Introduction
Chronic obstructive pulmonary disease (COPD) is a chronic inflammatory lung disease characterized by irreversible airflow obstruction and respiratory symptoms [1]. The clinical manifestations and natural progression of this disease can vary significantly across different phenotypes [2,3]. Emphysema, a major phenotype of COPD, is associated with significant clinical outcomes, including rapid lung function decline and increased mortality [4,5]. Consequently, extensive efforts have been made to analyze emphysema characteristics in COPD. As chest computed tomography (CT) represents the gold standard for diagnosing emphysema, radiological approaches have been actively pursued [6]. CT quantification, a critical effort in this context, is essential for elucidating COPD features and serves as a foundation for COPD-related radiomics [7]. This includes evaluating air trapping, airway branching, and small airway diameter. Among these, emphysema quantification based on low-attenuation areas (LAA) in CT, first introduced in 1988, is now commonly defined as areas below -950 hounsfield unit (HU) [8]. Quantitative measurements of emphysema have been reported to correlate with frequent exacerbations and poor clinical outcomes [9,10].
The Korea COPD Subgroup Study (KOCOSS) is a nationwide, prospective, observational, multicenter cohort study conducted throughout South Korea [11]. Since 2012, the cohort has collected chest CT scans from participating hospitals to assess the severity of emphysema. A major challenge in CT quantification, however, arises from the variability in CT equipment and protocols across different institutions. Hospitals employ scanners from various manufacturers, each with distinct settings, leading to differences in slice thickness and reconstruction kernels (sharp vs. smooth). This variability creates heterogeneity in imaging data, potentially hindering research and compromising the reliability of findings.
To address this issue, kernel conversion proves essential. Sharp kernels enhance fine details but also increase noise, whereas smooth kernels lessen noise but tend to blur details. These effects can significantly influence the quantification of emphysema by altering the visibility of LAA, a crucial factor in its evaluation. Kernel conversion helps minimize these discrepancies, enabling consistent and reliable analyses of imaging data across various centers, thus enhancing the accuracy of research outcomes.
In this study, our objective was to evaluate the impact of kernel conversion on enhancing the correlation between the emphysema index and clinical parameters in COPD.
Materials and Methods
1. Study population and dataset
The KOCOSS study encompassed participants aged ≥40 years with a physician-diagnosed COPD, exhibiting respiratory symptoms and a ratio of forced expiratory volume in 1 second (FEV1) to forced vital capacity (FVC) <0.7 post-bronchodilation. Exclusion criteria encompassed current asthma, recent major cardiovascular events, pregnancy, malignancies, other inflammatory diseases, and chronic steroid use for conditions other than COPD exacerbation. Of the 631 participants, 484 were eligible after excluding those with unsuitable slice thickness (n=133) or failure to meet the quantification criteria (n=14) (Figure 1).
2. Kernel conversion of computed tomography images
CT scans reconstructed with a sharp kernel were transformed to a smooth kernel using commercially available deep learning-based software, previously trained for automated kernel conversion (Aview COPD, Coreline Soft, Seoul, Korea). The software was trained using data from four manufacturers—Siemens (Munich, Germany), Philips (Amsterdam, the Netherlands), GE Medical Systems (Waukesha, WI, USA), and Toshiba (Tokyo, Japan)—each providing paired soft and sharp kernels. In the kernel transformation, a 2.5-D U-Net structure was employed, wherein the channel of each layer corresponds to the z-axis of the image in a U-Net with 2-D convolution layers [12,13]. The percentage of low-attenuation areas (%LAA-950) for the whole lung or specific regions was calculated and compared both before and after kernel conversion.
3. Study variables and outcomes
Baseline data included age, sex, body mass index, smoking status, and underlying conditions such as hypertension, asthma, diabetes, gastroesophageal reflux disease, myocardial infarction, osteoporosis, and heart failure. Patient-reported outcomes included the modified Medical Research Council (mMRC) score, COPD assessment test (CAT) scores, and St. George’s Respiratory Questionnaire for COPD (SGRQ-c) scores, as well as pulmonary function measurements such as FEV1, FVC, diffusion capacity of carbon monoxide (DLco), and the ratio of residual volume (RV)/total lung capacity (TLC), and the 6-minute walking distance. The primary outcome of this study was to compare the correlation between emphysema extent and the aforementioned clinical parameters before and after kernel conversion.
4. Statistical analysis
The t-test was utilized for continuous variables and the chi-square test for categorical variables. To evaluate the distribution of %LAA-950 before and after conversion, we employed Python's Seaborn library to create a histogram with a kernel density estimate line, indicating the mean value. Correlation analyses were conducted to compare the correlation between the emphysema index (%LAA-950) and other clinical variables before and after conversion. A two-sided p-value less than 0.05 was considered statistically significant. All statistical analyses were conducted using R version 4.0.3 (R Core Team 2020, R Foundation for Statistical Computing, Vienna, Austria).
5. Ethical considerations
This study complied with the Helsinki Declaration. Our study protocol was approved by the Ethics Committee of Konkuk University Medical Center (IRB number: 2012-03-015) and the Ethics Committee of each participating medical center. All data were provided anonymously, and all participants gave written informed consent prior to enrollment.
Results
1. Demographics of study population
Baseline characteristics of the study population are presented in Table 1. Of the 484 participants, 91.7% were male, with a mean age of 67.8±8.1 years. Current smokers comprised 21.3%, ex-smokers 64.9%, and never-smokers 13.8%. The mean values of mMRC, CAT, and SGRQ-c were 1.2± 0.9, 12.2±7.8, and 25.6±19.6, respectively. The mean FEV1 was 1.9±0.7 L (63.6%±19.9%), and the mean FVC was 3.5±0.9 L (83.7%±16.4%). Hypertension (38.6%), asthma (21.5%), and diabetes mellitus (17.2%) were the prevalent comorbidities. The distribution of COPD stages was as follows: stage 1 (21.1%), stage 2 (53.9%), stage 3 (21.5%), and stage 4 (3.5%).
2. Comparisons of the emphysema index between original and converted CT images
Figure 2 illustrates the histogram distribution, exhibiting skewness and kurtosis in both the original and converted datasets. Both skewness and kurtosis were mitigated after kernel conversion (original vs. converted: skewness 0.99 vs. 1.90; kurtosis 0.16 vs. 4.01). The mean %LAA-950 exhibited a significant difference between the original and converted datasets, with the latter demonstrating improved data variance (whole lung [12.6±11.0 vs. 8.8±11.9], right lung [12.9±11.6 vs. 9.1±12.6], and left lung [12.3±11.0 vs. 8.5±12.0], all p<0.01) (Table 2).
3. Correlation of the emphysema index in original and converted CT datasets with COPD parameters
Table 3 demonstrated the correlation analyses between %LAA-950 and COPD variables. FEV1 (%) and %LAA-950 exhibited a moderate negative correlation in both the original and converted datasets, which improved after kernel conversion (r=-0.35 in the original version vs. r=-0.41 in the converted version). In the case of RV/TLC, there was a moderate positive correlation, which was enhanced after kernel conversion (r=0.31 vs. 0.42). The correlations of mMRC, CAT, and SGRQ-c scores with LAA(%) in both datasets demonstrated weak positive correlations (mMRC: r=0.22 vs. 0.25, CAT: r=0.05 vs. 0.12, SGRQ-c: r=0.13 vs. 0.21). 6WMD exhibited a higher correlation with %LAA-950 in the converted dataset (r=-0.04 vs. 0.15). A moderate negative correlation was observed between DLco (%) and LAA (%) (r=-0.44 vs. -0.41). Subgroup analysis was conducted based on different CT devices (Supplementary Table S1). For images from the GE CT device, both pre- and post-kernel conversion values demonstrated statistically significant correlations with the emphysema index (all p<0.05). For the Siemens device, FEV1 (%), DLco (%), and mMRC showed statistically significant correlations with the emphysema index both before and after kernel conversion, with higher correlation coefficients observed. RV/TLC reached statistical significance only after kernel conversion. For Philips, DLco showed no difference in correlation between pre- and post-conversion values, while FEV1 (%) and RV/TLC exhibited statistically significant improvements in correlation after conversion.
Discussion
Efforts to address heterogeneity in CT devices and protocols have focused on varied strategies aimed at ensuring consistency. A widely explored method is kernel conversion, which employs specialized algorithms to standardize image texture and contrast across different reconstruction kernels [14,15]. Other approaches include harmonizing acquisition parameters (e.g., slice thickness, tube voltage, and current) and implementing standardized reconstruction techniques [16,17]. Such efforts are particularly crucial in multicenter studies where variations in CT protocols can introduce bias and affect the reliability of findings.
In CT imaging, sharp kernels and smooth kernels fulfill crucial roles in image reconstruction and interpretation [18]. Sharp kernels are designed to enhance edge definition, making them ideal for viewing structures with high contrast, such as bones and airways. These kernels excel at enhancing fine details but often escalate noise, which can obscure low-contrast areas, such as those affected by emphysema. Conversely, smooth kernels prioritize noise reduction and are more suitable for visualizing low-contrast structures. Through balancing image clarity with noise, smooth kernels are effective for detecting subtle pathologies in the lung, such as emphysema, proving indispensable for reducing variability in quantitative CT analysis [14].
We are the first to demonstrate the impact of CT kernel conversion on data from a South Korean COPD cohort, utilizing previously validated AI-based kernel conversion software [14,19]. In this study, we achieved an improved correlation of emphysema extent from different CT scanners across participating hospitals and associated COPD parameters. The emphysema index from the kernel-converted images revealed more significant findings compared to the original version.
It is well established that the reconstruction kernel significantly influences measurement variability in quantitative CT analysis of emphysema [12,20]. Smooth kernels are recognized for providing the most precise emphysema quantification, whereas sharp kernels are likely to overestimate the LAA-950 [19,21]. Consequently, for images already obtained with a high-frequency kernel, additional kernel conversion is necessary to assess the extent of emphysema accurately. Recent studies have investigated kernel conversion using deep learning techniques to address these issues [12,22-28]. These studies have shown that deep learning-based kernel conversion can enhance emphysema quantification across diverse CT protocols.
We demonstrated intriguing aspects of kernel conversion in emphysema quantification among COPD patients. Initially, we established uniform emphysema data values in the COPD cohort through kernel conversion. Owing to the nature of the multicenter prospective cohort study, heterogeneous CT image slices from various protocols were inevitably included. We successfully created a uniform, enhanced, and validated image database using kernel conversion. Additionally, we demonstrated that the converted values were consistent and exhibited a reduction in variance. Subsequently, we showed that values obtained through kernel conversion correlate more strongly with COPD parameters than unconverted images, suggesting that kernel conversion is more suitable for elucidating COPD parameters. Based on these findings, we foresee diverse applications in analyzing the relationship between radiologic biomarkers and COPD prognosis in the future. Finally, we observed that the correlation between kernel-converted values and COPD parameters varied by CT device. While some manufacturers showed minimal differences between pre- and post-conversion values, others displayed significant changes. This variation likely results from the differential use of sharp or smooth kernels across CT devices. Therefore, post-conversion image data are recommended for COPD research to ensure consistency across different devices.
This study had several limitations. First, emphysema quantification was performed using a single deep learning software. Further studies comparing more auto-segmentation methods could be beneficial, although the current method is widely accepted and has been validated in previous studies. Second, this study only analyzed cross-sectional data. Future research is necessary to evaluate the long-term prognosis and the influence of kernel conversion on the clinical applicability of emphysema quantification for COPD patients.
In conclusion, kernel conversion reduces the variability of CT images across different protocols and enhances the correlation with various clinical parameters in COPD patients from multiple centers. These findings suggest the potential for broader applications in establishing links between radiologic biomarkers and COPD prognosis.
PDF Links
PubReader
ePub Link
XML Download
Data Sharing Statement
Full text via DOI
Supplement
Print
Download Citation
