Biphasic quantitative CT to classify a residual volume-defined subclinical gas-trapping phenotype: development and internal validation of a multivariable model

Introduction

Small airway dysfunction (SAD) refers to a pathophysiological state characterized by expiratory airflow limitation in the small airways, resulting from early structural changes or decreased pulmonary elasticity, yet insufficient to meet the diagnostic criteria for obstructive ventilatory dysfunction (1). One of its commonly observed correlates is subclinical gas-trapping. Recent studies indicate that the prevalence of SAD among Chinese adults reaches as high as 43.5%, affecting over 400 million individuals, making it a significant public health issue deserving attention (2). Due to the small airways’ vast number, extensive surface area, and minimal resistance, early-stage SAD often develops without obvious clinical symptoms, and its insidious onset frequently leads to under-detection by routine assessments. Systematic detection beyond high-risk groups remains challenging (3).

Conventional spirometry has marked limitations in assessing the peripheral airways. It mainly reflects the average level of overall lung function and has difficulty characterizing the severity and spatial distribution of localized areas of dysfunction. Flow-based indices can only indirectly reflect resistance in the small to medium airways, are influenced by expiratory effort and the lung volume at the start of the maneuver, and have substantial within- and between-subject variability as reported in previous studies. In addition, because the lung parenchyma has a certain compensatory capacity, only when there is more than approximately 30% parenchymal destruction or about 75% small-airway obstruction are pulmonary function indices likely to become abnormal (4). Therefore, techniques capable of regionalized, phase-specific quantification are needed to capture early, occult abnormalities.

Paired inspiratory-expiratory quantitative computed tomography (qCT) directly samples density dynamics and expiratory low-attenuation burden at high spatial resolution. Derived metrics—such as the mean lung density ratio and difference (MLD_ex/in and MLD_ex-in) and low-attenuation fractions on inspiration or expiration [e.g., VI-950, VI-910, VI-856; VI denotes the fraction of lung voxels below the specified Hounsfield units (HU) threshold in the corresponding phase]—provide complementary information on impaired emptying and ventilation heterogeneity that is not available from spirometry (3,5,6). In parallel, qCT combined with machine-learning has shown promise in airway diseases, including chronic obstructive pulmonary disease (COPD) and asthma (7-9), but systematic work linking biphasic qCT indices to a residual volume (RV)-defined gas-trapping phenotype—and evaluating their joint performance within a parsimonious classification model—remains limited.

Accordingly, this study evaluates whether biphasic qCT metrics capture a subclinical gas-trapping phenotype operationalized by elevated RV expressed as percent predicted (RV%pred), and assesses the performance of a multivariable qCT-based classification model for early risk stratification. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-969/rc).

Methods

General information

This study included flight-major university students who underwent respiratory biphasic CT scans and pulmonary function tests at Air Force Medical Center between March 2022 and September 2022. This study was a cross-sectional observational study. The study was conducted at the Air Force Medical Center, a tertiary medical institution located in Beijing, China.

Inclusion criteria: (I) age 18–22 years; (II) male flight-major university students without a history of severe respiratory disease; (III) completion of inspiratory and expiratory biphasic CT scans and pulmonary function tests, with satisfactory quality control; and (IV) clear and complete CT images allowing effective lung fissure segmentation by post-processing software.

Exclusion criteria: (I) history of lung cancer or other space-occupying pulmonary lesions; (II) history of thoracic surgery or chest deformity; (III) diseases affecting lung analysis, such as extensive pulmonary infection, consolidation, atelectasis, substantial pleural effusion, or severe pulmonary interstitial fibrosis; (IV) co-existing respiratory diseases such as bronchial asthma, pulmonary tuberculosis, bronchiectasis, or pulmonary hypertension; (V) significant impairment of vital organs including the heart, liver, or kidneys; and (VI) definite history of occupational dust exposure.

Ultimately, 71 eligible male subjects were included, with an average age of 18.88±0.98 years and an average body mass index (BMI) of 21.98±1.39 kg/m². All participants were non-smokers and free from underlying diseases. Due to professional enrollment and training requirements, all participants were male flight-major university students. The sample size was based on feasibility considerations, determined by the number of eligible flight major university students available at Air Force Medical Center during the defined study period. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Ethics Committee of the Air Force Medical Center (approval No. 2024-68-PJ01), and written informed consent was obtained from all participants.

Pulmonary function tests

Pulmonary function tests were performed within 1 week before or after CT scanning, in accordance with the Guideline for Pulmonary Function Testing in Primary Care [2018] (10). All participants underwent routine pulmonary function testing and lung volume measurement using the single-breath (SB) gas dilution method. Examinations were conducted by certified respiratory technicians using the SensorMedics Vmax 299 system (SensorMedics Corporation, Yorba Linda, CA, USA). The testing procedure included maximal inspiration, breath-holding, and forced expiration maneuvers. SB dilution was selected for this young, non-smoking, highly compliant cohort and performed under strict laboratory quality control; while body plethysmography or multi-breath dilution are reference methods, SB is acceptable in healthy or mildly impaired subjects, reduces test burden, and may underestimate volumes only with marked obstruction—an unlikely physiology in this cohort. All tests met guideline acceptability/repeatability criteria; studies failing quality standards were excluded.

This study collected the following parameters: forced vital capacity (FVC), forced expiratory volume in 1 second (FEV₁), the FEV₁/FVC ratio, total lung capacity measured by SB dilution (TLC-SB), RV measured by SB dilution (RV-SB), and the carbon monoxide transfer coefficient adjusted for alveolar volume (DLCOcVA). Based on these, the following derived indices were calculated: the ratio of RV to TLC-SB (RV/TLC-SB), the ratio of functional residual capacity to TLC-SB (FRC/TLC-SB), as well as the ratios of measured to predicted values for each parameter (e.g., RV-SB measured/predicted, TLC-SB measured/predicted, FVC measured/predicted, etc.). The predicted values were derived using the pulmonary function reference equations published in 1988. Percent-predicted values (rather than z-scores) were used given the narrow age window (18–22 years) and the laboratory’s established calibration framework.

In this study, the RV-SB measured/predicted (i.e., RV%pred) was used as a stratification indicator, as it reflects dynamic changes in the proportion of residual gas throughout the respiratory cycle and has been shown to be associated with early SAD and gas-trapping. According to the Chinese Expert Consensus on Pulmonary Function Diagnosis in Adults [2022], RV%pred ≥120% is indicative of abnormal gas retention (11). Based on this clinical threshold, participants were divided into two groups: RV-elevated group (RV%pred ≥120%) and RV-normal group (RV%pred <120%), which were used for subsequent parameter analysis and predictive model development.

Chest biphasic CT scan

All participants underwent standardized respiratory training prior to scanning. Scans were performed using a Toshiba Aquilion ONE TSX-301A 320-slice spiral CT scanner (Toshiba Medical Systems Corporation, Otawara, Japan). Participants lay supine on the scanning table, with hands holding the head, entering head-first into the scanning aperture. Low-dose CT scans were conducted separately at deep inspiration (inspiratory phase) and forced expiration (expiratory phase), with breath-holding and without contrast agent injection.

Scanning ranged from the lung apex to lung base, covering the entire pulmonary parenchyma. Acquisition parameters were as follows: tube voltage 120 kV, tube current 30–40 mAs, collimator width 128 mm × 0.625 mm, pitch 1.0875, and slice thickness 5 mm. Biphasic images were reconstructed into 1-mm thin layers using a bone reconstruction algorithm, with a reconstruction matrix of 512×512 and a field of view (FOV) of 350–450 mm. All images were exported in Digital Imaging and Communications in Medicine (DICOM) format for subsequent quantitative analysis.

The radiologists and technicians performing CT image segmentation and quantitative analyses were blinded to pulmonary function test results and clinical grouping.

Lung segmentation algorithm

This study adopted an encoder-decoder network architecture and combined it with a recurrent cross-attention module (RCCA) (12) to achieve high-precision lung segmentation. The encoder part uses a pre-trained ResNet-50 (13) network to extract rich semantic features, the decoder part is responsible for precise positioning, and the RCCA module is used to obtain semantic dependencies between different categories to better understand the hierarchical structure of the lungs. The network input is nine consecutive slices, containing nine channels, namely the original images of nine consecutive slices and the whole lung segmentation labels of nine slices. The network output is the lung segmentation result of the central slice, including the whole lung category and the background category. The algorithm has been validated by in-house datasets and clinically applied (14).

Trachea segmentation algorithm

The proposed algorithm is a three-dimensional (3D) lung segmentation algorithm based on deep learning, which is a U-shaped network structure composed of a two-dimensional (2D) encoder module and a 3D decoder module, and can be adapted according to the characteristics of the data. In order to improve the receptive field of the model and fuse the global contextual information, we introduce the Atrous Spatial Pyramid Pooling module (15) into the model, which preserves both spatial and semantic information by fusing feature maps at different scales; moreover, the model uses PixelShuffle to recover the spatial resolution of the feature maps in order to preserve more spatial information in the up-sampling. In addition, the model uses PixelShuffle (16) to recover the spatial resolution of the feature maps during up-sampling in order to preserve finer features, thus obtaining a complete and continuous airway topology of the lungs. The algorithm has been validated by in-house datasets. First, after selecting the best model and threshold on the tuning set, the same configuration is used to test on the test set. The selection of test indicators is as follows. For the trachea, this algorithm uses the Dice score, which is the most commonly used in segmentation tasks, as the evaluation indicator.

Indicator calculation

Following segmentation of the inspiratory and expiratory CT images by lung and airway structures, nine quantitative metrics were derived using threshold-based segmentation and conventional computational methods:

• Mean lung density ratio (MLD_ex/in): calculated as the ratio of mean lung density (in HU) on expiratory CT to that on inspiratory CT, i.e., CT_exp/CT_insp.
• Mean lung density difference (MLD_ex-in): defined as the absolute difference in mean lung density between expiration and inspiration, i.e., CT_exp − CT_insp.
• Lung volume ratio (LV): defined as the ratio of expiratory to inspiratory lung volumes.
• Volume decrease ratio (VDR): calculated as (inspiratory lung volume − expiratory lung volume)/inspiratory lung volume.
• Relative volume change (RVC): The total lung volume was segmented within the attenuation range of −1,024 to −500 HU, and normal lung parenchyma was defined within −950 to −500 HU, excluding emphysematous regions. Lung tissue with attenuation values between −950 and −860 HU was extracted on both inspiratory and expiratory scans. Relative volume (%) was computed as: relative volume = [volume (−950 to −860 HU)/volume (−950 to −500 HU)] × 100. The respiratory variation (RVC_−950 to −860) was then defined as the difference in relative volume between expiration and inspiration.
• 15^th percentile lung density (Perc15): the CT attenuation value below which 15% of lung voxels are distributed in the voxel-based density histogram.
• Volume of inspiration <−950 HU (VI-950): volume of lung tissue with attenuation values below −950 HU during inspiration.
• Volume of inspiration <−910 HU (VI-910): volume of lung tissue with attenuation values below −910 HU during inspiration.
• Volume of expiration <−856 HU (VI-856): volume of lung tissue with attenuation values below −856 HU during expiration.

Subsequently, the correlations between these CT-derived quantitative parameters and pulmonary function metrics—particularly the RV-SB ratio (RV-SB actual/predicted; RV%pred)—were assessed. A multivariable prediction model was constructed to identify the most informative CT biomarkers for early risk stratification of SAD.

Statistical analysis

For quantitative data, the independent samples t-test or Mann-Whitney U test was used to compare the differences between groups. The Chi-squared test or Fisher’s exact test was used to compare categorical variables. Pearson correlation coefficient and Spearman’s rank correlation coefficient were used to evaluate the correlation between continuous variables according to different data distributions. All potential related variables were analyzed by univariate and multivariate analysis. All CT-derived quantitative predictors were treated as continuous variables without transformation. No standardization or normalization was applied prior to model building. Logistic regression models were used for prediction. Predictors were selected based on prior knowledge, univariate analysis results, and clinical relevance, without automated selection procedures such as stepwise regression or regularization methods. Model performance was internally validated using five-fold cross-validation. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, calibration curve, and decision curve analysis (DCA). The data analysis of this study used a variety of statistical methods, including significance analysis, correlation analysis, and logistic regression analysis. Data analysis was performed by using Python and R Statistical Software, and the statistical significance level was set at P<0.05.

Results

Baseline demographic and clinical characteristics of participants

Participants were stratified by the consensus threshold into an RV-normal group (n=48) and an RV-elevated group (n=23) (Table 1). The RV-elevated group was older than the RV-normal group [19.130±0.869 vs. 18.708±0.798 years; mean difference, 0.42 years; 95% confidence interval (CI): 0.00–0.84; P=0.047]. The absolute age difference was 0.42 years within a narrow 18–22-year window, which is unlikely to be clinically meaningful. Weight (63.696±5.269 vs. 65.917±5.896 kg, P=0.13) and height (174.478±4.273 vs. 173.688±4.203 cm, P=0.46) did not differ significantly between groups. P values were obtained from independent-samples t-tests.

Between-group differences in biphasic qCT metrics

Representative paired inspiratory-expiratory images illustrate qualitative differences between groups: relative to an RV-normal participant, an RV-elevated participant shows visually larger expiratory low-attenuation regions and a more muted inspiratory-expiratory density shift (Figure 1).

Figure 1 Representative biphasic CT images stratified by RV%pred. (A) RV-normal group (RV%pred <120%): inspiratory (left) and expiratory (right) images show only small, scattered expiratory low-attenuation areas (green), with a clear increase in parenchymal attenuation and visible lung-volume reduction on expiration. (B) RV-elevated group (RV%pred ≥120%): inspiratory (left) and expiratory (right) images demonstrate more extensive, confluent expiratory low-attenuation regions persisting on expiration (green), together with a smaller inspiratory-expiratory density shift—features consistent with a subclinical gas-trapping/ventilation-heterogeneity phenotype. Green overlays indicate expiratory low-attenuation regions automatically identified on expiratory images by threshold-based analysis. CT, computed tomography; HU, Hounsfield units; LV, lung volume ratio (expiration/inspiration); MLD_ex/in, mean lung density ratio (expiration/inspiration); MLD_ex-in, mean lung density difference (expiration − inspiration); Perc15, 15^th percentile lung density; RV, residual volume; RV%pred, residual volume percent predicted; RVC, relative volume change (−950 to −860 HU); VDR, volume decrease ratio (insp − exp)/insp; VI-950/VI-910/VI-856, inspiratory (−950 HU), inspiratory (−910 HU), and expiratory (−856 HU) low-attenuation lung volumes.

Between the RV-normal (n=48) and RV-elevated (n=23) groups, five of nine biphasic qCT metrics differed significantly (Figure 2; Table S1). MLD_ex/in was higher in the RV-elevated group {median 85.40 [interquartile range (IQR), 82.78–87.85]} than in the RV-normal group [median 80.39 (IQR, 75.70–85.08), P=0.005]. MLD_ex-in was lower in the RV-elevated group [median 124.0 (IQR, 102.5–146.5)] than in the RV-normal group [median 160.0 (IQR, 125.5–206.8), P=0.007]. Low-attenuation volumes were greater in the RV-elevated group for VL950 [median 1.14 (IQR, 0.71–1.48) vs. 0.62 (IQR, 0.34–1.14), P=0.01], VI-910 [median 21.27 (IQR, 13.48–33.26) vs. 13.13 (IQR, 7.35–23.06), P=0.03], and VI-856 [median 4.36 (IQR, 2.41–8.84) vs. 1.10 (IQR, 0.37–3.75), P<0.001]. By contrast, LV (mean 53.03±7.67 vs. 48.16±10.92) and VDR (mean 46.97±7.67 vs. 51.84±10.92) showed non-significant trends (both P=0.06), and differences in Perc15 [median −822.0 (IQR, −828.0 to −816.5) vs. −827.0 (IQR, −830.0 to −819.0), P=0.19] and RVC [median −67.37 (IQR, −71.94 to −54.94) vs. −62.10 (IQR, −71.64 to −56.40), P=0.69] were not statistically significant.

Figure 2 Intergroup differences of nine qCT parameters after grouping based on RV%pred (boxplots). Significant differences were observed between groups in parameters including MLD_ex-in, MLD_ex/in, VI-856, VI-910, and VI-950 (P<0.05). HU, Hounsfield units; LV, lung volume ratio (expiration/inspiration); MLD_ex/in, mean lung density ratio (expiration/inspiration); MLD_ex-in, mean lung density difference (expiration − inspiration); Perc15, 15^th percentile lung density; qCT, quantitative computed tomography; RV, residual volume; RV-SB, RV measured by SB dilution; RV%pred, residual volume percent predicted; RVC, relative volume change (−950 to −860 HU); SB, single-breath; VDR, volume decrease ratio (insp − exp)/insp; VI-950/VI-910/VI-856, inspiratory (−950 HU), inspiratory (−910 HU), and expiratory (−856 HU) low-attenuation lung volumes.

Collectively, the higher MLD_ex/in together with the lower MLD_ex-in and greater low-attenuation volumes indicate a smaller expiratory-inspiratory density shift and a higher low-attenuation burden in the RV-elevated group, consistent with the qualitative differences illustrated in Figure 1.

Correlation analysis between RV%pred and biphasic qCT metrics

Pairwise correlations were calculated between the RV-SB measured/predicted ratio (RV%pred) and prespecified biphasic qCT metrics (Figure 3, Table 2). RV%pred showed moderate positive correlations with VI-856 (expiration) (r=0.493, P=1.23×10⁻⁵), MLD_ex/in (r=0.457, P=6.12×10⁻⁵), and LV (r=0.415, P=3.23×10⁻⁴), and moderate negative correlations with MLD_ex-in (r=−0.447, P=9.46 ×10⁻⁵) and VDR (r=−0.415, P=3.23×10⁻⁴). A weaker but significant correlation was observed for VI-950 (inspiration) (r=0.254, P=0.03). Correlations with VI-910 (inspiration) (r=0.182, P=0.13), Perc15 (r=0.202, P=0.09), and RVC (r=0.088, P=0.47) were small and non-significant.

Figure 3 Heatmap of pairwise correlation coefficients (r) between biphasic qCT metrics and pulmonary function indices. Values denote r; colors encode the direction and magnitude of r. Act, actual; DLCOcSB, carbon monoxide transfer coefficient adjusted for SB; DLCOcVA, carbon monoxide transfer coefficient adjusted for alveolar volume; FEV1, forced expiratory volume; FRC, functional residual capacity; FVC, forced vital capacity; HU, Hounsfield units; LV, lung volume ratio (expiration/inspiration); MEF, maximal expiratory flow; MLD_ex-in, mean lung density difference (expiration − inspiration); MLD_ex/in, mean lung density ratio (expiration/inspiration); MMEF, maximal mid-expiratory flow; PEF, peak expiratory flow; Perc15, 15th percentile lung density; Pre, predicted; qCT, quantitative computed tomography; RV, residual volume; RV-SB, RV measured by SB dilution; RVC, relative volume change (−950 to −860 HU); SB, single-breath; TLC-SB, total lung capacity measured by SB dilution; VDR, volume decrease ratio (insp − exp)/insp; VI-950/VI-910/VI-856, inspiratory (−950 HU).

Overall, expiratory low-attenuation burden (VI-856) and density-dynamic indices (MLD_ex/in, MLD_ex-in) exhibited the closest coupling with the RV phenotype captured by RV%pred, whereas LV and VDR provided redundant, directionally opposite information by definition, and VI-950 showed a weaker association.

Model performance and feature importance analysis

A logistic regression model was built using prespecified biphasic qCT metrics (MLD_ex/in, MLD_ex-in, VI-950, VI-910, VI-856, LV, VDR, RVC, Perc15) to classify participants into the RV-elevated vs. RV-normal strata. Using five-fold cross-validation, the model showed moderate-to-good discrimination (Table 3): AUC was 0.862 in the training folds and 0.788 in the validation folds; corresponding accuracy was 0.761 and 0.732, specificity was 0.750 and 0.708, and sensitivity was 0.783 and 0.783, respectively.

Standardized coefficients (“relative weights”) are displayed in Figure 4 and detailed in Table 4. In terms of absolute contribution, the dominant predictors were VI-856 (expiration) (|weight| =1.00), MLD_ex/in (0.903), VI-950 (inspiration) (0.880), and MLD_ex-in (0.860), followed by VI-910 (0.463) and LV (0.274). RVC (0.303 in magnitude) and VDR (0.274) contributed modestly, whereas Perc15 had a minimal weight (0.082). The signs of the coefficients indicate directionality of association with the RV-elevated class (positive values increase, negative values decrease the predicted log-odds). Overall, density-dynamic indices (MLD_ex/in, MLD_ex-in) and low-attenuation metrics (particularly VI-856 and VI-950) carried the greatest information for stratification, which is consistent with the between-group differences and correlation patterns reported above.

Figure 4 Feature importance plot. The X-axis represents feature coefficients, indicating the contribution of each qCT parameter to the predictive model. Larger absolute values indicate stronger contributions to the predicted log-odds. Coefficient signs match Table 4. HU, Hounsfield units; LV, lung volume ratio (expiration/inspiration); MLD_ex/in, mean lung density ratio (expiration/inspiration); MLD_ex-in, mean lung density difference (expiration − inspiration); Perc15, 15^th percentile lung density; qCT, quantitative computed tomography; RVC, relative volume change (−950 to −860 HU); VDR, volume decrease ratio (insp − exp)/insp; VI-950/VI-910/VI-856, inspiratory (−950 HU), inspiratory (−910 HU), and expiratory (−856 HU) low-attenuation lung volumes.

The final logistic regression model for classifying RV-elevated vs. RV-normal was: logit{p(RV-elevated)} = −0.3689 + 0.7970 × MLD_ex/in + (−0.7597) × MLD_ex-in + (−0.8830) × VI-856 (expiration) + 0.7773 × VI-950 (inspiration) + 0.4088 × VI-910 (inspiration) + 0.2415 × LV + (−0.2415) × VDR + (−0.2673) × RVC + 0.0723 × Perc15. The predicted probability is obtained as p = 1 / {1 + exp[−logit(p)]}. Positive coefficients increase, and negative coefficients decrease, the log-odds of being classified as RV-elevated [note: coefficients are the unstandardized estimates corresponding to Table 4 (“coefficient”); “relative weight” values in Table 4 are standardized magnitudes for interpretability and do not enter the prediction formula].

Calibration and decision analysis

Calibration reliability curves based on 10 equal-width bins are shown in Figure 5. In the training set, the curve tracked the 45° reference line across most of the probability range, with mild underestimation at low predicted risks (<0.2) and slight overestimation near the mid-range (≈0.5–0.6). The validation set showed a broadly similar pattern, with modest departures in the mid–high range (≈0.4–0.8). Histograms indicate adequate dispersion of predicted risks. Overall, predicted probabilities were reasonably calibrated in both datasets.

Figure 5 Calibration plots (reliability curves) for the logistic regression model. Panels show calibration in (A) the training set and (B) the validation set, using 10 equal-width probability bins. The dotted 45° line denotes perfect calibration; the solid line with square markers shows the observed event fraction within each bin. Lower panels display the distribution of predicted probabilities (histograms). Overall, the model demonstrates reasonable calibration with mild departures at the extremes.

DCA (Figure 6) demonstrated that the model outperformed the treat-all and treat-none strategies across a clinically relevant range of thresholds, with higher net benefit for approximately 0.05–0.60 in the training set and 0.05–0.55 in the validation set (peaking near 0.55–0.60). These findings support the potential clinical utility of the model for risk stratification at moderate decision thresholds.

Figure 6 DCA of clinical net benefit. Net-benefit curves are plotted against decision threshold probabilities for the training (red) and validation (teal) sets. Gray and black lines represent “treat-all” and “treat-none” strategies, respectively. The model yields higher net benefit than both comparators across a clinically relevant threshold range (approximately 0.05–0.60 for training and 0.05–0.55 for validation). DCA, decision curve analysis; Val, validation set.

Discussion

This study evaluated whether respiratory biphasic qCT captures a subclinical gas-trapping/ventilation-heterogeneity phenotype corresponding to an elevated RV defined by consensus (RV%pred ≥120%). Three consistent observations emerged. First, despite broadly similar spirometric profiles, the RV-elevated group exhibited a higher expiratory low-attenuation burden (VI-856) and shifted density dynamics—higher MLD_ex/in and lower MLD_ex-in—together with greater inspiratory low-attenuation fractions (VI-950/VI-910). Second, RV%pred correlated moderately with VI-856 and with density-dynamic indices (MLD_ex/in and MLD_ex-in), indicating physiological coherence between CT-derived markers of air trapping and the RV phenotype. Third, a parsimonious qCT-based logistic model achieved reasonable discrimination (AUC =0.862 in training and 0.788 in validation), acceptable sensitivity/specificity balance, and good calibration with decision-curve net benefit across clinically relevant thresholds.

Small airways (<2 mm in diameter) are an early site of involvement across several chronic respiratory diseases (17-20). Because they contribute little to total airway resistance, early dysfunction is often clinically inapparent—hence the “silent zone”—and conventional spirometry, including maximal mid-expiratory flow (MMEF; also termed FEF25–75) and forced expiratory flow at 50% and 75% of FVC (FEF50, FEF75), has limited repeatability and sensitivity for subtle peripheral changes (21). This context motivates paired inspiratory-expiratory qCT, which can localize regional density dynamics and expiratory low-attenuation burden at a subsegmental/regional level.

Biphasic qCT indices probe complementary dimensions of small-airway physiology. MLD_ex/in and MLD_ex-in quantify the inspiratory-to-expiratory density excursion; a higher MLD_ex/in together with a lower MLD_ex-in indicates a blunted density swing, consistent with impaired emptying or early airway closure. The expiratory low-attenuation fraction VI-856 captures the lung volume that remains below −856 HU on expiration and serves as an imaging surrogate of regional gas-trapping. In this non-smoking cohort of young adults, inspiratory low-attenuation fractions (VI-950, VI-910) are more likely to reflect relatively low parenchymal density/hyperinflation at total lung capacity than established emphysema. In contrast, LV and VDR summarize global volume change and are more susceptible to breath-hold depth; RVC (−950 to −860 HU) tracks redistribution within a defined density band, and Perc15 characterizes the lower tail of the attenuation histogram. Collectively, the density-dynamic indices and the expiratory low-attenuation burden are the features most directly aligned with air-trapping behavior and with elevations in RV.

RV—the alveolar gas remaining after maximal expiration—provides a standardized readout of gas-trapping/hyperinflation when expressed as RV%pred, reflecting premature airway closure and prolonged regional time constants. In this well-coached, non-obstructed cohort, RV was measured using the SB gas-dilution method and analyzed as RV%pred, which we used as the stratification variable aligned with the consensus threshold for abnormal gas-trapping (RV%pred ≥120%) (11). Prior work indicates that RV-based indices are relatively sensitive to peripheral small-airway dysfunction (22). Consistent with this framework, in our dataset, RV%pred showed stronger correlations and between-group separability with qCT density-dynamic indices and expiratory low-attenuation burden than alternative lung-volume candidates (Table S2), supporting its role as the anchoring variable in this validation-oriented study.

Our results support both criterion validity—agreement with the consensus threshold for gas-trapping (RV%pred ≥120%)—and construct validity, evidenced by physiologically coherent patterns: higher VI-856, higher MLD_ex/in, and lower MLD_ex-in among participants with elevated RV. In a young cohort with largely unremarkable flow-volume spirometric indices (with RV-based stratification identifying a subset with elevated RV), biphasic qCT delineated a subclinical gas-trapping phenotype aligned with this pre-specified, consensus-based operational definition. Therefore, this model is suitable as a screening/risk-estimation aid for early risk stratification and monitoring. Whether this phenotype predicts future clinically meaningful small-airway dysfunction warrants longitudinal evaluation.

This work pairs density-dynamic and low-attenuation volumetric metrics within a single model, provides an explicit fitted equation and probability calculation to facilitate reproducibility and external validation, and demonstrates reasonable calibration and decision-analytic net benefit across thresholds of approximately 0.2–0.6, where clinical trade-offs are most relevant. Internal validity benefited from standardized breathing coaching and high participant compliance, reducing breath-level variability that can confound expiratory imaging.

This single-center study had a modest sample size and a narrow demographic (male aviation-major students aged 18–22 years), reflecting the program’s intake structure and the finite pool of eligible participants during the study period. This configuration facilitated stringent quality control for pulmonary function testing and biphasic CT acquisitions, thereby reducing measurement error and breath-level variability and yielding more precise within-cohort qCT-RV estimates, more stable cross-validation coefficients, and lower calibration noise; however, such homogeneity limits external generalizability. Second, RV was measured using the SB gas-dilution technique rather than body plethysmography or multi-breath dilution; while SB dilution can underestimate lung volumes in marked obstruction, this risk was minimal in our well-coached, non-obstructed cohort. To align with best practice and enhance comparability, future studies will incorporate plethysmographic and multi-breath measurements. Third, percent-predicted values were used instead of z-scores; to improve cross-study comparability, subsequent work will provide dual reporting (percent-predicted plus z-scores). Finally, the cross-sectional design limits inferences regarding temporal change and causality between qCT features and small-airway physiology. Multicenter longitudinal studies in larger and more diverse populations, with gold-standard physiological endpoints, are warranted to assess robustness, transportability, and clinical impact.

Conclusions

Respiratory biphasic qCT—particularly VI-856 and the MLD-based density-dynamic indices—aligns closely with an RV-defined gas-trapping phenotype. A simple qCT-based model yielded calibrated risk estimates and decision-curve net benefit, supporting its role in early risk assessment and individualized monitoring. Multicenter, longitudinal investigations across more diverse populations and with gold-standard physiological endpoints are warranted to test robustness, transportability, and clinical utility.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-969/rc

Data Sharing Statement: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-969/dss

Peer Review File: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-969/prf

Funding: This work was supported by the Key Project of the Air Force Logistics Department (No. BKJ23WS1J001).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-969/coif). J.X., C.H., and H.R. are from Hangzhou Deepwise & League of PHD Technology Co., Ltd. All authors report this work was supported by the Key Project of the Air Force Logistics Department (No. BKJ23WS1J001). The authors have no other conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. It was approved by the Ethics Committee of the Air Force Medical Center (approval No. 2024-68-PJ01). Written informed consent was obtained from all participants prior to their inclusion in the study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. van der Wiel E, ten Hacken NH, Postma DS, et al. Small-airways dysfunction associates with respiratory symptoms and clinical features of asthma: a systematic review. J Allergy Clin Immunol 2013;131:646-57. [Crossref] [PubMed]
2. Xiao D, Chen Z, Wu S, et al. Prevalence and risk factors of small airway dysfunction, and association with smoking, in China: findings from a national cross-sectional study. Lancet Respir Med 2020;8:1081-93. [Crossref] [PubMed]
3. McNulty W, Usmani OS. Techniques of assessing small airways dysfunction. Eur Clin Respir J 2014;
4. Cosio M, Ghezzo H, Hogg JC, et al. The relations between structural changes in small airways and pulmonary-function tests. N Engl J Med 1978;298:1277-81. [Crossref] [PubMed]
5. Lynch DA, Austin JH, Hogg JC, et al. CT-Definable Subtypes of Chronic Obstructive Pulmonary Disease: A Statement of the Fleischner Society. Radiology 2015;277:192-205. [Crossref] [PubMed]
6. Nambu A, Zach J, Schroeder J, et al. Quantitative computed tomography measurements to evaluate airway disease in chronic obstructive pulmonary disease: Relationship to physiological measurements, clinical index and visual assessment of airway disease. Eur J Radiol 2016;85:2144-51. [Crossref] [PubMed]
7. Galbán CJ, Han MK, Boes JL, et al. Computed tomography-based biomarker provides unique signature for diagnosis of COPD phenotypes and disease progression. Nat Med 2012;18:1711-5. [Crossref] [PubMed]
8. Verschakelen JA. Quantitative CT of the Lung to Study Asthma. Radiology 2022;304:460-1. [Crossref] [PubMed]
9. Wu Q, Guo H, Li R, et al. Deep learning and machine learning in CT-based COPD diagnosis: Systematic review and meta-analysis. Int J Med Inform 2025;196:105812. [Crossref] [PubMed]
10. Chinese Medical Association. Guideline for pulmonary function testing in primary care (2018). Chinese Journal of General Practitioners 2019;18:511-8.
11. Zhu L, Chen RC. Chinese experts' consensus on the standardization of adult lung function diagnosis. Journal of Clinical Pulmonary Medicine 2022;27:973-81.
12. Huang Z, Wang X, Wei Y, et al. CCNet: Criss-Cross Attention for Semantic Segmentation. IEEE Trans Pattern Anal Mach Intell 2023;45:6896-908. [Crossref] [PubMed]
13. He K, Zhang X, Ren S, et al. Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016:770-8.
14. Bai Y, Wang X, Zhou Z, et al. Pulmonary segments segmentation with hierarchical weak labels. In: 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI). 2023:1-5.
15. Chen LC, Papandreou G, Kokkinos I, et al. DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. IEEE Trans Pattern Anal Mach Intell 2018;40:834-48. [Crossref] [PubMed]
16. Shi W, Caballero J, Huszár F, et al. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016:1874-83.
17. Macklem PT. The physiology of small airways. Am J Respir Crit Care Med 1998;157:S181-3. [Crossref] [PubMed]
18. Burgel PR. The role of small airways in obstructive airway diseases. Eur Respir Rev 2011;20:23-33. [Crossref] [PubMed]
19. Hogg JC, Chu F, Utokaparch S, et al. The nature of small-airway obstruction in chronic obstructive pulmonary disease. N Engl J Med 2004;350:2645-53. [Crossref] [PubMed]
20. Calzetta L, Aiello M, Frizzelli A, et al. Small airways in asthma: from bench-to-bedside. Minerva Med 2022;113:79-93. [Crossref] [PubMed]
21. Zhao N, Wu F, Peng J, et al. Preserved ratio impaired spirometry is associated with small airway dysfunction and reduced total lung capacity. Respir Res 2022;23:298. [Crossref] [PubMed]
22. Stănescu D. Small airways obstruction syndrome. Chest 1999;116:231-3. [Crossref] [PubMed]