Precise Lung Density Quantification with a Physics-based CT Harmonizer

Abstract

Purpose:

To develop a physics-based image harmonization method that transforms images into a reference quality index of noise, spatial resolution, and lung volume and to evaluate its performance for improving reproducibility of lung density measurement.

Materials and Methods:

This retrospective analysis of Genetic Epidemiology of Chronic Obstructive Pulmonary Disease (COPDGene) study data included participants who underwent chest CT with full-dose (200 mAs) and reduced-dose (40–80 mAs) protocols during the same visit between November 2014 and July 2017. A harmonization algorithm was developed to reduce variations in lung density by adjusting spatial resolution, noise, and lung volume in sequence, aligning scans to a reference condition based on full-dose, soft-kernel reconstruction. The percentage of lung voxels less than −950 HU (LAA-950) and the 15th percentile of the lung density histogram (Perc15) were calculated from scans. Improvements in bias, limits of agreement, and reproducibility coefficients were evaluated, and harmonizer performance was compared with that of two existing techniques: volume-adjusted lung density (VALD) and median filtering followed by VALD (MF-VALD).

Results:

A total of 1159 participants (mean age ± SD, 65 years ± 9; 586 male) were studied. Across all imaging conditions, the harmonization technique improved the Perc15 reproducibility coefficient 4.8-fold, from 35.6 HU ± 0.7 to 7.4 HU ± 0.2. The harmonizer outperformed VALD and MF-VALD across the full-dose and reduced-dose scans, with reproducibility coefficients improving from 32.4 HU ± 0.7 before harmonization to 30.1 HU ± 0.3 with VALD, 10.3 HU ± 0.3 with MF-VALD, and 7.7 HU ± 0.2 with the harmonizer.

Conclusion:

This physics-based technique harmonized CT images to a reference quality index, improving the reproducibility of lung density metrics.

Clinical trial registration no. NCT00608764

Summary

A physics-based image harmonization method was developed and validated; it improved CT-based lung density reproducibility by aligning images to a reference quality index of noise, resolution, and lung volume.

CT is a noninvasive imaging modality that offers detailed insights into pathologic features, enabling precise characterization and therapeutic monitoring of various lung diseases, including emphysema (1). Despite the increased use of CT in disease assessment, robust quantification of lung density remains a critical challenge. This challenge stems from differences in lung volumes (2–4), CT equipment (5), radiation dose levels (6), and image reconstruction protocols (6) across acquisitions. Reproducibility coefficients, ranging from 7 HU in the standardized SPIROMICS (Subpopulations and Intermediate Outcome Measures in COPD Study) trial (7) to 18 HU in small cohorts (8), have been reported in previous studies, underscoring the magnitude of measurement variability. Such inconsistency limits reliable cross-sectional and longitudinal comparisons and hinders the design of clinical trials for emphysema therapy. Hence, image harmonization is critically needed to enable robust and reliable quantification of lung density across varying conditions.

Previous efforts have been made to reduce variations in image-based lung quantification. Volume inconsistencies have been addressed with models based on lung mass preservation (9), and reconstruction kernel effects have been mitigated with theoretical and deep learning methods (10–12). Generative adversarial networks have been applied to harmonize radiomics features (13) and other image-derived measures (14). Statistical models, such as linear mixed-effects models (15), and analytical adjustments for volume, noise, and bias (16) have also been explored. Although these methods have improved consistency, each is limited: Single-attribute corrections address only specific parameters, generative adversarial networks may not preserve clinically relevant biomarkers (17), and analytical techniques adjust metrics rather than images, restricting their utility for localized disease assessment and emerging technologies.

The purpose of this study was to build on previous techniques, develop a physics-based image harmonization method that transforms images into a reference quality index of noise, spatial resolution, and lung volume, and evaluate its performance for improving lung density measurement reproducibility.

Materials and Methods

Study Sample

This study was a retrospective analysis of the Genetic Epidemiology of Chronic Obstructive Pulmonary Disease (COPDGene) study. The COPDGene study is an ongoing multicenter observational study (October 2007 to present) conducted at 21 different centers across the United States in individuals with at least a 10 pack-year smoking history (www.clinicaltrials.gov [NCT00608764]) (18). Institutional review boards at all participating centers approved the COPDGene study, the study complied with the Health Insurance Portability and Accountability Act, and all participants provided written informed consent. S.S.P. was employed by Siemens Medical Solutions USA at the time of manuscript revision but not during the conduct of the study. All analyses were supervised by E.A.

Criteria for inclusion in this study were consecutive COPDGene study participants who underwent same-day full-dose (200 mAs) and reduced-dose (40–80 mAs) CT between November 2014 and July 2017. Exclusion criteria were the absence of data from any of the three phases (to ensure a stable cohort for future longitudinal analyses), missing scanner model information (required for the harmonization algorithm), unavailable modulation transfer function (MTF) measurements (because of missing phantom data), and image quality issues (ie, misaligned or atypical paired reconstructions). Study cohort selection based on inclusion and exclusion criteria is summarized in Figure 1.

Figure 1:

Flow diagram of patient selection. RD = reduced dose, MTF = modulation transfer function.

CT Acquisition

All participants underwent volumetric CT and were coached to perform full-inspiration and end-expiration breath holds (18). CT scans were acquired at two dose levels: a full-dose acquisition with a fixed tube current of 200 mAs and a reduced-dose acquisition leveraging dose modulation with tube current varying from 40 to 80 mAs (19). The acquisitions were reconstructed with one or more weighted filtered back-projection (WFBP) reconstruction kernels (ie, soft to sharp), resulting in a total of 3793 CT volumes (see Table S1 for details on scanner makes, models, and reconstruction kernels).

The data included an automatic lung segmentation (LungQ; Thirona) (20) and the calculation of imaging biomarkers of emphysema, including the 15th percentile of the lung density histogram (Perc15) and the percentage of lung voxels less than −950 HU (LAA-950). Predicted lung volumes were also calculated with an equation developed in the Multi-Ethnic Study of Atherosclerosis Lung Study (MESA) (21).

Effect of Changes in Image Quality and Lung Volume on Lung Density

Before harmonization, variations in image quality—specifically spatial resolution and noise—and lung volumes were quantified, and the respective effect of image quality on lung density was evaluated. To demonstrate this influence, spatial resolution was parametrized with the spatial frequency corresponding to the 50% magnitude (f₅₀) of the MTF, a scalar metric derived from the MTF curve, and compared across soft and sharp kernels for full-dose acquisitions (see Appendix S1, Table S1 for the definition of soft and sharp kernels and Fig S1 for the corresponding MTFs). Noise was measured with the global noise index metric (22) and compared across full-dose and reduced-dose acquisitions, reconstructed with soft kernels. Lung volume at total lung capacity was quantified through the use of segmented lung masks and compared across full-dose and reduced-dose scans.

Harmonization Method

In this study, a harmonization method was developed to improve the precision of lung density measurements by accounting for variations in spatial resolution, noise, and lung volume by transforming images to match reference image quality indexes and a predicted lung volume, and this method was then tested. The framework consisted of a preprocessing stage, followed by the application of three main modules in sequence. The preprocessing stage includes histogram extrapolation and voxel-based resampling. The main modules include spatial resolution matching, noise matching, and lung volume adjustment. The following paragraphs briefly summarize these components (see Appendix S2 for further details):

Histogram extrapolation addresses the truncation artifact from densities that are lower bound at −1024 HU, a prominent artifact observed in low-dose settings (23).

Voxel-based resampling compensates for the variability in voxel size across data from the same patient.

Spatial resolution matching reduces variability across reconstruction kernels through the use of a physics-based image filtering technique based on MTFs as the filter function (10), with filtering applied to spatial-frequency contents below the Nyquist limit. The sharp kernel reconstructions were adjusted to match soft reconstructions, assigned as reference, given the superior accuracy of lung density biomarkers with soft kernels (6,24).

Noise matching addresses the noise variations across reduced-dose and full-dose scans. Median filtering, a conventional denoising technique, has been successfully applied in previous quantitative imaging studies (19) and therefore was used as a basis for the filtering approach, referred to as the noise-adaptive median filtering (NAMF) approach. This method was designed to apply an initial denoising step using the two-dimensional median filter (25), followed by an adjustment of the reduced-dose scan with the residual image (ie, difference between denoised and original image) to match the noise level of the full-dose scan.

Volume adjustment reduces the influence of lung volume variability and is performed by using a sponge model operating on the MESA-predicted lung volumes to calculate the volume-adjusted lung density (VALD) values (19). Volume adjustment can be applied with the following equation:

$$H{U}_{i,\mathit{\text{volume}}-\mathit{\text{adjusted}}}=\left(\frac{\mathit{\text{Volum}}{e}_{\mathit{\text{original}}}}{\mathit{\text{Volum}}{e}_{\mathit{\text{predicted}}}}\times (H{U}_{i,\mathit{\text{original}}}+1000)\right)-1000,$$

where $H{U}_{i,\mathit{\text{volum}}e-\mathit{\text{adjusted}}}$ is the Hounsfield unit value at voxel $i$ in the volume-adjusted scan, $H{U}_{i,\mathit{\text{original}}}$ is the Hounsfield unit value at voxel i in the original scan, $\mathit{\text{Volum}}{e}_{\mathit{\text{original}}}$ is the lung volume obtained from the original CT scan, and $\mathit{\text{Volum}}{e}_{\mathit{\text{predicted}}}$ is the MESA-predicted lung volume. The MESA-predicted lung volume can be calculated with the equation developed in a previous study (21).

Together, these modules harmonize image quality and account for key sources of variability in lung density measurements.

Performance Assessment of Harmonization

Before application to patient data, the harmonization algorithm was tested on an imaging dataset from 20 anthropomorphic virtual human models with emphysema (26). Use of this dataset allowed an objective validation and assessment of the algorithm in a controlled environment without the confounding interplay of lung volume and noise across repeated examinations of the same patient (see Appendix S3 for further details on the virtual dataset settings and Tables S2 and S3 and Figs S2–S5 for related results [27–31]).

After the validation on virtual data, the performance of the harmonization algorithm was evaluated on the COPDGene dataset under changes in radiation dose (full-dose and reduced-dose) and reconstruction kernel (soft and sharp). The reference condition was considered the full-dose acquisition reconstructed with the soft kernel, and the remaining imaging conditions were transformed to that reference. The paired soft and sharp kernel reconstructions of reduced-dose and full-dose scans enabled targeted evaluation of resolution and noise matching components of the harmonization. The performance of the harmonizer was evaluated in a tiered fashion for first the spatial resolution matching; then the spatial resolution and noise matching; and then spatial resolution matching, noise matching, and volume adjustment together.

In addition to the evaluation noted above, the applied NAMF denoising technique was assessed on its own against established denoising methods (bilateral, Gaussian, median, Wiener, and nonlocal means filters), with an added lung volume adjustment to account for variations in lung volumes between full-dose and reduced-dose scans. The applied denoising technique (NAMF) followed by VALD (NAMF-VALD) was compared with the traditional VALD method and with median filtering followed by VALD (MF-VALD). All comparisons were made in terms of Perc15 and LAA-950 biomarkers.

Statistical Analysis

Comparisons of image quality (spatial resolution and noise), long volume measurements, and their influence on Perc15 measurements across imaging conditions (reconstruction kernels and dose levels) were assessed using paired t tests (SciPy package in Python, scipy.stats.ttest_rel). Agreement in lung density across protocols before and after harmonization was evaluated using parameters from the Bland-Altman plots, namely bias (mean of differences) and limits of agreement (1.96 times the SD of differences). The deviation of bias from zero was analyzed using paired $t$ tests. Measurement reproducibility, calculated before and after harmonization, denoted the threshold below which differences between repeated measurements on the same patient, under different conditions, should fall with 95% probability. Measurement reproducibility was calculated using the total within-subject SD method as follows:

$$\mathit{\text{Reproducibility}}\mathit{\text{Coefficient}}=2.77\times s_w$$

where $s_w$ is the pooled within-subject SD across all participants. The pooled within-subject variance $\left({\mathit{s}}_w^2\right)$ was computed from the individual within-subject variances for each subject $i$ as:

$${\mathit{s}}_w^2=\frac{1}{n}\sum _{i=1}^n{\mathit{s}}_i^2$$

$${\mathit{s}}_i^2=\frac{1}{m-1}{\sum _{j=1}^m\left({\mathit{x}}_{ij}-\overline{{\mathit{x}}_i}\right)}^2$$

where $x_{ij}$ denotes the $j$th measurement for subject $i$ and $\overline{{\mathit{x}}_i}$ is the mean for $m$ measurements per subject:

$$\overline{{\mathit{x}}_i}=\frac{1}{m}\sum _{j=1}^m{\mathit{x}}_{ij}$$

where $i$ indexes participants, $j$ indexes repeated measurements, $m$ is the number of measurements per subject, and $n$ is the total number of participants. The reproducibility coefficient was summarized as the bootstrap mean ± SD, obtained from 5000 bootstrap samples to characterize its sampling variability. Statistical significance was defined as a two-sided $P$ value less than .05. S.S.P. performed the statistical analyses.

Results

Sample Characteristics

A subset of 1431 participants enrolled in the COPDGene study underwent full-dose and reduced-dose acquisitions between November 2014 and July 2017. Exclusion factors included lack of longitudinal scans $(n=198)$, unknown scanner model $(n=33)$, unknown modulation transfer function measurements $(n=3)$, and image quality issues $(n=38)$. Therefore, the sample evaluated in this study included 1159 participants (mean age ± SD, 65 years ± 9; 586 male participants, 573 female participants), self-identifying as non-Hispanic White (739 participants [63.8%]) or African American (420 participants [36.2%]) for inclusion in the COPDGene study. Participant demographic characteristics are presented in Table 1.

Table 1:

Demographic and Clinical Characteristics of the COPDGene Participants

Characteristic	COPDGene Participants (FD: $n=1159$; RD: $n=1159$)
Age (range) (y)	65 ± 9 (47–87)
BMI	29.3 ± 6.5
Sex
Male	586 (50.6)
Female	573 (49.4)
Race
African American	420 (36.2)
White	739 (63.8)
Visit GOLD stage
PRISm	178
GOLD 0	528
GOLD 1	111
GOLD 2	193
GOLD 3	81
GOLD 4	32
Unknown	36
Smoking history
Current	517 (44.6)
Former	625 (53.9)
Never	17 (1.5)
Imaging biomarkers
Perc15 (HU)
FD	−912.6 ± 30.0
RD	−926.0 ± 29.8
VALD (HU)
FD	87.3 ± 23.2
RD	73.1 ± 22.6
LAA-950 (%)
FD	8.1 ± 7.8
RD	9.3 ± 7.8
TLC (L)
FD	5.3 ± 1.4
RD	5.3 ± 1.4
Functional parameters
DLCO percentage predicted GLI (%)	81.3 ± 20.6
FEV1 percentage predicted (%)	81.3 ± 22.9
FEV1/FVC	0.70 ± 0.13

Note.—Data are presented as numbers, with percentages in parentheses, or as the means ± SDs. BMI = body mass index (calculated as weight in kilograms divided by height in meters squared), COPDGene = Genetic Epidemiology of Chronic Obstructive Pulmonary Disease, Dlco = diffusing capacity of the lungs for carbon monoxide, FD = full-dose, FEV₁ = forced expiratory volume in 1 second, FVC = forced vital capacity, GLI = Global Lung Function Initiative, GOLD = Global Initiative for Obstructive Lung Disease, LAA-950 = percentage of voxels in the lungs less than −950 HU, Perc15 = 15th percentile of the lung density histogram, PRISm = preserved ratio inspired spirometry, RD = reduced-dose, TLC = total lung capacity, VALD = volume-adjusted lung density.

Effect of Changes in Image Quality and Lung Volume on Lung Density

Spatial resolution

The spatial resolution differed between the soft and sharp kernels (f_50-soft, 0.35 mm⁻¹ ± 0.03; f_50-sharp, 0.54 mm⁻¹ ± 0.07; $P<.001$). This discrepancy resulted in a difference in Perc15 measurements between soft and sharp kernels (Perc15_soft, −910.1 HU ± 31.4; Perc15_sharp, −925.6 HU ± 29.2; $P<.001$). Box plots of MTF f₅₀ and Perc15 values are presented in Figure S6A–6B.

Noise

The noise level was different between the full-dose and reduced-dose scans (global noise index for full dose, 25.5 HU ± 7.3; global noise index for reduced dose, 57.0 HU ± 12.0 $(P<.001)$. This discrepancy resulted in a difference in Perc15 measurements between full-dose and reduced-dose scans (Perc15_FD, −926.0 HU ± 29.8, Perc15_RD, −912.6 HU ± 30.0; $P<.001$). Box plots of global noise index and Perc15 values are presented in Figure S6C–6D.

Lung volume

A comparison of lung volumes at total lung capacity between the full-dose and reduced-dose dose scans is shown in Figure S6E (reduced dose, 5287 mL ± 1405; full-dose, 5300 mL ± 1396; $P=.16$). The bias ± limits of the agreement was −12.0 mL ± 594.4.

Harmonizations

Spatial resolution

The influence of the spatial resolution matching module on the harmonization of paired reconstruction kernels from a clinical scan, demonstrated in Figure 2, resulted in improved consistency of the lung density histogram. Across all cases, spatial resolution matching improved consistency of lung density measurements, shown using Bland-Altman plots (Fig 3A–3B for Perc15 and Fig S7A–SB for LAA-950) and reproducibility metrics. The Perc15 reproducibility coefficient improved from 35.6 HU ± 0.7 to 18.6 HU ± 1.0. Measurements for Perc15 and LAA-950 imaging biomarkers are shown in Tables 2 and S4, respectively.

Figure 2:

CT images in a 79-year-old man without pre-existing emphysema who formerly smoked. CT scan reconstructed with (A) the sharp kernel and (B) the soft kernel and (C) the harmonized image obtained using the spatial resolution matching technique. Difference between the sharp and soft kernel reconstructed images (D) before and (E) after spatial resolution matching, with color bars demonstrating attenuation difference in Hounsfield units. (F) The line profile as shown by the yellow, blue, and green dotted lines on the sharp and soft kernel reconstructed images and the harmonized image, respectively. (G) The corresponding lung density histogram.

Figure 3:

Bland-Altman plot for the 15th percentile of the lung density histogram (Perc15) imaging biomarker (A) before harmonization; (B) after spatial resolution matching; (C) after spatial resolution and noise matching; and (D) after spatial resolution, noise, and volume adjustment. Bias and limits of agreement are shown as blue and red dotted lines, respectively. The Perc15 bias ± limits of agreement improved from −16.3 HU ± 24.8 to −7.7 HU ± 18.8 after spatial resolution matching; 1.5 HU ± 14.4 after spatial resolution and noise matching; and 1.3 HU ± 6.4 after spatial resolution, noise, and volume adjustment.

Table 2:

Perc15 Reproducibility Coefficients, Biases, Limits of Agreement, and $P\text{Value}$ for Each Comparison between the Input and the Reference Imaging Condition

				Bias ± LoA (HU)
SRM	NM	VA	Reproducibility Coefficient (HU)	Mean ± 1.96 SD	Median (2.5th, 97.5th Percentile)	$P\text{Value}$
No	No	No	35.6 ± 0.7	−16.3 ± 24.8	−14.0 (−49.0, −1.75)	<.001
Yes	No	No	18.6 ± 1.0	−7.7 ± 18.8	−6.0 (−28.0, 1.0)	<.001
Yes	Yes	No	16.4 ± 1.1	1.5 ± 14.4	0.0 (−9.0, 16.0)	<.001
Yes	Yes	Yes	7.4 ± 0.2	1.3 ± 6.4	0.0 (−3.0, 9.0)	<.001

Note.—“Yes” refers to “performed,” and “No” refers to “not performed.” Reproducibility coefficients are presented as means ± SDs. Metrics and the paired $t$ test were calculated between the input and the reference imaging condition, before harmonization and after each harmonization step. The paired $t$ test evaluated whether the mean difference (ie, bias) was significantly different from zero. The reference condition corresponds to full-dose and soft-kernel reconstructions, and the input includes all other imaging settings (ie, reduced-dose scans, sharp-kernel reconstructions, or both). LoA = limits of agreement (defined as the interval within which approximately 95% of the differences between the two quantifications [the input and the reference imaging condition]) are expected to fall, NM = noise matching, Perc15 = 15th percentile of the lung density histogram, SRM = spatial resolution matching, VA = lung volume adjustment per equation developed in the Multi-Ethnic Study of Atherosclerosis Lung Study.

Noise level

The influence of the noise matching module on the clinical CT scans of an individual with full-dose and reduced-dose acquisition protocols is shown in Figure 4. Across all cases, noise matching improved the consistency of lung density measurements, shown using Bland-Altman plots (Fig 3C for Perc15 and Fig S7C for LAA-950) and reproducibility metrics. The Perc15 reproducibility coefficient improved from 18.6 HU ± 1.0 to 16.4 HU ± 1.1.

Figure 4:

CT images in a 56-year-old man with pre-existing mild centrilobular emphysema at baseline who currently smoked. (A) CT scan imaged with a reduced-dose (RD) acquisition protocol and (B) full-dose (FD) acquisition protocol. (C) The harmonized image obtained with the noise-adaptive median filtering technique. (D) Corresponding lung density histogram.

Lung volume

Despite adjustment for image quality metrics, lung density histograms may still differ because of variations in lung volume across repeated acquisitions (Fig 5). Across all cases, this variation can be reflected especially in the right tail of the Perc15 Bland-Altman plots before lung volume adjustment (Figs 3A–3C and 6A), where larger differences between input and reference conditions are observed. Lung volume adjustment improved agreement between scans, resulting in more consistent lung density histograms (Fig 5; individual example), more stable lung density measurements (Fig 3D for Perc15 and Fig S7D for LAA-950; across all cases), and improved reproducibility metrics overall. The Perc15 reproducibility coefficient improved from 16.4 HU ± 1.1 to 7.4 HU ± 0.2 after application of the lung volume adjustment module.

Figure 5:

Lung density histograms in a 63-year-old woman across reduced-dose (RD) and full-dose (FD) acquisitions with different lung volumes (FD, 3631 mL; RD, 3213 mL). The importance of lung volume adjustment after noise matching when lung volume is inconsistent across scans is shown.

Figure 6:

Bland-Altman plots for the 15th percentile of the lung density histogram (Perc15) imaging biomarker (A) before harmonization, (B) with volume-adjusted lung density (VALD), (C) median filtering and VALD (MF-VALD), and (D) noise-adaptive median filtering and VALD (NAMF-VALD). Bias and limits of agreement are shown as blue and red dotted lines, respectively. The Perc15 bias ± limits of agreement improved from −13.3 HU ± 19.3 to −14.4 HU ± 10.6 using VALD, 3.4 HU ± 7.8 using MF-VALD, and 2.4 HU ± 6.1 using NAMF-VALD.

Overall, the Perc15 bias ± limits of agreement improved from −16.3 HU ± 24.8 to 1.3 HU ± 6.4 after harmonization.

Comparison against existing approaches

A comparison of the proposed denoising technique (ie, NAMF) against established denoising filters across a range of their parameters is shown in Table S5. The results demonstrated more accurate performance of the proposed denoising technique.

Bland-Altman plots for the Perc15 imaging biomarker compared the proposed denoising technique (NAMF) followed by volume adjustment (NAMF-VALD), against VALD and median filtering followed by VALD (MF-VALD), as shown in Figure 6. The reproducibility coefficient improved from 32.4 HU ± 0.7 before harmonization to 7.7 HU ± 0.2 when the VALD-NAMF approach was used, a 3.9-fold improvement over the VALD technique (to 30.1 HU ± 0.3) and a 1.3-fold improvement over the MF-VALD technique (to 10.3 HU ± 0.3). Results for the Perc15 imaging biomarker are shown in Table 3.

Table 3:

Perc15 Reproducibility Coefficients, Biases, Limits of Agreement, and $P\text{Value}$ for Each Comparison between Reduced-Dose and Full-Dose Scans Reconstructed with the Soft Kernel

NM	VA	Reproducibility Coefficient (HU)	Bias ± LoA (HU)
			Mean ± 1.96 SD	Median (2.5, 97.5 Percentile)	$P\text{Value}$
No	No	32.4 ± 0.7	−13.3 ± 19.3	−13.0 (−31.0, 3.15)	<.001
No	Yes	30.1 ± 0.3	−14.4 ± 10.6	−14.0 (−26.0, −5.0)	<.001
Yes (median filter)	Yes	10.3 ± 0.3	3.4 ± 7.8	3.0 (−4.0, 11.0)	<.001
Yes (NAMF)	Yes	7.7 ± 0.2	2.4 ± 6.1	2.0 (−4.0, 9.0)	<.001

Note.—“Yes” refers to “performed,” and “No” refers to “not performed.” Reproducibility coefficients are presented as means ± SDs. Metrics and the paired $t$ test were calculated for each comparison between the input and the reference imaging conditions. The paired $t$ test evaluated whether the mean difference (ie, bias) was significantly different from zero. The reference and input conditions correspond to full-dose and reduced-dose scans, both reconstructed with the soft kernel. The reconstruction kernel was held constant to ensure a fair comparison with other approaches that cannot accommodate variations in kernel reconstruction. LoA = limits of agreement (defined as the interval within which approximately 95% of the differences between the two quantifications [input and the reference imaging condition]) are expected to fall, NAMF = noise-adaptive median filtering, NM = noise matching, Perc15 = 15th percentile of the lung density histogram, VA = lung volume adjustment per equation developed in the Multi-Ethnic Study of Atherosclerosis Lung Study.

Discussion

Robust CT-based lung density quantification remains a major challenge, particularly given the variability in scanners, reconstruction settings, radiation dose levels, and lung volumes. This variability creates a critical need for reliable harmonization approaches. Previous image-based approaches have largely addressed these issues individually without providing a comprehensive solution. The purpose of this study was to build on previous techniques and develop a physics-based image harmonization method that transforms images into a reference quality index of noise, spatial resolution, and lung volume and to evaluate its performance for improving lung density measurement reproducibility.

We focused on clinical data from the COPDGene dataset, acquired with the use of full-dose and reduced-dose protocols and reconstructed with soft and sharp WFBP reconstruction kernels. The proposed harmonization framework improved the Perc15 reproducibility coefficient from 35.6 HU ± 0.7 to 7.4 HU ± 0.2 under changes in dose and reconstruction kernel settings, a 4.8-fold improvement. The harmonizer outperformed the VALD technique by 3.9-fold and the MF-VALD technique by 1.3-fold across full-dose and reduced-dose scans reconstructed with the soft kernel, improving the reproducibility coefficient from 30.1 HU ± 0.3 (VALD) and 10.3 HU ± 0.3 (MF-VALD) to 7.7 HU ± 0.2. This highlights the superior performance of our approach when reconstruction kernels remain unchanged; moreover, our approach can address variations in kernel settings if needed, unlike VALD and MF-VALD. This performance improvement is clinically important because reducing the reproducibility coefficient to clinically acceptable thresholds is essential for detecting true disease progression, particularly in light of reported lung density declines of −7.0 g/L (−7.0 HU) and −4.5 g/L (−4.5 HU) over successive 5-year intervals in participants from the COPDGene cohort with pre-existing emphysema who currently smoked (15).

The proposed CT-based harmonization approach is potentially important for two main reasons. First, a more robust CT biomarker of emphysema could increase statistical power in cross-sectional multicenter epidemiologic studies that evaluate genetic and environmental contributors to emphysema (32,33). Second, the current lack of a sensitive biomarker is a major disincentive to pharmaceutical manufacturers wishing to embark on clinical trials of novel treatments for emphysema. Upon subsequent validation of our harmonization method in longitudinal cohorts, a harmonized metric of CT number may improve the sensitivity for longitudinal change, thereby increasing statistical power and potentially decreasing trial duration (34). In our study, we selected emphysema biomarkers (ie, Perc15 and LAA-950) primarily as an illustrative example, given their established role in the COPDGene dataset. However, the proposed harmonization algorithm is not restricted to emphysema and can be extended to other lung density–based biomarkers, highlighting the broader potential impact of our method on lung imaging research.

Our study had some limitations. First, we did not incorporate a bias-correction module in the harmonization approach, which may limit applicability to longitudinal scans on the same scanner. Accordingly, we did not evaluate discrimination of clinical trajectories—progressors versus nonprogressors—or comparisons with existing analytical methods (16). Future work may incorporate image-based bias estimation and longitudinal analyses to extend evaluation beyond Perc15 and LAA-950 and support regional progression mapping when combined with registration. Second, nonimaging metrics, such as pulmonary function measures, were not assessed because this was outside the primary scope of the study; integrating them in longitudinal studies may provide insights on cross-modality relationships. Third, we focused on WFBP reconstructions, consistent with standard COPDGene protocols, and did not investigate modern iterative or deep learning–based reconstructions, which vary across vendors and introduce nonlinear noise and bias characteristics. Nonetheless, the generalizability of the proposed framework was demonstrated across different kernels, and the framework could be extended to evolving CT technologies (eg, photon-counting CT) through corresponding MTF measurements. Future work may investigate the generalizability of the harmonizer to modern reconstruction techniques through methods that address CT noise and bias non-uniformity or MTF curves that capture image nonlinearities. Last, the cohort was derived from a COPDGene substudy in which full-dose and reduced-dose scans were obtained during a single visit, which may introduce selection bias. Additional studies using larger and more diverse populations are warranted to assess generalizability.

In conclusion, our image harmonization pipeline reduced variations in lung density measurements. Compared with the traditional approach of MF-VALD, the proposed technique, tested on COPDGene study cases, demonstrated improved performance. Robust quantifications, enabled by such harmonization techniques, may enable more precise assessments of lung disease severity. Future research may be focused on the longitudinal utility of such a harmonization framework for providing local assessments of disease progression.

Supplementary Material

Supplemental material: Supplemental material is available at Radiology: Cardiothoracic Imaging online.

Supplemental material is available for this article.

Key Points.

• In a retrospective analysis of 1159 Genetic Epidemiology of Chronic Obstructive Pulmonary Disease (COPDGene) study participants who underwent same-day chest CT at full and reduced doses, harmonization improved the 15th percentile of the lung density histogram (Perc15) reproducibility coefficient 4.8-fold, from 35.6 HU ± 0.7 to 7.4 HU ± 0.2.

• The harmonizer outperformed volume-adjusted lung density (VALD) and median filtering followed by VALD 3.9-fold and 1.3-fold, improving the Perc15 reproducibility coefficient from 30.1 HU ± 0.3 and 10.3 HU ± 0.3, respectively, to 7.7 HU ± 0.2.

Abbreviations

COPDGene

Genetic Epidemiology of Chronic Obstructive Pulmonary Disease

f₅₀

spatial frequency corresponding to the 50% magnitude

LAA-950

percentage of voxels in the lungs less than −950 HU

MESA

Multi-Ethnic Study of Atherosclerosis Lung Study

MTF

modulation transfer function

NAMF

noise-adaptive median filtering

Perc15

15th percentile of the lung density histogram

VALD

volume-adjusted lung density

WFBP

weighted filtered back-projection

Data sharing:

Data generated or analyzed during the study are available from the corresponding author by request.