The COPDGene Study is a multi-center observational investigation designed to analyse the genetic and epidemiologic factors associated with COPD . This study recruited 10,364 subjects from 21 institutions, with inclusion and exclusion criteria as described in . All subjects underwent spirometry and CT imaging of the chest at full inspiration (TLC) and relaxed expiration (FRC) . This research protocol has obtained institutional review board approval at every site and written informed consent was provided by all enrolees. For this work, the COPDGene study provided 366 scans from 183 subjects, obtained at one institution with two Siemens scanners, and 372 scans from 186 subjects, acquired at six institutions with three GE scanners. Characteristics of the patients included in our study are provided in Tables 1 and 2.
Whole-lung volumetric multi-detector CT was obtained with scanners from two different manufacturers: Siemens and GE. All scans were reconstructed applying filtered back projection (FBP) with two kernels: a standard one and a sharp one. For both manufacturers, CT scans were acquired at full inspiration without contrast medium at 120 KVp tube energy and 200 mAs effective dose. The reconstruction field-of-view was configured per patient to encompass the widest diameter of the lungs.
In the Siemens group, a Siemens Definition scanner (n = 122, 64 × 0.6 mm detector configuration, 1.1 pitch) or a Siemens Definition AS+ scanner (n = 61, 128 × 0.6 mm detector configuration, 1.0 pitch) were used. All images were reconstructed with two kernels: b31f (standard) and b45f (sharp), using a reconstruction field-of-view ranging from 260 to 410 mm and a 512 × 512 matrix, yielding a pixel size between 0.5 and 0.8 mm. The slice thickness used was 0.75 mm with an interval of 0.5 mm.
In the GE group, scans were taken using a GE LightSpeed 16 scanner (n = 105, 16 × 0.625 mm detector configuration, 1.375 pitch), a GE LightSpeed VCT scanner (n = 71, 16 × 0.625 mm detector configuration, 1.375 pitch) or a GE LightSpeed Pro 16 scanner (n = 10, 16 × 0.625 mm detector configuration, 1.375 pitch). Scans were reconstructed with STANDARD (standard) and BONE (sharp) kernels. All subjects underwent CT imaging with 0.625 mm slice thickness with an interval of 0.625 mm. The reconstruction field-of-view ranged from 260 to 500 mm and a 512 × 512 matrix, resulting in a pixel size between 0.5 and 1 mm.
Six virtual cohorts were constructed for analysis: four cohorts using scans reconstructed with a single kernel (b31f Siemens (n = 183), b45f Siemens (n = 183), STANDARD GE (n = 186), BONE GE (n = 186)) and two mixed cohorts. The single kernel cohorts allow direct comparison between standard and sharp kernels. They were used to illustrate the variation in emphysema scoring due to the reconstruction kernels, and to demonstrate that normalization reduces this variation while maintaining correlations with lung function parameters. The first mixed cohort contained all scans reconstructed with the standard kernel on the Siemens and GE scanners (n = 369). This mixed cohort represents a typical situation in multi-center studies in which different sites have different scanners and the most similar reconstruction settings are chosen. The second mixed cohort (n = 369) was constructed to contain scans with both standard and sharp kernels from both manufacturers: from the Siemens group, the b31f reconstruction was chosen for 92 randomly selected subjects, and the b45f reconstruction for the remaining 91 subjects. From the GE group, the STANDARD kernel was selected for 93 randomly selected subjects and the BONE kernel for the remaining 93 subjects. This mixed cohort simulates a possible situation in which scans from different studies are combined, or scans are retrospectively collected from different centers with different acquisition protocols.
Pulmonary function tests
Spirometry was performed using an EasyOne spirometer (ndd Medical Technologies, Andover, MA). The mean time between CT imaging and spirometry was 10.45 days (range 0–98).
Quantification of emphysema was performed using CIRRUS Lung 13.08 (http://cirrus.diagnijmegen.nl, Diagnostic Image Analysis Group, Nijmegen, The Netherlands; Fraunhofer MEVIS, Bremen, Germany). As a first step, the lungs were automatically segmented using methods based on region growing and morphological operations . All segmentations were visually checked and corrected when needed in CIRRUS Lung. The percentage of lung affected by emphysema was quantified in terms of ES, using -950 HU as an attenuation threshold.
To rule out possible variations in ES due to segmentation differences in data reconstructed with different filter kernels, we segmented the lungs in the image reconstructed with the standard kernel, and used this segmentation for both reconstructions.
To reduce the variability in measured ES, the proposed method changes the appearance of data obtained with various reconstruction kernels so that it will have similar characteristics as a chosen reference reconstruction. The rationale behind the proposed normalization is the fact that filter kernels affect spatial resolution and image noise of the reconstructed data: sharper reconstruction kernels will preserve higher spatial frequencies but increase image noise .
The proposed normalization decomposes the scan into frequency bands and alters the energy in each band according to the average energies observed in the set of scans reconstructed with a reference kernel. This will reduce image noise to a similar level. In this study, we selected Siemens b31f as the reference kernel. For a detailed description of the algorithm, see Appendix 1.
Statistical analysis was performed using statistical software (IBM SPSS Statistics, Release 20.0; IBM Corp, Armonk, NY). Medians and inter-quartile ranges were computed for the non-normally distributed ES. Differences in ES for two reconstructions of the same scan were computed as ES in scans reconstructed with the sharp kernel minus ES in scans reconstructed with the standard kernel. Mean and standard deviations were computed for the normally distributed differences. The Bland-Altman approach  was used for analysis, with 95 % confidence intervals as limits of agreement.
Spearman correlation coefficients were calculated to evaluate the correlation between ES and forced expiratory volume in 1 second (FEV1) and FEV1-to-forced vital capacity (FVC) ratio. The statistical significance of the difference between correlation values was evaluated using the statistical test described by Steiger , using the R statistical analysis package (R Foundation for Statistical Computing, Vienna, Austria (URL http://www.R-project.org/).