Automated CT‐based measurements of radial and longitudinal expansion of airways due to breathing‐related lung volume change

Abstract

Background

Respiratory function is impaired in chronic obstructive pulmonary disease (COPD). Automation of multi‐volume CT‐based measurements of different components of breathing‐related airway deformations will help understand multi‐pathway impairments in respiratory mechanics in COPD.

Purpose

To develop and evaluate multi‐volume chest CT‐based automated measurements of breathing‐related radial and longitudinal expansion of individual airways between inspiratory and expiratory lung volumes.

Methods

We developed a method to compute breathing‐related airway deformation metrics and applied it to total lung capacity (TLC) and functional residual capacity (FRC) chest CT scans. The computational pipeline involves: (1) segmentation of airways; (2) skeletonization of airways; (3) labeling of anatomical airway segments at TLC and FRC; and (4) computation of radial and longitudinal expansion metrics of individual airways across lung volumes. Radial expansion (∆CSA) of an airway is computed as the percent change of its cross‐sectional area (CSA) between two lung volumes. Longitudinal expansion (∆L) of an airway is computed as the percent change in its airway path‐length from the carina between lung volumes. These measures are summarized at different airway anatomic generations. Agreement of automated measures with their manually derived values was examined in terms of concordance correlation coefficient (CCC) of automated measures with those derived using manual outlining. Intra‐class correlation coefficient (ICC) of automated measures from repeat CT scans (n = 37) was computed to assess repeatability. The method was also applied to a set of participants from the Genetic Epidemiology of COPD (COPDGene) Iowa cohort, distributed across COPD severity groups (n = 4 × 60).

Results

The CCC values for the automated ∆CSA measure with manually derived values were 0.930 at the trachea, 0.898 at primary bronchi, and greater than 0.95 at pre‐segmental and segmental airways; these CCC values were consistently greater than 0.95 for ∆L at all airway generations. ICC values for repeatability of ∆CSA were 0.974, 0.950, 0.943, and 0.901 at trachea, primary bronchi, pre‐segmental, and segmental airways, respectively; these ICC values for ∆L were 0.973, 0.954, and 0.952 at primary bronchi, pre‐segmental, and segmental airways, respectively. ∆CSA values were significantly reduced (p < 0.001) with increasing COPD severity at each of primary bronchi, pre‐segmental, and segmental airways. Significantly lower ∆L values were observed for moderate (p = 0.042 at pre‐segmental and p = 0.037 at segmental) and severe (p = 0.019 at pre‐segmental and p < 0.001 at segmental) COPD groups as compared to the preserved lung function group. Body mass index (BMI) and smoking status were found to significantly associate with ∆CSA at segmental airways (r = 0.17 and −0.19, respectively; significance threshold = 0.13), while age and sex were significantly associated with ∆L (r = −0.21 and −0.17, respectively); COPD severity was significantly associated with both ∆CSA and ∆L (r = −0.35 and −0.22, respectively).

Conclusion

Our CT‐based automated measures of breathing‐related radial and longitudinal expansion of airways are repeatable and in agreement with manually derived values. Automation of different airway mechanical biomarkers and their observed significant associations with age, sex, BMI, smoking, and COPD severity establish an effective tool to investigate multi‐pathway impairments of respiratory mechanics in COPD and other lung diseases.

1. INTRODUCTION

Chronic obstructive pulmonary disease (COPD), a progressive lung disease characterized by airflow limitation, is a major global health concern. ¹ In COPD, lung function including airflow and gas exchange is compromised, and multiple pathways of lung function impairment have been established in COPD that may alter lung respiratory mechanics differentially. ² , ³ Emphysema ⁴ , ⁵ is the destruction of air sacs and lung tissues disrupting the airway‐parenchymal coupling ¹³ and altering parenchymal mechanics. ⁶ , ⁷ Similarly, chronic airway inflammation causes airway constriction, wall thickening, and bronchitis that can also drive changes in airway stiffness, collapsibility, and parenchymal coupling. ⁸ , ⁹ Moreover, these mechanical changes can lead to air trapping that manifests as hyperinflated lungs ¹⁰ at end‐expiratory lung volumes leading to permanently expanded rib cage or barrel chest ¹¹ , ¹² , ¹³ and diaphragm flattening ¹⁴ , ¹⁵ that impairs elastic recoil needed for effective breathing. These findings suggest multi‐pathway effects of emphysema, bronchitis, air trapping, and hyperinflation on respiratory mechanical impairments of airways and their parenchymal coupling in COPD. Plopper et al. ¹⁶ have explained the tethering of breathing‐related radial and longitudinal expansion of airways within the lung and their unique relationships with respiratory mechanics of different pulmonary anatomic structures. Therefore, decoupling the two different breathing‐related deformation components of airways will facilitate quantitative characterization of multi‐pathway respiratory mechanical impairments of airways and their parenchymal coupling with aging, smoking, COPD, and other lung diseases.

There are several nationwide established studies investigating COPD, ¹⁷ , ¹⁸ , ¹⁹ , ²⁰ which have generated large repositories of longitudinal data involving inspiratory and expiratory chest CT imaging, questionnaires, demographics, as well as biological and clinical outcomes. These databases along with established computational methods have been widely applied to investigate emphysema, ²¹ , ²² air trapping, ²³ , ²⁴ airway counts and morphology, ²⁵ , ²⁶ , ²⁷ and multi‐volume determined metrics of functional small airway disease (fSAD) ²⁸ in COPD. Several computational methods measuring airway morphological features such as wall area percentage (WA%), lumen cross‐sectional area (CSA), wall thickness, and model‐estimated airway wall thickness at an airway internal perimeter of 10 mm (Pi10) have been established and applied in research studies related to COPD and other lung diseases. ²⁵ , ²⁷ Haghighi et al. performed imaging‐based cluster analysis of CT‐based structural and physiological variables, including those related to airway tree morphology, and observed that the derived clusters are attributed with unique clinical characteristics. ²⁹ Kim et al. found that CT‐based metrics of airway wall thickness and emphysema demonstrated an inverse relationship when assessed in severe COPD. ²² Kirby et al. found a significant decline in total airway counts was associated with an increase in COPD GOLD status. ²⁶ , ³⁰ Shim et al. applied a CT‐based method involving human interaction and correction to compute percentage change in airway lumen area between inspiratory and expiratory lung volume or “delta‐lumen” and observed that the delta‐lumen measure correlated with severity and number of exacerbation in asthmatic patients. ³¹ These findings on descriptive biomarkers offer little insight into respiratory mechanical impairments in COPD. An automated and reliable method for quantitative respiratory mechanical biomarkers characterizing different features of breathing‐related airway mechanics will be effective in studying mechanical impairments in COPD and other lung disease.

In this study, we develop and validate a multi‐volume chest CT‐based automated method to quantify breathing‐related radial and longitudinal expansion of individual airways. Prior work by Shim et al. required time‐consuming manual evaluation and comparison at equivalent airway branch labels across inspiratory and expiratory lung volumes. ³¹ Automation will enable time‐efficient and repeatable research studies investigating multi‐pathway alterations of airway respiratory mechanics in COPD and other lung diseases, their associations with pathological conditions and disease severity, and impacts of airway mechanical alterations in disease progression and clinical outcomes. Moreover, decoupling of respiratory airway deformation into radial and longitudinal expansion components is novel, and it will help understand multi‐pathway respiratory impairments in COPD or other lung diseases.

Our overall computational pipeline integrates previously validated airway segmentation and labeling algorithms ³² , ³³ and computes CSA and path‐length matched to specific airway branches in inspiratory (total lung capacity—TLC) and expiratory (functional residual capacity—FRC) lung volumes up to the segmental level. Synthesis of different approaches that makes it feasible to automate multi‐volume CT‐based airway deformation measurements is another important advancement presented in this paper. Agreements of automated measures of breathing‐related radial and longitudinal expansion of airways between TLC and FRC with their manually derived values were evaluated. Also, the repeatability of these automated measures was examined. Finally, the method was applied to a subset of the Genetic Epidemiology of COPD (COPDGene) ¹⁷ Iowa cohort population to examine associations of different airway mechanical biomarkers with age, sex, body mass index (BMI), smoking status, and COPD severity.

2. MATERIALS AND METHODS

A block diagram of the image processing pipeline for computing the radial expansion metric at the left medial bronchi (LMB) is illustrated in Figure 1. Specifically, the pipeline involves the following steps: (1) CT‐based airway segmentation and skeletonization, (2) airway branch labeling, (3) delineation of lumen volume over a local orthogonal slab around the mid‐airway skeletal voxel, (4) computation of airway lumen CSA at a specific airway branch, and (5) computation of the radial expansion metric as the percent change in CSA at the specific matching branch between inspiratory and expiratory lung volumes. A similar pipeline is applied for computation of the branch‐level longitudinal expansion metric by replacing Steps 3 to 5 as follows. After airway branch labeling, geodesic distance between carina, the location of tracheal bifurcation, and the end of a target branch is computed. Finally, longitudinal expansion at a target branch is computed as the percent change in its path‐length measure between inspiratory and expiratory lung volumes. Note that the path‐length between carina and the end of a specific branch represents the geodesic length of the bronchial pathway from carina to the end of that branch. It may be further clarified that branch‐level longitudinal expansion characterizes along‐the‐path stretching of the entire bronchial pathway from the carina to the end of the specific branch.

FIGURE 1.

Inspiratory and expiratory chest CT‐based computation of breathing‐related radial expansion of individual airways. Airway lumen cross‐sectional area (CSA) is computed at a matching branch from inspiratory and expiratory chest CT scans. The radial expansion at a specific branch is computed as the percent change in its airway lumen CSA between inspiratory and expiratory lung volumes.

2.1. Chest CT datasets

The following two datasets were used for our evaluative experiments.

Data_rep (n = 37) consists of TLC and FRC CT scans, which were used for evaluation of repeatability and agreement with manually defined measures. This dataset includes three CT scans for each of 37 participants. For each participant, baseline TLC and FRC CT scans were acquired, and one‐third or repeat CT scan either at TLC or FRC was acquired. Prior to the repeat scan, the participant was removed from the scanner, asked to stand, and then repositioned back onto the same scanner table after a stopgap of 3–5 min. Participants were randomly selected for a repeat TLC or a repeat FRC scan, which resulted in 20 participants with repeat TLC scans and 17 participants with repeat FRC scans. ³⁴ Participants in this dataset were selected over an age range of 20–90 years. Other inclusion criteria were current or ex‐smokers with BMI < 32 kg/m² and weight < 100 kg, and normal pulmonary function tests (PFTs).

Data_group (n = 240) consists of randomly selected participants from the Iowa cohort of the COPDGene study (ClinicalTrials.gov: NCT00608764) at their first follow‐up visits with matching TLC and FRC chest CT scans. Following the COPDGene study design, participants in this dataset were between the ages of 45 and 80 years and current or ex‐smokers with at least 10 pack‐years of smoking history at baseline visits. For our study, participants were selected from different COPD severity groups based on their Global Initiative for Chronic Obstructive Lung Disease (GOLD) status defined using forced expiratory volume in 1 s (FEV1) and its ratio to forced vital capacity (FEV1/FVC) measures. Specifically, COPD severity groups were defined as follows: (1) preserved lung function (GOLD 0), (2) mild COPD (GOLD 1 and preserved ratio impaired spirometry (PRISm), (3) moderate COPD (GOLD 2), and (4) severe COPD (GOLD 3 and 4). Sixty (30 males) participants were randomly selected from each COPD severity group.

There was no overlap between the participants in Data_rep and Data_group. For both datasets, their parent studies were approved by the University of Iowa Institutional Review Board, and the usage of the data under the current study was Health Insurance Portability and Accountability Act (HIPAA) compliant. Written and informed consents were obtained from all participants. All participants from Data_group were scanned on a Siemens (Forchheim, Germany) Definition Flash 128 scanner. Twenty‐eight participants from Data_rep were scanned on a Siemens Sensation 64 scanner, and the others were scanned on the Siemens Definition Flash 128 scanner using the same CT imaging protocol. ³⁴ Details of CT acquisition protocols for both datasets are summarized in Table 1.

TABLE 1.

Chest CT acquisition protocols for the two datasets: Data_rep (n = 37) and Data_group (n = 240).

	Dataset
	Datarep	Datagroup
Number of participants (n)	37	240
Tube voltage (kVp)	120	120
Tube current (mAs)	110	200 for TLC 50 for FRC
Pitch	1.0	1.1
Slice thickness (mm)	0.75	0.75
Slice spacing (mm)	0.5	0.5
Pixel size (mean ± SD) (mm)	0.62 ± 0.07	0.64 ± 0.08
Array size	512 × 512	512 × 512
Reconstruction kernel	B35f	B40d

For a specific dataset, same parameters were used for both total lung capacity (TLC) and functional residual capacity (FRC) CT scans except for the tube current for Data_group.

Abbreviation: SD standard deviation.

2.2. CT‐based airway lumen segmentation

A previously published airway segmentation algorithm ³³ , ³⁵ combining deep learning (DL) and multi‐parametric freeze‐and‐grow (FG) methods is applied to segment the airway tree from a chest CT scan; see Figure 2. The DL module was implemented using a modified three‐dimensional (3‐D) U‐Net ³⁶ , ³⁷ with three pooling and three deconvolutional layers and was trained to generate a voxel‐level airway lumen likelihood map from an input CT image over a 64 × 64 × 64 voxel patch. These patches of the airway lumen likelihood map are stitched together to reconstruct the likelihood map over the entire CT image space. The DL module used in the previous algorithm ³³ was trained only on TLC CT scans from the SubPopulations and InteRmediate Outcome Measures in COPD Study (SPIROMICS), ¹⁸ which includes CT images from different study sites using different scanners. Therefore, the DL module was retrained using both TLC and FRC scans of 50 participants from COPDGene Phase 2 and transfer learning. ³⁸ , ³⁹ These 50 participants were uniformly distributed over COPD severity and sex and were exclusive from Data_group as well as Data_rep used for evaluative experiments. The iterative FG module is applied to the airway lumen likelihood map to segment the airway lumen volume. Starting with a conservative threshold parameter, the FG method adopts iterative parameter relaxation to grow and capture finer details while using forbidden regions and connectivity freezing around leakages.

FIGURE 2.

Illustrations of airway lumen segmentation from a chest CT scan. Airway lumen volume was segmented using a previously validated automated algorithm. ³³ , ³⁵ (a–c) Segmentation results from a chest CT scan at total lung capacity: (a) segmented airway lumen volume; (b) airway lumen regions (red) overlaid on a coronal CT slice; (c) same as (b) on an axial slice. CT contrast settings: level = −450 Hounsfield units (HU); window = 1200 HU.

The airway tree skeleton, a single‐voxel thin centerline representation, is computed using centralized geodesic paths and local scale‐based pruning approaches. ³³ , ⁴⁰ The previously published airway segmentation method was previously validated in terms of branch detection accuracy. In this study, we use the initial segmentation of lumen volume to delineate lumen boundaries in local orthogonal planes and evaluate its performance.

2.3. Airway branch labeling

Standardized anatomical labels of airway branches ⁴¹ up to the segmental level are shown in Figure 3. Airways in the first three generations, shown in red, follow a fixed branching pattern; local variabilities are observed beyond this generation. Our algorithm applies a rule‐based tree traversal approach to locate the first three generations of branches and adopts a two‐stage neural network (NN) classifier beyond third generation to detect segmental branches.

FIGURE 3.

Illustration of anatomical airway branches used for analysis. Trachea (blue), primary bronchi (green), and pre‐segmental (yellow) as well as segmental (red) airway branches at different lobes are labeled.

A single classifier is used for Stage 1 classification, and the primary goal of this classifier is to differentiate valid anatomic branches from insignificant topological branches often caused by variations in airway branching patterns. A fully connected NN with three hidden layers consisting of 256, 512, and 1024 neurons was used to implement the Stage 1 classifier. Rectified linear unit (ReLu) activation was adopted for hidden layers, while sigmoid activation was applied to the output layer, which generated the likelihood of a given branch being an anatomical one. Multiple NN classifiers are used for Stage 2 classification identifying specific anatomical segmental airway branches. Only valid anatomic branches recognized at Stage 1 classification are used at this stage. One NN classifier is dedicated for each lobe. Let us consider a specific lobe, say, the right upper lobe. The Stage 2 classifier for the specific lobe was trained using valid branches in the right upper lobe beyond the branch RULB. During the training process, three distinct class labels were assigned for the three target segmental bronchi RB1, RB2, RB3, while a fourth label was assigned to all other valid anatomic branches in the lobe including those in the sub‐ and sub‐sub‐segmental levels. Similar to the Stage 1 classifier, a fully connected NN with three hidden layers consisting of 128, 256, and 512 neurons was used to implement each lobe‐specific Stage 2 classifier. ReLu activation was used for hidden layers, while softmax activation was applied to the output layer to predict class output.

Both Stage 1 and 2 classifiers were separately trained for TLC and FRC CT scans using manual airway labeling as reference. The same 50 participants, used to train the DL module of airway segmentation, were used to re‐train these two classifiers. For both Stage 1 and 2 classifiers, hierarchical features are used to represent a candidate branch as a feature vector, which includes features from its parent and grandparent branches as well as all those from all sibling and immediate child branches. The following geometric and topologic features are used: (1) topological branch generation number, (2) Euclidean and geodesic branch length, (3) projected branch length along each image coordinate axis, (4) path‐length from the carina, (5) Euclidean distance of the branch to the carina along each image coordinate axis, and (6) branch angle with each image coordinate axis. Additionally, the number of siblings and children are used resulting in a feature vector of length “’90.” A preliminary version of the airway branch labeling method with its validation was previously reported in a conference paper. ³²

2.4. Computation of lumen cross‐sectional area and path‐length at individual airways

Lumen CSA and path‐length at individual airway branches are independently computed at TLC and FRC. At either lung volume, lumen CSA at a specific airway is computed over the middle half of its centerline representation (Figure 4). First, a B‐spline representation of the digital centerline of the specific airway is computed, the middle half is located, and a predefined number of uniform sample points are defined on the middle half. At each sample point $s$, a two‐dimensional (2‐D) digital image is computed on the plane orthogonal to the local centerline. The image is generated at an isotropic resolution of 0.25 × 0.25 mm², and lumen boundary is detected on the image using 72 radial lines emanating from the sample point $s$ at 5° angular separation. A locally adaptive half‐max‐based threshold is used to locate lumen boundary on individual radial lines. On a given orthogonal digital image, the values of minimum and maximum CT intensity are determined at the 5th and 95th percentile intensity values, respectively, over a dilated region of the initial lumen segmentation obtained during airway segmentation described in Section 2.2. The amount of dilation is defined using local airway scale computed as the largest distance transform value in the initial lumen segmentation on the specific orthogonal plane. This step of lumen boundary delineation on an orthogonal plane was previously presented in our conference paper. ⁴²

FIGURE 4.

Computation of airway lumen cross‐sectional area (CSA) at an airway branch. (a) Selection of lumen region for CSA computation over the middle half of centerline representation of a target airway branch. (b) Polygonal representation of lumen boundaries on locally orthogonal planes at sample points on the centerline and lumen volume between successive sample points using a polyhedral representation. The curve segment ${\mathit{\alpha }}_{\mathit{l}}$ of the B‐spline centerline representation between two successive sample points and associated lumen volume ${\mathit{v}}_{\mathit{i}}$ are shown. Lumen CSA at the target branch is computed as ${\sum }_{\mathit{i}=1}^{\mathit{n}-1}{\mathit{v}}_{\mathit{i}}/((\mathit{n}-1)\times {\mathit{\alpha }}_{\mathit{l}})$, where $\mathit{n}$ is the number of sample points and ${\mathit{\alpha }}_{\mathit{l}}$ is the fixed geodesic distance between two successive sample points.

After locating lumen boundary points along individual radial lines, a polygonal representation of the lumen boundary is obtained on each orthogonal plane (Figure 4). The lumen volume $v_i$ between two orthogonal planes at two successive sample points is defined as a polyhedron defined by joining matching polygonal vertices of the lumen boundary at two successive orthogonal planes. Finally, the lumen CSA at the airway branch is computed as ${\sum }_{i=1}^{n-1}{v}_i/((n-1)\times {\alpha }_l)$, where $n$ is the number of sample points on the central half of the airway centerline and ${\alpha }_l$ is the fixed geodesic distance between two successive sample points. The path‐length of an airway branch is computed as the geodesic distance along the smoothened airway centerline from the tracheal bifurcation at the carina to the end of that airway branch.

2.5. Breathing‐related radial and longitudinal expansion metrics for airways

Radial and longitudinal expansion of individual airways between inspiratory and expiratory lung volumes are measured to characterize breathing‐related mechanical properties of airways. Specifically, a deformation metric is defined as the percent change in its values at the matching airway branch between inspiratory (TLC) and expiratory (FRC) lung volumes. Let $b$ denote a specific anatomic airway branch and ${\mathrm{CSA}}_{\mathrm{ins}}(b)$ and ${\mathrm{CSA}}_{\exp }(b)$ denote the airway lumen CSA at $b$ computed from inspiratory and expiratory chest CT scans, respectively. The radial expansion at $b$, denoted by ΔCSA(b), is computed as:

$$\mathrm{\Delta }\mathrm{CSA}(b)=\frac{{\mathrm{CSA}}_{\mathrm{ins}}(b)-{\mathrm{CSA}}_{\exp }(b)}{{\mathrm{CSA}}_{\mathrm{ins}}(b)}\times 100\%$$ (1)

Similarly, the longitudinal expansion at $b$, denoted by $\mathrm{\Delta }\mathrm{L}(b)$, is computed as:

$$\mathrm{\Delta }\mathrm{L}\left(b\right)=\frac{{\mathrm{L}}_{\mathrm{ins}}\left(b\right)-{\mathrm{L}}_{\exp }\left(b\right)}{{\mathrm{L}}_{\mathrm{ins}}\left(b\right)}\times 100\%,$$ (2)

where ${\mathrm{L}}_{\mathrm{ins}}(b)$ and ${\mathrm{L}}_{\exp }(b)$ are the path‐lengths at $b$ computed from inspiratory and expiratory chest CT scans, respectively.

2.6. Experiments

Experimental methods were designed to evaluate: (1) agreement of automated measures of CSA and path‐length with respective manual measures at TLC and FRC, (2) agreement of automated measures of ΔCSA and ΔL with their manual measures, and (3) repeatability of automated measures of CSA and path‐length at TLC and FRC and of ΔCSA and ΔL. Also, repeatability of automated measures of CSA, path‐length, ΔCSA, and ΔL was compared with repeatability of manually derived measures. These evaluative experiments were performed on Data_rep. Data_group was used for comparative analysis of breathing‐related airway deformation metrics between different COPD groups and to examine their associations with age, sex, BMI, smoking status, and COPD severity. All computational experiments were performed on a server computer with Intel Xeon Gold 6420 CPU and a 32 GB NVIDIA Tesla V100 graphics card. Materials and methods involved in our experiments are described in the following.

2.6.1. Manual definition of different measures

Manual inputs were used to generate reference values for different metrics. Specifically, manual outlining of airway lumen boundaries on locally orthogonal 2‐D digital images were performed using 3‐D Slicer ⁴³ to obtain references values for ∆CSA at individual branches. For each branch, five locally orthogonal planes and corresponding 2‐D images, uniformly distributed over the middle half of a branch, were used to derive the reference CSA value at a specific branch. Reference values of airway path‐lengths were obtained by manually identifying individual anatomic branches and locating the two centerline branching points defining the beginning and ending of a specific branch. It was performed using the tools provided within ITK‐SNAP. ⁴⁴

Both manual and fully automated methods were applied to compute radial and longitudinal expansion metrics at 25 anatomical airway branches up to the segmental level (Figure 3), and summary measurements were obtained for four anatomic tree generations: (1) trachea; (2) primary bronchi (right medial bronchus (RMB) and LMB); (3) four pre‐segmental branches; and (4) nine segmental branches in each lung. Note that ∆L cannot be reliably computed for the trachea since the visible extent of the trachea in the CT images was based on the image field‐of‐view (FOV), which was set to contain all parts of the lungs while maximizing in‐plane resolution.

2.6.2. Data analysis

Static CSA and breathing‐related radial expansion measures of different airways at each of the trachea, primary bronchi, pre‐segmental, and segmental levels were summed, and the performance of these cumulative measures was evaluated. Also, the summary statistics, including mean and standard deviation (SD) of the cumulative CSA and ΔCSA at each of the trachea, primary bronchi, pre‐segmental, and segmental levels, were estimated. Average values of airway path‐length and breathing‐related longitudinal expansion measures of different airways were computed at each of the primary bronchi, pre‐segmental, and segmental levels and used for performance evaluation as well as summary statistics. Agreement of a target airway metric with its manual measurement was computed as the concordance correlation coefficient (CCC) between computer‐generated and manual reference values of the metric. Additionally, agreement of automated airway segmentation with manual outlining was evaluated in terms of Dice scores, and the agreement of automated anatomic airway labeling with manual labeling was examined. For assessment of repeatability, intraclass correlation coefficients (ICCs) of computer‐generated values of a metric derived from repeat CT scans in Data_rep were computed, and the performance was compared with repeatability of manual outlining‐based measures. Summary statistics of ΔCSA and ΔL metrics including their means and SDs were computed for different participant groups. Also, Pearson correlation coefficients of ΔCSA and ΔL metrics with age, sex, BMI, smoking status, and COPD severity were examined; the significance threshold for this correlation analysis was 2/√240 or 0.13 as n = 240. Unpaired t‐tests were performed for comparison between two different participant groups. The statistical significance level was set at p < 0.05.

3. RESULTS AND DISCUSSION

Demographics of both datasets used in this study are summarized in Table 2. The learning phase of the DL network used for computation of airway lumen likelihood map was completed in 33 h (3 h/epoch and 11 epochs to converge). Training of Stage 1 and 2 NN classifiers used for airway branch labeling required 4 and 6 h, respectively. The mean ± SD of Dice scores of agreements of automated airway segmentation with manual outlining were 0.995 ± 0.013 and 0.985 ± 0.020 for TLC and FRC CT scans, respectively. Automated anatomic labeling the trachea and pre‐segmental airways were in full agreement with manual labeling, while agreement was observed to be 96.1% for segmental airways. For each participant, the end‐to‐end image processing pipeline required approximately 15 min to complete the computation breathing‐related radial and longitudinal expansion metrics of airways.

TABLE 2.

Demographics, smoking history, lung function, and chronic obstructive pulmonary disease (COPD) severity metrics of participants in different datasets included in this study.

Characteristics	Datarep	Datagroup
No. of participants	37	240
Demographic characteristics
Age (years), mean ± SD	30.5 ± 11.1	70.3 ± 7.9
Female sex, no. (%)	16 (43)	120 (50)
Body mass index, mean ± SD	<32	28.1 ± 5.7
Smoking history
Pack‐years a smoking, mean ± SD	–	51.1 ± 26.2
Never‐smokers	21 (57)	0 (0)
Ex‐smokers	16 (43)	175 (73)
Current smokers		65 (27)
Spirometry (postbronchodilator) results
FEV1/FVC, mean ± SD	–	0.6 ± 0.2
% predicted FEV1, mean ± SD	–	71.5 ± 25.2
COPD severity b
Preserved lung function, no. (%)	–	60 (25)
Mild COPD, no. (%)	–	60 (25)
Moderate COPD, no. (%)	–	60 (25)
Severe COPD, no. (%)	–	60 (25)

Specific information not available for a dataset is indicated with “−.”

^a Defined by the American Thoracic Society as the number of packs of cigarettes smoked every day multiplied by the total number of smoking years.

^b COPD is defined as postbronchodilator forced expiratory volume in 1 s (FEV1) to forced vital capacity (FVC) ratio of < 0.7 by pulmonary function tests. Preserved lung function consists of participants with Global Initiative for Chronic Obstructive Pulmonary Disease (GOLD) 0, mild COPD consists of participants with GOLD 1 or preserved ratio impaired spirometry (PRISm), moderate COPD consists of participants with GOLD 2, and severe COPD consists of participants with GOLD 3 or 4.

3.1. Repeatability and agreement with manually defined measures

Summary statistics of the agreement of automated measures of static airway metrics at both inspiratory and expiratory lung volumes and breathing‐related airway deformation metrics between the two lung volumes with manually derived measures are presented in Table 3. As observed in the table, the cumulative static CSA increases with the anatomic generation for both inspiratory and expiratory lung volumes. For example, in the inspiratory lung volume, the mean cumulative CSA values at the trachea, medial bronchi, pre‐segmental, and segmental airways are 274.0, 320.0, 338.5, and 348.1 mm², respectively. These findings are consistent with established knowledge from histologic studies that individual airway diameter decreases nearly exponentially with the generation number, while the cumulative CSA of airways increases. ⁴⁵ Similar observations are found for the expiratory lung volume. Mean and SD values of observed ∆CSA at different airway generations are reported. The mean of ∆CSA at the trachea was particularly lower than that at other generations. More importantly, the variability or SD of observed ∆CSA was increased at pre‐segmental and segmental generations. As expected, airway path‐length increased with the anatomic generation for both inspiratory and expiratory lung volumes. Specifically, at the inspiratory lung volume, the mean path‐length values at primary bronchi, pre‐segmental, and segmental airways are 68.8, 123.2, and 147.2 mm, respectively, and these values at the expiratory lung volume were 65.2, 109.9, and 127.3 mm, respectively. Also, ∆L increased with airway generation with the observed mean values being 5.6%, 10.9%, and 13.7% at primary bronchi, pre‐segmental, and segmental airways.

TABLE 3.

Summary statistics (mean and standard deviation (SD)) of observed values of single volume airway metrics and breathing‐related airway deformation metrics between two lung volumes using our fully automated computerized methods and their concordance correlation coefficients (CCCs) with manually derived reference values of respective metrics (n = 37)

Measure	Trachea mean ± SD (CCC)	Primary bronchi mean ± SD (CCC)	Pre‐segmental mean ± SD (CCC)	Segmental mean ± SD (CCC)
CSA (mm2) at inspiration	274.0 ± 72.5 (0.971)	320.0 ± 114.1 (0.969)	338.5 ± 123.0 (0.968)	348.1 ± 88.3 (0.961)
CSA (mm2) at expiration	228.6 ± 52.8 (0.971)	236.3 ± 96.0 (0.970)	259.4 ± 85.4 (0.967)	266.0 ± 81.7 (0.958)
ΔCSA (%)	15.7 ± 4.6 (0.930)	26.1 ± 4.0 (0.898)	27.3 ± 7.1 (0.951)	23.1 ± 15.7 (0.952)
Path‐length (mm) at inspiration		68.8 ± 10.3 (0.972)	123.2 ± 9.9 (0.972)	147.2 ± 16.0 (0.973)
Path‐length (mm) at expiration		65.2 ± 10.8 (0.972)	109.9 ± 13.8 (0.972)	127.3 ± 15.1 (0.972)
ΔL (%)		5.6 ± 3.2 (0.958)	10.9 ± 4.4 (0.963)	13.7 ± 1.9 (0.963)

Abbreviations: CSA = cross‐sectional area; ΔCSA = radial expansion of airways; ΔL = longitudinal expansion of airways.

High CCC values were observed between computer‐generated and manually derived values of the static CSA measure at every airway generation in both inspiratory and expiratory lung volumes. Similar findings were observed for static path‐length measures at different airway generations in both inspiratory and expiratory lung volumes. Also, for ∆L, the CCC between computer‐generated and manually derived values was consistently greater than 0.95 at every airway generation. For ∆CSA, the CCC between computer‐generated and manually derived values was greater than 0.95 at each of pre‐segmental and segmental airway generations. Bland‐Altman plots for errors in ∆CSA and ∆L are illustrated in Figure 5. In these plots, differences between manual and computer‐generated values of ∆CSA or ∆L are color‐coded by their airway generation. As observed in these plots, no shift or gradient bias was observed in the difference values. No divergence in difference values was noted with increasing values of manual measurements. Also, no obvious association of difference values was observed between difference values and airway generation. Magnitudes of differences for ∆CSA were relatively higher than those for ∆L. Specifically, the 95% confidence interval for ∆CSA and ∆L were [−2.69%, 2.56%] and [−0.18%, 0.18%], respectively.

FIGURE 5.

Bland‐Altman plots of differences of fully automated computerized measures of CT‐based breathing‐related airway deformation metrics from manually derived reference measures. (a) Results of Bland–Altman analysis for breathing‐related radial expansion (ΔCSA) of airways at different anatomical generations. (b) Same as (a) for longitudinal expansion (ΔL) of airways. Airway deformation metrics were computed between total lung capacity and functional residual capacity lung volumes.

Repeatability of static and breathing‐related deformation metrics of airways computed at different airway generations are presented in Figure 6 for both manual and computerized methods. Specifically, the ICC of different metric values derived from the baseline and same‐day repeat scans at different airway generations are illustrated in the figure. No notable difference in ICC is observed for computerized and manually derived repeat‐scan values of corresponding airway metrics. In other words, the repeatability of the computerized method is similar to that of the refence method using manual outlining. ICC of all computerized as well as manual measures were greater than 0.9, which is considered as excellent performance for repeatability. ⁴⁶ Additionally, at each airway generation, it was observed that both manual and computer‐generated measures have excellent repeatability (> 0.9) at both lung volumes and all airway generations. For computer‐generated measures of ∆CSA, the observed ICC values at trachea, primary bronchi, pre‐segmental airways, and segmental airways were 0.974, 0.950, 0.943, and 0.901, respectively. In general, ICC values for airway length‐related metrics were greater than CSA‐related metrics at matching airway generations. ICC values of the computerized ∆L metric at the primary bronchi, pre‐segmental airways, and segmental airways were 0.973, 0.954, and 0.952, respectively. Bland‐Altman plots for signed difference of observed ∆CSA (or, ∆L) values between baseline and same day repeat scans for individual participants at different airway generations are illustrated in Figure 7. As observed in Figure 7, for the ∆CSA metric, magnitudes of repeat‐scan differences are greater at higher airway generations. This observation in Figure 7 is consistent with reducing ICC values with increasing airway generations as illustrated in Figure 6. This divergence in magnitudes of repeat‐scan differences with increasing airway generations is relatively small for ∆L.

FIGURE 6.

Plots of intraclass correlation coefficients (ICCs) for chest CT‐based static and breathing‐related deformation metrics of individual airways. (a) ICCs of single volume‐derived and breathing‐related deformation metrics of airway cross‐sectional area (CSA) at different anatomic generations for manual and computerized methods. (b) Same as (a) but for airway path‐length (L).

FIGURE 7.

Bland–Altman plots of differences between repeat measures of CT‐based breathing‐related airway deformation metrics using fully automated computerized measures and manually derived measures. (a) Results of Bland–Altman analysis for breathing‐related radial expansion (ΔCSA) of airways at different anatomical generations. (b) Same as (a) for longitudinal expansion (ΔL) of airways. Airway deformation metrics were computed between total lung capacity and functional residual capacity lung volumes.

3.2. Airway deformation metrics for different COPD groups

Results of COPD group comparison for radial and longitudinal expansion metrics at different airway generations are presented in Table 4. Specifically, mean and SD of ∆CSA and ∆L metrics observed at different airway generations for participants in different COPD groups are presented. Also, the p‐value of unpaired t‐test and the effect size for each of mild, moderate, and severe COPD groups, as compared to the preserved lung function group, are reported for each airway generation. As observed in the table, ∆CSA values were significantly reduced (p < 0.001) with increasing COPD severity at each of primary bronchi, pre‐segmental, and segmental airways. At the trachea, ∆CSA values for moderate and severe COPD groups were significantly less than observed ∆CSA values for the preserved lung function group. No significant difference in ∆CSA between mild COPD and preserved lung function groups were observed at the trachea. Compared to the preserved lung function group, the highest effect size was observed at the pre‐segmental airways for each of the mild, moderate, and severe COPD groups. Our preliminary findings using inspiratory and expiratory chest CT‐based automated measure of ∆CSA are consistent with histologic findings that increased airway obstruction, associated with COPD, leads to reduced radial expansion of airways during inhalation. ⁸ , ⁹

TABLE 4.

Comparison of radial and longitudinal expansion airway deformation metrics among chronic obstructive pulmonary disease (COPD) severity groups (n = 4 × 60).

		Preserved lung function	Mild COPD p‐value, effect size	Moderate COPD p‐value, effect size	Severe COPD p‐value, effect size
ΔCSA (%)	Trachea	18.3 ± 7.3	19.0 ± 10.2 p = 0.367, 0.090	11.7 ± 11.9 p = 0.012, 0.64	9.4 ± 15.7 p < 0.001, 0.69
	Primary bronchi	25.4 ± 7.3	22.5 ± 10.0 p < 0.001, 0.33	15.1 ± 11.2 p < 0.001, 0.96	10.1 ± 16.1 p < 0.001, 1.05
	Pre‐segmental	26.0 ± 6.9	19.6 ± 9.6 p < 0.001, 0.74	15.6 ± 11.5 p < 0.001, 0.97	9.9 ± 16.3 p < 0.001, 1.09
	Segmental	23.1 ± 7.4	16.3 ± 9.7 p < 0.001, 0.36	13.4 ± 11.6 p < 0.001, 0.53	10.3 ± 15.5 p < 0.001, 0.78
ΔL (%)	Primary bronchi	5.0 ± 1.3	5.1 ± 3.1 p = 0.401, 0.043	4.9 ± 2.3 p = 0.282, 0.054	3.2 ± 2.6 p = 0.079, 0.81
	Pre‐segmental	10.9 ± 1.2	10.8 ± 3.3 p = 0.364, 0.039	10.0 ± 1.5 p = 0.042, 0.63	7.8 ± 1.6 p = 0.019, 1.48
	Segmental	9.7 ± 1.0	9.8 ± 3.3 p = 0.071, 0.082	9.4 ± 1.6 p = 0.037, 0.22	5.6 ± 2.1 p < 0.001, 1.57

All participants were ex‐smokers with a smoking history of at least 10 pack‐years.

Abbreviations: CSA = cross‐sectional area; ΔCSA = radial expansion of airways; ΔL = longitudinal expansion of airways.

p‐values present comparisons with preserved lung function group. p‐values in bold are statistically significant (p < 0.05).

COPD is defined as postbronchodilator forced expiratory volume in 1 s (FEV1) to forced vital capacity (FVC) ratio of < 0.7 by pulmonary function tests. Preserved lung function consists of participants with Global Initiative for Chronic Obstructive Pulmonary Disease (GOLD) 0, mild COPD consists of participants with GOLD 1 or preserved ratio impaired spirometry (PRISm), moderate COPD consists of participants with GOLD 2, and severe COPD consists of participants with GOLD 3 or 4.

At pre‐segmental and segmental airway generations, values of ∆L were reduced with increasing COPD severity, and the moderate and severe COPD groups had significantly less ∆L as compared to the preserved lung function group. No significant difference in ∆L between mild COPD and preserved lung function group was observed at any airway tree generation. No significant COPD group difference in ∆L was observed at primary bronchi. The ∆CSA metric was able to detect increased airway obstruction at the early stage of COPD and showed significant difference between mild COPD and preserved lung function groups. The ∆L metric was able to detect reduction in longitudinal airway expansion at moderate and severe COPD groups.

Based on the results of our correlation analysis (Figure 8), BMI and smoking status were significantly associated with ∆CSA at segmental airways (r = 0.17 and −0.19, respectively), while age and sex were significantly associated with ∆L (r = −0.21 and −0.17, respectively); COPD severity was significantly associated with both ∆CSA and ∆L (r = −0.35 and −0.22, respectively); however, the magnitude of association was greater for ∆CSA. These findings related to associations of breathing‐related radial and longitudinal expansion of airways with BMI, sex, age, and smoking status are novel and significant in COPD‐related research because BMI, sex, age, and smoking are important risk factors in COPD. Specifically, BMI is known to be to be associated with poor outcomes in COPD and impoverished quality of life. ⁴⁷ Smoking is a major risk factor of COPD. ¹ After reaching peak lung function at the age of 20–25 years, the respiratory system undergoes progressive decline with aging, including increased airway destruction and loss of elastic recoil, ⁴⁸ and these age‐related impairments are accelerated in COPD. ⁴⁹ Greater longitudinal expansion of airways in males as compared to females is consistent with the known dependence of COPD susceptibility on sex. ¹ These results are suggestive to different pathways of associations of airway deformation metrics with demographic, smoking, and lung health data. Our method will play significant roles in future studies investigating multi‐pathway interactions of airway deformation biomarkers with respiratory mechanics of the diaphragm and chest wall as well as changes in lung health in terms of air trapping and emphysema.

FIGURE 8.

Pearson correlation values of multi‐volume chest CT‐based breathing‐related radial (ΔCSA) and longitudinal (ΔL) expansion of airways with age, sex, body mass index (BMI), smoking status, and chronic obstructive pulmonary disease (COPD) severity. CSA = cross‐sectional area. Red color indicates negative association and green a positive association. ^aCOPD severity is considered as an ordinal variable where “0” represents participants with preserved lung function (COPD global obstructive lung disease (GOLD) status 0), “1” represents mild COPD (GOLD 1 and preserved ratio impaired spirometry (PRISm)), “2” represents moderate COPD (GOLD 2), and “3” represents severe COPD (GOLD 3 or 4).

A limitation of this study is that agreement with manual measures and repeatability of the methods were evaluated largely in a healthy subgroup, while feasibility is assessed in much more severe patients with COPD. In future studies, it will be worth assessing the performance of the methods in disease groups.

4. CONCLUSION

A multi‐volume chest CT‐based automated method has been presented to measure breathing‐related radial and longitudinal expansion of individual airways between inspiratory and expiratory lung volumes. Experimental results based on human chest CT scans show that the fully automated method is repeatable and computationally efficient, and the derived measures are in high agreement with manually defined measures. Also, the results of application of the method to multi‐volume chest CT data of a subset of participants (n = 240) from the COPDGene Iowa cohort show that both radial and longitudinal expansion of airways between inspiratory and expiratory lung volume reduce with COPD severity. Moreover, differences in associations of radial and longitudinal expansion biomarkers of airways with demographic, smoking, and lung health parameters are suggestive to multi‐pathway impairments of respiratory mechanics of airways. Experimental results demonstrate the feasibility of our automated method for application to large population‐based thoracic research and clinical studies investigating different pathways of alterations of airway respiratory mechanics and their impacts on overall lung function as well as disease severity and progression.

CONFLICT OF INTEREST STATEMENT

Punam K. Saha has received grants from the National Institutes of Health (NIH). Alejandro P. Comellas has received grants from the NIH and the Bowers Emphysema Research Fund at the University of Iowa and is a paid consultant for GlaxoSmithKline, Eli Lilly, and AstraZeneca. Eric A. Hoffman has received grants from the NIH and American Lung Association; is a participant (unpaid) on Siemens photon counting CT advisory board; and is founder and shareholder of VIDA Diagnostics, a company commercializing lung image analysis software developed, in part, at the University of Iowa. Sean B. Fain has received grants from NIH and the American Lung Association (ALA) and serves as a scientific advisor and receives grant support from Polarean Inc, Siemens Healthineers and GE Healthcare for development of pulmonary CT and MRI. Kung‐Sik Chan and Syed Ahmed Nadeem have no competing interests.

Nadeem SA, Comellas AP, Chan K‐S, Hoffman EA, Fain SB, Saha PK. Automated CT‐based measurements of radial and longitudinal expansion of airways due to breathing‐related lung volume change. Med Phys. 2025;52:2316–2329. 10.1002/mp.17592