Introduction
Coronary artery calcium (CAC) is important for cardiovascular risk determination in asymptomatic individuals [
1]. CAC is visualized with cardiac computed tomography (CT) and quantified using the Agatston score [
2]. Furthermore, an Agatston score of zero is proven to be a strong negative predictor of future cardiovascular events [
3]. This, in turn, indicates the importance of accurate detection and subsequent quantification of small calcified lesions.
One important factor influencing CAC quantification is the type of image reconstruction used in CT [
4]. Over the last decade advanced reconstruction techniques such as hybrid iterative reconstruction (HIR) and model-based iterative reconstruction (MBIR) became available for CT [
5]. These reconstruction algorithms reduce image noise, and therefore allow for a decrease in radiation dose while maintaining image quality equal to traditional filtered back projection (FBP) [
6,
7]. Previous studies have shown a good agreement in Agatston scores between FBP and HIR and MBIR [
8‐
10]. However, it was also shown that HIR resulted in decreased Agatston scores for small and/or low density lesions [
9]. Similarly, MBIR resulted in decreased detection of small calcifications [
8].
Recently, one of the main CT manufacturers introduced a new deep learning-based reconstruction (DLR) technique. DLR improves image quality by applying a deep learning network trained on pairs of high-dose, advanced MBIR and HIR images [
11] and prevents image quality degradation and ‘plastic-like’ appearance of the image [
12]. As previously shown with low dose acquisitions, DLR outperforms MBIR in terms of noise reduction which may potentially allow for further radiation dose reduction beyond current levels [
11,
13]. However, the influence of this novel image reconstruction technique on CAC detection and quantification is unknown.
As previously noted, the detection of CAC, resulting subsequently in zero or non-zero Agatston scores, is of utmost importance for correct risk stratification. Because small or low-density CAC can resemble image noise and HIR, MBIR, and DLR all decrease image noise, these CT reconstruction techniques may impact the detection of very small or low-density CAC. This is even more important for acquisitions at a reduced radiation dose [
14]. As previously shown, risk classification was underestimated up to 50% for CAC scores from IR images acquired at reduced radiation dose [
4]. Consequently, the Society of Cardiovascular Computed Tomography recommends further evaluation of reconstruction techniques before clinical implementation [
15]. Therefore, we designed a phantom study in which we aimed to investigate the influence of four reconstruction methods (FBP, HIR, MBIR, and DLR) on static and dynamic CAC detectability and quantification for standard and reduced radiation dose. Subsequently, we verified the effect of all four image reconstruction techniques on CAC quantification and risk classification in a patient study.
Discussion
The main finding of the phantom part in the present study is that detection of small calcifications at routine (100%) radiation dose is reduced up to 22% depending on the used reconstruction algorithm. Furthermore, this trend was even more pronounced on reduced radiation dose scans. For CAC quantification, our dynamic phantom study showed no clinically relevant differences in Agatston score based on reconstruction algorithm for the routine radiation dose protocol. The patient study showed excellent agreement between FBP and HIR, MBIR, and DLR, with only a small number of risk reclassifications, although MBIR resulted in significantly higher Agatston scores.
To the best knowledge of the authors, this study is the first to systematically assess the influence of all reconstruction techniques currently available for one vendor on CAC detection and quantification. Compared to FBP all reconstruction methods reduced CAC detection, except in the case of the small chest phantom at full dose level. Both IR techniques as well as DLR reduce image noise [
11]. The, in general, reduced CAC detectability in comparison with FBP for these reconstruction techniques might therefore be explained by erroneous identification of CAC containing voxels as noise. Furthermore, as presented in this study, decreased detectability may be due to reduced HU peaks in small calcifications. This behavior will, of course, be more pronounced at reduced tube current and increased patient size due to increased noise levels, as also shown in this study. As a result, HIR, but especially MBIR and DLR may miss small calcifications and improperly classify patients into the zero Agatston score risk group. However, based on our patient study, none of the patients was incorrectly assigned to the zero Agatston score group.
Independent of the reconstruction method, for medium and large density calcifications, the Agatston score increased with velocity, while for small density calcification, Agatston score decreased. This finding is in line with previous results of van der Werf et al. [
19,
25] and Groen et al. [
26] and might be explained by motion blurring. Due to motion blurring, the number of voxels above 130 HU increases in medium and large density calcifications, which increases the Agatston score. In low density calcifications, in turn, the number of voxels above 130 HU decreases, which decreases the Agatston score.
As we know from the CONFIRM registry, small calcifications visually detected on CCTA scans in patients previously assigned to the zero Agatston score risk group, increased risk of major adverse cardiac events [
27]. Therefore, detectability of small calcifications plays a crucial role in further patient management. Importantly, when reduced tube currents were used, detectability of small calcifications decreased, especially for MBIR and DLR. Our hypothesis is that this can be explained by the need for increased noise suppression by these reconstruction algorithms. Therefore, based on these detectability insert results we assume that patients might be misclassified into the zero Agatston score risk group when a reduced radiation dose protocol is used. Future patient studies with more small calcifications should verify this.
Additionally, at routine tube current level, the current study did not show relevant differences between reconstruction methods in terms of Agatston scores. However, when the tube current was decreased to 50%, Agatston score of low density calcifications acquired from the large dynamic phantom deviated from the standard measurement [
2]. Therefore, as also underlined in SCCT guidelines [
15], caution should be taken in terms of radiation dose reduction by decreasing tube current, especially in combination with iterative reconstruction methods. Nevertheless, the Agatston score of medium and high density calcification did not differ from baseline, when radiation dose was reduced by 50%. Similar findings were presented by Choi et al. who applied 75% dose reduction with comparable image quality [
8].
The patient study showed that only the Agatston score measured from MBIR differed significantly from the reference Agatston score based on FBP. When considering patients with a zero Agatston score as defined by the reference method, MBIR classified one patient as a nonzero Agatston score, thereby increasing the risk classification. However, similar results were presented before, with 17% of cases reclassified into higher risk group, including 8% of patient misclassified as non-zero Agatston scores [
8]. One explanation for this behaviour might be the impact of the edge enhancement algorithm, whereby more pronounced CAC edges increase overall Agatston scores. Also, the Bland–Altman limits of agreement of MBIR compared with FBP were almost twice as large as the limits of HIR or DLR compared with FBP. Nevertheless, overall statistical agreement in risk classification was excellent for all reconstruction methods. Similar findings were presented by Szilveszter et al. and Tang et al., who showed that despite lower Agatston score based on HIR or MBIR, the effect on cardiovascular risk stratification was modest [
10,
28]. Nevertheless, clinicians should bear in mind that a change in cardiovascular risk classification influences further patient management, including initiation of lipid-lowering therapy [
29]. Therefore, the small discrepancy between FBP, MBIR, HIR, and DLR, may bring long-term consequences for patients.
Importantly, for our patient group, none of the patients was reclassified as a false negative. Currently, both American and European guidelines use CAC scoring as an additional tool not only for patient risk classification, but also for guiding statin and aspirin therapy [
30]. Therefore, the lack of CAC measurement reproducibility and its dependency on different reconstruction methods, may affect patient management and outcome [
23]. Based on patients results from our study and using FBP as reference, the most accurate calcium scoring was achieved when HIR or DLR was used, in terms of correct patient risk classification.
This study has several limitations. First, while our systematic analysis included both a static and dynamic phantom as well as a patient study, we only included a small number of patients (n = 50). Moreover, only twelve patients (24%) presented with Agatston score between 0 and 10, which is the most susceptible group in terms of calcium detectability. However, the results give a good indication of the differences between the reconstruction techniques and validate our phantom results. A larger patient study is needed to verify these results in all patient risk categories. Second, we acquired data from one vendor. Therefore, a multivendor study analyzing the influence of different reconstruction methods on calcium detectability, quantification, and risk stratification is certainly needed. Third, all patients were scanned with the standard protocol. Therefore, the effect of decreased radiation dose could not be evaluated in patients. Fourth, the D100 insert is a static insert. Thus, we were not able to acquire dynamic detectability phantom data. However, due to the decrease in detectability, even in a static situation, care should be taken when using non-FBP reconstructions for detecting CAC with this CT system.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.