Cross-Domain Knowledge Transfer from Expert-Annotated Gated CT via Synthetic Ungated CT Improves Coronary Artery Calcium Scoring on CT Attenuation Correction Scans

Abstract

Background:

Coronary artery calcium (CAC) is an established measure of coronary atherosclerosis from computed tomography (CT). While deep learning (DL) can quantify CAC from non-dedicated CT, the accuracy is limited by image quality.

Purpose:

We derived and validated a novel method for DL CAC segmentation on ultra-low dose CT attenuation correction (CTAC) scans that is trained with synthetic low-dose, ungated images.

Materials and Methods:

Models were trained using one center and externally tested in two other centers. Synthetic, ungated CT scans were generated so that expert segmentations from dedicated CAC scans could be used as ground truth for perfectly registered synthetic images through knowledge adaptation (KAD-CAC). We evaluated agreement between CAC scoring methods vs expert readers on a per-patient and per-vessel basis, as well as associations with the primary outcome of death or myocardial infarction (MI).

Results

The DL models were externally tested on 5969 patients with a median age of 64 (IQR 56 – 73), of whom 50.2% were male. The KAD-CAC model had higher Cohen’s kappa K (0.86, 95% CI 0.85 – 0.87) compared to previous convolutional LSTM model (K 0.78, 95% CI 0.76 – 0.80, p<0.01), or models trained with only gated images (K 0.81, 95% CI 0.80 – 0.82, p<0.01). Net reclassification improvement for CAC stratified risk of death or MI, was greatest for the KAD-CAC model over baseline including age, sex, hypertension, diabetes, dyslipidemia, family history, smoking, stress total perfusion deficit, and left ventricular ejection fraction.

Conclusion:

We use paired synthetic ungated scans to transfer expert gated CAC annotations into the ungated domain, resulting in substantially better vessel-level CAC scoring and improved risk stratification.

Summary Statement:

By transferring information from expert-annotated gated CT to synthetic ungated CT, a knowledge-adapted deep learning model substantially improved coronary artery calcium assessment on low-dose attenuation correction CT.

Introduction

Coronary artery disease (CAD) is the leading cause of morbidity and mortality worldwide (1). Accurate risk stratification is critical for guiding medical therapy for CAD. Coronary artery calcification (CAC) is an established marker of coronary atherosclerosis (2). The presence of CAC is associated with an increased risk of future cardiovascular events in both symptomatic and asymptomatic patients (3, 4). In addition, the vascular distribution of CAC has been shown to have important prognostic implications beyond the global CAC score (5, 6). CAC was originally described as a measure on dedicated, ECG-gated, computed tomography (CT) scans. However, there is now extensive literature demonstrating that CAC can also be evaluated on low dose lung cancer screening (7), and ungated CT attenuation correction (AC) imaging (8–10).

Deep learning (DL) has emerged as an efficient method to automate quantification of CAC on CT. Previous studies have shown the ability of a variety of DL models to identify CAC (11). It was previously demonstrated that a convolutional Long Short-Term Memory model (cLSTM) can accurately quantify CAC (12, 13). However, the low-dose CT AC scans acquired with SPECT and PET have relatively high image noise. A typical radiation dose associated with a CTAC scan is 0.2 to 0.3 mSv, while a typical radiation dose for a dedicated CAC scan is 1 to 3 mSv (14). This is further compounded by the lack of cardiac gating which results in artifacts related to coronary motion. As a result, it can be difficult to accurately identify small lesions and correctly assign vessel-specific CAC even for expert clinicians. Training DL models with dedicated solutions for these technical challenges has the potential to drastically improve the clinical utility of CAC scores.

The aim of this study was to develop and externally test a transformer-based segmentation architecture with a cross-domain knowledge transfer strategy for CAC from ultra-low dose CTAC scans. The approach includes generation of synthetic, low-dose ungated images so that model training can utilize expert segmentations from dedicated gated CAC scans as ground truth while also learning to discern image noise and motion artifacts from low-dose ungated CT. We evaluated agreement with expect CAC segmentations as well as risk stratification in an external population.

Materials and methods

Study Population

In this multi-center study, we used imaging and clinical data from three centers to train and test a knowledge adaptation-based approach for quantifying vessel specific CAC scores. This included internal training (n=2058) and validation (n=308) on ECG gated non-contrast cardiac CT from Cedars Sinai Medical Centre (Los Angeles, US). External testing was performed using low-dose, ungated CTAC imaging (n=5969) from patients undergoing SPECT/CT myocardial perfusion imaging at either Yale (New Haven, US) or the University of Calgary (Calgary, Canada)(15). Only the first study for each patient was included in the analysis. Patients with previous myocardial infarction (MI), percutaneous coronary intervention or coronary artery bypass grafting were excluded. This study was approved by the institutional review board at each site and was performed in accordance with the Declaration of Helsinki.

CT Image acquisition and image analysis

Detailed information regarding image acquisition in the study populations has been published previously (8, 16, 17). Additional details are available in the supplement.

Expert quantification of CAC was performed using dedicated software (QPET Suite, Cedars Sinai Medical Centre, Los Angeles). CAC scores were calculated on a per-vessel basis using the Agatston method for both ECG gated CT in training and CT AC imaging in test(18). The ground truth expert derived CAC scores were segmented by experienced technologists and clinicians with >10 years’ experience. CAC in the left main (LM), left anterior descending (LAD) coronary artery and diagonal arteries was assigned to the LM/LAD territory. CAC in the left circumflex (LCX) coronary artery and obtuse marginal branches was assigned to the LCX territory. CAC in the right coronary artery (RCA) and acute marginal branches was assigned to the RCA territory. CAC in the posterior descending artery and posterior lateral branches was assigned to the RCA territory, unless it could be established from the images that the circulation was left dominant.

Deep learning model architecture and training

An end-to-end DL framework was developed to enable accurate CAC quantification from synthetic ungated CT attenuation correction (CTAC) scans by leveraging anatomical knowledge from gated CAC scans. The proposed method (KAD-CAC) uses a 3D transformer-based segmentation architecture with a cross-domain knowledge transfer strategy to transfer coronary artery anatomical information from gated to ungated domains(19). The overall approach is designed to mimic expert reader behavior by exploiting three-dimensional spatial context across contiguous CT slices.

Network Architecture

The proposed framework follows the architecture shown in Figure 1. Additional details are available in the supplement. To ensure that the knowledge adaptation from gated scans benefits the overall model training, we compared against training the 3D Swin UNeTR with 1) gated data only and 2) synthetic ungated data only (Supplemental Figure 1).

Figure 1. Overview of model training and testing.

The model was trained using a cohort with paired gated and ungated computed tomography (CT) scans via a knowledge distillation strategy. We developed a synthetic (fake) ungated CT, with the model trained to segment CAC from synthetic ungated imaging with expert segmentations from gated chest CT as the ground truth. End-to-end model includes cost function as a weighted sum of gated segmentation loss, ungated segmentation loss and domain aware contrastive loss from embeddings Z_gated, Z_ungated obtained through a linear projector g. The network encoder f is shared during training. The red box indicates a zoom into the region as sliding window-based input and output volumes are generated.

Creation of Synthetic Ungated CTAC Volumes

Direct voxel-wise pairing between separately acquired gated CAC scans and ungated CTAC scans is challenging due to respiratory and cardiac motion, differences in breath-hold status, and partial volume effects(20), which can lead to misclassification of CAC (Supplemental figure 2). To address this limitation, paired synthetic ungated CTAC volumes were generated from gated CAC scans using a noise model derived from real ungated CTAC data. Creation of the paired synthetic ungated CTAC volumes enables use of the CAC annotations from ECG-gated CT scans to be used as the ground truth for perfectly co-registered synthetic ungated CT (Figure 2). Expert segmentation of ECG-gated CT scans has improved reproducibility compared to ungated, low-dose CTAC imaging (19). A blinded evaluation of images was used to evaluate the visual comparability of synthetic ungated imaging to actual ungated imaging. Additional details are available in the supplement.

Figure 2. Synthetic CTAC creation pipeline:

Ungated CTAC and gated CT scan pairs used to estimate noise using alternating direction method of multipliers (ADMM) -based penalized weighted least squares total variation (PWLS-TV). Noise statistics from ungated CTAC scans were then computed and injected into the gated volumes, with a scaling factor applied to match the standard deviation observed in real ungated data for noise characteristics to match. For CT intensity characteristics to match, we augmented the data with CT-like Gaussian noise rather than simple white noise. White noise was first spatially correlated using a Gaussian filter to match the characteristic grain size imposed by CT reconstruction kernels and slice thickness. The noise amplitude was then modulated as a function of Hounsfield Units HU, with higher variance in air and lung regions and lower variance in soft tissue and bone, reflecting the physics of CT image formation through colored noise. Voxels corresponding to calcified lesions (≥130 HU) were preserved to prevent artificial degradation of CAC regions. Sigma defines the standard deviation in different sections.

External testing

The DL model was tested on external ungated data from two sites which were not used in any way for model training. Per-vessel CAC scores for the expert ground truth on the ungated CTACs and DL models were categorized into four groups (0, 1–100, 101–400 and >400 Agatston units [AU]). Examples of DL and expert segmentations are shown in Figure 3.

Figure 3. Vessel specific deep learning (DL).

coronary artery calcium (CAC) scoring in patients with 0, 1–100, 101–400 and >400 Agatston units (AU) CAC assessed on CT attenuation correction (CTAC) by expert ground truth annotation and DL. Red –left main coronary artery, blue –left anterior descending artery, green –left circumflex artery

Clinical outcomes

In addition to evaluating CAC score agreement, we evaluated associations with death or myocardial infarction (MI). Our secondary outcome was incidence of major adverse cardiovascular events (MACE), which included death, MI, admission for unstable angina, or revascularization. Incidence of death was ascertained through the national death index in patients from Yale University and Alberta Vital statistics at the University of Calgary. Incidence of MI and admission for unstable angina was adjudicated by experienced clinicians at each site using standard criteria (17).

Statistical Analysis

Details are available in the supplement.

Results

Study population

The DL model was externally tested on 5969 patients with a median age of 64 (IQR 56 – 73), of whom 2995 (50.2%) were male. Characteristics of the training and external testing sets are shown in Table 1. A summary of CAC scores using the various models is shown in Supplemental Table 1. Training and internal validation summary of CAC scores in outlined in Supplemental Table 2.

Table 1.

Population characteristics in the training and external testing datasets. Continuous variables shown as median (interquartile range)

	Overall N=8335	Training and Validation N=2366	External testing N=5969	P-value
Age	66 (57, 74)	70 (62, 78)	64 (56, 73)	<0.001
Male (%)	4160 (49.9%)	1165 (49.2%)	2995 (50.2%)	0.44
Body mass index (kg/m2)	28.7 (25, 33.3)	27.6 (24.2, 32)	29.2 (25.4, 33.8)	<0.001
Hypertension	5285 (63.4%)	1744 (73.7%)	3541 (59.3%)	<0.001
Diabetes mellitus	2160 (25.9%)	719 (30.4%)	1441 (24.1%)	<0.001
Dyslipidemia	4235 (50.8%)	1413 (59.7%)	2822 (47.3%)	<0.001
Family history of CAD	2069 (24.8%)	374 (15.8%)	1695 (28.4%)	<0.001
Current smoker	1076 (12.9%)	188 (7.9%)	888 (14.9%)	<0.001

Validation of synthetic CTAC imaging

Likert scale visual analysis on a set of 300 images (including 100 gated, ungated and synthetic CT attenuations scans) revealed Bang BI for gated images of 0.58 (95% CI: 0.42–0.74), indicating substantial recognition of gated images. For ungated images, the BI was 0.79 (95% CI: 0.70–0.87), indicating strong recognition of ungated appearance. Only three equivocal ratings (1%) were recorded overall. Gated images were predominantly rated as gated-like (79%), whereas ungated images were primarily rated as ungated-like (79%) but exhibited greater variability. Synthetic ungated images were almost uniformly rated as ungated or ungated-like (97%), with minimal gated-like or equivocal ratings. (Supplemental Figure 3). Overall, these results suggest that the synthetic ungated images preserve the perceptual characteristics of ungated motion, remaining far closer to real ungated images than to gated images, while exhibiting reduced heterogeneity and fewer ambiguous appearances.

Across all gated CAC and ungated CTAC cases, the synthetic CTAC images closely reproduced the noise characteristics of actual CTAC scans. Noise magnitude agreement was strong, with near-unity noise standard-deviation ratios at the voxel level (median 1.03, IQR 0.93–1.13) and at the patch level in whole-body regions (median 0.94, IQR 0.86–1.05) as well as within a cardiac-focused central region (median 0.95, IQR 0.83–1.05). Patch-based distributional comparison demonstrated limited local differences, with median Kolmogorov–Smirnov statistics of 0.046 (IQR 0.037–0.066) globally and 0.048 (IQR 0.032–0.071) in the cardiac region, and corresponding patch-level Wasserstein distances of 6.36 HU and 6.10 HU respectively. Global HU intensity profiles showed substantial overlap between real and synthetic ungated CTAC volumes, with low Jensen–Shannon divergence (0.08). Stratified intensity analysis indicated stronger agreement for high-attenuation voxels (≥130 HU) than for lower-HU regions, consistent with explicit preservation of calcification-relevant structures during synthesis.

Assessment of DL CAC measures

Assessment of vessel-specific KAD-CAC scoring took 20.6 s ± 2.8 per scan using a graphics processing unit (GeForce RTX 3090, AMD Ryzen 9 5950X 16-core Processors, 64 GB RAM).

The correlation between all DL CAC measures and expert CAC scoring was excellent. There was higher kappa for the KAD-CAC model (Cohen’s kappa 0.86, 95% CI 0.84 – 0.88) compared to the cLSTM model trained on gated scans (Cohen’s kappa K 0.78, 95% CI 0.76 – 0.80), Swin UNeTR trained on gated scans (K 0.81, 95% CI 0.80 – 0.82, p<0.001), and Swin UNeTR trained on synthetic ungated scans(K 0.84, 95% CI 0.83 – 0.85) with all pairwise differences remaining statistically significant after false discovery rate (FDR) correction (adjusted p < 0.01). Concordance with expert CAC scores was higher for the KAD-CAC model compared to the other DL models in all coronary territories (results in Supplemental Table 3). Per-vessel concordance matrices are shown in Supplemental Figure 4–7. Linear weighted Cohen’s Kappa showed significant improvement in all vessels except LAD when compared to the KAD- CAC model. The results for all models are summarized in Figure 4.

Figure 4. Agreement between DL and expert ground truth.

coronary artery calcification (CAC) for the all-vessel score, left main (LM), left anterior descending (LAD), LM/LAD, left circumflex (LCX), and right coronary artery (RCA) for all patients undergoing attenuation correction (AC) CT. p<0.01 for all comparisons to KAD-CAC except for LAD. ICC- Inter Class Correlation. cLSTM- convolutional long, short term memory network trained on gated data.

For the identification of any CAC on CTAC (identified by expert segmentation), the KAD-CAC model demonstrated excellent diagnostic performance on a global and per-vessel basis. Sensitivity, specificity, negative predictive value (NPV), and positive predictive value (PPV) were 94%, 89%, 95%, and 88% for the LM/LAD territory; 84%, 73%, 71%, and 86% for the LCX territory; and 80%, 84%, 70%, and 90% for the RCA territory, respectively. For total CAC, the KAD-CAC model demonstrated 92% sensitivity, 87% specificity, 88% NPV, and 91% PPV.

Associations with outcomes

During a median follow-up of 813 days (IQR 498 – 1185), 450 patients experienced either death (total n=349) or MI (total n=122). Incidence of death or MI as a function of CAC category using the KAD-CAC model is shown in Supplemental Figure 8. Compared to patients with CAC 0, patients with CAC 1–100 (unadjusted hazard ratio [HR] 2.56, 95% CI 1.90 – 3.46, p<0.001), CAC 101–400 (unadjusted HR 3.61, 95% CI 2.66 – 4.90, p<0.001) and CAC >400 (unadjusted HR 5.38, 95% CI 4.08 – 7.10, p<0.001) were at increased risk of death or MI. The risk associated with CAC categories in unadjusted and adjusted analyses for each method of CAC scoring are shown in Table 2. Kaplan-Meier curves stratified by expert segmentation CAC categories are shown in Supplemental Figure 9. The risk associated with discordant classification of CAC 0 was also assessed, with a summary shown in Supplemental Figure 10.

Table 2:

Associations with death or myocardial infarction. The multivariable model included age, sex, medical history (hypertension, diabetes, dyslipidemia), family history of CAD, smoking, stress total perfusion deficit and left ventricular ejection fraction.

		Unadjusted HR (95% CI)	p-value	Adjusted HR (95% CI)	p-value
Expert	CAC 0	Reference		Reference
Increase in LR chi 38.5	CAC 1 – 100	2.04 (1.53 – 2.73)	<0.001	1.61 (1.20 – 2.18)	0.002
	CAC 101 −400	3.00 (2.25 – 4.00)	<0.001	1.98 (1.45 – 2.69)	<0.001
	CAC>400	4.36 (3.36 – 5.66)	<0.001	2.47 (1.83 – 3.28)	<0.001
KAD-CAC	CAC 0	Reference		Reference
Increase in LR chi 60.1	CAC 1 – 100	2.56 (1.90 – 3.46)	<0.001	2.11 (1.56 – 2.87)	<0.001
	CAC 101 −400	3.61 (2.66 – 4.90)	<0.001	2.54 (1.84 – 3.51)	<0.001
	CAC>400	5.38 (4.08 – 7.10)	<0.001	3.14 (2.30 – 4.28)	<0.001
Swin UNeTR trained on ungated data	CAC 0	Reference		Reference
Increase in LR chi 58.1	CAC 1–100	2.41 (1.80 – 3.24)	<0.001	1.99 (1.47 – 2.68)	<0.001
	CAC 101 −400	3.29 (2.44 – 4.43)	<0.001	2.30 (1.68 – 3.14)	<0.001
	CAC>400	5.16 (3.96 – 6.72)	<0.001	3.03 (2.25 – 4.09)	<0.001
Swin UNeTR trained on gated data	CAC 0	Reference		Reference
Increase in LR chi 53.7	CAC 1 – 100	2.42 (1.82 – 3.22)	<0.001	1.92 (1.43 – 2.57)	<0.001
	CAC 101 −400	3.11 (2.32 – 4.17)	<0.001	2.19 (1.61 – 2.98)	<0.001
	CAC>400	4.78 (3.70 – 6.17)	<0.001	2.78 (2.09 – 3.72)	<0.001
Old cLSTM	CAC 0	Reference		Reference
Increase in LR chi 50.1	CAC 1 – 100	2.15 (1.59 – 2.91)	<0.001	1.86 (1.37 – 2.52)	<0.001
	CAC 101 −400	3.40 (2.50 – 4.62)	<0.001	2.32 (1.68 – 3.20)	<0.001
	CAC>400	4.96 (3.73 – 6.58)	<0.001	2.89 (2.11 – 3.95)	<0.001

CAC – coronary artery calcium, HR – hazard ratio, LR – likelihood ratio.

Categorical NRI for death or MI is shown in Figure 5. Categorical NRI was highest for the KAD-CAC model (categorical NRI 0.157, 95% CI 0.103 – 0.210) over a baseline model including age, sex, hypertension, diabetes, dyslipidemia, family history, smoking, stress total perfusion deficit, and left ventricular ejection fraction. The positive NRI was 0.051 (95% CI 0.001 to 0.102) and the negative NRI was 0.106 (95% CI 0.088 to 0.123). The categorical NRI for expert CAC annotated on CTACs was compared with the KAD-CAC model using a paired bootstrap (5,000 samples) to account for correlation. The difference in categorical NRI was 0.089 (95% CI 0.051 to 0.125; p <0.001). The difference in categorical NRI was also significant when compared to the model trained with synthetic ungated data (p=0.004) and gated data (p=0.012). Associations with MACE are shown in Supplemental Table 4 and Supplemental Figure 11.

Figure 5. Categorical Net-reclassification improvement (NRI).

for CAC through different methods with respect to prediction of death or myocardial infarction. The the baseline model included age, sex, hypertension, diabetes, dyslipidemia, family history, smoking, stress TPD, and LVEF as covariates * - p<0.05 compared to all other models. Expert CAC -scoring performed on ungated data. cLSTM model trained on gated data only

Case Examples

Supplemental Figure 12 shows improvement in the LM identification on CTAC imaging from a female patients in her 80s with 3 vessel involvement. The expert LM CAC score was 41 and KAD-CAC LM score was 49. The cLSTM model assigned the LM lesion to the LAD and LCx territories, with LM score of 0.

Additional case examples are shown in Supplemental Figure 13 and Supplemental Figure 14. Graphical abstract is included in Figure 6.

Figure 6. Graphical abstract.

Vessel-specific coronary artery calcium (CAC) from attenuation correction CT (CTAC) patients in a large multi-center registry with Deep Learning can be performed accurately, rapidly, and provides important prognostic information for death or myocardial infarction.

Discussion

In this study, we have developed a DL model which is capable of vessel specific CAC assessment and demonstrated that it can rapidly and accurately assess CAC in a large, multi-center external dataset. In order to account for image noise and motion artifacts common in CTAC imaging, we developed a method for synthetic low-dose, ungated CT images from dedicated CAC scans. By training with synthetic CTAC images the model benefits from both expert segmentations from high-quality gated studies to utilize as the ground truth while also learning the image noise and motion artifact patterns of CTAC imaging. The DL model showed strong to excellent agreement and concordance with expert CAC segmentation, both on a total and per-vessel basis. This approach outperforms models trained without data augmentation as well as models trained only on gated or synthetic ungated imaging. DL CAC, by all methods, was an independent predictor of death or MI. Furthermore, the risk reclassification was superior for the proposed KAD-CAC model compared to the other DL models or expert segmentations on ungated CTACs when evaluated in an external population. The proposed KAD-CAC model could be applied in clinical practice to provide rapid and accurate CAC quantification from ungated CT chest imaging.

Automated approaches for CAC segmentation have the potential to dramatically improve cardiac risk stratification (2) and guide therapeutic decisions (21, 22). While this information can be obtained from dedicated studies, many chest CT scans performed for other reason could be leveraged for this information opportunistically (7–10). These studies are typically performed without cardiac gating, and with lower radiation exposure (resulting in increased image noise), which can make it challenging to identify small lesions (23). To ensure these limitations were reflected in the training cases, we developed synthetic CTAC images which included both motion artifact and increased image noise (with similar signal to noise as actual ungated imaging). The synthetic images had noise characteristics and an overall appearance that closely approximated actual ungated imaging. The main advantage of using synthetic CTAC images is that they are perfectly registered with the higher-quality, gated CT images (which can then be used as the ground truth segmentation). Previous studies have demonstrated that CAC scores from dedicated ECG-gated CT Agatston have lower interobserver variability compared to CAC scores from low-dose, ungated CT imaging (24). Even expert physicians may struggle with differentiating some CAC lesions from noise on CTAC imaging; whereas this task is relatively simple from dedicated gated imaging. As a result of this novel training approach, KAD-CAC may identify some CAC lesions which are only readily apparent on dedicated CT scans but are difficult to identify on CTAC imaging alone. This technical advancement of the KAD-CAC model, and external testing in a large population from two sites, suggest that the model should be generalizable to a large variety of ungated chest CT scans.

While many previous studies have focused solely on total CAC score as the measure of agreement, we also evaluated results on a per-vessel level. This agreement helps ensure that the CAC identified by the DL is the same CAC lesion identified by expert segmentation. However, to date this agreement has only been evaluated in a few studies (25–28). Winkel et al tested a convolutional neural network on 262 ECG-gated CT scans and showed a 94% accuracy for the identification of the correct branch, but vessel specific CAC scores were not assessed (29). In the present study, we demonstrated that the KAD-CAC model demonstrated superior per-vessel agreement with expert segmentations across all vascular territories compared to the other DL models. While this agreement helps support the overall accuracy of the model’s segmentations, it may also provide additional prognostic insights. In non-gated lung cancer screening CT de Vos et al used location specific CAC, along with self-reported cardiovascular risk factors, to predict cardiovascular disease mortality with an AUC of 0.76 (30). In previous work (28), CAC>400 in the left main was associated with the highest risk of MACE. Accordingly, vessel-level accuracy may represent an important step toward clinically actionable, lesion-specific CAC assessment.

The true test for any imaging marker is whether it can provide robust risk stratification to guide therapeutic decisions (31). In this regard, we evaluated associations with death or MI. As expected, we identified a higher risk of death or MI with increasing CAC score for all DL models and expert segmentation. While the differences in risk stratification were small, the KAD-CAC model had the greatest improvement in model fit. Additionally, categorical NRI was significantly higher for the KAD-CAC model compared to expert CAC segmentation as well as compared to the other DL methods. As mentioned above, the unique training approach for KAD-CAC allows it to benefit from expert annotations from dedicated gated images. As a result, and one potential reason for improved risk stratification, KAD-CAC may be able to differentiate small CAC lesions from noise on CTAC imaging. In support of this possibility is the finding that risk in patients classified as CAC 0 by the KAD-CAC model (but not by expert segmentation) was lower compared to patients classified as CAC 0 by expert segmentation (but not by the KAD-CAC model). These results suggest that classification of CAC 0 by the KAD-CAC model could provide reassurance for patients where there is uncertainty regarding cardiovascular risk.

Study Limitations

This study has a few important limitations. We utilized expert segmentation of CTAC images as the reference standard when assessing agreement with DL scores. However, there remains some uncertainty regarding our ability to accurately identify CAC on CTAC scans. In particular, due to image noise and cardiac motion there will be voxels with inherent uncertainty in classification. We excluded patients with known coronary artery disease including previous revascularization and therefore did not evaluate how well the DL models differentiate stents or surgical clips from CAC. However, it is unlikely that the extent of CAC would significantly shift the cardiovascular risk of those patients since they would all typically be considered high risk.

Conclusions

We developed a transformer-based segmentation architecture with cross domain knowledge transfer strategy for CAC segmentation. More specifically, we generated synthetic ungated images from paired gated CAC scans so that the model learns from both perfectly co-registered high-quality expert segmentations and the image noise and motion characteristics of CTAC imaging. The KAD-CAC model provided better agreement with expert annotations compared to DL models trained without synthetic AC images or models trained using either the gated or ungated images alone. Furthermore, this approach provides superior risk stratification for death or MI compared to expert segmentation or other DL approaches.

Key Results:

In external testing using 5969 patients from two sites, the proposed model had improved agreement with expert segmentation and provided the greatest net reclassification improvement for CAC stratified risk of death or MI over baseline including age, sex, hypertension, diabetes, dyslipidemia, family history, smoking, stress total perfusion deficit, and Left Ventricular Ejection Fraction.

List of abbreviation

AI

artificial intelligence

CAC

Coronary Artery Calcium

CTAC

computed tomography attenuation correction

MPI

myocardial perfusion imaging

PET

positron emission tomography

SPECT

single-photon emission computed tomography

cLSTM

convolutional Long Short-Term Memory

Funding Statement

This research was supported by grant R35HL161195 from the National Heart, Lung, and Blood Institute/ National Institutes of Health (NHLBI/NIH) (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.

Data Availability

To the extent allowed by data sharing agreements and institutional review board protocols, data and codes used in this manuscript will be shared upon written request.

Code Availability:

cLSTM code code is publicly available under a Creative Commons BY-NC license at https://doi.org/10.5281/zenodo.10632288. SWIN-UNetR code available: https://monai.readthedocs.io/en/0.9.1/_modules/monai/networks/nets/swin_unetr.html