Generative models for reproducible coronary calcium scoring

Abstract

Purpose: Coronary artery calcium (CAC) score, i.e., the amount of CAC quantified in CT, is a strong and independent predictor of coronary heart disease (CHD) events. However, CAC scoring suffers from limited interscan reproducibility, which is mainly due to the clinical definition requiring application of a fixed intensity level threshold for segmentation of calcifications. This limitation is especially pronounced in non-electrocardiogram-synchronized computed tomography (CT) where lesions are more impacted by cardiac motion and partial volume effects. Therefore, we propose a CAC quantification method that does not require a threshold for segmentation of CAC. Approach: Our method utilizes a generative adversarial network (GAN) where a CT with CAC is decomposed into an image without CAC and an image showing only CAC. The method, using a cycle-consistent GAN, was trained using 626 low-dose chest CTs and 514 radiotherapy treatment planning (RTP) CTs. Interscan reproducibility was compared to clinical calcium scoring in RTP CTs of 1662 patients, each having two scans. Results: A lower relative interscan difference in CAC mass was achieved by the proposed method: 47% compared to 89% manual clinical calcium scoring. The intraclass correlation coefficient of Agatston scores was 0.96 for the proposed method compared to 0.91 for automatic clinical calcium scoring. Conclusions: The increased interscan reproducibility achieved by our method may lead to increased reliability of CHD risk categorization and improved accuracy of CHD event prediction.

Keywords: calcium scoring; computed tomography; cycle-consistent generative adversarial network; generative models; reproducibility.

Fig. 1

An image containing CAC is decomposed into an healthy tissue image without CAC and an image containing only CAC.

Fig. 2

Examples of two RTP scans of one breast cancer patient, (a) one with BH and (b) one without BH, and (c) an NLST scan of one participant.

Fig. 3

Schematic example of the weak training labels per axial slice indicating whether CAC is present in the (a) slice. (b) Indicated region of the lesion with assigned coronary artery label.

Fig. 4

Schematic overview of the proposed CycleGAN that translates the images from the domain of images with CAC (CAC domain) to the domain of images without CAC (No CAC domain) and back. In the pathway that removes CAC, a CAC map is predicted by the generator (${\mathrm{G}}_{\mathrm{R}}$) and subtracted from the image to obtain a synthetic image without CAC. In the pathway that synthesizes CAC the CAC map predicted by the generator (${\mathrm{G}}_{\mathrm{S}}$) is added to the image. Synthetic images are compared to real examples by discriminators (D).

Fig. 5

$F1$-score for per scan CAC detection and per scan absolute relative difference for concordant positive scans (${\mathrm{\Delta }}_{\mathit{\text{R}}}\text{P}\text{os}$) on the validation set.

Fig. 6

The results of the ablation study for the method with and without heart segmentation (HS ±) and with and without slice classification (CL ±). We evaluated the (a) detection performance with sensitivity, FPR and $F1$-score. We evaluated the (b) interscan reproducibility with absolute relative difference in CAC mass for all pairs (${\mathrm{\Delta }}_R$ All) and concordant positive pairs (${\mathrm{\Delta }}_R$ Pos). For comparison, the performance of clinical manual and clinical automatic calcium scoring are shown with horizontal lines.

Fig. 7

Examples of lesions, visual assessment of the CAC lesion area, calcium segmentation using the clinical definition, corresponding synthetic images without CAC and segmentation of CAC with the proposed method and following the clinical protocol. Lesions with excessive motion are indicated with the blue box.

Fig. 8

Bland–Altman plots for adjusted Agatston scores of the proposed and automatic clinical method. For comparison purposes the adjusted Agatston scores of the proposed method are scaled to the range of scores of the clinical method. We can note that since the same linear scaling is used for all scores, this does not influence the agreement. About 95% limits of agreement are indicated by dashed lines. Because the errors tend to increase with increasing CAC, regression for nonuniform differences was used to model the variation of the absolute differences in scan pairs. To calculate the 95% limits of agreement, the predicted absolute differences were multiplied by $1.96\times {(\pi /2)}^{1/2}$, because the absolute differences have a half-normal distribution.