This study includes three data sets. First, a data set of labeled low-dose chest NCCT scans from the National Lung Screening Trail (NLST) was used. The NLST enrolled 53,454 current or former heavy smokers aged 55–74 in the United States (26). In our previous study, a set of 1,687 baseline chest NCCT scans was selected (16). This set was designed to be diverse with respect to scanner model and reconstruction algorithm. The selected scans were acquired on 13 different scanner models in 31 hospitals. These chest NCCT scans were acquired with breath hold after inspiration and using a tube voltage 120 or 140 kVp, depending on the subjects weight. Scans were reconstructed to 0.49–0.98 mm in-plane resolution, 1–3 mm slice thickness, and 0.6–3 mm increment. For our work, all scans were resampled to 3 mm slice thickness and 1.5 mm increment, following earlier studies (16).

Second, a mixed set of labeled cardiac NCCT and unlabeled CCTA scans was used. Specifically, 200 labeled cardiac NCCT scans were acquired in clinical patient workup at University Medical Center Utrecht, The Netherlands (19, 27) and 200 unlabeled CCTA scans were acquired at Amsterdam University Medical Center location University of Amsterdam, The Netherlands. The cardiac NCCT scans were acquired with a Philips Brilliance iCT 256 scanner, with ECG synchronization and 120 kVp tube voltage. Scans were reconstructed to 0.29–0.49 mm in-plane resolution, 3 mm slice thickness, and 1.5 increment. The CCTA scans were acquired with a Siemens Somatom Force CT Scanner, with ECG synchronization and 70–120 kVp tube voltage. Scans were reconstructed to 0.22–0.46 mm in-plane resolution, 0.6 mm slice thickness, and 0.4 mm increment.

Third, a data set of labeled 313 CCTA scans from Amsterdam University Medical Center location University of Amsterdam, The Netherlands was used to evaluate the CAC detection on the target domain (CCTA test set). These CCTA scans were acquired with the Siemens Somatom Force CT Scanner, with ECG synchronization and 70–120 kVp tube voltage. Scans were reconstructed to 0.19–0.77 mm in-plane resolution, 0.6–1 mm slice thickness, and 0.4 mm increment.

Manual reference labels of CAC were available from previous studies for the low-dose chest NCCT scans in the NLST data set(16) and the cardiac NCCT in the mixed set (19). The labeling was performed semi-automatically: all regions of ≥ 3 adjacent voxels with a CT value above 130 HU were shown as overlay. An observer manually identified lesions and labeled them according to their anatomical location, i.e., left anterior descending artery (LAD), left circumflex artery (LCX), or right coronary artery (RCA) (19). Given that chest CT without ECG synchronization does not allow visualization of the left main (LM) artery, CAC in the LM was labeled as LAD. Examples of chest NCCT slices and manual reference annotations are shown in the Supplementary Figure S1.

For the 200 CCTA scans in the mixed set, reference labels of CAC were not available. Hence, for the CCTA scans from the CCTA test set, CAC was manually annotated with a semi-automated method as either LAD, LCX, or RCA. This was done using an in-house developed software designed in MevisLab 3.2 (28). In agreement with manual labeling in NCCT, CAC in the LM was labeled as LAD. Because the standard 130 HU threshold for CAC detection in NCCT can not be used in CCTA, we used scan specific thresholds, following earlier studies (25, 29). For this, a region of interest (ROI) defined by a bounding box with a size around 35 × 36 × 44 voxels in the ascending aorta at the level of the origin of the left coronary artery was manually selected. Subsequently, the mean mean_ROI and standard deviation STD_ROI from the CT values of the voxels within the ROI were used to compute a scan specific threshold mean_ROI+3STD_ROI. Using this threshold, each coronary artery calcification was manually identified by a mouse click on the lesion. Subsequently, all connected voxels in the lesion above the scan specific threshold were marked as CAC in LAD, LCX, or RCA using 3D connected component labeling considering six-voxel connectivity. Examples of CCTA slices and manual reference annotations are shown in the Supplementary Figure S2.

In this study, NCCT scans (both chest and cardiac) are considered the source domain and CCTA scans are representing the target domain. The NCCT scans with CAC annotations from the NLST data set were used to train the CAC detection network on the source domain. The mixed set of labeled cardiac NCCT (source domain) and unlabeled CCTA (target domain) was used to train our unsupervised domain adaptation method. The labeled CCTA scans (target domain) in the CCTA test set were only used to evaluate the CAC detection on the target domain. The description of data sets and their usage are illustrated in Table 1.

Unsupervised domain adaptation aims to transfer a model trained with data from a source domain with labels Ds=(Xsi,Ysi)i=1..ns to a target domain without labels Dt=(Xti)i=1..nt, where D represents domain, X represents images and Y represents labels. As proposed by Dou et al. (8), we use an adversarial training strategy to adapt the CNN to the target domain. In our application, a large set of chest NCCT scans with CAC labels is available, and hence, we aim to transfer the knowledge from NCCT to CCTA for CAC scoring. Therefore, the CAC scoring CNN trained with labeled low-dose chest NCCT scans is transfered to CCTA using adversarial domain adaptation.

We used our previous CAC scoring method described by Lessmann et al. (16) that has been trained and evaluated with a large set of low-dose chest NCCT data. The method consists of two sequential convolutional neural networks (CNN). The first CAC scoring CNN detects CAC candidates and labels them according to their anatomical location, i.e., as CAC in LAD, LCX, or RCA. The second CNN reduces the number of false positive detections. In our current work, only the first CNN is used to transfer knowledge obtained by training the network with NCCT to enable application in CCTA data using unsupervised domain adaptation.

To adapt the CAC detection network(16) from the source domain to the unlabeled target domain, we aim to align the distributions of extracted features from the two domains following the work by Dou et al. (8). For this, we divide the CAC detection network into two parts: a feature generator G(·) and a classifier C(·), as shown in Figure 1. The G(·) maps input images into feature representations in the latent space and the C(·) predicts the output class from the feature representations. The early layers of the network which are used for feature extraction are mostly related to the domain, while the deeper layers are mostly task-specific and learn semantic-level features for conducting the predictions (8, 30). Hence, we adapt the feature generator G(·) trained with NCCT to enable application in CCTA with adversarial domain adaptation, and we reuse the classifier C(·) as originally trained.

To enable adversarial learning, we design a discriminator D(·) to identify whether the features are from the source domain or the target domain. While the feature generator G(·) aims to extract features with similar distributions for both domains, the D(·) discriminates between the two domains (Figure 2). The adversarial loss based on the differences in feature distribution between the two domains is formulated as:

where G(·) is optimized to minimize the adversarial loss, and D(·) is optimized to maximize the same loss. The generator G(·) is optimized based on the objective function calculated from the discriminator D(·), which can lead to an incorrect optimization forgetting the classification task. That means the features extracted by the trained G(·) can fool the D(·). However, these features are not beneficial for the final classification task C(G(·)). For cross domain learning with paired data, the alignment loss in feature space, such as L1(G(x_s), G(x_t)) or L2(G(x_s), G(x_t)), can be used as a constraint for the generator optimization (31). For cross-domain learning with unpaired training data as in our case, such an alignment loss in feature space can not be used as a constraint for the generator optimization. In this work, the images were not registered to a common space either. Instead, as proposed in the work by Chen et al. (4), we use a classification loss in the source domain Ds as constraint to stabilize the training and avoid catastrophic forgetting.

The classification loss is formulated as:

where L_CE is the cross-entropy loss, X_s and Y_s are the images and the corresponding reference labels on the source domain. During training, the D(·) is trained to maximize the objective of Ladv, while the G(·) is optimized to minimize the objective of Ladv and Lcls. These are formulated as:

where α is a hyper-parameter for balancing the two loss terms. It is set to 2.0 in this work, based on a grid search strategy.

To identify CAC lesions, 3D connected component labeling is performed from the detected voxels and the scan specific threshold (25, 29). To remove potential false positive detections, detected lesions smaller than 1 mm³ are discarded as those are likely noise voxels. Similarly, detected lesions larger than 500 mm³ are discarded as those exceed the expected CAC volume (27). In addition, lesions detected outside the heart are discarded. For this, the heart volume is defined by segmentation of cardiac chambers, as described by Bruns et al. (32) which was trained with CCTA scans of 12 patients scanned for transcatheter aortic valve implantation (SOMATOM Force, Siemens, 70–120 kVp, 310–628 mAs, in-plane resolution 0.31–0.61 mm, slice thickess 0.31–0.61 mm, slice increment 0.45 mm). No additional changes or fine tuning for the data in this current study was performed. Subsequently, the segmentation of cardiac chambers was dilated by a sphere as a structuring element with diameter of 10.0 mm to ensure the heart wall and coronary arteries are included in the segmentation.

To evaluate the performance of CAC scoring on CCTA, the volume-wise and lesion-wise performance was determined by comparing automatically detected CAC with the manually annotated reference. Since the typically used Agatston score (11) is not applicable for CAC quantification in CCTA, the volume score was used. The evaluation was performed for total CAC and separately for CAC in LAD, LCX, and RCA. Both the volume-wise and lesion-wise performance was evaluated using sensitivity, false-positive (FP) rate, and F1 score (16). The agreement of calcium volume and number of lesions between the automatic detection and the reference labels was determined with Spearman correlation coefficients. Finally, the agreement between automatic volume scores and manual reference volume scores was assessed by examining Bland-Altman plots including 95% limits of agreement. Since errors tend to increase with increasing CAC volume, the variation of absolute differences between automatic and manual scores was modeled using regression for nonuniform differences(33). Because the absolute differences have a half-normal distribution, the modeled absolute differences were multiplied by 1.96 × (π/2)^0.5 to obtain the 95% limits of agreement.

First, we retrained the two-stage CNNs for CAC detection (16) with the labeled chest NCCT data as the source domain. For this, the 1,687 NCCT scans in the NLST data set were randomly divided into 60% training set (1,012 scans), 10% validation set (169 scans), and 30% test set (506 scans). As originally reported (16), during the training, categorical cross-entropy was used as loss function, Adam was used as optimizer with a learning rate of 5 × 10⁻⁴. The first CNN was trained with three orthogonal (axial, sagittal and coronal) patches of 155×155 pixels and the second CNN with three orthogonal patches of 65×65 pixels (16). Randomized patch extraction was used as augmentation for training.

Next, to stabilize adversarial training in the unsupervised domain adaptation, the generator was initialized with the weights of the CAC scoring model trained with the chest NCCT data from the NLST dataset. The unsupervised domain adaptation method was trained with the mixed dataset of labeled cardiac NCCT data from source domain and unlabeled CCTA data from target domain. When performing unsupervised domain adaptation with mixed data containing labeled cardiac NCCT and unlabeled CCTA scans the method achieved sensitivity of 0.78 in CCTA (Table 2). For comparison, the sensitivity of 0.53 was achieved when unsupervised domain adaptation was performed with mixed data containing labeled chest NCCT and unlabeled CCTA scans. Labeled cardiac NCCT data was chosen because these scans resemble CCTA scans more than chest NCCT. Unlabeled CCTA were used as unlabeled data from the target domain. To obtain a reliable discriminator, the discriminator was solely pretrained for 1,000 iterations first. Thereafter, the generator and discriminator were optimized together by training alternately. Specifically, the generator was optimized one iteration after every 20 iterations of the discriminator, according to the heuristic rules of training a Wasserstein GAN (34). Following the standard for adversarial training (34, 35), the discriminator was kept in a compact space. To enforce this constraint, the weights were clipped between [−0.1, 0.1]. The RMSProp optimizer was used to optimize the discriminator with a learning rate of 5 × 10⁻⁴, and the generator with a learning rate of 5 × 10⁻⁵, respectively (36). The optimal hyperparameters were determined by grid search. The adversarial learning was trained for 200 epoch. The networks were implemented in PyTorch (37). All the training was trained on NVIDIA GeForce RTX 2080 Ti.

To establish the performance of the CNN adapted from NCCT to CCTA, the network was evaluated with the 313 labeled CCTA test scans. The adapted CNN obtained an average volume-wise sensitivity of 0.78, an average FP volume per scan of 73.9 mm³ and an F1-score of 0.41. After the FP reduction, the proposed method achieved an average volume-wise sensitivity of 0.80 with an average FP volume per scan of 19.8 mm³, and F1 of 0.66. There were 36 patients without CAC but with FP detected by the proposed method, with an average FP volume per scan of 40 mm³. The Spearman correlation between automatically detected and reference CAC volume was 0.73. The Bland-Altman plots comparing automatically detected CAC volume with manually annotated reference are illustrated in Figure 3.

Coronary CT angiography slices and corresponding automatic CAC detections for two outliers cases (marked orange in Figure 3) are shown in Figures 4a,b. In addition, two representative cases from the labeled CCTA test set are shown in Figures 4c,d. For lesion-wise evaluation, the proposed method achieved an average sensitivity of 0.79 and FP lesion per scan of 1.06. The correlation between the number of automatically detected and manually annotated reference lesions was 0.69.

To establish whether our retraining of the original CAC scoring network on the source domain led to adequate performance, the CAC scoring network was evaluated on NLST test set (Section CAC scoring on CCTA) and compared with the originally reported results (16). Results are listed in Table 2 (column 3 showing NCCT results). Our retained network obtained a sensitivity of 0.89, an average FP volume of 73.6 mm³ per scan and F1 of 0.66. The sensitivity is in agreement with the results (0.84 - 0.91) reported in the original work (16), while the originally reported FP rate (40.7–62.8 mm³) and therefore F1 (0.84–0.89) slightly outperform our results.

To evaluate the performance of the two-stage CAC scoring networks trained on NCCT to CCTA, the trained CNNs was directly applied to CCTA test scans without adversarial domain adaptation learning. This led to an average sensitivity of 0.41, an average FP volume per scan of 139.7 mm³, and F1 of 0.16 (Table 2, column 7 showing the CCTA results). Subsequently, adding FP reduction led to an average sensitivity of 0.43, an average FP volume of 0.58 mm³ and F1 of 0.41. Note that FP reduction stage slightly improved the sensitivity as the region-growing algorithm (38) used to define the lesions from the voxels detected by the CNN may improve lesion segmentation and lead to better agreement with manual reference that used the region-growing algorithm to define CAC lesions.

To investigate the benefit of using the adversarial loss and classification loss for domain adaptation, and FP reduction, additional experiments were performed. The proposed method obtained a volume-wise sensitivity of 0.80, average FP volume per scan of 19.8 mm³, and F1 of 0.66. Without FP reduction, the volume-wise sensitivity decreased to 0.78, average FP volume per scan increased to 64.5 mm³ and consequently, F1 score decreased to 0.41. Furthermore, removing the classification loss Lcls from the objective function resulted in the volume-wise sensitivity of 0.68, average FP volume per scan of 25.8 mm³, and F1 of 0.49. Finally, as described above, removing the adversarial loss Ladv (i.e., without adversarial domain adaptation learning) led to sensitivity of 0.41, FP volume of 139.7 mm³ per scan, and F1 of 0.16. Detailed results are listed in Table 2 columns 4–7.

The performance of the proposed method was compared with previously published methods that use deep learning for CAC scoring in CCTA scans (22–25). Wolterink et al. (25) proposed a method that employed paired CNNs for CAC scoring. The first CNN was used to identify CAC-like voxels and the second CNN was used to reduce CAC-like negatives. Fischer et al. (22) proposed a method that firstly detected the coronary artery centerlines and then identified CAC in cross-sectional images along the detected centerlines using long short-term memory (LSTM). In the study by Liu et al. (23), a vessel focused 3D CNN was proposed for CAC detection. The coronary arteries were firstly extracted and straightened volumes were reformed along the coronary arteries. Thereafter, a CNN was used for CAC detection. The results as reported in the original work are listed in Table 3. These demonstrate that our unsupervised method achieved competitive performance. Given that the original implementations of these earlier studies are not publicly available, the compared methods the results should be used as indication only.