Evidence-Based Investigation of Coronary Calcium Score in Cardiac Computed Tomography

Abstract

This study aimed to verify whether increased body mass index (BMI) increases the noise in computed tomography (CT) images due to heightened effective thickness, impacting calcium scores. Calcium scores were measured in 30 sets of images from normal weight patients. Calcium scores were also measured in 30 sets of images from hypothetical overweight and obese patients, generated by extracting the noise from overweight and obese patients, respectively, and inserting it into the images of normal weight patients. In addition, a phantom study was performed using three calcium phantoms with intensities below the threshold of 130 Hounsfield units and three calcium phantoms with intensities above this threshold. Calcium scores were measured in the absence and presence of a bolus at the heart level to simulate an obese patient. All calcium scores were measured by three radiologists. In the patient study, the total calcium scores of the hypothetical overweight and hypothetical obese groups were 14.93% (p = 0.014) and 22.19% (p = 0.012) higher than those of the normal weight group. In the phantom study, the total calcium score of the six calcium phantoms without a bolus was 1.61% higher at a tube voltage of 120 kV than at 100 kV, and 12.06% higher at a slice thickness of 1 mm than at 3 mm. The total calcium score of the six calcium phantoms with a bolus was 0.13% higher at a tube voltage of 120 kV than at 100 kV, and 14.76% higher at a slice thickness of 1 mm than at 3 mm. These results can be used as a reference to train automated calcium scoring programs on effective thickness through deep learning to reduce calcium score errors caused by increased BMI.

1. Introduction

Computed tomography (CT) is one of the most important imaging modalities in current clinical practice, owing to its ability to provide diagnostic information quickly and non-invasively. It is used in research, as well as in the clinic, and can diagnose tumors, bleeding, infections, and fractures, and assess bone condition [1]. Advances in CT now allow the detection of stem cells labeled with gold nanoparticles [2]. Advances in CT equipment have also made it possible to obtain accurate images of coronary arteries. More recently, the development of multi-detector CT (MDCT) technology and various software packages has made it possible to diagnose coronary artery stenosis. Among the cardiac examinations performed using CT, the calculation of the coronary artery calcification score helps to determine the degree of coronary artery stenosis by quantifying calcification of the coronary arteries [3]. The advent of dual-source CT using two X-ray tubes and detectors allows image reconstruction with only a 90-degree rotation, thereby improving the temporal resolution two-fold with the same X-ray tube rotation speed as that of MDCT [4]. This feature reduces the required examination time and allows excellent image acquisition, even at a heart rate of 70 beats/min or above [5,6].

Coronary images can be analyzed for calcification using coronary calcium quantification software, which determines the location of the calcification detected in the axial image, as well as its volume and mass. The results can then be compared with those of age- and sex-matched individuals. Coronary artery calcification is an indicator of atherosclerosis [7] and serves as an early warning sign of coronary artery disease. Primary care physicians can therefore use calcification scores to evaluate the risk of developing coronary artery disease [8,9,10]. Additionally, calcification can reflect disease progression or regression, and can therefore be quantified to assess treatment effectiveness. Previous research has indicated a robust association between coronary atherosclerotic burden and coronary heart and cardiovascular diseases [11,12].

Any intensity exceeding 130 Hounsfield units (HU) in a CT image is considered calcium [13]. However, intensities resulting from noise and artifacts can also exceed this threshold. Distinguishing artifacts from calcium is straightforward owing to their distinct patterns and shapes. By contrast, noise is random and can be challenging to distinguish from calcium; this determination relies on the subjective judgment of the assessor. Hence, the calcium score cannot be definitively attributed solely to calcium [14].

Quantum noise, which contributes to the noise found in radiographic images, arises from increased radiation attenuation due to effective thickness [15,16,17]. Consequently, radiographic images of overweight and obese patients may exhibit Poisson noise, which may be quantified as calcium if its intensity surpasses 130 HU. Hence, the calcium score may be less accurate in overweight and obese patients than in patients of an average weight. Obesity is strongly correlated with coronary heart disease and cardiovascular risk factors; obese patients are therefore at a higher risk of coronary artery calcification [18,19,20,21].

Accurately discerning noise in images from overweight and obese patients is crucial for obtaining precise calcium scores. This study sought to confirm that an elevated BMI is associated with increased noise levels in CT images. Noise levels obtained from the CT scans of patients who were overweight and obese were compared with those of patients of a normal weight.

2. Materials and Methods

2.1. Patients

This study included patients who underwent coronary artery CT calcium examinations between 1 March 2023 and 30 April 2023. The following patients were excluded from the study: (1) pediatric patients; (2) patients whose calcium measurements were obscured by beam hardening artifacts; (3) patients who had undergone a valve procedure, owing to the presence of additional artifacts; and (4) patients who were unable to raise their hands above their heads, as this increased the effective thickness of the heart. This retrospective study was approved by the Institutional Review Board of Severance Hospital (4-2023-1221), which waived the need for informed consent owing to the retrospective design of the study. Patients were divided into three groups according to body mass index (BMI): normal weight, BMI of 19–23 kg/m²; overweight, BMI of 24–26 kg/m²; and obese, BMI of 27–39 kg/m². In World Health Organization guidelines, a BMI over 25 is considered overweight and a BMI over 30 is considered obese (https://www.who.int/health-topics/obesity#tab=tab_1). The criteria for classification were based on WHO standards. Boundaries were set for each item, and BMI areas were divided as above in this experiment. The normal weight patient group comprised 30 patients (12 males and 18 females), selected to ensure an even distribution among the five levels of the Agatston calcium score risk category, with a mean age of 65.7 (range, 24–87) years. The overweight group comprised three patients (two females and one male), with a mean age of 65.6 (range, 57–75) years. The obese group comprised 15 patients (seven males and eight females), with a mean age of 57.6 (range, 31–83) years.

2.2. CT Image Acquisition

CT was performed using a Somatom Definition Force (Siemens Healthineers, Erlangen, Germany). The scan parameters used for patients are shown in

Table 1. CT scanning parameters for the patient and phantom studies.

Parameter	Patient Study	Phantom Study
kVp	100	120, 100
mAs	80	80
Scan time (s)	0.14	0.14
Rotation time (s)	0.25	0.25
Slice thickness/increment (mm)	3.0/1.5	3.0, 1.0/1.5
Field of view (mm)	300	300
Reconstruction kernel	Sa 36	Sa 36
Scan mode	Sequence	Sequence
Direction	Craniocaudal	Craniocaudal
Window	Mediastinum	Mediastinum

. The scanning range for the calcium score was from the lung apex to the costophrenic angle. Patients were placed in the supine position with feet first on the gantry, in the isocenter of the gantry, and with both arms raised above the head clear of the examination area. The electrocardiogram was attached to the left shoulder, right shoulder, left torso, and right torso. The scanning method was prospective, using the electrocardiogram-gating method, with scans performed at 75% of the R-R interval, which was mid-diastolic for those with heart rates below 70 bpm, and at 45% of the R-R interval, which was end-systolic for those with heart rates above 70 bpm.

The scan parameters used for phantoms are shown in Table 1: tube voltage, 120 or 100 kV; tube current, 80 mA; scan time, 0.14 s; rotation time, 0.25 s; slice thickness/increment, 3.0 or 1.0/1.5 mm; direction, craniocaudal; field of view, 300 mm; reconstruction kernel, Sa 36; and window, mediastinum.

2.3. Image Analysis

Many mathematical models have been developed to represent degradation in radiologic images [22], of which the standard noise model is the most commonly used. However, it is limited in its ability to distinguish between Gaussian and Poisson noise. The Poisson–Gaussian mixture model addresses these limitations. It involves dividing the image into roughly uniform regions, estimating the mean and standard deviation for each region, and then using these local estimates to determine the global implicit functional relationship between image intensity and noise standard deviation through regression. In this study, we adopted this model because it better reflects the noise degradation characteristics of radiographic images. The Poisson–Gaussian mixture model is described by Equation (1):

$$y(x)=p(x)+\eta (p(x))\delta , {\eta }^2(p(x))=\alpha p(x)+{\beta }^2,$$

(1)

where $x$ is the pixel position, and $y$ and $p$ are the degraded and ideal images without the noise component, respectively; $\eta (·)$ denotes the standard deviation of the noise distribution, and $\delta$ is the zero-mean independent random noise and standard deviation equal to one; $\alpha$ is the Poisson signal-dependent component, and $\beta$ is the standard deviation of the Gaussian signal-dependent component. Among the noise level estimation methods used to predict $\alpha$ and $\beta$, the patch-based noise parameter prediction method has proven to be effective and accurate [23,24,25]. A homogeneous patch is crucial for predicting noise levels because it is assumed that any variation within the patch is solely due to noise. First, homogeneous patches were extracted from the image using the non-parametric test Kendall’s tau rank correlation [26,27]. Kendall’s tau was calculated as shown in Equation (2):

$$\tau ={\displaystyle \frac{C-D}{C+D^{\prime }}}$$

(2)

where $C$ is the number of concordant pairs and $D$ is the number of discordant pairs. $\tau$ is in the range of −1 to 1. When each variable has the same relative order in a comparison, it forms a concordant pair, and Kendall’s tau is particularly useful when the sample size is small or the data show high levels of similarity, extracting only the most similar patches to calculate the noise level. In this study, the homogeneous patches were 16 × 16 pixels. The smaller the value of $\tau$, the more homogeneous the patch. The homogeneous patches can be detected by averaging the four-directional $\mathrm{\tau }$-values. The obtained homogeneous patches can be expressed in terms of the mean ($\mu$) and standard deviation ($\sigma$). Second, it can be modeled as shown in Equation (3) to determine the optimal noise levels of $\alpha$ and ${\beta }^2$:

$$f^{\mathrm{*}}=\underset{f\in R^{+}}{\mathrm{argmin}}{‖Uf-V‖}_1, f=\left(\begin{array}{c}\alpha \\ {\beta }^2\end{array}\right), U=\left(\begin{array}{cc}{\mu }_1& 1\\ {\mu }_2& 1\\ ⋮& ⋮\\ {\mu }_n& 1\end{array}\right), V=\left(\begin{array}{c}{\sigma }_1^2\\ {\sigma }_2^2\\ ⋮\\ {\sigma }_n^2\end{array}\right),$$

(3)

where $f$ denotes the matrix of noise level parameters that satisfies the feasible sets $f\in R^{+}$, and $U$ and $V$ are the mean matrix of $n$ × 2 and variation matrix of $n$ × 1, respectively. The minimization problem can be solved using a preconditioned primal-dual algorithm [28]. This allows us to finally calculate $\alpha$, which corresponds to the Poisson noise, and ${\beta }^2$, which represents the Gaussian noise.

Table 2 presents the $\alpha$ and ${\beta }^2$ of all overweight and obese patients according to Equations (1) and (3). The mean and standard deviation values of $\alpha$ and ${\beta }^2$ in overweight patients were 0.19 $\pm$ 0.03 and 8.36 $\pm$ 0.13, respectively. The same values in obese patients were 0.26 $\pm$ 0.08 and 11.41 $\pm$ 1.17, respectively. The imnoise and poissrnd functions in the MATLABTM toolbox (R2022a; MathWorks, Inc., Natick, MA, USA) were then used to generate 30 sets of hypothetical overweight and 30 sets of hypothetical obese CT images for quantitative evaluation by adding noise components to CT images from 30 normal weight patients using the pre-acquired noise level parameters. Example CT images are presented in Figure 1.

Table 2. Noise level descriptive statistics of CT images from overweight and obese patients.

Overweight	$\mathbf{Poisson}-\mathbf{Dependent} \mathbf{Parameter} \left(\mathit{\alpha }\right)$	$\mathbf{Gaussian}-\mathbf{Dependent}\mathbf{ }\mathbf{Parameter} \left({\mathit{\beta }}^2\right)$
Patient 1	0.21	8.49
Patient 2	0.16	8.35
Patient 3	0.19	8.23
Mean $\pm$ SD	0.19 $\pm$ 0.03	8.36 $\pm$ 0.13
Median	0.19	8.35
Minimum, maximum	(0.16, 0.21)	(8.23, 8.49)
95% CI	(0.12, 0.25)	(8.03, 8.68)
Obese	$\mathbf{Poisson}-\mathbf{Dependent} \mathbf{Parameter} \mathbf{\left(}\mathit{\alpha }\right)$	$\mathbf{Gaussian}-\mathbf{Dependent} \mathbf{Parameter} \left({\mathit{\beta }}^2\right)$
Patient 1	0.25	11.36
Patient 2	0.19	10.29
Patient 3	0.27	10.42
Patient 4	0.33	12.27
Patient 5	0.27	10.67
Patient 6	0.23	11.10
Patient 7	0.36	14.04
Patient 8	0.39	12.60
Patient 9	0.33	12.42
Patient 10	0.38	12.57
Patient 11	0.16	11.31
Patient 12	0.15	9.74
Patient 13	0.17	10.22
Patient 14	0.21	11.62
Patient 15	0.17	10.59
Mean $\pm$ SD	0.26 $\pm$ 0.08	11.41 $\pm$ 1.17
Median	0.25	11.31
minimum, maximum	(0.15, 0.39)	(9.74, 14.04)
95% CI	(0.21, 0.30)	(10.77, 12.06)

Table 2. Noise level descriptive statistics of CT images from overweight and obese patients.

Overweight	$\mathbf{Poisson}-\mathbf{Dependent} \mathbf{Parameter} \left(\mathit{\alpha }\right)$	$\mathbf{Gaussian}-\mathbf{Dependent}\mathbf{ }\mathbf{Parameter} \left({\mathit{\beta }}^2\right)$
Patient 1	0.21	8.49
Patient 2	0.16	8.35
Patient 3	0.19	8.23
Mean $\pm$ SD	0.19 $\pm$ 0.03	8.36 $\pm$ 0.13
Median	0.19	8.35
Minimum, maximum	(0.16, 0.21)	(8.23, 8.49)
95% CI	(0.12, 0.25)	(8.03, 8.68)
Obese	$\mathbf{Poisson}-\mathbf{Dependent} \mathbf{Parameter} \mathbf{\left(}\mathit{\alpha }\right)$	$\mathbf{Gaussian}-\mathbf{Dependent} \mathbf{Parameter} \left({\mathit{\beta }}^2\right)$
Patient 1	0.25	11.36
Patient 2	0.19	10.29
Patient 3	0.27	10.42
Patient 4	0.33	12.27
Patient 5	0.27	10.67
Patient 6	0.23	11.10
Patient 7	0.36	14.04
Patient 8	0.39	12.60
Patient 9	0.33	12.42
Patient 10	0.38	12.57
Patient 11	0.16	11.31
Patient 12	0.15	9.74
Patient 13	0.17	10.22
Patient 14	0.21	11.62
Patient 15	0.17	10.59
Mean $\pm$ SD	0.26 $\pm$ 0.08	11.41 $\pm$ 1.17
Median	0.25	11.31
minimum, maximum	(0.15, 0.39)	(9.74, 14.04)
95% CI	(0.21, 0.30)	(10.77, 12.06)

Figure 1. Examples of hypothetical overweight and obese images generated from normal weight patient CT images using the noise level parameters, $\alpha$ and ${\beta }^2$. Noise level parameters were obtained from experimental images of overweight and obese patients.

Figure 1. Examples of hypothetical overweight and obese images generated from normal weight patient CT images using the noise level parameters, $\alpha$ and ${\beta }^2$. Noise level parameters were obtained from experimental images of overweight and obese patients.

2.4. Phantom Study

An adult ATOM Phantom (Model No. 701-G; CIRS, Norfolk, Virginia, USA) was used in this study. A tissue-equivalent bolus was used to mimic the effective thickness of obese patients, the thickness of which was determined using abdominal CT data from 60 patients. The measurements included a slice thickness of 3.0 mm and the widest view of the kidney on axial images. The anterior–posterior and lateral lengths were calculated from the center of the abdomen. Equation (4) shows the calculation of effective thickness, indicating that the average effective thickness for patients with a BMI in the normal range was 237 mm, whereas for those in the obese range it was 271 mm, representing a 1.15-fold increase. In the experiment, because the bolus could only be raised on the stomach, the effective thickness of obese patients was reproduced by assuming that the size of the $\sqrt{LAT}$ is constant and obtaining the size of $\sqrt{AP}$ that can increase the effective thickness of the phantom 1.15-fold. Consequently, the average effective diameter of the phantom used to simulate an obese patient was calculated as 3.0 cm, and images were acquired with the bolus positioned at the heart level of the phantom.

$$Effective Diameter=\sqrt{AP}\times \sqrt{LAT},$$

(4)

Figure 2 shows the calcium phantoms, which were prepared by placing each substance in a polypropylene microtube (diameter, 14 mm; length, 40 mm), with three phantoms less than 130 HU and three phantoms more than 130 HU. The phantom representing an average of 20 HU (C1) was prepared by diluting 0.1 cc contrast agent (Xenetix; Laboratories Guerbet, Villepinte, France) in 5.0 cc normal saline. The phantom representing an average of 50 HU (C2) was prepared by diluting 0.2 cc contrast agent in 5.0 cc normal saline. The phantom with an average of 100 HU (C3) was prepared by diluting 0.3 cc contrast agent in 5.0 cc normal saline. Phantoms more than 130 HU were based on the average calcium value observed in 15 coronary artery CT scans, ranging from 303 to 505 HU. The phantom with an average of 360 HU (C4) was made from seashells, the phantom with an average of 420 HU (C5) was made of chalk, and the phantom with an average of 480 HU (C6) was made of graphite.

2.5. Calcium Scoring

Experienced radiologists (with 14, 5, and 1 years of experience) were blinded to all clinical information. The non-contrast heart image was subjected to calcium scoring using syngo.via software version VB40 (Siemens Healthineers). This software highlights areas with an intensity exceeding 130 HU, allowing identification. Anatomical parts with high HU, such as vertebrae and ribs, are also colored. The shape of areas exceeding 130 HU in the coronary artery was examined to identify calcium. The mass and volume of the examined areas were measured, and the calcium score calculated.

A total of 30 sets of calcium images from 30 normal weight patients were analyzed, along with 30 sets of virtual images from hypothetical overweight patients, created by extracting noise from calcium images of overweight patients and inserting them into the images from the 30 normal weight patients, and 30 sets of virtual images from hypothetical obese patients created by extracting noise from calcium images of obese patients and inserting them into the images from the 30 normal weight patients. Calcium scoring was performed at four locations: left main trunk (LM), left coronary artery (LAD), circumflex artery (CX), and right coronary artery (RCA), on the 90 sets of images by three radiologists (Figure 3). The Agaston score (requiring voxels > 130 HU) was assigned to each patient using syngo.via software. In addition, the average value of the calcium score of the three radiologists was categorized by Agatston calcium score risk: 0 to 1, Normal; 1 to 10, Minimal; 11 to 100, Mild; 101 to 400, Moderate; and 400 or more, Severe [29].

Six calcium phantoms were examined, once without a bolus and once with a bolus, and the acquired images were scored by three radiologists (Figure 4). All patient and phantom images were renamed, stored as numbers, and evaluated randomly so that the radiologists were blinded to the image information.

2.6. Study Designs

Calcium scores were assessed in the 90 image sets, comprising 30 sets of calcium images from normal weight patients, 30 sets of simulated overweight patient images, and 30 sets of simulated obese patient images. The calcium score measured in the images of normal weight patients was used as a reference and compared with the calcium scores of hypothetical overweight and obese patients. In addition, the distribution of calcium scores from normal weight patients across the five levels of Agatston calcium score risk was used as a reference to identify any changes in hypothetical overweight and obese patients.

The calcium scores from CT images of phantoms with a bolus were also compared to those of phantoms without a bolus. The bolus increased the thickness of the skin surface to reduce the radiation dose reaching the deeper layers, resulting in CT images with different noise characteristics and reflecting the situation in patients with an increased BMI.

2.7. Statistical Analysis

SPSS software version 21.0 (IBM Corp., Armonk, NY, USA) was used to test for between-group significance at a 95% confidence interval. The threshold value of 130 HU was employed when performing calcium scoring; above this threshold, the subjective assessment of the radiologist was required to determine the presence of calcium. As scoring was performed by three radiologists with different levels of professional experience, the Intraclass Correlation Coefficient (ICC) was calculated to assess the reliability of the calcium score. Diagnostic performance (sensitivity, specificity, accuracy, positive predictive value [PPV], and negative predictive value [NPV]) was determined at various calcium scores for hypothetical overweight and obese patients. Paired t-tests were performed to analyze the difference in total calcium scores between normal weight, hypothetical overweight, and hypothetical obese patients, averaged by three radiologists. The paired t-tests indicated the change in calcium scoring in the hypothetical overweight or obese patient groups by referencing the calcium scoring of normal weight patients. In addition, the calcium scores of the LM, LAD, CX, and RCA measured by the three radiologists were expressed as Agatston calcium score risk levels, and the kappa value was calculated to test the agreement between the radiologists. Statistical significance was set at p < 0.05.

3. Results

3.1. Patient Study

Very high agreement was observed among the radiologists, with ICCs of 1 [0.999, 1] and 0.998 [0.996, 0.999] for the hypothetical overweight and obese patient images, respectively (p < 0.001).

The total calcium scores of the hypothetical overweight and obese groups were significantly higher than those of the normal weight group, with increases of 14.93% (p = 0.014) and 22.19% (p = 0.012), respectively (Figure 5a). In addition, diagnostic performance was lower in the hypothetical obese group (sensitivity, 100%; specificity, 72.72%; accuracy, 90.00%; PPV, 86.36%; NPV, 100%) than in the hypothetical overweight group (sensitivity, 100%; specificity, 81.81%; accuracy, 93.33%; PPV, 90.47%; NPV, 100%). The kappa value of the total calcium score was calculated to confirm the agreement of the values classified into the five levels of the Agaston calcium score risk category. It was 0.660 [0.597, 0.688] in the hypothetical overweight and obese patient images, showing high agreement among the three radiologists (p < 0.001).

When the average calcium score measured by the three radiologists was categorized into the five levels of the Agatston calcium score risk, five hypothetical overweight patients showed an increased risk level compared with normal weight patients: two from Normal to Minimal, two from Minimal to Mild, and one from Mild to Moderate (Figure 5b). Eight hypothetical obese patients showed an increased risk level compared with normal weight patients: three from Normal to Minimal, three from Minimal to Mild, and two from Mild to Moderate. Three cases increased in both the hypothetical overweight and obese groups, two in the hypothetical overweight group only, and five in the hypothetical obese group only.

3.2. Phantom Study

The total calcium score for the six calcium phantoms was 2% higher with the bolus than without the bolus (Figure 6a). For the three calcium phantoms below 130 HU, the total calcium scores were all 0 without the bolus, however, with the bolus, a calcium score of 1.9 was measured at C3. Of the three calcium phantoms above 130 HU, C4 increased by 0.56%, C5 by 1.01%, and C6 by 10.54% with the bolus compared to without the bolus (Figure 6b).

Decreasing the tube voltage from 120 to 100 kV with a slice thickness of 3 mm decreased the total calcium score for the six calcium phantoms without a bolus by 1.61%. For the three calcium phantoms below 130 HU, all values were zero at 120 kV, but a calcium score of 3.1 was measured for C3 at 100 kV. Among the three calcium phantoms above 130 HU, C4 decreased by 0.54%, C5 by 2.84%, and C6 by 1.86% at 100 kV compared to 120 kV. Decreasing the tube voltage from 120 to 100 kV with a slice thickness of 3 mm decreased the total calcium score for the six calcium phantoms with a bolus by 0.13%. For the three calcium phantoms below 130 HU, a calcium score of 1.9 was measured in C3 at 120 kV, whereas calcium scores of 14.8 and 0.7 were measured in C3 and C2, respectively, at 100 kV. Of the three calcium phantoms above 130 HU, C4 decreased by 1.10%, C5 by 0.96%, and C6 by 2.76% at 100 kV compared to 120 kV (Figure 7).

Decreasing the slice thickness from 3 to 1 mm at a tube voltage of 120 kV increased the total calcium score for the six calcium phantoms without a bolus by 12.06%. For the three calcium phantoms with values below 130 HU, all calcium scores were 0 at 3 mm, however, at 1 mm, a calcium score of 6.35 was measured at C3. For the three calcium phantoms above 130 HU, C4 increased by 17.02%, C5 by 7.15%, and C6 by 23.02% at 1 mm compared with 3 mm. Decreasing the slice thickness from 3 to 1 mm at a tube voltage of 120 kV increased the total calcium score for the six calcium phantoms with a bolus by 14.76%. For the three calcium phantoms below 130 HU, a calcium score of 1.9 was measured in C3 at 3 mm; calcium scores of 20.3, 2.8, and 0.8 were measured in C3, C2, and C1, respectively, at 1 mm. For the three calcium phantoms above 130 HU, C4 increased by 18.72%, C5 by 10.24%, and C6 by 19.74% at 1 mm compared with 3 mm (Figure 8).

Figure 7. (a) Total calcium scores of calcium phantoms with and without a bolus at tube voltages of 100 and 120 kV. (b) Total calcium scores of the calcium phantoms C4, C5, and C6 with and without a bolus at tube voltages of 100 and 120 kV.

Figure 7. (a) Total calcium scores of calcium phantoms with and without a bolus at tube voltages of 100 and 120 kV. (b) Total calcium scores of the calcium phantoms C4, C5, and C6 with and without a bolus at tube voltages of 100 and 120 kV.

Figure 8. (a) Total calcium scores of calcium phantoms with and without a bolus at slice thicknesses of 3 and 1 mm. (b) Total calcium scores of the calcium phantoms C4, C5, and C6 with and without a bolus at slice thicknesses of 3 and 1 mm.

Figure 8. (a) Total calcium scores of calcium phantoms with and without a bolus at slice thicknesses of 3 and 1 mm. (b) Total calcium scores of the calcium phantoms C4, C5, and C6 with and without a bolus at slice thicknesses of 3 and 1 mm.

4. Discussion

With recent advances in deep learning technology, numerous studies have reported improvements in automated calcium scoring [30,31,32]. Deep learning technology uses the traditional threshold of 130 HU, above which a region is considered to contain calcium. However, the potential for noise risks introducing errors into automated calcium scoring, necessitating the accurate identification of the causes of noise in CT images.

The results of the present study suggest that an increase in the BMI of a patient leads to greater effective thickness, subsequently elevating noise levels in CT images and influencing calcium scores. Several previous studies have also reported an association between BMI and noise [33,34,35]. Based on these previous studies, we excluded other variables affecting noise by extracting the noise from images of overweight and obese patients and inserting it into the images of normal weight patients to generate hypothetical overweight and obese patient images. This allowed us to compare the calcium scores of hypothetical overweight and obese patients to those of normal weight patients, with the increased noise caused only by increased BMI, and other factors kept constant. It was thus confirmed that noise increased along with BMI, which resulted in an increase in calcium scores. The findings of this study indicate that the impact of noise can be minimized by utilizing a training model based on BMI when implementing automated measurements. Furthermore, our phantom study allowed us to manipulate effective thickness, tube voltage, and slice thickness, which all impact noise, to observe the corresponding changes in calcium scores. An increase in tube voltage permits the transmission of a greater number of X-rays to the detector when examining a range of effective thicknesses with a constant mAs. This may also enhance the image quality. Furthermore, a reduction in slice thickness increases image noise. Currently, there are fewer advantages to using thin slices because the increase in noise can impact the accuracy of manual diagnoses, however, it is advantageous to use thin slices that provide more anatomical information when using automated measurement. We therefore investigated the effect of CT settings on noise levels for calcium scoring.

The calcium scores for hypothetical overweight and obese patients were increased compared to normal weight patients, with a more notable rise observed in hypothetical obese patients. Additionally, diagnostic performance was poorer in hypothetical obese patients than in hypothetical overweight patients, with the reduced specificity in hypothetical obese patients suggesting that greater effective thickness results in images with more noise, affecting calcium scores. ICC and kappa values indicated very high agreement among the three radiologists, and the increase in calcium scores was consistent across all radiologists. Previous studies have measured the calcium score when the patient was irradiated according to the guidelines and without considering BMI, and compared it to the calcium score of the low-dose image [36,37]. However, the results of this study showed that an increase in effective thickness caused by an increase in BMI can increase the uncertainty of the calcium score; the higher the BMI, the greater the uncertainty. Therefore, efforts are needed to reduce this error in the calcium score.

The Agatston calcium score risk level was increased in the hypothetical obese group compared with the hypothetical overweight group. The occurrence of changes from Mild to Moderate in both groups implied that noise in overweight and obese patients may introduce inaccuracies into the calcium score, although there were no changes at higher levels. This inaccuracy can affect the interpretation, possibly leading to unnecessary additional examinations.

The results of the phantom study showed a tendency for the calcium scores to be higher with a bolus than without, which is consistent with the results of the patient study. The calcium scores of the phantoms below 130 HU erroneously indicated the presence of calcium, as in the patient study. The calcium scores of phantoms above 130 HU showed a tendency to increase in the presence of a bolus. In addition, calcium scores tended to increase as the applied tube voltage increased. However, the rate of increase decreased as the effective thickness increased. This suggests that using a tube voltage of 120 kV may result in a smaller change in the calcium score due to effective thickness. In addition, thinner slices were associated with higher calcium scores, regardless of effective thickness. Slice thickness had a greater impact than tube voltage on calcium scores. When obtaining automated calcium score measurements, more detailed anatomical information can be collected from thinner slices [38], while simultaneously minimizing the impact of noise through denoising techniques. Furthermore, as the effective thickness increases, higher tube voltages can be employed to reduce the uncertainty associated with calcium scoring. With ongoing technological advancements, it is to be expected that the precision of calcium scoring will be enhanced through the aforementioned methods, particularly in the context of automated calcium scoring, which employs deep learning technology.

This study has some limitations. First, the noise used to generate hypothetical overweight and obese patient images in the patient study may have been caused by factors other than effective thickness. This is because we could not arbitrarily increase the effective thickness of a single patient, and it is against radiation ethics to irradiate a single patient multiple times. Second, the study does not include demographic and body composition data, such as age, sex, and body fat distribution. The current simple calcium scoring method is limited by its inability to fully take into account pathophysiologic information. Finally, only one calcium measurement program was used. Different programs vary in the measurement of calcium, which may lead to differences in calcium scores. Evaluation of these different programs would increase the generalizability of the results of this study.

5. Conclusions

This study found that an increase in BMI reflects an increase in effective thickness, and the associated increased noise affects calcium scores. We expect that these results can be used as a reference for the development of automated calcium scoring using deep learning technology, to reduce the error in calcium scores due to effective thickness. Rather than simply performing calcium scoring based on a threshold of 130 HU, the proposed method can minimize calcium score variations by classifying CT images according to BMI when training the program. Calcium scoring performed by a program trained on BMI should show a reduced impact from noise.