Assessing the consistency of CT-based ventilation imaging under noise reduction processing with simulated quantum noise using a nonrigid alveoli phantom

Background: Previous studies have reported that quantum noise inherently present in CT images hinders the generation of CT-based ventilation image (CTVI), while quantum noise reduction approaches that do not affect CTVI have not yet been reported.

Aims: The purpose of this study was to evaluate the impact of noise reduction preprocessing on the accuracy and robustness of CTVI in relation to quantum noise present in CT images.

Methods and material: To reproduce the quantum noise, Gaussian noise (SD: 30, 80, 150 HU) was added to each inhalation and exhalation CT image. CTVI_ref and CTVI_noise was generated from CT_ref and CT_noise. A median filter and the noise reduction by the CNN were also applied to the CT image, which contained the quantum noise, and CTVI_med and CTVI_cnn was created in the same manner as CTVI_ref. We evaluated whether the regions classified as high, middle, or low in CTVI_ref were accurately represented as high, middle, or low in CTVI_noise, CTVI_med and CTVI_cnn. Additionally, to evaluate the ventilation function of each voxel, we compared two-dimensional histograms of CTVI_ref, CTVI_noise, CTVI_med and CTVI_cnn.

Statistical analysis used: Cohen's kappa coefficient and Spearman's correlation were used to assess the agreement between CTVI_ref and each of the following: CTVI_noise, CTVI_med, and CTVI_cnn.

Results: CTVI_cnn significantly improved categorical consistency and voxel-level correlation of CTVI, particularly under high-noise conditions (150 HU), outperforming both CTVI_noise and CTVI_med.

Conclusions: CNN-based denoising effectively improved the accuracy and robustness of CTVI under quantum noise.

1 Introduction

A variety of imaging modalities exist to assess pulmonary ventilation. Examples include computed tomography (CT), dual-energy CT, magnetic resonance imaging, single-photon emission computed tomography (SPECT), and positron emission tomography (PET). These techniques can accurately assess the three-dimensional (3D) distribution of ventilatory function in a patient's lungs (1–3). In radiation therapy, treatment plans that utilize routine CT imaging and deformable image registration (DIR) to generate CT-based ventilation images (CTVIs) have demonstrated the ability to avoid areas of high ventilatory function within the lungs (4). The approach aims to reduce the dose administered to regions with elevated ventilatory function, creating a more targeted and personalized treatment strategy. This approach not only enhances the accuracy of predicting adverse lung events but also contributes to a more effective reduction in the occurrence of such events during treatment (5, 6). CT scans are used as part of routine radiation treatment procedures for most lung cancer patients and can provide additional functional information about the patient without requiring additional functional imaging equipment or methods. Treatment planning with CTVI is a practical, high-resolution, cost-effective, and time-saving approach that can be performed based on four-dimensional (4D) or expiratory and inspiratory CT images (4–9).

Studies are currently underway to validate the accuracy of CTVI. Radionuclide imaging is widely used to assess pulmonary function and is considered the standard of choice for assessing other functional imaging modalities (7, 8). Recent studies have demonstrated that CT-based and SPECT ventilatory function imaging have good spatial measurement accuracy and correlation (9, 10). In addition, clinical trials demonstrate that radiotherapy using CTVI significantly reduces dose to ventilated lung regions (NCT02528942, NCT02308709, NCT02843568) (11).

Accurate assessment of pulmonary ventilation function is crucial for using CTVI in treatment planning. Small changes in DIR parameters have been reported to cause large relative changes in the CTVI (12). The study noted that DIR-based images may not show accurate ventilatory function even when the spatial accuracy of the deformations is acceptable using target registration error (TRE). The quantum noise in CT images does not significantly affect the accuracy of DIR but may hinder the generation of accurate CTVI (13). A nonrigid alveoli phantom was developed to evaluate the CTVI, based on the assumption that an accuracy validation phantom is required to investigate the causes of these obstacles and improve CTVI accuracy (14). However, various problems related to CTVI methods have not yet been solved. It has been demonstrated that there is a significant difference in the CTVI produced via DIR when different DIR parameters are used, even after meeting the tolerance for DIR accuracy with this phantom (15).

CT images inevitably contain quantum noise, owing to the nature of x-ray images. It is desirable to use high-resolution CT images to create the CTVI. CTVI is used for both treatment planning and tracking pulmonary ventilation function using cone-beam CT acquired during treatment. Therefore, the accuracy of the CTVI must be independent of the CT image quality. To achieve this, the image quality must be improved using noise reduction and image correction techniques. Consequently, it could improve the accuracy of treatment planning and patient outcomes. Therefore, CTVI plays a crucial role in radiotherapy treatment planning, and it is desirable to improve its reliability and robustness through various methods.

In recent years, in addition to conventional filtering techniques such as median and Gaussian filters (16), image denoising methods using artificial intelligence (AI) have also been increasingly utilized in the field of medical imaging. In particular, deep learning methods based on convolutional neural networks (CNNs) have attracted attention as they can suppress noise while preserving structural details (17–19). Such AI-based preprocessing techniques are being explored as potential means to enhance the robustness and reproducibility of CT-based functional imaging, including CTVI.

In this study, we used a nonrigid alveoli phantom with ventilation functionality, which we developed as the world's first quality control tool for CTVI. We investigated the effect of preprocessing using both a conventional median filter and a deep learning-based denoising model on the accuracy and robustness of CTVIs. The purpose of this study is to evaluate how preprocessing methods, including AI-based denoising, affect the quality of CTVI, and to clarify their potential to improve robustness and accuracy in clinical applications.

2 Subjects and methods

2.1 CT datasets

The expiratory and inspiratory CT images were acquired using a 16-row detector CT scanner (Aquilion LB, Toshiba Medical Systems, Otawara, Japan). Image resolution was set to 0.78 × 0.78 × 3 mm, and a helical scan protocol was used. The scan parameters were set to 120 kVp, 300 mA, rotation time of 0.5 s, and slice thickness of 3.0 mm. The nonrigid alveoli phantom comprised an acrylic cylinder filled with polyurethane foam simulating alveoli, a polyurethane membrane simulating the diaphragm, a metal rod with piston function simulating respiratory muscles, and a polyurethane tube simulating the airway (14). Various motion patterns can be programmed to simulate breathing patterns of various frequencies. Additionally, airflow can be controlled by pressure changes in the vessel owing to diaphragm movement. The phantom was placed horizontally and adjusted to align with the longitudinal axis of the CT system. The respiratory cycle of the phantom was set to 10 s. The normal respiratory cycle is approximately 4 s; however, to focus only on quantum noise, the respiratory cycle of the phantom was set at which the motion artifact was as small as possible.

2.2 Simulation of quantum noise and noise reduction by the median filter

Additional noise was applied to the CT images to simulate the quantum noise in a simplified manner. The amplitude of quantum noise can be mathematically approximated by a Gaussian distribution (20). In this study, a Python script was developed to add noise with varying standard deviation [0–150 Hounsfield units (HU)] based on a normal distribution. Quantum noise was added using this script to a set of three pairs of expiratory and inspiratory images (30, 80, and 150 HU) to simulate the quantum noise in the CT images, as shown in Figure 1. We developed a script to calculate the median filter for the CT images with simulated quantum noise and fit it to all simulated noise. The filter used a kernel size of 3 × 3.

Figure 1

Figure 1. Comparison of exhale CT images simulating different noise levels.

2.3 Noise reduction by the deep learning-based denoising model

We constructed a denoising model based on a two-dimensional U-Net architecture (24) (Figure 2), with a network depth of 3 and an initial number of filters set to 32. The input to the model was a noisy image, and the output was the corresponding denoised image. To enable fair comparison with conventional filtering methods, no normalization was applied to the pixel values. Each image had a resolution of 512 × 512 pixels, and 131 slices were used per subject. For training, 14 types of Gaussian noise with standard deviations of 10, 20, 50, 60, 70, 100, 110, 120, 130, 140, 170, 180, 190, and 200 HU were added to clean images. For validation, noise levels of 40, 90, and 160 HU were used, and for testing, levels of 30, 80, and 150 HU were selected. The model was trained using the Adam optimizer with a batch size of 8 for up to 500 epochs. The L1 norm loss is defined as shown in Equation 1:

$l(x,y)={\displaystyle \frac{1}{N}\sum _{n=1}^N|x_n-y_n|,}(1)$

Figure 2

Figure 2. The denoising model based on a two-dimensional U-Net architecture.

where $x_n$ and $y_n$ denote the predicted and ground-truth pixel values, respectively, and N is the total number of pixels. For the validation dataset, the clean images were used as the ground truth, and the model yielding the lowest loss between the output and the clean images was selected for final testing. All training and evaluation were performed on a workstation equipped with an Intel Core i9-9920X 3.5 GHz twelve-core processor, 32 GB RAM, and an NVIDIA GeForce RTX 2080 Ti GPU running Ubuntu 22.04.4 LTS with NVIDIA Driver 535.183.01, CUDA 12.1, and cuDNN 8.9.7.29-1.

2.4 Deformable image registration

In this study, the inhalation image was deformed to match the reference expiration image. Deformations were performed on a set of ten pairs of expiratory and inspiratory images: the reference image CT_ref without additional noise, CT_noise with noise (30, 80, and 150 HU), and CT_med and CT_cnn, which are denoised versions of CT_noise using a median filter and the CNN model, respectively. Deformable image registration was performed using NiftyReg (version 1.4.2), a free and open-source software package for non-rigid image registration. NiftyReg uses a B-spline-based free-form deformation algorithm, which estimates the transformation between moving and reference images by optimizing a normalized mutual information while applying smoothness constraints. The DIR parameters used for these deformations were the optimal parameters reported in a previous study (15): “bending-energy penalty term,” introduced in the cost function to smooth deformations; “max number of iterations,” which affects the computation time; “number of levels to perform,” which refers to the number of optimization calculations; and “Jacobian-based penalty term,” which penalizes large local volume changes and prevents folding (21). The deformation was performed in four steps, following the deformation strategy previously reported as optimal in earlier studies (15). Each step was visually checked, and if the deformation was over-deformed, the deformation step was omitted. The deformation vector field was obtained at each step. It was input at the next step and integrated for each deformation. The CT scans in this study were performed in one imaging session and the phantom outline was not moving; therefore, no rigid registration was performed before the deformation process.

2.5 CT-based ventilation imaging

The sum of the deformation vector field acquired for each of the seven paired sets was converted to a Jacobian determinant to obtain the respective CTVI_ref, CTVI_noise (30, 80, and 150 HU), and CTVI_med (30, 80, and 150 HU), and CTVI_cnn (30, 80, and 150 HU). The DIR-based Jacobian metric was developed by Reinhardt et al. and is a measure of spatial volume change; it ensures that local volume changes do not alter the signal throughout the volume (22). The Jacobian determinant was calculated for each voxel in the phantom using (Equation 2).

$\begin{array}{r}\begin{array}{c}\mathrm{Jacobian}\mathrm{determinant}(\mathit{x},\mathit{y},\mathit{z})\\ =\left|\begin{array}{ccc}1+{\displaystyle \frac{\mathrm{\partial }{\mathit{u}}_{\mathit{x}}(\mathit{x},\mathit{y},\mathit{z})}{\mathrm{\partial }\mathit{x}}}& {\displaystyle \frac{\mathrm{\partial }{\mathit{u}}_{\mathit{x}}(\mathit{x},\mathit{y},\mathit{z})}{\mathrm{\partial }\mathit{y}}}& {\displaystyle \frac{\mathrm{\partial }{\mathit{u}}_{\mathit{x}}(\mathit{x},\mathit{y},\mathit{z})}{\mathrm{\partial }\mathit{z}}}\\ {\displaystyle \frac{\mathrm{\partial }{\mathit{u}}_{\mathit{y}}(\mathit{x},\mathit{y},\mathit{z})}{\mathrm{\partial }\mathit{x}}}& 1+{\displaystyle \frac{\mathrm{\partial }{\mathit{u}}_{\mathit{y}}(\mathit{x},\mathit{y},\mathit{z})}{\mathrm{\partial }\mathit{y}}}& {\displaystyle \frac{\mathrm{\partial }{\mathit{u}}_{\mathit{y}}(\mathit{x},\mathit{y},\mathit{z})}{\mathrm{\partial }\mathit{z}}}\\ {\displaystyle \frac{\mathrm{\partial }{\mathit{u}}_{\mathit{z}}(\mathit{x},\mathit{y},\mathit{z})}{\mathrm{\partial }\mathit{x}}}& {\displaystyle \frac{\mathrm{\partial }{\mathit{u}}_{\mathit{z}}(\mathit{x},\mathit{y},\mathit{z})}{\mathrm{\partial }\mathit{y}}}& 1+{\displaystyle \frac{\mathrm{\partial }{\mathit{u}}_{\mathit{z}}(\mathit{x},\mathit{y},\mathit{z})}{\mathrm{\partial }\mathit{z}}}\end{array}\right|\end{array}\end{array}(2)$

where u_x, u_y, and u_z are the x, y, and z components of u, respectively. Jacobian determinant measures the expansion and contraction at position (x, y, z) in the image. When Jacobian determinant is greater than one, local tissue expansion is present, and when Jacobian determinant is less than one, local tissue contraction is present. Jacobian determinant is a relative measure of ventilatory functionality on a voxel-by-voxel basis within the lung.

2.6 Evaluation of spatial deformation accuracy by DIR

Twenty-five landmarks were manually placed by an experienced medical physicist in a volume near the pulmonary vessels and bronchi in a nonrigid alveoli phantom (Figure 3). The target displacement error, that is, the displacement of a landmark due to respiratory motion, was measured as the Euclidean distance between the exhalation and inhalation images. The Euclidean distance was calculated using the formula shown in Equation 3:

$\begin{array}{c}\sqrt{{({\mathit{x}}_{\mathit{r}}-{\mathit{x}}_{\mathit{t}})}^2+{({\mathit{y}}_{\mathit{r}}-{\mathit{y}}_{\mathit{t}})}^2+{({\mathit{z}}_{\mathit{r}}-{\mathit{z}}_{\mathit{t}})}^2},\end{array}(3)$

Figure 3

Figure 3. Twenty-five landmark setups placed in a volume near the pulmonary vessels and bronchi in a nonrigid alveoli phantom.

where (x_r, y_r, z_r) and (x_t, y_t, z_t) are the landmark coordinates of the reference and target images, respectively. To evaluate the spatial accuracy of DIR with added noise, we used the Euclidean distance between the corresponding landmarks defined in the expiratory and deformed inspiratory images to calculate the Euclidean distance, which is denoted as TRE. TRE represents the spatial 3D distance discrepancy. When the deformed image perfectly matches the reference expiratory image (Euclidean distance = 0), TRE is equal to zero. Relative spatial accuracy was evaluated and compared with the spatial accuracy of the reference noiseless DIR.

2.7 Global consistency analysis using kappa statistics

To evaluate the clinical consistency of CTVI for treatment planning, each voxel in the CTVI was categorized into three regions—high, middle, and low ventilation—by evenly dividing the range of ventilation values in CTVI_ref. This categorization reflects a typical clinical scenario where high-ventilation regions are avoided during irradiation. The same classification thresholds were applied to all other CTVIs, including CTVI_noise (30, 80, and 150 HU), CTVI_med (30, 80, and 150 HU), and CTVI_cnn (30, 80, and 150 HU). Although the absolute ventilation values may differ, consistency was defined as the regions categorized as high, middle, or low in CTVI_ref being similarly categorized in the compared CTVIs. To quantify consistency, the proportion of voxels in each test CTVI that retained the same categorical label (high, middle, or low) as in CTVI_ref was calculated. The degree of agreement was assessed using Cohen's kappa coefficient.

2.8 Voxel-based local evaluation using 2D histograms and spearman correlation

To evaluate the consistency of local ventilation function in each voxel, a two-dimensional (2D) histogram was constructed by plotting the Jacobian determinant value of each voxel in CTVIref against the corresponding value in CTVI_noise (30, 80, and 150 HU). Similarly, 2D histograms were created for CTVI_med (30, 80, and 150 HU) and CTVI_cnn (30, 80, and 150 HU), which were generated by denoising CT_noise using a median filter and the CNN model, respectively. All histograms were generated based on voxel-wise spatial correspondence with CTVI_ref, enabling direct comparison of local ventilation values. Spearman's rank correlation coefficients were calculated from the 2D histograms to evaluate the consistency between each CTVI and the reference.

3 Results

3.1 Evaluation of DIR spatial deformation accuracy

Figure 4 presents a comparison of TRE values, indicating the spatial accuracy of DIR between CT images containing noise and CT image pairs with noise removed using the median filter and CNN-based denoising. The average displacement between the expiratory and inspiratory CT images was 14.59 ± 6.42 mm. The mean TRE values of the 25 landmarks were 1.39 ± 0.89 mm (maximum 2.95 mm) for CT_ref. The mean TRE values of the 25 landmarks were 1.22 ± 0.65 mm (maximum 2.54 mm), 0.71 ± 0.45 mm (maximum 1.89 mm), and 1.10 ± 0.83 mm (maximum 2.71 mm) for CT_noise (30, 80, and 150 HU). The mean TRE values for CT_med were 1.78 ± 0.71 mm (maximum 2.93 mm), 1.77 ± 0.78 mm (maximum 2.94 mm), and 1.42 ± 0.65 mm (maximum 2.80 mm) at 30, 80, and 150 HU, respectively. The mean TRE values for CT_cnn were 1.34 ± 0.00 mm (maximum 2.55 mm), 1.28 ± 0.00 mm (maximum 2.82 mm), and 1.22 ± 0.00 mm (maximum 2.45 mm) at 30, 80, and 150 HU, respectively. TREs for all conditions, including noisy and denoised images, remained within 3 mm. When comparing mean TRE values between CT_noise and denoised images at each noise level, TREs were higher in CT_med, while CT_cnn yielded values similar to or slightly higher than CT_noise.

Figure 4

Figure 4. Comparison of TRE values was performed for cT_noise, cT_med, cT_cnn, and cT_ref at different noise levels (30, 80, and 150 HU), where cT_ref represents the DIR results based on CT images without added quantum noise. In this context, movement indicates the displacement caused by respiratory motion between the expiratory and inspiratory phases.

3.2 Evaluation of CTVIs

3.2.1 Visual assessment of CTVIs

Figure 5 shows a visual comparison of CTVIs, including CTVI_ref, CTVI_noise, CTVI_med, and CTVIcnn at different noise levels. The visual assessment reveals that, compared to CTVI_ref, the location of high-functioning regions near the diaphragm remains consistent. However, additional high-functioning regions not observed in CTVI_ref are present, and the resolution of ventilatory function distribution is reduced. The visual assessment indicates that CTVI_med (30 HU) is closer to CTVI_ref than the corresponding CTVI_noise shown in Figure 5, suggesting that noise reduction improves the visual accuracy of CTVI under lower noise conditions. CTVI_cnn appear visually closer to CTVI_ref than CTVI_noise at all noise levels, indicating that noise reduction improves the visual accuracy of CTVI. Among the three noise levels, CTVI_cnn at 150 HU shows a particularly notable improvement over CTVI_noise, suggesting that CNN-based denoising enhances the visualization of ventilation distribution under high noise conditions.

Figure 5

Figure 5. CTVI_noise, CTVI_med and CTVI_cnn at different noise levels. High-functioning regions are shown in red, while regions with decreased lung ventilation are displayed in blue.

3.2.2 Global consistency analysis using kappa statistics

To evaluate the clinical utility of CTVI for treatment planning, ventilation values in CTVI_ref were evenly divided into three regions: high, middle, and low ventilation. The same categorization was applied to CTVI_noise, CTVI_med and CTVI_cnn at each noise level. Consistency was defined as regions categorized as high, middle, or low in CTVI_ref being similarly categorized in CTVI_noise, CTVI_med and CTVI_cnn, regardless of absolute ventilation values. If the categorizations were perfectly consistent, the percentage of voxels correctly matching high, middle, and low regions would be 33.3% for each category. Figure 6 summarizes the percentage of voxels that correctly matched high, middle, and low regions between CTVI_ref and CTVI_noise. For example, at 30 HU noise, only 11.32%, 12.64%, and 12.22% of voxels in the high, middle, and low regions of CTVI_ref were correctly identified in CTVI_noise, respectively, with notable mismatches such as 14.36% of voxels categorized as low in CTVI_ref being misclassified as high in CTVI_noise. In contrast, CTVI_med at 30 HU showed substantially improved agreement, with 21.26%, 22.81%, and 23.18% correctly matching the high, middle, and low regions of CTVI_ref, respectively. Additionally, at 80 HU noise, CTVI_med also demonstrated an improvement in consistency compared to CTVI_noise. The percentages of voxels correctly categorized as high, middle, and low in CTVI_med were 21.24%, 20.02%, and 22.34%, respectively, indicating a better agreement with CTVI_ref compared to CTVI_noise, where the corresponding percentages were only 12.15%, 12.89%, and 12.07%. This result highlights that the noise reduction processing improved clinical consistency not only at 30 HU but also at 80 HU noise levels. In addition, CTVI_cnn showed consistent improvement across all noise levels. Notably, at 150 HU, the consistency between CTVI_cnn and CTVI_ref was significantly higher than that of CTVI_noise. The percentage of correctly matched voxels was 26.23%, 28.91%, and 26.85% for high, middle, and low ventilation regions, respectively, compared to only 12.03%, 13.25%, and 12.67% for CTVI_noise. These results highlight that CNN-based denoising particularly improves clinical consistency under high-noise conditions. Cohen's kappa coefficients further quantified the agreement between CTVI_ref and both CTVI_noise and CTVI_med. Table 1 presents these results, showing that kappa values for CTVI_noise were 0.043, 0.057, and 0.069 at noise levels of 30 HU, 80 HU, and 150 HU, respectively. For CTVI_med, the kappa values were significantly higher at lower noise levels (0.51 at 30 HU and 0.45 at 80 HU), but the consistency diminished at 150 HU (0.083). CTVI_cnn exhibited consistently high kappa values across all noise levels, with 0.60 at 30 HU, 0.51 at 80 HU, and 0.73 at 150 HU, indicating superior categorical agreement with CTVI_ref compared to both CTVI_noise and CTVI_med. These results support the effectiveness of CNN-based denoising in preserving clinically relevant ventilation patterns, even under high noise conditions.

Figure 6

Figure 6. Proportion of categorized voxels across high, middle, and low ventilation regions for CTVI_noise, CTVI_med and CTVI_cnn compared to CTVI_ref. The horizontal axis represents CTVI_ref, while the vertical axis shows the three ventilation categories (high, middle, and low) for CTVI_noise, CTVI_med and CTVI_cnn. Darker colors indicate higher agreement in classification.

Table 1

Table 1. Cohen's kappa coefficients of CTVIs for CTVI_ref.

3.2.3 Voxel-based local evaluation using 2D histograms and spearman correlation

To assess the consistency of local ventilation function, a voxel-by-voxel comparison between CTVI_ref and both CTVI_noise, CTVI_med and CTVI_cnn was conducted using 2D histograms (Figure 7). These histograms demonstrate that the distribution improves as it approaches y = x, indicating greater consistency with CTVI_ref. At a noise level of 30 HU, the histogram of CTVI_med shows a marked improvement in consistency compared to CTVI_noise, as the distribution is more concentrated along y = x. However, at noise levels of 80 and 150 HU, no significant improvement was observed. In contrast, CTVI_cnn demonstrated a different trend: while there was no notable improvement in the 2D histograms at 30 and 80 HU, the histogram at 150 HU showed greater alignment with the y = x line, suggesting improved voxel-level consistency. Table 2 presents the Spearman correlation coefficients for these comparisons. The results show that CTVI_med exhibits better correlation with CTVI_ref than CTVI_noise at lower noise levels, but the correlation decreases as the noise level increases. Despite this, a statistically significant trend toward improvement is observed (<.0001). The Spearman correlation coefficients for CTVI_cnn were 0.59, 0.65, and 0.83 at 30, 80, and 150 HU, respectively. This indicates that although CTVI_cnn did not improve correlation at lower noise levels, it showed a substantial improvement at 150 HU (<.0001).

Figure 7

Figure 7. Two-dimensional histograms of CTVI_noise, CTVI_med and CTVI_cnn at different noise levels with respect to CTVI_ref. The horizontal axis represents CTVI_ref, while the vertical axis shows the voxel-based lung ventilation of CTVI_noise, CTVI_med and CTVI_cnn. Perfect agreement is indicated by the histogram distribution along y = x.

Table 2

Table 2. Spearman correlation coefficients of CTVIs for CTVI_ref.

4 Discussion

Despite the extensive body of research on CTVI (4–10), few studies have employed phantoms capable of replicating human-like ventilation in the context of clinical applications. While it is well known that noise significantly affects CT image analysis, the impact of noise reduction preprocessing on the accuracy and consistency of CTVI remains insufficiently explored. In this study, quantum noise was simulated in CT images using a nonrigid alveoli phantom designed to mimic lung motion.

In this study, the spatial accuracy of the DIR was evaluated by simulating noise levels from 0 to 150 HU. For all CT_ref, CT_noise, CT_med and CT_cnn, the deformation accuracy was within 3 mm of the tolerances given in TG-132 (23), regardless of the noise level. Although previous studies have investigated noise levels of 200 HU (13), in the initial experiments of this study, the noise level of 200 HU resulted in over-deformation due to DIR for the same deformation parameters, and an accurate CTVI could not be established. This result suggests that noise levels above 200 HU significantly affect the deformation accuracy of the DIR. High noise levels may lead to errors in the DIR algorithm. Therefore, the results of this study suggest that the DIR technique has sufficient accuracy for generating CTVI at quantum noise levels up to 150 HU. Although all TRE values were within the TG-132 tolerance of 3 mm, CT_med and CT_cnn exhibited slightly higher TREs compared to CT_noise. One possible explanation is that the denoising process may have smoothed out anatomical features critical for deformable registration, resulting in slightly reduced precision. Alternatively, mild quantum noise may have enhanced local contrast in CT_noise, unintentionally aiding DIR alignment. However, these differences remained within the clinically acceptable margin and are unlikely to affect the final CTVI outcome.

Table 1; Figure 6 illustrate the impact of quantum noise and the application of preprocessing filters on the consistency of CTVI from a clinical perspective. When using CTVI for treatment planning, lung ventilation is categorized into three levels—high, middle, and low—and treatment plans are designed to avoid high-function regions. In CTVI_ref, approximately 33% of the lung ventilation is classified into each category. Ideally, in cases of accurate classification, the relationships between CTVI_ref, CTVI_noise, CTVI_med and CTVI_cnn should result in high-high, middle-middle, and low-low matches approaching 33%. Focusing on CTVI_noise, the maximum agreement across all noise levels was only 13.25%, indicating significant misclassification of lung ventilation when quantum noise is present. In contrast, CTVI_med achieved over 20% agreement in all ventilation categories at noise levels below 80 HU. A particularly important observation is the proportion of regions classified as high in CTVI_ref but misclassified as low. This proportion was kept below 4.3% at its maximum. In addition, CNN-based denoising further improved the consistency of CTVI at all noise levels. Notably, CTVI_cnn achieved the highest agreement with CTVI_ref, particularly under high-noise conditions. At 150 HU, CNN showed the greatest improvement in both categorical agreement and voxel-wise correlation (κ = 0.73, Spearman ρ = 0.83), outperforming both CTVI_noise and CTVI_med. These results indicate that CNN-based denoising has strong potential to enhance the robustness of CTVI, even under clinically challenging noise conditions.

Table 2; Figure 7 focus on the voxel-level accuracy of CTVI, demonstrating the local effects of quantum noise and the application of preprocessing filters on CTVI accuracy using two-dimensional histograms and Spearman correlation. The results of this study confirmed that as noise levels increase, the accuracy of CTVI decreases, proving that quantum noise is a significant factor that hinders the accuracy of CTVI. However, CNN-based denoising yielded stronger improvements in correlation, especially at 150 HU, suggesting it is a more robust solution in high-noise environments.

Interestingly, while Cohen's kappa coefficients and Spearman correlation coefficients generally showed consistent trends, some discrepancies were noted. For example, at 150 HU, CTVI_med yielded a relatively high Spearman correlation (ρ = 0.61) but a low kappa value (κ = 0.083), suggesting that voxel-wise rankings were preserved even though many values crossed categorical thresholds. Conversely, at 30 HU, CTVI_cnn showed a high kappa (κ = 0.60) despite having a lower Spearman correlation (ρ = 0.59), indicating that category-level agreement was strong, while voxel value variations limited rank correlation. These findings emphasize that categorical and continuous metrics capture different aspects of agreement, and highlight the need for using both to comprehensively evaluate CTVI accuracy.

Limitations of this study include the lack of comparison with vendor-provided denoising techniques and the absence of hybrid preprocessing strategies. Future work should explore combining CNN-based and conventional filtering approaches and testing these methods in patient datasets.

5 Conclusion

This study quantitatively evaluated the effect of preprocessing on the accuracy and robustness of the CTVI using a nonrigid alveoli phantom with ventilation functionality, developed as the world's first quality control tool for CTVI. We demonstrated that quantum noise significantly impairs the accuracy and consistency of CTVI. While median filtering was shown to be a simple and effective method for mitigating this effect, CNN-based denoising provided superior performance, particularly under high-noise conditions. These findings suggest that both conventional and AI-based preprocessing approaches contribute to improving the quality of CTVI, with deep learning methods offering strong potential to enhance robustness and accuracy in clinical applications.

Data availability statement

The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding author.

Author contributions

SM: Conceptualization, Investigation, Methodology, Writing – original draft. HF: Methodology, Project administration, Writing – review & editing. KY: Data curation, Methodology, Writing – review & editing. NK: Data curation, Formal analysis, Investigation, Validation, Writing – review & editing. MT: Formal analysis, Investigation, Writing – review & editing. HN: Formal analysis, Investigation, Writing – review & editing. SM: Data curation, Methodology, Writing – review & editing. FT: Investigation, Methodology, Writing – review & editing. TF: Supervision, Writing – review & editing.

Funding

The author(s) declare that financial support was received for the research and/or publication of this article. This work was supported by JSPS KAKENHI Grant Number JP22K15891.

Acknowledgments

The authors would like to thank Editage (https://www.editage.jp) for the English language review.

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declare that no Generative AI was used in the creation of this manuscript.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

1. Suga K. Technical and analytical advances in pulmonary ventilation SPECT with xenon-133 gas and tc-99m-technegas. Ann Nucl Med. (2002) 16:303–10. doi: 10.1007/BF02988614

PubMed Abstract | Crossref Full Text | Google Scholar

2. van Beek EJR, Wild JM, Kauczor H-U, Schreiber W, Mugler JP 3rd, de Lange EE. Functional MRI of the lung using hyperpolarized 3-helium gas. J Magn Reson Imaging. (2004) 20:540–54. doi: 10.1002/jmri.20154

PubMed Abstract | Crossref Full Text | Google Scholar

3. Harris RS, Schuster DP. Visualizing lung function with positron emission tomography. J Appl Physiol (1985). (2007) 102:448–58. doi: 10.1152/japplphysiol.00763.2006

PubMed Abstract | Crossref Full Text | Google Scholar

4. Latifi K, Forster KM, Hoffe SE, Dilling TJ, van Elmpt W, Dekker A, et al. Dependence of ventilation image derived from 4D CT on deformable image registration and ventilation algorithms. J Appl Clin Med Phys. (2013) 14:4247. doi: 10.1120/jacmp.v14i4.4247

PubMed Abstract | Crossref Full Text | Google Scholar

5. Yamamoto T, Kabus S, von Berg J, Lorenz C, Keall PJ. Impact of four-dimensional computed tomography pulmonary ventilation imaging-based functional avoidance for lung cancer radiotherapy. Int J Radiat Oncol Biol Phys. (2011) 79:279–88. doi: 10.1016/j.ijrobp.2010.02.008

PubMed Abstract | Crossref Full Text | Google Scholar

6. Yaremko BP, Guerrero TM, Noyola-Martinez J, Guerra R, Lege DG, Nguyen LT, et al. Reduction of normal lung irradiation in locally advanced non-small-cell lung cancer patients, using ventilation images for functional avoidance. Int J Radiat Oncol Biol Phys. (2007) 68:562–71. doi: 10.1016/j.ijrobp.2007.01.044

PubMed Abstract | Crossref Full Text | Google Scholar

7. Eslick EM, Kipritidis J, Gradinscak D, Stevens MJ, Bailey DL, Harris B, et al. CT ventilation imaging derived from breath hold CT exhibits good regional accuracy with galligas PET. Radiother Oncol. (2018) 127:267–73. doi: 10.1016/j.radonc.2017.12.010

PubMed Abstract | Crossref Full Text | Google Scholar

8. Kipritidis J, Tahir BA, Cazoulat G, Hofman MS, Siva S, Callahan J, et al. The VAMPIRE challenge: a multi-institutional validation study of CT ventilation imaging. Med Phys. (2019) 46:1198–217. doi: 10.1002/mp.13346

PubMed Abstract | Crossref Full Text | Google Scholar

9. Castillo R, Castillo E, McCurdy M, Gomez DR, Block AM, Bergsma D, et al. Spatial correspondence of 4D CT ventilation and SPECT pulmonary perfusion defects in patients with malignant airway stenosis. Phys Med Biol. (2012) 57:1855–71. doi: 10.1088/0031-9155/57/7/1855

PubMed Abstract | Crossref Full Text | Google Scholar

10. Yamamoto T, Kabus S, Lorenz C, Mittra E, Hong JC, Chung M, et al. Pulmonary ventilation imaging based on 4-dimensional computed tomography: comparison with pulmonary function tests and SPECT ventilation images. Int J Radiat Oncol Biol Phys. (2014) 90:414–22. doi: 10.1016/j.ijrobp.2014.06.006

PubMed Abstract | Crossref Full Text | Google Scholar

11. Yamamoto T, Kabus S, Bal M, Keall PJ, Moran A, Wright C, et al. Four-dimensional computed tomography ventilation image guided lung functional avoidance radiation therapy: a single-arm prospective pilot clinical trial. Int J Radiat Oncol Biol Phys. (2022) 115:1144–54. doi: 10.1016/j.ijrobp.2022.11.026

PubMed Abstract | Crossref Full Text | Google Scholar

12. Castillo E, Castillo R, Vinogradskiy Y, Guerrero T. The numerical stability of transformation-based CT ventilation. Int J Comput Assist Radiol Surg. (2017) 12:569–80. doi: 10.1007/s11548-016-1509-x

PubMed Abstract | Crossref Full Text | Google Scholar

13. Latifi K, Huang T-C, Feygelman V, Budzevich MM, Moros EG, Dilling TJ, et al. Effects of quantum noise in 4D-CT on deformable image registration and derived ventilation data. Phys Med Biol. (2013) 58:7661–72. doi: 10.1088/0031-9155/58/21/7661

PubMed Abstract | Crossref Full Text | Google Scholar

14. Miyakawa S, Tachibana H, Moriya S, Kurosawa T, Nishio T, Sato M. Design and development of a nonrigid phantom for the quantitative evaluation of DIR-based mapping of simulated pulmonary ventilation. Med Phys. (2018) 45:3496–505. doi: 10.1002/mp.13017

Crossref Full Text | Google Scholar

15. Miyakawa S, Tachibana H, Moriya S, Kurosawa T, Nishio T. Evaluation of deformation parameters for deformable image registration-based ventilation imaging using an air-ventilating non-rigid phantom. Phys Med. (2018) 50:20–5. doi: 10.1016/j.ejmp.2018.05.016

PubMed Abstract | Crossref Full Text | Google Scholar

16. Gonzalez RC, Woods RE. Digital Image Processing. 4th ed. Boston, MA: Pearson (2018).

Google Scholar

17. Chen H, Zhang Y, Kalra MK, Lin F, Chen Y, Liao P, et al. Low-dose CT with a residual encoder-decoder convolutional neural network (RED-CNN). IEEE Trans Med Imaging. (2017) 36:2524–35. doi: 10.1109/TMI.2017.2715284

PubMed Abstract | Crossref Full Text | Google Scholar

18. Zhang K, Zuo W, Chen Y, Meng D, Zhang L. Beyond a Gaussian denoiser: residual learning of deep CNN for image denoising. IEEE Trans Image Process. (2017) 26:3142–55. doi: 10.1109/TIP.2017.2662206

PubMed Abstract | Crossref Full Text | Google Scholar

19. Yang Q, Yan P, Zhang Y, Yu H, Shi Y, Mou X, et al. Low-dose CT image denoising using a generative adversarial network with Wasserstein distance and perceptual loss. IEEE Trans Med Imaging. (2018) 37(6):1348–57. doi: 10.1109/TMI.2018.2827462

PubMed Abstract | Crossref Full Text | Google Scholar

20. Herman GT. Image reconstruction from projections. In: The Fundamentals of Computerized Tomography. New York: Academic Press (1980). p. 141–2.

Google Scholar

21. Modat M. Lung registration using the NiftyReg package. Med Image Anal Clin Grand Chall. (2010):33–42.

Google Scholar

22. Reinhardt JM, Ding K, Cao K, Christensen GE, Hoffman EA, Bodas SV. Registration-based estimates of local lung tissue expansion compared to xenon CT measures of specific ventilation. Med Image Anal. (2008) 12:752–63. doi: 10.1016/j.media.2008.03.007

PubMed Abstract | Crossref Full Text | Google Scholar

23. Brock KK, Mutic S, McNutt TR, Li H, Kessler ML. Use of image registration and fusion algorithms and techniques in radiotherapy: report of the AAPM radiation therapy committee task group No. 132. Med Phys. (2017) 44:e43–76. doi: 10.1002/mp.12256

PubMed Abstract | Crossref Full Text | Google Scholar

24. Ronneberger O, Fischer P, Brox T. U-Net: convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. Springer International Publishing (2015). p. 234–41.

Google Scholar