Emphysema Quantifications With CT Scan

Abstract

Background

CT scan has notable potential to quantify the severity and progression of emphysema in patients. Such quantification should ideally reflect the true attributes and pathologic conditions of subjects, not scanner parameters. To achieve such an objective, the effects of the scanner conditions need to be understood so the influence can be mitigated.

Research Question

How do CT scan imaging parameters affect the accuracy of emphysema-based quantifications and biomarkers?

Study Design and Methods

Twenty anthropomorphic digital phantoms were developed with diverse anatomic attributes and emphysema abnormalities informed by a real COPD cohort. The phantoms were input to a validated CT scan simulator (DukeSim), modeling a commercial scanner (Siemens Flash). Virtual images were acquired under various clinical conditions of dose levels, tube current modulations (TCM), and reconstruction techniques and kernels. The images were analyzed to evaluate the effects of imaging parameters on the accuracy of density-based quantifications (percent of lung voxels with HU < −950 [LAA-950] and 15^th percentile of lung histogram HU [Perc15]) across varied subjects. Paired t tests were performed to explore statistical differences between any two imaging conditions.

Results

The most accurate imaging condition corresponded to the highest acquired dose (100 mAs) and iterative reconstruction (SAFIRE) with the smooth kernel of I31, where the measurement errors (difference between measurement and ground truth) were 35 ± 3 Hounsfield Units (HU), −4% ± 5%, and 26 ± 10 HU (average ± SD), for the mean lung HU, LAA-950, and Perc15, respectively. Without TCM and at the I31 kernel, increase of dose (20 to 100 mAs) improved the lung mean absolute error (MAE) by 4.2 ± 2.3 HU (average ± SD). TCM did not contribute to a systematic improvement of lung MAE.

Interpretation

The results highlight that although CT scan quantification is possible, its reliability is impacted by the choice of imaging parameters. The developed virtual imaging trial platform in this study enables comprehensive evaluation of CT scan methods in reliable quantifications, an effort that cannot be readily made with patient images or simplistic physical phantoms.

FOR EDITORIAL COMMENT, SEE PAGE 1001

Take-home Points.

ResearchQuestion: How do CT scan imaging parameters affect the accuracy of emphysema-based quantifications and biomarkers?

Results: The most accurate imaging condition corresponded to the highest acquired dose (6.5 mGy-CTDI_vol) and iterative reconstruction (SAFIRE) with the smooth kernel of I31.

Interpretation: The results highlight that although CT scan quantification is possible, its reliability is impacted by the choice of imaging parameters and can be comprehensively evaluated using virtual imaging trials.

COPD is one of the leading causes of death worldwide.¹ This disease is characterized by bronchitis and emphysema, with variable severity and distribution in individuals and across patient populations.^2,3 The heterogeneity of COPD cannot be fully captured using conventional diagnostic methods, that is, spirometry tests, respiratory symptoms, and history of exposure to smoking or environmental factors.^4,5 CT scan is an in vivo tool that provides morphological and functional information about the lungs. Prior studies have demonstrated the potential of CT scan in quantitation and stratification of the disease.^{5, 6, 7, 8, 9}

Although promising, CT scan quantifications can be affected by the scanner parameters and conditions.^{10, 11, 12, 13} For this quantification to reflect the true attributes of the patient, these effects need to be comprehensively understood and optimized. Such assessments cannot be fully done using patient images because they lack the underlying ground truth. In addition, patients cannot be imaged repeatedly across multiple conditions because of radiation dose concerns. Alternative assessment of the effects via physical phantoms is further limited because such phantoms do not represent the diverse anatomy and pathology of real patients. Alternatively, these imaging trials can be explored using realistic simulations of targeted patients and imaging scanners, a process known as in silico or virtual imaging trials (VITs).¹⁴

Although VITs have been developed and used for many years across various medical imaging applications,¹⁴ they have not yet been applied in thoracic CT scan applications, especially COPD quantifications. The purpose of this study was to develop a framework to enable VITs in COPD imaging and further investigate the effects of CT scan imaging parameters on the accuracy of imaging biomarkers.

Study Design and Methods

The overall framework of this study is shown in Figure 1. In summary, we developed the first suite of computational, anthropomorphic phantoms with emphysema abnormalities. The developed phantoms were input to a validated, scanner-specific CT scan simulator (DukeSim), acquiring virtual images at various clinical imaging conditions. The virtual images were studied to evaluate the effects of imaging parameters on the accuracy of density-based COPD quantifications across varied patients. These steps are further described in the following sections.

Figure 1.

Overall framework of the study. A library of virtual patients with emphysema were developed and virtually imaged with scanner-specific conditions across various imaging parameters. The resulting images were compared to the ground truth of the patient models in terms of COPD imaging biomarkers.

Computational, Anthropomorphic Phantoms

A total of 20 models with emphysema were developed in this study. The anatomic components of the models were created in three separate modeling steps: (1) body habitus, (2) lung intra-organ structures, and (3) emphysematous regions. For the body habitus, we used the Extended Cardiac-Torso (XCAT) phantoms, which are a library of anthropomorphic phantoms based on real human CT scan images.¹⁵ The phantoms include patient-specific models of the major organs and structures in the body. XCAT phantoms have been used in numerous medical imaging research studies to evaluate and optimize technologies in terms of radiation dose and image quality.^14,15 Originally, the XCAT phantoms did not include detailed models of lungs. The lung nonparenchyma (airways, arteries, and veins) and parenchyma structures were added using previously developed mathematical models that were anatomically informed.^16,17 The models included detailed tree structures of airways, arteries, and veins as well as models of secondary pulmonary lobules for the parenchymal regions.

We built emphysema models and incorporated them in the XCAT phantoms. The emphysema models were created based on real CT scan images of patient images. Twenty CT scan cases were randomly chosen from the COPDGene Phase I dataset⁵ (with institutional review board approval) to include diverse models of emphysema with varying size, patterns (eg, centrilobular, paraseptal, and panlobular), density, and morphology. The CT scan cases were acquired at the end-inspiration phase. For each CT scan volume, the emphysema regions were manually segmented by an expert medical imaging scientist. The segmentations were done slice by slice for the whole lung by delineating areas that had lower attenuations compared with the healthy regions of the lung parenchyma. The boundaries of the segmentation masks were jagged because of the spatial resolution limitation of the reconstructed CT scan images. Therefore, the emphysema masks were fit with polygon mesh surfaces and smoothed. The smoothed emphysema surfaces were incorporated into the lungs of 10 XCAT phantoms, creating 20 human models with variable anatomy and emphysema abnormalities. Each XCAT phantom had two different diseased models of emphysema, enabling us to study different emphysema and lung density conditions within the same body habitus.

To place the emphysema model within a given XCAT phantom, we imported the polygon mesh surfaces for the segmented lung and emphysema from the COPDGene case into the Rhinoceros 3D modeling software. The target XCAT lung surface for the desired phantom was also imported. Affine transformation was applied to the patient lung and emphysema surfaces to initially match them to the phantom lung, using the translation and scaling operations within Rhinoceros. The CageEdit function of Rhinoceros was then used to non-rigidly deform the patient surfaces to match the target lung volume. CageEdit allows the surfaces to be deformed, using a bounding box as a control object. The bounding box is defined by a 3D matrix of control points, which can be moved, causing the portions of the surfaces in the vicinity to deform. With this function, the patient surfaces were manually fine-tuned to match the target phantom lung. This ultimately resulted in the emphysema surfaces fitting within the new lung volume of the target XCAT phantoms. The final XCAT models with emphysema were transformed to a voxelized format with an isotropic voxel size of 0.25 mm for CT scan simulation purposes.

The materials for the phantoms were defined by assigning density and elemental compositions to each organ and structure. For the healthy organs and structures, both density and elemental compositions were set based on the International Commission for Radiation Units report 46.¹⁸ For the emphysema regions, the elemental compositions were defined as lung parenchyma, and the density was set based on measurements from the COPDGene CT scan cases to provide diverse density values across the library of modeled patients. For each CT scan case, the average Hounsfield Unit (HU) value was measured in the segmented emphysema mask. To minimize potential overlaps of other tissues, the segmented masks were eroded by a cubic morphological element with a volume of 2 × 2 × 2 mm³.The corresponding density was derived using the relationship between HU, density, and linear attenuation coefficient:

$$H{U}_{emphysema}=\mathrm{1,000}\times \frac{\rho \times {\left(\raisebox{1ex}{$\mu $}\!\left/ \!\raisebox{-1ex}{$\rho $}\right.\right)}_{emphysema}-{\mu }_{water}}{{\mu }_{water}-{\mu }_{air}}\text{,}$$

where $\rho$ is density (g/cm³), $\raisebox{1ex}{$\mu $}\!\left/ \!\raisebox{-1ex}{$\rho $}\right.$ is the mass attenuation coefficient (defined as lung parenchyma for emphysema, cm²/g), and $\mu$ is the linear attenuation coefficient (1/cm). Each XCAT phantom had one density assigned to the whole emphysema regions. The same approach was also used to assign lung density for the lung parenchyma for each case.

Image Acquisition

The developed COPD patient models were imaged with a validated, scanner-specific CT scan simulator (DukeSim¹⁹) across various imaging parameters. DukeSim is a hybrid CT scan simulator that generates projection images of a computational phantom, in a voxelized format, using a ray-tracing module (for primary signal) and a Monte-Carlo module (for scatter signal).¹⁹ DukeSim accounts for various scanner-specific features, including poly-energetic source spectrum, bowtie filter, focal spot wobbling, tube current modulation, poly-energetic detector response, and beam hardening correction.¹⁹ The simulator is validated to provide close representation of experimental clinical data in terms of both dosimetry and image quality across multiple phantoms, scanner models, and imaging settings.^{19, 20, 21, 22}

In this study, DukeSim was set to model the acquisition geometry and source and detector physics of a clinical CT scanner (Siemens Definition Flash). The emphysema models were virtually imaged using DukeSim with imaging parameters typically used in chest CT scans. The projection images were acquired at a tube voltage of 120 kV, a Siemens body filter, and a pitch of 1.0. To study the effects of radiation dose on biomarker quantifications, the human models were imaged at three dose levels (20, 60, and 100 mAs, ie, 1.3, 3.9, and 6.5 mGy Volume CT Dose Index [CTDI_vol]), to provide low-, medium-, and high-dose conditions. The acquisitions were done with and without tube current modulation (TCM). TCM was modeled using a scanner-specific approach.²¹ To simulate TCM, coronal and sagittal CT scan localizers (aka scout or topogram images) were acquired with scanner-specific conditions (ie, beam collimation and tube current). The localizer images were used to calculate varying attenuations across the longitudinal direction at both coronal and sagittal angles. The attenuations were interpolated between the sagittal and coronal angles to model the angular modulation. The interpolated attenuations were converted into mA vectors using

$$mAs\left(p\right)={mAs}_{ref}{\left(\frac{A\left(p\right)}{A_{ref}}\right)}^a\text{,}$$

where $p$ refers to the projection number, ${mAs}_{ref}$ is the reference mAs, $A\left(p\right)$ is the interpolated attenuations for all projection angles, $A_{ref}$ is the reference attenuation, and $a$ is the modulation strength. The ${mAs}_{ref}$ was set to match the fixed tube current dose levels. The $A_{ref}$ was set to 6.1, to mimic a typical chest scan. A medium modulation strength level was simulated, setting the $a$ to 0.524 for attenuations smaller than $A_{ref}$, and 0.442 for attenuations larger than $A_{ref}$.²¹

The final projection images were reconstructed using a vendor-specific reconstruction software (ReconCT, Siemens Healthineers). To evaluate the effects of reconstruction algorithms and kernels, the reconstructions were done using both weighted filtered backprojection (wFBP) and iterative reconstruction methods at multiple kernels. The iterative reconstruction algorithm used in this study was Sinogram Affirmed Iterative (SAFIRE). The reconstruction kernels were B31f, B46f, and B70f for wFBP and I31f, I46, and I70 for SAFIRE, representing smooth, medium, and sharp kernels. All images were reconstructed at 0.75-mm slice thickness with a slice interval of 0.5 mm, an in-plane pixel size of 0.48 mm, and a field of view of 500 mm.

Data Analysis

The reconstructed CT scan images were compared against the corresponding reference ground truth phantoms in terms of mean absolute error (MAE), and lung histogram-based imaging biomarkers. The MAE was calculated once in the whole lung parenchyma (called lung MAE) and once in only the emphysema regions (called emphysema MAE). To calculate the reference ground truth values, linear attenuation coefficients at the effective energy were assigned to each material of the phantoms for all the voxels. The effective energy of the acquisition (ie, the single energy yielding the same CT scan measurements as the poly-energetic spectrum) depends on the imaging conditions (eg, X-ray spectrum, beam collimation, bowtie filter conditions). Its value (67 keV) was calculated in the beam hardening correction step.¹⁹ The reference linear attenuation coefficients were converted to the unit of CT number (Hounsfield Unit, HU).

The histograms of the lung parenchyma regions were measured in the CT scan images and compared against the ground truth values in terms of three histogram features: mean histogram, LAA-950 (percent of lung voxels with HU < −950), Perc15 (15^th percentile of lung histogram HU). These LAA-950 and Perc15 were chosen because prior studies have shown their predictive capabilities in COPD assessments.⁶ For each CT scan dataset and imaging biomarker, the accuracy of the histogram features were derived by subtracting the measured biomarker with the reference, ground truth value in the corresponding phantom. Furthermore, to compare against previous studies on physical phantoms,^{23, 24, 25} median HU was measured in the lung parenchyma of the reconstructed CT images across all the acquired imaging conditions. MAE compares the CT scan images with the ground truth map on a voxel-by-voxel basis, and then it averages the errors. The imaging biomarkers and median HU, however, reflect descriptive statistics of the histogram of the lung as a whole.

The quantified metrics (ie, lung MAE, emphysema MAE, mean HU, LAA-950, Perc15, median HU) were evaluated across all the studied imaging conditions. Paired t tests were performed to explore statistical differences between any two imaging conditions (total of 630 paired comparison combinations). The null hypothesis was that the difference between the measurements from two imaging conditions had a zero-mean distribution. A Bonferroni correction was applied to account for the multiple comparisons, dividing the significance level (.05) by the total number of paired comparisons (ie, 630 combinations). The statistical analysis was done using MATLAB (MathWorks Inc.).

Results

Computational, Anthropomorphic Phantoms

Figure 2 shows three examples of CT scan images that were used to inform the emphysema models and their corresponding segmented emphysema masks, demonstrating that the emphysema models included various shapes and patterns. Figure 3 shows a collage of the developed 20 phantoms. The phantoms included diverse renditions of emphysema with various size, pattern, morphology, and density. Table 1 shows a summary of the attributes of the phantoms (eg, sex, age, BMI, volume, and density of the lung) and the Fleischner Society visual scores of the paraseptal and centrilobular emphysema of the COPDGene cases that these phantoms were modeled based on, reflecting the visual estimate of the disease severity.²⁶ This table demonstrates that the used phantoms were diverse in terms of the volume and density of the total lung and emphysematous regions. The total lung and emphysema volumes were 5.3 ± 1.24 L and 0.49 ± 0.56 L, respectively. The defined ground truth HU values of the lungs and emphysema were −880.0 ± 37.9 and −969.2 ± 22.7, respectively.

Figure 2.

Top row, CT scan images of three COPDGene cases. Bottom row, segmented masks of the emphysema regions, highlighted in blue.

Figure 3.

Collage of the 20 computational phantoms developed in this study. Ten anatomically variable XCAT models were used, each containing two different renditions of emphysema. XCAT = Extended Cardiac-Torso.

Table 1.

A Summary of the Phantom Attributes Used in This Study

Phantoms	Sex	Age, y	BMI	Lung Volume, L	Lung Density, HU	Emphysema Volume, L	Emphysema Density, HU	Perc15, HU	LAA-950, %	Emphysema Visual Scorea
XCAT 1-a	M	58	36.1	6.71	−891	0.266	−965	−917	3.93	Trace centr. mild paras.
XCAT 1-b	M	58	36.1	6.71	−815	0.029	−950	−862	0.44	Mild centr. subs. paras.
XCAT 2-a	F	27	18.6	3.53	−863	0.031	−974	−898	0.90	Trace centr. not paras.
XCAT 2-b	F	27	18.6	3.53	−936	1.425	−983	−983	40.8	Destr. centr. not paras.
XCAT 3-a	M	31	22.7	7.42	−878	0.005	−998	−909	0.06	Not centr. not paras.
XCAT 3-b	M	31	22.7	7.42	−898	0.549	−996	−919	7.51	Conf. centr. not paras.
XCAT 4-a	F	36	18.2	5.51	−834	0.148	−945	−874	0.0	Mod centr. not. paras.
XCAT 4-b	F	36	18.2	5.51	−883	0.315	−984	−908	5.83	Mild centr. subs. paras.
XCAT 5-a	M	38	21.9	4.14	−852	0.041	−944	−889	0	Trace centr. not. paras.
XCAT 5-b	M	38	21.9	4.14	−844	0.016	−955	−884	0.40	Mod. centr. mild. paras.
XCAT 6-a	M	48	28.5	5.83	−848	0.118	−985	−884	2.04	Mod. centr. mild paras.
XCAT 6-b	M	48	28.5	5.83	−929	1.647	−992	−992	28.7	Destr. centr. subst. paras.
XCAT 7-a	F	49	35.5	6.09	−902	0.378	−974	−923	6.43	Mod. centr. subst. paras.
XCAT 7-b	F	49	35.5	6.09	−915	1.118	−978	−978	19.0	Conf. centr. not paras.
XCAT 8-a	F	51	22.3	5.57	−861	0.249	−940	−894	0	Not centr. subst. paras.
XCAT 8-b	F	51	22.3	5.57	−921	0.755	−987	−934	13.6	Conf. centr. subst. paras.
XCAT 9-a	F	59	24.1	3.95	−918	1.474	−976	−976	37.3	Destr. centr. subst. paras.
XCAT 9-b	F	59	24.1	3.95	−880	0.045	−961	−910	1.16	Mod. centr. not paras.
XCAT 10-a	M	36	24.4	4.35	−922	1.241	−985	−985	28.5	Mod. centr. subst. paras.
XCAT 10-b	M	36	24.4	4.35	−811	0.038	−910	−859	0	Trace centr. mild paras.
Average	…	43.3	25.2	5.31	−880	0.049	−969	−919	9.83	…
SD	…	10.9	6.12	1.25	37.9	0.567	22.7	42.6	13.5	…

In the visual scoring, the centrilobular emphysema was categorized as none, trace, mild, confluent, and destructive. The paraseptal emphysema was categorized as none, mild, and substantial. Cent. = centrilobular; conf. = confluent; destr. = destructive; LAA = percent of lung voxels with HU < −950; mod. = moderate; paras = paraseptal; subs. = substantial; XCAT = Extended Cardiac-Torso.

^aThe emphysema subtype is based on visual scoring of CT scan images provided by the COPDGene study.

Qualitative Demonstrations

Figure 4 shows three simulated CT scan images of three different phantoms (XCAT 2-b, XCAT 6-b, XCAT 1-a) as examples, demonstrating the qualitative realism of the reconstructed images. Figure 5 shows simulated images of a phantom (XCAT 6-b) acquired at two dose levels (20 and 100 mAs) and reconstructed with wFBP and SAFIRE with two different kernels (B31 and B70 for wFBP, I31 and I70 for SAFIRE). Qualitatively, this figure demonstrates the major effects of dose level and reconstruction settings on the noise characteristics and HU values. As expected, the images acquired at the lower dose and reconstructed with wFBP and sharper kernels were noisier, whereas images with sharper kernels had better spatial resolution.

Figure 4.

Simulated CT scan images of three XCAT phantoms (left to right: XCAT 2-b, XCAT 6-b, XCAT 1-a), imaged with a constant tube current of 60 mAs and reconstructed using the weighted filtered backprojection algorithm with a kernel of B46. XCAT = Extended Cardiac-Torso.

Figure 5.

An XCAT phantom (XCAT 6-b) imaged at two dose levels (20 and 100 mAs), and four reconstruction settings. “B” refers to kernels for weighted filtered backprojection and “I” refers to iterative reconstruction. “31” and “70” are kernel sharpness, where the larger number refers to a sharper kernel. XCAT = Extended Cardiac-Torso.

Figure 6 shows the histogram of the lung HU values for two XCAT models (XCAT 2-b and XCAT 4-b) with different extents of emphysema. The figure includes the ground truth histogram before the imaging as well as the histogram after the subject’s being imaged at various dose levels and reconstructed with various kernels. The plots demonstrate that the studied imaging parameters (mAs and reconstruction algorithms) distort the ground truth histogram with the magnitude and shape of this distortion (eg, skewness) influenced by the imaging parameters and patient attributes. This figure qualitatively showed that the histograms were wider and more skewed to the left in the conditions with higher noise magnitude (ie, lower mAs, sharper kernels, and filtered backprojection reconstructions).

Figure 6.

A, Lung HU histograms of the XCAT 2-b. B, Lung HU histograms of the XCAT 4-b. The ground truth histogram is in red. The blue are the histograms of the CT images across different dose levels and reconstruction conditions. In the first row, the dose varies while reconstruction settings are constant. The second and third rows show the effects of weighted filtered backprojection reconstruction and iterative reconstructions, respectively, across three kernel sharpness. HU = Hounsfield Units; XCAT = Extended Cardiac-Torso.

Quantitative Results

Figure 7 shows the quantitative evaluation of the lung MAE, emphysema MAE, the accuracy of mean lung HU, LAA-950, and Perc15, and the measured median lung HU values across all the studied imaging parameters. The results are reported in terms of average and SD across all subjects. In general, the heatmaps in Figure 7 highlight that CT scan quantifications are not generally consistent across different imaging parameters.

Figure 7.

Lung MAE (A), emphysema MAE (B), accuracy of lung mean HU (C), LAA-950 (D), Perc15 (E), and lung median HU (F) across different dose (y-axis) and reconstruction settings (x-axis). The plots on the left are the average values across all patients, and the ones on the right are the corresponding standard deviations. Accuracy is defined as the difference between measurements and their corresponding ground truth value. LAA-950 = percent of lung voxels with HU < −950; MAE = mean absolute error; HU = Hounsfield Units; TCM = tube current modulations.

The lung MAE heatmaps show errors between 57 to 175 HU and 54 to 144 HU in the lung parenchyma and emphysema regions, respectively. Several factors contribute to this error, including noise magnitude, presence of scatter, and the beam hardening effect. The errors in the emphysema regions were lower than the whole lung parenchyma. The reason could be that the emphysema regions had lower density values with uniform materials and the error increase in the denser regions and edge voxels (Fig 8).

Figure 8.

Left, Ground truth rendition of an XCAT phantom (XCAT 9-a). Middle, the corresponding virtual CT scan image acquired with fixed 100 mAs and reconstructed at B31. Right. The difference between the CT and ground truth.  The ground truth and CT histograms are shown in the bottom. XCAT = Extended Cardiac-Torso.

The lowest errors were observed in the highest dose (6.5 mGy-CTDI_vol, 100 mAs) and iterative reconstruction with kernel of I31. Figure 9 shows the difference of lung MAE between the studied imaging parameters while other parameters were set to their optimum values. Without tube current modulation and at the kernel of I31, the lung MAE improved by 3.5 ± 1.9 (average ± SD) HU, and 4.2 ± 2.3 (average ± SD) HU, when mAs was increased from 20 to 60 and from 20 to 100 mAs, respectively (Fig 9A). Tube current modulation did not contribute to a systematic improvement of lung MAE (Fig 9B). The sharpest kernel had the worst lung MAE across all cases. However, at the optimum fixed dose of 100 mAs, the difference between the lung MAE for I46 and I31 were not found statistically significant (Fig 9C). In addition, at the fixed dose of 100 mAs, the iterative reconstruction kernel of I31 improved the lung MAE by 2.0 ± 0.8 (average ± SD), compared with B31 images (Fig 9D).

Figure 9.

A-C, Box-plots of differential lung MAE comparing each studied parameter while setting the other parameters at their optimum values. A, Compares dose levels without TCM and I31 reconstruction kernel (B) with and without tube current at 100 mAs and I31 reconstruction kernel. C, Compares kernel sharpness at 100 mAs, with iterative reconstruction. D, Compares iterative and filtered backprojection at 100 mAs, without TCM. ∗Refers to comparisons when the P value of the paired t-test was smaller than the significance level (0.05/630). and “n.s.” refers to comparisons which found not significant. HU = Hounsfield Units; MAE = mean absolute error; TCM = tube current modulations.

Similar to the MAE results, the heatmaps in Figure 7C-7E demonstrate the variable measurements of mean Lung HU, LAA-950, and Perc15. The mean of the lung histogram was overestimated across all imaging conditions, showing a positive bias (compared with the ground truth) mainly because of the presence of scatter.²⁷ This bias was found to be lower in the noisier conditions (ie, lower dose, filtered backprojection, and sharper kernels), mainly because the histogram of the lungs was skewed to the left in those conditions. This same effect caused an increase in LAA-950 and a decrease in Perc15 measurements with a decrease in radiation dose and increase in the kernel sharpness.^12,23,28

The accuracy of the histogram features (ie, mean HU, LAA-950, Perc15) were not necessarily optimum at the expected imaging parameters of highest dose, iterative reconstruction, and smooth kernel. This is because the measured histogram has a positive bias, whereas noisier conditions skew the histogram to the left. The combination of these two effects could cause a coincidental improvement in the accuracy of some histogram measurements.

Table 2 summarizes the difference of all the measurements between the studied imaging parameters while other parameters were set to their optimum values. This table also specifies the paired conditions where the differences in the measurements were found statistically significant. In general, this table demonstrates that changes in radiation dose and reconstruction algorithm and kernel could result in significant difference in density-based measurements, whereas the measurements were quite robust to the presence or absence of tube current modulations. Furthermore, for the smoother kernels of “I31” and “I46,” the differences between the measurements were found insignificant, except for the mean lung HU and Perc15.

Table 2.

Paired Differences Between Measurements at Various Imaging Parameters

Imaging Settings	Compared Parameters	Lung MAE	Emph. MAE	Mean lung	LAA-950	Perc15	Median Lung	Fixed Parameters
Dose	100-60 mAs	−0.7 ± 0.4a	−0.8 ± 0.4a	0.1 ± 0.0a	−0.0 ± 0.2	1.7 ± 1.5	−0.2 ± 0.8	No TCM, I31
	100-20 mAs	−4.2 ± 2.4a	−4.3 ± 1.7a	0.6 ± 0.2a	−0.9 ± 1.5	9.2 ± 5.7a	−0.5 ± 2.3
	60-20 mAs	−3.5 ± 1.9a	−3.4 ± 1.3a	0.4 ± 0.2a	−0.8 ± 1.3	7.6 ± 4.3a	−0.2 ± 1.7
TCM	No TCM - TCM	−0.2 ± 0.4	−0.4 ± 0.5	0.1 ± 0.1	0.1 ± 0.1	0.7 ± 1.7	−0.2 ± 0.7	100 mAs, I31
Recon Kernel	I31-I46	0.9 ± 1.0	−0.2 ± 0.9	2.7 ± 0.3a	−0.1 ± 0.4	3.4 ± 2.5a	0.8 ± 1.2	100 mAs, no TCM
	I31-I70	−13.8 ± 10.2a	−7.8 ± 8.8	12.1 ± 3.1a	−8.6 ± 5.7a	44.5 ± 14.9a	6.1 ± 4.5a
	I46-I70	−14.7 ± 9.3a	−7.6 ± 8.1	9.4 ± 2.8a	−8.6 ± 5.5a	41.1 ± 12.7a	5.3 ± 3.3a
Recon algorithm	I31-B31	−2.0 ± 0.8a	−1.2 ± 0.8a	−0.9 ± 0.1a	−0.1 ± 0.6	3.8 ± 2.6a	−1.1 ± 1.1	100 mAs, no TCM

Emph = emphysema; LAA = percent of lung voxels with HU < −950; MAE = mean absolute error; TCM = tube current modulation. The units are all in HU except LAA-950, which is in %.

^aDenotes the conditions where the difference was found to be statistically significant.

Discussion

The severity and progression of COPD can potentially be characterized in a patient-specific manner using quantitative CT scan. CT scan quantifications have variable accuracy depending on how the images are acquired and processed. This study comprehensively explored the effects of imaging parameters (dose level, tube current modulation, and reconstruction algorithm and kernel) on CT scan quantifications for COPD imaging using the novel approach of virtual imaging trials. We developed the first suite of human models with diverse emphysema abnormalities and patient attributes. The human models were imaged using a validated, scanner-specific simulation platform. This tool enabled us to study the imaging effects against the ground truth that was defined digitally in the human models.

Our findings demonstrated that the quantification accuracies were quite variable when different dose levels and reconstruction settings were used. The MAE values showed that images acquired with high dose, iterative reconstructions, and smoother kernels could provide more accurate quantifications, likely because they are less affected by image noise. MAE was quite similar with and without tube current modulation, because both were conditioned at similar reference dose levels.

The findings in this paper were in line with those of previous studies. In a study, Sieren et al²⁴ imaged a physical phantom (COPDGene2 test-object) at various dose levels and reconstructed with both iterative (SAFIRE) and weighted filtered back-projection algorithms. In some conditions close to this study (CTDI_vol of 1.58-5.97 mGy and reconstruction kernels of I30 and B35), they observed less than 1 HU difference in median lung HU, which is in line with the results in this study (Fig 9E, Table 2). Rodriguez et al²³ imaged a similar phantom and demonstrated that use of a sharp kernel and low radiation dose could change Perc15 of the “lung foam” insert up to 23 HU. Although the scanner model and the reconstruction settings were different from those in our study, the trends and order of changes are within the same range. Our findings agree with a study from Yuan et al²⁸ in terms of LAA-950. They observed that low dose increases the LAA-950 by approximately a 5% difference at a standard reconstruction kernel. Chong et al²⁹ and Hochhegger et al³⁰ have shown that sharper kernels increase the variability in the LAA-950 measurements (Fig 8, Table 2). Several studies have also shown that density-based quantifications were less variable when higher dose and iterative reconstructions (compared with wFBP) were used.^{25,31, 32, 33} Although these findings agree with our results, they did not evaluate how accurate the measurements are because they did not have access to the ground truth. We were able to quantify this accuracy by comparing the measurements against the digital ground truth of our computational phantoms.

The density of the lungs in the CT scan images was different from the ground truth (Figure 6, Figure 7, Figure 8). The presence of scatter results in a positive bias to the histogram, which causes an overestimation in the density measurements,²⁷ suggesting a need to quantify the magnitude of scatter across different scanner architectures (eg, dual-source vs single source) and to minimize its effect through improved hardware design or scatter correction processing algorithms. The presence of noise increases the variation of the HU values, which resulted in a broadened lung HU histogram. This degradation of the histogram could be asymmetrical and skewed to more negative HU values^12,23 (Fig 6).

The VIT platform developed in this study can further be used toward more reliable and accurate CT scan quantifications in COPD imaging for both prospective and retrospective studies. For prospective studies, VITs may explore the optimized imaging settings for more accurate and consistent measurements across various scanner makes and models, including the more advanced technologies such as photon-counting CT scan. For retrospective studies, VITs may explore harmonization techniques to equate quantifications that are measured with diverse imaging settings and scanner makes and models.

This study has certain limitations. The emphysema models in this study included diverse sizes, shapes, patterns, and densities, informed by CT scan images. Although this diversity is necessary to run such virtual imaging trials, the emphysema models could be more accurate in terms of modeling specific subtypes (eg, centrilobular, paraseptal, panlobular). Such models require better understanding of the shapes, texture, and pathology-based definition of densities. Future studies may explore modeling these subtypes more accurately for other imaging applications such as texture- and subtype-based imaging biomarkers. We focused on a single but common scanner model and explored four major imaging parameters. Future studies may investigate the effects of current and emerging scanner models and settings. Although our study included subjects with variable lung volume and density, we did not directly evaluate the effects of inspiration level in the accuracy of biomarker quantifications. Such a study requires computational subjects with realistic respiration models and may be explored in future studies.

This study was focused on density-based features in patients with emphysema. The potential relationships between emphysema and pulmonary airways and vessels were not considered. Future studies may create such computational phantoms, creating emphysema and nonparenchyma abnormalities (eg, bronchitis) with potential applications in other COPD-related imaging biomarkers such as airway and vasculature measurements. We developed 20 human models with various patient attributes and emphysema abnormalities. The generalizability of this study may be explored and further improved with a larger patient cohort and more diverse severity and subtypes of emphysema. A larger patient cohort would also enable systematic evaluations of patient attribute effects on CT scan quantifications.

Interpretation

The developed virtual imaging trial platform in this study enables comprehensive evaluation of CT scan technologies and quantifications that are not possible using patient images or simplistic physical phantoms. The virtual trial study conducted in this work demonstrated that the accuracy of density-based quantifications is affected by radiation dose, reconstruction algorithms, and reconstruction kernels. Images with high dose, iterative reconstructions, and smooth kernels were more accurate. The quantification accuracies were also variable across subjects. Smaller patients and patients with lower densities had more accurate quantifications. Our results highlight that although CT scan quantification is insightful, its reliability cannot be assumed without careful consideration in the choices for CT scan imaging parameters.

Funding/Support

This study was funded by National Institutes of Health (R01HL155293 and P41EB028744).

Financial/Nonfinancial Disclosures

The authors have reported to CHEST the following: E. A. and W. P. S.: Unrelated to this study, they have collaborations with Metis Health Analytics. D. L.: Receives support from the NHLBI (U01 HL089897, U01HL089856). E. S.: Unrelated to this study, collaborations with GE, Siemens, Imalogix, 12Sigma, SunNuclear, Nanox, Metis Health Analytics, Cambridge University Press, and Wiley and Sons. None disclosed (G. J.).