1. Introduction
Chronic obstructive pulmonary disease (COPD) is a complex lung disease that is characterized by the destruction of lung tissue and persistent airflow obstruction [
1]. COPD causes breathing difficulties resulting in pulmonary emphysema, which can contribute to the development of additional respiratory complications [
2]. These findings highlight the critical importance of early diagnosis for COPD. Lung computed tomography (CT) analysis is a key tool for assessing lung conditions and diagnosing COPD. This diagnostic approach leverages the contrast in CT numbers between air and lung tissues, distinguishing air spaces from damaged lung tissue [
3,
4,
5]. By calculating the ratio of functional lung tissue to air spaces, lung CT image analysis helps diagnose COPD and evaluate its severity. The Lung Density Committee of the Quantitative Imaging Biomarkers Alliance (QIBA) has established official CT number thresholds to differentiate lung tissue from air spaces for diagnosing COPD [
3]. However, the technique has limitations that restrict its ability to accurately demonstrate air-tissue contrast and produce reliable diagnostic results. A primary challenge is the inconsistency of lung density measurements, which are significantly influenced by the CT scanner manufacturer and the scanning protocols employed. In lung CT image analysis studies, images are often acquired using varying dose levels and reconstruction kernels, depending on the patient’s condition and the available scanners [
4]. These variations affect the stability of lung tissue visibility and CT number measurements, thereby reducing the diagnostic accuracy of lung CT images.
In recent years, several approaches have been proposed to address the limitation of lung CT analysis, aiming to mitigate the effects of lung density variations caused by different acquisition protocols and increase CT number consistencies. Several studies have employed the ComBat CT number standardization algorithm, which evaluates X-ray dose and acquisition protocols to scale the CT numbers based on various statistical models [
6,
7]. However, the ComBat technique requires parameter analysis prior to the CT number standardization, and the Lung Density Committee provides the scaling information for limited CT scanners and scanning protocols. Chen-Mayer et al. introduced lung CT density standardization method based on the Martinez’s CT number-electron density relation [
8]. The method employs the COPD gene2 phantom, which contains various foams within the lung density range, as references (3 foams, air, and water) for CT number standardization. The study demonstrated the effectiveness of the method using 22 CT images acquired with different scanning protocols across four CT scanners. However, using water as a reference decreases the reliability of the standardization process. Water’s density and corresponding CT number can fluctuate due to the external influences, including bubble formation induced by CT scanner vibrations, precipitate accumulation from corrosion within the containing cylinder, and other environmental factors such as pressure and temperature. As illustrated in
Figure 1 and
Table 1, the average CT number of water, which is expected to be 0 Hounsfield Unit (HU) shows inconsistency even when the same scanner and scanning protocol are used.
In this study, we alternatively used a solid material instead of the water as references to improve the performance and reliability of the CT number standardization method. To assess the performance of the proposed method and applicability across various CT scanners and scanning protocols, we compared the standardization performance of the proposed method with the original approach. To explore the standardization performance for broader applicability, we used 234 CT images of the COPD gene2 phantom acquired from various CT scanners and scanning protocols, and analyzed the impact of the scanning parameters, such as dose level, tube potential and reconstruction algorithm.
3. Results
To assess the performance of the proposed standardization method and investigate the effect of each protocol on the standardization performance, we calculated the mean and standard deviation of the CT numbers across various scanners and protocols, both before and after standardization. We labeled the stages as “Raw” (before standardization), “Internal Cal.” (after internal calibration), and “Standardization” (after monochromatic energy mapping) in this section. A reduction in the standard deviation of the CT numbers reflects a reduction in CT number variation and the effectiveness of the standardization process. To visualize CT-number distributions and compare variations at each step of standardization, box plots of CT numbers are provided using the reference foams and the lung foam.
To evaluate the overall effectiveness of the proposed standardization method and the impact of using water as a reference, we statistically analyzed the CT numbers of reference foams 2 and 3, as well as the lung foam, across all 234 datasets.
As shown in
Table 4, using water as one of the standardization references increases the standard deviation of the reference foams and the lung foam during the internal calibration step. This might be caused due to the CT number uncertainty of the water as a standardization reference. Since water serves as the calibration basis during the internal calibration, the CT number fluctuation of water propagates to other reference foams and the lung foam during the calibration. On the other hand, alternating the reference foam 1 as a basis prevents the error during the internal calibration for all foams, as demonstrated in
Table 5. The standard deviation of CT numbers for each foam decreases at each step, demonstrating the effectiveness of the proposed method. Additionally, narrowing the standardization density range from [0, 1000] kg/m
3 to [0, 321.0] kg/m
3 improves the validity of the linearity assumption between electron density and CT number, resulting in enhanced performance, as illustrated in
Figure 3.
Although the overall standardization analysis confirmed the effectiveness of the proposed method, the dose level during the CT scan has a significant impact on its performance. As shown in
Table 6, the standardization results for each dose level indicate that performance varies with the noise level in the CT images. While the mean CT number at each dose level remains unaffected, the standard deviation tends to increase under low-dose conditions. At 1.0 mGy, the elevated noise level amplifies CT number variations for the foams in the COPD gene2 phantom, diminishing their reliability as reference materials.
Table 7 provides an analysis of standardization performance across different reconstruction algorithms and kernels. Sharp kernels in FBP enhance tissue contrast but also increase noise levels, leading to amplified CT number variation for the references and reduced standardization performance. As shown in
Table 7, FBP with sharp kernels exhibited the highest standard deviation and the lowest reduction rate after standardization due to the noise enhancement. In contrast, IR demonstrated superior performance in CT number standardization. By reducing image noise during reconstruction, IR achieves a similar effect to increased scanning doses, minimizing image noise and CT number variation in the reference foams. Notably, at lower dose levels, CT images reconstructed with IR showed better standardization performance than those reconstructed with FBP, largely due to IR’s effective noise reduction capabilities. However, IR compromises image visibility by excessively smoothing low-level features and reducing the standardization performance. Demonstrated in
Table 7, IR with 40% regularization strength reduced the CT number standard deviation from 1.56 to 0.31, whereas 100% regularization achieved a more modest reduction from 1.47 to 0.46.
To further investigate the impact of tube potential on CT number standardization performance, we categorized the data by tube potential and standardized the CT numbers of the lung foam. As shown in
Table 8, higher tube potentials resulted in lower standard deviations of the CT numbers for the lung foam, indicating improved standardization performance. Increasing the tube potential reduces CT number variation by enhancing the signal-to-noise ratio in CT images and minimizing photon starvation as the X-rays pass through the scanned object.
However, increasing the tube potential does not always lead to better standardization performance. As shown in
Table 9, at a dose level of 1.0 mGy, the reduction in CT number standard deviation is greater at 80 kVp compared to 100 kVp. Similarly, at 1.5 mGy, 100 kVp demonstrated the best performance for CT number standardization, reducing the standard deviation to 0.74 from 2.22. This variation occurs because tube potential and tube current are inversely proportional at a fixed dose level. Increasing the tube potential decreases the X-ray flux from the source, while low tube potential can result in radiation saturation due to insufficient photon energy. On the other hand, low tube current increases noise levels due to a reduced number of photons.
To optimize CT number standardization performance, maintaining a balance between tube potential and current is critical. This balance will vary, depending on the total CTDI used for scanning and patient-specific characteristics, such as size, density, and tissue composition.
4. Discussion
CT numbers can vary, depending on the scanner and scanning protocol, making it essential to follow a specific protocol to obtain reliable CT images for COPD diagnosis [
20,
21,
22]. However, this requirement limits the ability to establish universally accepted CT number thresholds for COPD. To overcome this challenge, Chen-Mayer introduced a standardization method based on scanner-dependent parameters to reduce variations in CT lung density measurements. Unlike traditional diagnostic approaches, this method allows for the establishment of protocol-independent thresholds for COPD diagnosis.
In this study, we expanded the analysis to include a wider range of scanners and protocols by examining 234 CT images of the COPD gene2 phantom. This broader evaluation enabled us to assess the effectiveness of the standardization method across diverse scanners and scanning protocols, including variations in CTDI, tube potential, and reconstruction algorithm settings. Building on Chen-Mayer’s CT number standardization method, we excluded water from the reference list to narrow the effective density range and eliminate the variability associated with liquid materials. This modification improved the reliability of the reference materials and reduced CT number variation in the images, leading to enhanced standardization performance. As a result, the standard deviation of CT numbers for the lung foam decreased to 0.97 from 2.12, demonstrating consistent performance across a broad range of scanning protocols.
In conventional studies and clinical workflows, water is frequently used as a reference material for imaging calibration and CT number standardization. However, our findings indicate that the CT number of water is prone to variability due to several factors, including bubble formation induced by CT scanner vibrations, precipitate formation from corrosion within the water-containing cylinder, and other environmental effects. To improve the robustness of the standardization process, we substituted water with foam 1, a solid reference material included in the COPD gene2 phantom.
The use of foam 1 not only mitigated CT number fluctuations but also enhanced the consistency and accuracy of the standardization by constraining the calibration density range to [321.0 1000] kg/m3. This narrower range helps minimize approximation errors associated with the parameter simplification in the Martinez formula, which models the relationship between CT number and electron density through a single attenuation parameter, . While reducing the density range improves estimation precision, the performance gain eventually saturates, and the calibration range must still encompass the density of the target material. Therefore, foam 1 was chosen as a practical and effective reference material for this study.
Additionally, the effects of the dose level, tube potential and reconstruction kernel on the performance of the standardization procedure are analyzed as follows:
(1) CT number variation increases at low dose levels due to elevated image noise, which negatively impacts the performance of the standardization process under such conditions. However, applying noise reduction techniques can mitigate this performance degradation. For instance, the noise-suppressing capabilities of IR improved standardization performance, even at low dose levels.
(2) In general, increasing the tube potential reduces CT number variation by enhancing the signal-to-noise ratio in CT images and minimizing photon starvation as X-rays pass through the patient. However, at a fixed dose level, higher tube potential decreases the X-ray flux, which can impact image contrast and noise levels. CT numbers also vary, depending on the tube potential of the X-ray source. In the diagnosis of COPD using CT lung density measurements, the choice of tube potential varies based on dose level and patient-specific thoracic characteristics, such as size, density, and tissue composition. Based on the results, it is crucial to strike a balance between tube potential and X-ray current to optimize the performance of CT number standardization.
(3) CT image sharpness and noise levels vary, depending on the reconstruction techniques used. FBP with a sharp kernel exhibited notable degradation in standardization performance due to increased image noise accompanying contrast enhancement. At low dose levels, CT images reconstructed with IR demonstrated better standardization performance than those reconstructed with FBP, primarily because of IR’s effective noise reduction capabilities. However, excessive regularization strength in IR can compromise image visibility in clinical CT images, potentially reducing diagnostic accuracy.
To implement the proposed method to clinical CT images, several validations must be considered. We found the strong dependency of the standardization performance on CT noise levels, and the noise-reducing characteristics of IR increase the performance of the proposed method. Given the higher image noise from low-dose scanning in clinical environments, noise reduction techniques are expected to have a more pronounced effect on improving standardization performance. However, to ensure the robustness of the CT number standardization and corresponding diagnostic performance, it is important to optimize the denoising strength in order to prevent the reduction in CT number contrast, considering the complex anatomical tissue structure. In future work, we plan to investigate the impact of advanced noise reduction techniques, such as Block-Matching 3D filtering (BM3D) and machine learning-based denoisers, on CT number standardization.
The primary objective of the proposed method is to standardize CT numbers of lung tissues to improve the diagnostic accuracy of COPD using lung CT images. As the method only requires reference materials scanned under the same imaging protocols, it can be seamlessly integrated into existing clinical CT imaging and diagnostic workflows without altering current acquisition procedures. The reference materials may be miniaturized and integrated onto the patient table, thereby facilitating the application of the proposed method directly to patient CT images without necessitating an additional reference phantom scan. With further clinical validation using patient CT data, the method holds promise for incorporation into CT image reconstruction systems and diagnostic software for COPD assessment.
In conclusion, we proposed and evaluated a novel method for standardizing the CT number of lung foam in the COPD gene2 phantom. The proposed method reduces diagnostic dependency on CT scanners and scanning protocols, enhancing reliability in distinguishing lung tissue from air pockets.