 1)	<0.001
History of CAD	1,443 (33%)	208 (76%)	235 (8.6%)	<0.001
Clinical CAC score	0 (0, 220)	-	106 (0, 523)	<0.001
CAC score category				<0.001
CAC = 0	685 (25%)	-	685 (25%)
CAC 1–100	662 (24%)	-	662 (24%)
CAC 101–400	586 (21%)	-	586 (21%)
CAC >400	804 (29%)	-	804 (29%)

Statistics presented: median (inter quartile range), n (%); Abbreviations: BMI – body mass index; CABG – coronary artery bypass grafting, CAC – coronary artery calcification; CAD – coronary artery disease; PCI -percutaneous coronary intervention, MFR – myocardial flow rate, SDS – summed difference score

Risk stratification by DL-CTAC scores

Kaplan-Meier curves for MACE risk stratified by DL-CTAC score categories in the entire testing cohort as well as the univariable proportional Cox model hazard ratios (HR) are presented in Figure 3A. Patients with DL-CTAC score >0 were more likely to experience MACE than patients with no CAC identified by DL in CTAC maps (log rank p-value = 0.016 for CAC=0 vs. CAC 1–100, <0.0001 for remaining comparisons). HR for MACE increased with each category, with significantly higher risk in every DL-CTAC score category in comparison with DL-CTAC score of 0. Patients with DL-CTAC score >400 were at 3.2 (95% confidence interval [CI]: 2.7, 3.79) times higher risk of experiencing MACE in comparison with those with DL-CTAC score of 0. After adjusting for myocardial ischemia and flow reserve, the incremental pattern of HR was preserved with significantly increased risk in low (DL-CTAC score 1–100, HR 1.25, 95% CI:1.01,1.56, p=0.043), medium (DL-CTAC score 101–400, HR 1.74, 95% CI: 1.42,2.13; p<0.001) and high scores (DL-CTAC >400, HR 2.45, 95% CI: 2.06,2.92, p<0.001). Detailed results of multivariable Cox analysis are given in Supplementary Table 5. In the whole external cohort (n=4331) DL-CTAC score of 0 had a negative predictive value of 82% for future MACE.

Figure 3. Outcome analysis for CAC scores from deep learning segmentations of PET CTAC maps in testing cohort (left) and subset with no CAD history (right).

Top: Kaplan – Meier curves for MACE risk by CAC score categories. P-values for Kaplan-Meier curves using log-rank tests. Bottom: univariate MACE-risk analysis using proportional hazards Cox model; CAC – coronary artery calcium; CAD -coronary artery disease; CTAC – computed tomography attenuation correction; MACE – major adverse cardiac event

The analysis was repeated in a subset of patients with no history of CAD (n=2888), where similar results with significantly increased MACE risk in each predicted DL-CTAC score category were seen (Figure 3B). DL-CTAC score of zero had a negative predictive value of 83% in the subset with no prior CAD.

Comparison to clinical CAC scores

Figure 4 shows Kaplan-Meier curves in the subset of 2737 (63% of the cohort) patients with available same-day standard CAC scans, for CAC scores derived by human experts from clinical CAC scans as well as for DL-CTAC. Similar risk stratification of DL-CTAC is shown despite the full automation and the difference in the image quality between DL-CTAC and standard CAC scans. Similar results in the subset of patients with no history of prior CAD (n=2502) with available same-day standard CAC scans are shown in Figure 5. For the higher CAC categories, significant differences in MACE rates were observed for both standard and DL-CTAC scores in all CAC categories> 0. Univariable Cox revealed incrementally increasing HR in each of CAC score categories with significantly higher MACE risk in the low (CAC score 1–100) moderate (CAC score 101–400) and high (CAC score >400) categories regardless of whether prior CAD cases were included.

Figure 4. Comparison of outcomes by expert reader standard CAC score from ECG-gated CT (grey left) vs. score from deep learning from PET CTAC maps (blue right).

Top: Kaplan – Meier curves for MACE risk by CAC score categories. P-values for Kaplan-Meier curves using log-rank tests. Bottom: univariate MACE risk analysis using proportional hazards Cox model; CAC – coronary artery calcium; CAD- coronary artery disease; CTAC – computed tomography attenuation correction; MACE – major adverse cardiac event

Figure 5. Comparison of outcomes by expert reader CAC score from ECG-gated CT (grey left) vs. score automatically by deep learning from PET CTAC maps (blue right) without known CAD.

Top: Kaplan – Meier curves for MACE risk by CAC score categories. P-values for Kaplan-Meier curves using log-rank tests. Bottom: univariate MACE risk analysis using proportional hazards COX model. CAC – coronary artery calcium; CAD- coronary artery disease; CTAC – computed tomography attenuation correction; MACE – major adverse cardiac event

Negative predictive values of clinical CAC =0 vs. DL-CTAC CAC =0 were not significantly different at 85% and 83%, respectively (p=0.19). In patients with no history of CAD the negative predictive values of clinical CAC =0 vs. DL-CTAC CAC =0 were 86% and 84%, respectively (p=0.15).

The overall NRI showed no improvement by clinical CAC scores as compared to DL-CTAC scores (−0.017, 95% CI −0.11, 0.07). The event and non-event NRIs were 0.1 (95% CI −0.02 – 0.16) and −0.11 (95% CI −0.15, −0.07), respectively. Similarly, after excluding cases with prior CAD, NRI was −0.005 (95% CI −0.11, 0.09), and the event and non-event NRIs were 0.12 (95% CI 0.05, 0.19) and −0.12 (95% CI −0.17, −0.08), respectively.

Examples of ECG-gated CAC scans with CAC lesions annotated by expert observers in standard clinical software alongside CTAC scans and automatic DL-calcium definition are presented in Figure 6. Additional cases are shown in Supplementary Figures 3 and 4. Calcium score categories obtained using the DL framework from CTAC maps exhibited good agreement with standard CAC scores quantitatively assessed on gated CAC scans. The linearly weighted Cohen's Kappa was 0.62 (95% CI 0.6, 0.64). The concordance matrices showing numbers of concordant and discordant pairs in each CAC category for both rest and stress scan-preferred strategy are shown in Supplementary Figure 5. Scatter plots showing the correlation between DL-CTAC, and CAC scores are shown in Supplementary Figure 6.

Figure 6. Example predictions of the deep learning on CTAC maps and expert annotations on ECG-gated CAC scans in challenging cases.

A – the model classifies lesions correctly in LM and LAD as CAC (red and blue on CAC scan, red on CTAC scan) and correctly tells CAC apart from aortic calcifications (light blue on CAC scan, green on CTAC scan). B – the model correctly segments CAC in RCA and correctly differentiates from aortic calcifications. C – the model correctly scores a small lesion in RCA. D – the model correctly ignores large mitral calcification; LM – left main coronary artery; LAD – left anterior descendent coronary artery; RCA – right coronary artery

Despite comparable risk stratification there were several cases where DL-CTAC score was zero, while standard CAC score was > 0 (false negatives) as well as where standard score was zero but DL- CTAC score was not (false positives) (Supplementary Figures 5 and 6). We provide individual examples of such cases in Supplementary Figure 7. These inaccuracies occurred because of partial volume effects or attribution of CAC as non-coronary calcifications (for ‘false negatives’) and scoring noise or non-coronary opacities as CAC (for ‘false positives’).

Added value of DL-CTAC to MPI PET

Proportional hazard Cox model hazard ratios for MACE, adjusted for myocardial flow ratio and SDS are shown in Table 3. In patients without prior CAD, all CAC score categories, both by DL-CTAC and clinical CAC scores were associated with a significantly higher risk of future MACE independently from PET MPI findings.

Table 3.

Hazard ratios or major adverse cardiac events adjusted for myocardial blood flow reserve and perfusion for standard CAC scores and DL-CTAC scores in patients with no prior CAD and available CAC scans (n=2502).

Category	Standard CAC score			DL-CTAC CAC score
CAC = 0	—	—	—	—	—	—
CAC 1–100	1.34	1.02, 1.75	0.037	1.30	1.00, 1.67	0.046
CAC 101–400	1.56	1.19, 2.05	0.001	1.72	1.34, 2.20	<0.001
CAC >400	2.59	2.02, 3.31	<0.001	2.28	1.81, 2.86	<0.001

Abbreviations: CI – confidence interval; CAC - coronary calcium score; CTAC – computed tomography attenuation correction; MFR- myocardial flow rate; SDS – summed difference score

Discussion

This is the first study to show the prognostic value of CAC scores obtained fully automatically and rapidly by a novel DL approach from low-dose ungated CTAC maps in a large PET/CT patient cohort. Importantly, negative predictive value for MACE based on CAC scores obtained from standard-dose gated CT scans did not differ significantly and standard scores did not provide significant reclassification improvement over the fully automatic DL-CTAC scores. Furthermore, the associations of these scores with increased MACE were independent of ischemia or flow measures. After adjusting for standard PET variables, the DL-CTAC scores provided incremental prognostic information in multivariable analysis, highlighting the clinical relevance of performing CAC imaging as part of cardiac PET/CT acquisitions.

Consistent data have shown the strong prognostic value of CAC assessment in asymptomatic individuals. Assessment of CAC during SPECT MPI scans has been shown to add to perfusion in risk assessment(5) and improve assessment of pretest likelihood of CAD(15), thereby contributing to diagnostic accuracy and leading to changes in preventive medications and beneficial changes in patienťs adherence to medication recommendations. In a substantial proportion of centers performing PET MPI, separate ECG-gated CAC scans are not performed due to added cost and radiation, as well as lack of reimbursement. Our study demonstrates that DL-CTAC scores provide similar risk stratification to CAC scores from dedicated ECG-gated scans. In some cases, it might be preferable to acquire a higher-quality scan, at the cost of slightly increased radiation exposure and additional time, to quantify CAC more precisely. However, the ability of CAC assessments to be made from CTAC scans, obtained with all PET/CT procedures provides the opportunity to take advantage of the known benefits of additional CAC information in a wide group of patients with no modifications to study protocols.

In previous studies, agreement of visually estimated CAC scores from CTAC with standard CAC scores was good(16). Quantitative CAC scores from PET CTAC were shown to correlate fairly well with standard CAC scores(12,17) (linearly weighted Kappa up to 0.79 [95% CI 0.73–0.89] for agreement between human readers on dedicated CAC scans vs. stress CTAC scans)(12). However, CAC scoring in low-dose ungated scans represents a challenge for both human readers as well as for the AI system, as the high noise levels, partial volume effect, and motion artifacts(18) affect the appearance of CAC lesions(16). This inherent aspect of CAC scoring in CTAC limits the agreement with standard CAC scores. Nevertheless, despite of these limitations, the ability of CTAC CAC to stratify the risk as well as standard scores is reassuring. Due to the cumbersome, time-consuming manual annotation and high observer variability in low-quality scans, quantitative CAC scoring is not routinely performed in CTAC maps. In contrast, our DL approach could allow routine and instant automatic quantification of these scans on any standard workstation.

Recent studies have reported automatic quantification systems for CAC scoring using DL (12,13,19–24). One of the latest investigations showed that on multiple large cohorts CAC scores derived automatically from various chest-CT scans(24), however that model was not trained in ungated free-breathing low-dose CTAC scans. The expert manual CAC scoring in CTAC scans is itself not well-established and is not routinely performed. Only few studies from one group (12,13,23) have evaluated the use of fully automated DL for automatic CAC scoring in low-dose ungated CTAC maps, the reported computation was 2–7 minutes as compared to under 6 seconds per scan with our model. Additionally, the convLSTM approach has shown to outperform recent U-Net based model in terms of speed and memory consumption in contrast CT angiography plaque(25) and lung segmentation(26).

Only one study reported the prognostic value of CAC scores automatically obtained from PET CTAC scans(23) with simplified (two-level) CAC categorization. A unique aspect of our study is the head-to head comparison in a large cohort of fully automatic scoring of low-dose (typical radiation dose of 0.2–0.3mSv) CTAC maps with clinical CAC scores from gated scans (that are associated with up to an order of magnitude higher typical radiation dose) in the same patients.

Notably, aside from MPI studies, our developments could be used also for non-cardiac PET. In the US alone, there are over 2 million PET scans performed annually(27). Indeed, a very recent study has shown that CAC scores acquired automatically for all patients undergoing radiotherapy for breast cancer provide valuable prognostic information(22). As our data represents the lowest possible quality of CTAC obtained in clinical practice for attenuation correction (low dose, non-ECG gated, shallow normal breathing), our technique may perform even better for CAC detection with ECG-gated, breath-held CT. Nevertheless, our results demonstrate that DL-CAC scores, obtained even from low-quality CTAC scans acquired for attenuation correction, could be applied broadly to improve risk stratification following any cardiac PET/CT or thoracic PET/CT scan.

Our DL approach could potentially provide complementary stratification of cardiac risk for millions of patients undergoing PET for oncological purposes. In this context, the fact that the risk stratification in all-comers was similar to that in the population without prior CAD is an additional advantage.

Study Limitations

Our study has several limitations. Our population size was insufficient to evaluate the subset of patients with very low CAC scores (CAC 1–10) as a separate group; however, previous studies have suggested that even these scores may be associated with increased cardiovascular risk. We used clinical CAC scores stored in the patients' records performed by several expert readers during clinical evaluation; however, the inter-observer variability of standard gated CAC scoring is low. While testing was performed in a strict external regimen, separate from the sites used for model creation, further validation is needed to verify generalizability to other protocols or scanners. Motion artifacts can affect CAC scoring, which is particularly unavoidable with ungated CTAC scans. While the overall agreement between DL-CTAC and the standard CAC scan scores was good, the high number of discordant pairs in low calcium score categories indicate that DL-CTAC scores do not allow for definite exclusion of the presence of any calcification. However, similar MACE risk discrimination for manual standard dose ECG-gated CAC scoring and automatic DL-CTAC scores confirms the clinical validity our of development.

In our study, we present all the analyses in the general population of patients undergoing PET MPI as well as in the subset of patients with no history of prior CAD. While our model, similar to other state-of-the-art algorithms capable of automatic CAC scoring (24), may classify stent on CTAC scan as CAC (Supplementary Figure 3C), it also has the ability to correctly reject stents and other non-CAC opacities (Supplementary Figure 4), because stents were included in training data as unmarked regions by the expert observer. As our DL model includes a heart segmentation step, it will automatically reject CAC lesions outside heart boundaries, so calcifications in grafts would normally be ignored. However, further experienced reader supervision may be warranted, especially in cases with stents or other complicated situations. Further studies are needed to evaluate the incremental value of DL-CTAC scores in patients with a history of CAD.

Conclusions

CAC scores obtained automatically in a few seconds by DL from low-dose ungated CTAC scans predict MACE similarly to standard clinical CAC scoring by expert readers from dedicated ECG-gated CT scans. DL-CTAC scores can be obtained instantly for all patients undergoing PET/CT with no modifications in the study protocol. The addition of routine CAC scoring with CTAC scans could lead to improved diagnosis, risk stratification, and disease management and could influence lifestyle recommendations.

Central illustration. Comparison of outcomes by expert reader standard CAC score from ECG-gated CT (grey left) vs. score automatically by deep learning from PET CTAC maps (blue right).

Top: Kaplan – Meier curves for MACE risk by CAC score categories. P-values for Kaplan-Meier curves using log-rank tests. Bottom: univariate MACE risk analysis using proportional hazards Cox model; CAC – coronary artery calcium; CAD- coronary artery disease; CTAC – computed tomography attenuation correction; MACE – major adverse cardiac event

Clinical perspectives.

Competency in medical knowledge:

CAC scores derived automatically by deep learning from low-dose, ungated CTAC scans, routinely acquired for PET/CT when compared head-to-head with scores obtained by expert readers from dedicated, ECG-gated CAC scans provide similar risk stratification, but without any additional radiation dose or time.

Competency in patient care

For every patient undergoing PET/CT scan that includes the chest region, CAC scores can be automatically obtained by deep learning. These scores identify patients at high risk of adverse cardiovascular events that can benefit from intensified medical therapy.

Translational outlook

Our model allows routine automatic CAC scoring in CTAC scans for any PET/CT scan. Given the lack of reimbursement of dedicated CAC scanning as a part of PET/CT workflow, our method could offer a cost-effective solution that provides additional CAC information from all PET/CT studies, including studies performed for non-cardiology indications.

Abbreviations:

CAC

coronary artery calcium

CAD

coronary artery disease

CTAC

computed tomography attenuation correction

DL

deep learning

ECG

electrocardiogram

IQR

inter-quartile range

MACE

major adverse cardiovascular event

MPI

myocardial perfusion imaging

PET/CT

positron emission tomography/ computed tomography

SPECT

single photon emission computed tomography