The accuracy of noise measurements in patient CT images

Abstract

Objective.

Noise magnitude is one of image quality indicators in computed tomography (CT) for assessing imaging performance, and the CMS (Centers for Medical and Medicaid Service) recently included a measure of noise magnitude in terms of “global noise”. Despite its importance, no standard method currently exists to measure noise magnitude in patient images. Theoretically, the most accurate approach is to assess voxel value variation across repeated images, the so-called ensemble noise. Obviously, such a method is not ethically feasible in actual patients. To surmount this impasse, we deployed virtual imaging techniques to benchmark three noise magnitude calculation methods against two gold standard ensemble noise measures across 36 imaging conditions.

Methods.

Over 1800 virtual image datasets were generated from imaging an American College of Radiology (ACR) phantom and Extended Cardiac-Torso (XCAT) human models using a validated, scanner-specific CT simulator (DukeSim). The ACR phantom was imaged under 36 different imaging conditions defined by combinations of chest and abdominopelvic protocols, three dose levels, three reconstruction kernels, and both Filtered Back Projection and Iterative Reconstruction algorithms. At the same conditions, XCAT models were repeatedly imaged 50 times. Noise magnitudes in the ACR phantom were calculated in 5 circular ROIs. In patients, noise was measured in air surrounding the body and in soft tissues by applying HU<−900 and −300≤HU≤100 thresholds, respectively. Per each imaging condition, measured noise magnitudes were compared against ensemble noise in soft tissue, liver, and lungs.

Results.

Across all imaging conditions, noise measurements in the ACR phantom and in air surrounding the patient underestimated ensemble noise by approximately 60% and 50%, respectively. In contrast, soft tissue-based noise measurements were closer to the gold standard with median differences between −8% and +4%.

Conclusions.

This study introduced a virtual imaging-based framework to benchmark clinical CT noise metrics against ensemble noise measures. Virtual imaging enabled objective comparison of different noise magnitude calculation methods in large, realistic populations simulating clinical conditions. Noise measured in air cannot represent soft tissues noise. The results validated soft tissue-based noise measurements as a reliable surrogate to inform protocol design, technology assessment, and equitable healthcare reimbursement.

3. Introduction

In diagnostic X-ray imaging, the only reason why patients are exposed to minimal stochastic cancerogenic risks induced by ionizing radiation is to generate images from which valuable diagnostic information may be obtained (1). The quality of such information is qualitatively assessed by the diagnosticians when reading the images. Although a qualitative image evaluation can be timely at the point of care, it does not provide an objective evidence of the actual diagnostic suitability of the image. In computed tomography (CT), such an ephemeral evaluation of the image quality conflicts with the measurement of the radiation risks which can rely on a myriad of well-established metrics and methods, all with different degrees of appropriateness (2, 3). As a result, although the assessment of CT procedures and technologies, the justification of the exam, and the design of optimization actions should rely on an appropriate balance between radiation dose and image quality, they are largely dominated by the urge to reduce radiation burden, often overlooking the final procedure outcome evaluation in terms of the quality of the images (4).

To close this gap, and inspired by international guidelines from the International Commission on Radiological Protection (ICRP) and the International Atomic Energy Agency (IAEA) (5, 6), several research groups developed different methods to quantitatively measure image quality in CT images both in phantoms (in phantasma) and in patients (in vivo) (7–10). These image quality measurements ensure reproducibility and robustness and have been deployed to evaluate attributes such as noise magnitude, spatial resolution, and signal in terms of Hounsfield Unit (HU). In this milieu, noise magnitude measurements recently hit the headlines after a ruling of the Centers for Medical and Medicaid Service (CMS) proposing an ambiguous “global noise” figure as a metric of noise magnitude for diagnostic computed tomography exams in adults (11). The mandate implied that such a metric can be used as a reliable gauge of “inadequate image quality” as a basis of healthcare reimbursement.

Despite the implication of the CMS mandate and the ambiguity of its definition of noise, currently there exists no fully reliable, singular method to measure noise magnitude in CT images. Indeed, there has been a myriad of noise magnitude measures from those in single size uniform phantoms, in multi-sized phantoms (12), in patient images considering the air surrounding the patients (8), in patient images considering only soft tissue (7), and in individual organs of patients (13). Also, noise is not uniform across a CT dataset – aggregating a value across that diversity is a task of its own. This heterogeneity can lead to drastically different results, even when analyzing same datasets, biasing the evaluation of radiological procedures, protocol design, and healthcare reimbursement.

Therefore, it is necessary to objectively compare different approaches to estimate noise magnitude against a gold-standard method. The most accurate approach to measure noise magnitude is to ascertain the so-called ensemble noise (N_ensemble) by scanning a patient multiple times, and sampling noise of each pixel across the ensemble of images. This approach requires repeated imaging, an ethically unfeasible prospect for patient imaging (8). Moreover, current quality assurance guidelines for CT table positioning reproducibility allow a tolerance of ±1 mm which may exceed the image pixel size, making it impossible to accurately sample the same pixel across repeated acquisitions (24). As a result, it is unknown which method provides the closest representation of noise magnitude to the ensemble noise gold standard.

To overcome this impasse, in this study, we estimated ensemble noise deploying Virtual Imaging (in silico) techniques and used it as a gold standard to compare three noise magnitude calculation methods, including measuring noise from phantom data, and from patient data in the air outside patients and in the soft tissue. The analysis determined to what extent different noise measurements can accurately represent the true noise magnitude in patient images.

4. Materials and methods

4.1. Image acquisition

A total of 1800 virtual patient and 18 phantom virtual CT image datasets, one for each imaging condition described below, were included in this study. The patient data was used to assess in silico noise measurement techniques from patient images whereas the phantom data was used to evaluate in phantasma noise measurements.

The patient cases included Chest and Abdominopelvic human models representing a typical adult with BMI at the 50^th percentile of the US population (XCAT; male; 67 years old; BMI: 28.2 kg/m², water equivalent diameter: 29.5 cm) generated by segmenting patient images using nonuniform rational B-spline as described by Segars et al. (14, 15); the phantom data were based on a computational ACR phantom (16). The virtual scans were done using a validated CT simulator (DukeSim) (14, 15, 17, 18). DukeSim encompasses a GPU-accelerated approach that combines ray-tracing and Monte Carlo methods to generate a final sinogram incorporating both primary and scatter signals. It accurately replicates various scanner-specific characteristics, including geometry, bowtie-filtered source spectrum, automatic tube current modulation, flying focal spot wobbling, anti-scatter grid, electronic noise, and poly-energetic detector response. Additionally, DukeSim has been integrated with manufacturer-specific reconstruction software and has been validated across various scanner models and imaging conditions (17, 19–22).

The virtual ACR phantom was imaged with DukeSim simulating a clinical CT scanner (Somatom Definition Flash, Siemens Healthineers, Erlangen, Germany) at 120 kV, pitch of 1, and three different dose levels (50, 100 and 150 mAs, corresponding to CTDI_vol of 2.9, 5.7, and 8.6 mGy, respectively). The obtained datasets were reconstructed with a 5.0 mm slice thickness and a 500 mm field of view, using an offline reconstruction prototype (ReconCT 15.0.35098.0, Siemens Healthineers), with three different reconstruction kernels (Br32f, Br46f, and Br62f), with both weighted filtered back projection (FBP) and iterative reconstruction (IR) algorithms (ADMIRE, level 3 across all imaging conditions), to generate a total of 18 datasets. Previous studies showed that DukeSim can simulate CT images with image quality attributes similar to real CT images. In particular, noise magnitude was compared in ACR and Mercury Phantom (Duke University, Durham, NC) both in energy-integrating detector and photon-counting detector CT for different doses and reconstruction kernels (17, 23).

4.2. Noise measurement methods

Following the AAPM TG 233 methodology, a medical physicist with 4 years of experience measured the noise for each dataset from five ROIs, with size covering approximating 1% of the phantom area, placed at the center, and at 12, 3, 6, and 9 o’clock of each slice over the uniform section of the imaged phantom, as shown in Figure 1 (24). Noise magnitude (N_ACR) was obtained by averaging the standard deviation of pixel values across each ROI location and for 19 slices. N_ACR standard deviation and coefficient of variation were analogously calculated across the 19 slices.

Figure 1:

ROI placements (in red) for noise magnitude measurement on the virtual ACR phantom.

The two virtual patients were also imaged using DukeSim with the same conditions used to acquire the ACR phantom images. Each scan included 39 image slices and was repeated 50 times to enable stable ensemble noise estimations. The sinogram images were then reconstructed with a 5.0 mm slice thickness, a matrix size of 1024 × 1024 pixels, and a 500 mm field of view using ReconCT. Each sinogram data was reconstructed six times, utilizing three different kernels (Br32f, Br46f, and Br62f) and both FBP and IR algorithms, generating a total of 1800 studies (Figure 2.a). Although FBP and IR exhibit different relationships between the input projection data and the reconstructed images, both were included to reflect clinically relevant reconstruction conditions.

Figure 2.

Abdominopelvic (top) and Chest (bottom) image reconstructed with FBP (a); with GNI threshold in red (b); and with N_AIR threshold in red (c).

Using a previously published automated methods (7), for each imaging condition and in each repeated scan, noise magnitudes were calculated in soft tissues as global noise index (GNI), and in the air surrounding the patient (N_AIR) by applying thresholds of −300≤HU≤100 and HU<−900, respectively (Figure 2.b–c). In particular, for each slice, an ROI (30 pixels × 30 pixels) is identified around each pixel, generating a standard deviation and a histogram of standard deviation map across the slice. Noise of that slice is computed as the value corresponding to the peak of the histogram. Finally, the noise values for the study were computed as the mean of the values calculated in each slice. Standard deviation and coefficient of variation across the 50 repetitions were also calculated for each noise magnitude metric and imaging condition.

4.3. Enslemble noise measurement

The ensemble noise was calculated in the same GNI soft tissue areas (N_ensemble), and in the liver and in the lungs (N_organ) for abdominopelvic and chest studies, respectively (Figure 3). In particular, for each slice and for every pixel within the segmented area, the HU of that pixel was measured across the 50 repeated images. The standard deviation of these 50 measurements defined the ensemble noise of that pixel. The slice ensemble noise was obtained by averaging the pixel-level ensemble noise across all the pixels in the slice. Lastly, the final ensemble noise (N_ensemble and N_organ) was then calculated by averaging the slice ensemble noise across all the slices. Because the HU variation for a given pixel across the 50 repetitions cannot be attributed to anatomical or morphological changes in the patient or phantom, this variation reflects only photon statistics and electronic noise within the simulation framework. Therefore, N_ensemble and N_organ can be regarded as gold standard for quantifying noise magnitude. Standard deviation and coefficient of variation values were also calculated across the 50 repetitions for each noise magnitude metric and imaging condition.

Figure 3.

Human model (XCAT) liver (left) and lungs (right) organ masks applied to calculate N_organ.

The analysis was performed separately for each set of dose level, kernel, and reconstruction algorithm. Median noise magnitude values from different methods were compared in terms of percentage difference. N_ACR, N_AIR and GNI were compared to the ensemble noise calculated in soft tissue (N_ensemble) and to the ensemble noise calculated in the liver and in the lungs (N_organ) for abdominopelvic and chest studies. Lastly, for each metric, the relative deviation from N_ensemble and N_organ was calculated across variations in mAs, reconstruction algorithm, and kernel, and the maximum absolute percentage error with respect to N_ensemble and N_organ was reported across all parameter combinations.

5. Results

Table 1 and Table 2 summarize the median N_ACR, N_AIR, GNI, N_ensemble, and N_organ for chest and abdominopelvic images per each kernel, mAs value, and reconstruction algorithm involved in the study. As expected, the noise values decreased with the increase in radiation dose and increased with the kernel sharpness. Moreover, because the ACR phantom is smaller than the virtual patients, the noise magnitude in the phantom is consistently lower than the values measured in patients for constant tube current.

Table 1.

Median N_ACR, N_AIR, GNI, N_ensemble, and N_organ for chest images per each kernel, mAs value, and reconstruction algorithm involved in the study. Units are HU.

Chest
				FBP					IR
		NACR	NAIR	GNI	Nensemble	Norgan	NACR	NAIR	GNI	Nensemble	Norgan
	50 mAs	4.9	4.1	7.4	7.7	7.0	3.2	3.2	6.0	6.7	6.7
	100	3.5
Br32f	mAs		3.0	5.4	5.4	5.0	2.3	2.4	4.4	4.7	4.7
	150	2.9
	mAs		2.5	4.5	4.4	4.0	1.9	2.1	3.7	3.9	3.9
	50 mAs	13.6	14.3	21.5	21.9	19.6	8.4	9.5	14.4	16.4	15.6
	100	9.6
Br46f	mAs		10.0	15.2	15.4	13.8	5.9	6.7	10.2	11.5	11.1
	150	7.8
	mAs		8.1	12.5	12.6	11.6	4.7	5.5	8.4	9.4	9.2
	50 mAs	45.4	44.0	102.8	98.0	86.4	24.6	30.2	57.7	57.5	52.6
Br62f	100	31.8
	mAs		35.9	74.4	69.1	61.3	17.1	25.4	41.0	40.7	37.2
	150	25.8
	mAs		32.0	61.6	56.4	50.1	13.7	22.8	34.0	33.2	30.5

Table 2.

Median N_ACR, N_AIR, GNI, N_ensemble, and N_organ for abdominopelvic images per each kernel, mAs value, and reconstruction algorithm involved in the study. Units are HU.

Abdomen
		FBP					IR
		NACR	Nair	GNI	Nensemble	Norgan	NACR	NAIR	GNI	Nensemble	Norgan
	50 mAs	4.9	4.8	10.8	10.8	12.1	3.2	3.8	8.9	9.5	10.3
	100
Br32f	mAs	3.5	3.4	7.8	7.6	8.5	2.3	2.8	6.3	6.7	7.2
	150
	mAs	2.9	2.9	6.4	6.2	7.0	1.9	2.3	5.2	5.4	5.8
	50 mAs	13.6	17.1	32.2	30.1	33.9	8.4	11.5	22.0	22.5	24.2
	100
Br46f	mAs	9.6	12.0	22.6	21.3	23.8	5.9	7.9	15.4	15.9	16.9
	150
	mAs	7.8	9.6	18.5	17.4	19.4	4.7	6.5	12.5	13.0	13.7
	50 mAs	45.4	51.5	143.5	131.4	149.3	24.6	33.8	84.5	77.7	86.6
	100
Br62f	mAs	31.8	40.2	103.6	92.5	105.0	17.1	27.4	60.0	54.9	60.8
	150
	mAs	25.8	35.8	86.1	75.4	85.5	13.7	24.8	48.9	44.8	49.4

N_ACR and N_AIR both underestimated the gold standards N_ensemble and N_organ across all scanning conditions for both chest and abdominopelvic exams (Figure 4 and 5). The difference between N_ACR and N_ensemble was −54% and −61% for FBP and IR studies, respectively; and the difference between N_ACR and N_organ was −54% and −60% for FBP and IR. The difference between N_AIR and N_ensemble was −46% and −49% for FBP and IR studies; the difference between N_AIR and N_organ was −49% and −51% for FBP and IR, respectively.

Figure 4.

Boxplots of the relative percentage differences between N_ACR, N_AIR, GNI and N_ensemble, and N_organ for FBP and IR studies considering all kernels and mAs values. The red line represents the median value, and the blue boxes represent the 1st and 3rd quartiles.

Figure 5.

Heatmap of the relative percentage differences across mA values between N_ACR, N_AIR, GNI and N_ensemble, and N_organ for each protocol, kernel, and reconstruction algorithm.

In contrast, Global noise index (GNI) was closer to gold standard N_ensemble: +4% in FBP and −3% in IR images. Moreover, GNI values showed a difference of +3% and −8% when compared with N_organ measured in FBP and IR images, respectively. Tables 3 and 4 report standard deviation and coefficient of variation for N_ACR, N_AIR, GNI, N_ensemble, and N_organ for chest and abdominopelvic images per each kernel, mAs value, and reconstruction algorithm. Differences in standard deviation and coefficient of variation magnitude reflect the different methods used to calculate the associated noise values.

Table 3a.

Standard deviation values for N_ACR, N_AIR, GNI, N_ensemble, N_organ for chest images per each kernel, mAs value, and reconstruction algorithm involved in the study. Units are HU.

Chest
		FBP					IR
		NACR	NAIR	GNI	Nensemble	Norgan	NACR	NAIR	GNI	Nensemble	Norgan
	50 mAs	0.141	0.009	0.026	1.27	2.09	0.092	0.006	0.016	1.14	1.73
	100	0.095
Br32f	mAs		0.006	0.015	0.896	1.45	0.063	0.005	0.026	0.801	1.20
	150	0.060
	mAs		0.004	0.007	0.732	1.18	0.041	0.005	0.015	0.655	0.970
	50 mAs	0.324	0.043	0.058	3.51	5.46	0.178	0.033	0.042	2.76	4.23
Br46f	100	0.208
	mAs		0.024	0.042	2.49	3.80	0.140	0.012	0.033	1.95	2.90
	150	0.157
	mAs		0.027	0.037	2.06	3.15	0.083	0.010	0.023	1.60	2.34
	50 mAs	1.336	0.086	0.240	14.3	24.3	0.708	0.062	0.158	8.80	14.9
	100	0.812
Br62f	mAs		0.076	0.243	10.0	16.9	0.437	0.058	0.110	6.20	10.4
	150	0.615
	mAs		0.069	0.185	8.19	13.7	0.275	0.033	0.102	5.08	8.49

Table 4a.

Coefficient of variation values for N_ACR, N_AIR, GNI, N_ensemble, N_organ for chest images per each kernel, mAs value, and reconstruction algorithm involved in the study.

Chest
		FBP					IR
		NACR	NAIR	GNI	Nensemble	Norgan	NACR	NAIR	GNI	Nensemble	Norgan
	50 mAs	0.029	0.002	0.004	0.156	0.267	0.029	0.002	0.003	0.160	0.236
Br32f	100 mAs	0.027	0.002	0.003	0.156	0.263	0.028	0.002	0.006	0.160	0.231
	150 mAs	0.021	0.002	0.002	0.156	0.261	0.022	0.003	0.004	0.160	0.229
Br46f	50 mAs	0.024	0.003	0.003	0.152	0.251	0.021	0.003	0.003	0.159	0.246
	100 mAs	0.022	0.002	0.003	0.152	0.248	0.024	0.002	0.003	0.160	0.237
	150 mAs	0.020	0.003	0.003	0.154	0.249	0.018	0.002	0.003	0.161	0.233
	50 mAs	0.029	0.002	0.002	0.140	0.254	0.029	0.002	0.003	0.146	0.258
Br62f	100 mAs	0.026	0.002	0.003	0.139	0.250	0.026	0.002	0.003	0.145	0.255
	150 mAs	0.024	0.002	0.003	0.139	0.249	0.020	0.001	0.003	0.145	0.254

The maximum error compared to N_ensemble across differentials due to mAs, reconstruction type, and kernel was 30% for N_ACR, and N_AIR, 16% for GNI, and 5% for N_organ. The maximum error compared to N_organ across differentials due to mAs, reconstruction type, and kernel was 30% for NACR and NAIR, 13% for Nensemble, and 5% for Norgan.

6. Discussion

Through the application of virtual imaging techniques, we performed an objective and unbiased comparison of different noise magnitude estimation methods against the gold standard ensemble noise. The noise measured in the ACR phantom and in the air surrounding the patient largely underestimated the ensemble noise measured in soft tissues and in the organ of interest. Such methods can be instrumental in the comparison of different scanning techniques or in the evaluation of scanner performance, as there is a linear correlation between N_AIR and GNI, but they were not introduced to represent noise magnitude in patient tissues, as confirmed in this study. Conversely, the global noise index (GNI) in soft tissue represented the closest surrogate to ensemble noise, affirming the validity of soft tissue-based noise measurements to inform protocol design and technology assessments (7).

Regarding the comparison between ACR phantom and patient-based measurements, it is important to consider differences in object attenuation. The ACR phantom has a diameter of 20 cm, whereas the water-equivalent diameter of the virtual patient model is approximately 29.5 cm. Since all simulations were performed under a fixed tube current, absolute noise values are not directly comparable between these two configurations without accounting for differences in attenuation. To this end, Jadick et al. showed strong agreement between real and images simulated with DukeSim in terms of noise magnitude (20). Moreover, they calculated noise magnitude in the 30 cm section of the Mercury Phantom, corresponding to a 29 cm water equivalent diameter, using DukeSim to simulate same scanner and dose levels considered in the present study (20). Their simulation employed a slice thickness of 0.6 mm and the Br40d reconstruction kernel. Under the simplifying assumption that noise magnitude scales inversely with the square root of slice thickness, and without accounting for potential effects of reconstruction kernel or inter-slice noise correlation, the noise values reported by Jadick et al. can be adjusted to the 5 mm slice thickness used in the present study. Specifically, the corresponding noise levels were equivalent to circa 18 HU, 13.8 HU, and 9.7 HU for the three dose levels. These values fall between the noise levels we measured in our virtual patient using the Br32f and the Br46f kernels, supporting the consistency of the simulation results.

Among other strategies to measure image quality features, namely contrast and resolution (9, 10), methods to measure noise are more technologically and clinically mature. Indeed, a few applications to assess noise magnitude have been already implemented in clinical practice (7, 25–27). For this reason, when it came to choosing an image quality measurand to assess the performance of diagnostic computed tomography exams in adults, the Centers for Medical and Medicaid Service proposed the noise magnitude (11). This CMS measure applies to payments for hospitals and clinicians: the associated data collection and reporting begun in January 2025, impacting the reimbursements starting in January 2027. Because of its critical role in the implementation of this new provision, it is essential that noise measurements are consistent, robust, reproducible, and standardized across institutions.

To that end, a commissioned panel from the American Association of Physicists in Medicine (AAPM) recently published a list of issues and ambiguities related to the CMS measure (28). In particular, the AAPM panel highlighted that: “It is unclear how the measurement of calculated CT global noise is performed given the lack of an official standard and limited documentation in the specification”. Specifically, the experts emphasized how peer-reviewed literature currently recognized two main methods to measure noise magnitude in patient CT images: the noise measured in soft tissues and the noise measured in the air surrounding the patient. Our study showed that the noise in air (N_AIR) can underestimate the gold standard N_ensemble by as much as 62%, whereas the GNI absolute average difference with N_ensemble is 5%. Therefore, the new CMS rule cannot be implemented without referencing a single method to calculate noise magnitude or without resolving how values obtained with different methods can be unbiasedly compared.

This issue underscores the growing importance of image quality in assessing the diagnostic value of CT examinations. Historically, CT performance metrics focused almost exclusively on radiation dose, as before the implementation of automated current and kV modulation techniques, dose once served as a proxy for image quality. Moreover, in vivo image quality assessment required unavailable computational resources. However, both the International Commission on Radiological Protection (ICRP) and the International Atomic Energy Agency (IAEA) now emphasize that optimization must integrate radiation dose and image quality (6, 29). Because image quality is directly correlated with diagnostic detection (30, 31) and to the patient benefit, it is important to assess it in clinically relevant image areas, as shown in this study.

This study has some limitations that should be acknowledged. First, it considered virtual images generated with the simulator configured to replicate a clinical CT scanner. It is known that different vendors implement different strategies to balance radiation output and image quality (12, 24). Future studies can apply the presented methodology to include different CT models. Moreover, concerning the choice of simulating only adult chest and abdominopelvic clinical scenarios with three radiation outputs and images reconstructed with FBP and IR and three kernels at 5 mm slice thickness, it is reasonable to suppose that the specific anatomies and scanning parameters do not affect the trends reported in the study. Future comparisons including different patient anatomical regions for pediatric and adult patients, and different imaging techniques, as well as different range of tissues within a patient CT scan (i.e.: bones), can be performed to confirm this assumption. Finally, we note that there is more to image quality than noise alone. Caution is in order not to presume a singular assessment of noise magnitude can reflect image quality which includes a myriad of other features including noise texture, resolution, and contrast (28, 32).

7. Conclusion

Soft tissue-based noise measurements, particularly the global noise index (GNI), provide more accurate, robust, and reproducible representation of the gold standard ensemble noise in CT imaging, compared to measuring noise in air or phantom datasets. As healthcare policies increasingly mandate image quality evaluation alongside radiation dose, establishing a consistent methodology for noise assessment is essential to ensure reliable procedure optimization and equitable reimbursement practices.

Table 3b.

Standard deviation values for N_ACR, N_AIR, GNI, N_ensemble, N_organ for abdominopelvic images per each kernel, mAs value, and reconstruction algorithm involved in the study. Units are HU.

Abdomen
		FBP					IR
		NACR	NAIR	GNI	Nensemble	Norgan	NACR	NAIR	GNI	Nensemble	Norgan
	50 mAs	0.141	0.010	0.026	1.54	1.80	0.092	0.008	0.026	1.16	1.37
	100	0.095
Br32f	mAs		0.006	0.032	1.07	1.26	0.063	0.005	0.028	0.793	0.942
	150	0.060
	mAs		0.005	0.022	0.869	1.02	0.041	0.004	0.001	0.635	0.751
	50 mAs	0.324	0.043	0.067	4.29	4.89	0.178	0.038	0.048	2.76	3.08
	100	0.208
Br46f	mAs		0.038	0.038	2.97	3.41	0.140	0.016	0.028	1.89	2.10
	150	0.157
	mAs		0.023	0.040	2.41	2.77	0.083	0.013	0.030	1.52	1.69
	50 mAs	1.34	0.070	0.244	19.3	21.6	0.708	0.052	0.126	10.6	11.2
	100	0.812
Br62f	mAs		0.064	0.132	13.4	15.1	0.437	0.055	0.094	7.25	7.72
	150	0.615
	mAs		0.050	0.106	10.9	12.3	0.275	0.044	0.079	5.84	6.25

Table 4b.

Coefficient of variation values for N_ACR, N_AIR, GNI, N_ensemble, N_organ for abdominopelvic images per each kernel, mAs value, and reconstruction algorithm involved in the study.

Abdomen
		FBP					IR
		NACR	NAIR	GNI	Nensemble	Norgan	NACR	NAIR	GNI	Nensemble	Norgan
	50 mAs	0.029	0.002	0.002	0.143	0.151	0.029	0.002	0.003	0.124	0.136
	100	0.027
Br32f	mAs		0.002	0.004	0.141	0.151	0.028	0.002	0.004	0.121	0.134
	150	0.021
	mAs		0.002	0.003	0.140	0.150	0.022	0.002	0.001	0.119	0.132
	50 mAs	0.024	0.003	0.002	0.143	0.148	0.021	0.003	0.002	0.124	0.130
	100	0.022
Br46f	mAs		0.003	0.002	0.140	0.147	0.024	0.002	0.002	0.120	0.127
	150	0.020
	mAs		0.002	0.002	0.139	0.147	0.018	0.002	0.002	0.119	0.125
	50 mAs	0.029	0.001	0.002	0.147	0.148	0.029	0.002	0.002	0.137	0.133
	100	0.026
Br62f	mAs		0.002	0.001	0.145	0.148	0.026	0.002	0.002	0.133	0.130
	150	0.024
	mAs		0.001	0.001	0.144	0.147	0.020	0.002	0.002	0.132	0.130

Take home points.

• Soft tissue-based noise measurements, particularly the global noise index (GNI), provide more accurate, robust, and reproducible representation of the gold standard ensemble noise in CT imaging, compared to measuring noise in air or phantom datasets. Soft tissue-based noise measurements is a close surrogate to inform protocol design and technology assessment.

• Virtual imaging techniques enabled an unbiased comparison of different noise magnitude calculation methods in large and realistic populations simulating clinical conditions. As healthcare policies increasingly mandate image quality evaluation alongside radiation dose, establishing a consistent methodology for noise assessment is essential to ensure reliable procedure optimization and equitable reimbursement practices.

Data statement

The authors declare that they had full access to all the data in this study and the authors take complete responsibility for the integrity of the data and the accuracy of the data.