Introduction
Bronchial parameters are increasingly being investigated for use in the characterisation of pulmonary diseases such as chronic obstructive pulmonary disease (COPD) [
1]. A potential benefit of developing robust bronchial parameters is the early detection of pulmonary disease. For example, screening for lung cancer with computed tomography (CT) may offer the opportunity for the evaluation of “off-target” organ systems such as the heart, bronchi, and vasculature [
2]. While bronchial parameters could be used for the evaluation of pulmonary disease, their use is limited by the man-hours necessary for (manual) measurements. This step is further complicated by the low dose of screening CT scans, which can result in a worse image quality with more noise. Due to this, the development of reliable automated methods for CT bronchial parameter measurement is a necessary step.
To calculate bronchial parameters, most methods require segmenting and measuring the airway lumen and wall from chest CT scans. Segmentation of the airway lumen is challenging, due to the complex structure of the airway tree and the small size of most branches. Recently, deep learning methods for automatic segmentation of the airway lumen have achieved success [
3‐
7]. Segmentation of the airway walls in the smaller branches is even more demanding, due to its small thickness and low contrast between the wall, lumen, and surrounding parenchyma. The thickness of the wall may fall below the scanner resolution, therefore lacking the stark contrast available in the larger airways. Airway wall segmentation has received less attention; currently, there are no automatic methods to obtain this directly from the CT scan without an initial seed placement or lumen segmentation [
8]. Instead, the airway wall can be obtained as an additional refinement step using, for example, full-width at half-maximum [
9], phase congruency [
10], or optimal-surface graph-cut methods [
11].
To evaluate early biomarkers of respiratory disease on low-dose chest CT scans, we built an automated pipeline for segmenting and quantifying the airway lumen and wall. We did this by combining two validated, open-source methods, for obtaining the airway lumen and wall segmentations, respectively. While previous studies have evaluated AI on lumen segmentations, we could not identify studies that have assessed their reproducibility when also measuring the airway wall in a fully automated way. Furthermore, this is the first combination of these 3D-Unet and 3D optimal-surface graph-cut methods for fully automated bronchial parameter evaluation. We aim to quantify the repeatability of this pipeline on low-dose chest CT. Subsequently, we computed the bronchial parameter measurements. We tuned this pipeline for the low-dose chest CT scan protocol and investigated its reproducibility using short-term repeated scanning.
Discussion
In this study, we built an automated pipeline for low-dose chest CT scans to obtain segmentations of the airway lumen and wall by combining two open-source methods. The resulting segmentations yielded automated quantitative bronchial parameters. Repeated scans showed moderate to good reproducibility (R2 > 0.6) of bronchial parameters down to the 6th generation. The Bland–Altman analysis showed no systematic bias and narrow limits of agreement for Pi10 and WAP, but wider for LA, demonstrating a lower variability in summary parameters like Pi10 and WAP compared to the direct measurement of LA.
The use of low-dose CT scans for lung cancer screening provides the opportunity to screen for other early diseases such as COPD, bronchiectasis, and cardiac disease, which may influence lung cancer risk and/or prognosis. Automated bronchial parameter measurement can enable the screening of large cohorts in a reasonable timeframe with good reliability. Furthermore, fully 3D segmentation can be readily useful in clinical tasks such as virtual bronchoscopy or surgical planning. However, for bronchial parameters, it is hard to determine whether the airways are normal or abnormal. The number of never-smokers in bronchial parameter research is typically very small [
27]. Combined with heterogenous bronchial parameter methodology, it is unclear what quantitatively defines “normal” airways on low-dose CT and by which bronchial parameter. This study demonstrated a wider variability in measurements for LA than Pi10 or WAP. While this could in part concern variability or error due to methodology, additional factors like seasonal changes, smoking, or illness before a scan could result in true differences. Pi10 averages many branches, while WAP includes wall thickness in its calculation and so could be more resistant than LA to localised variations in measurements. Our pipeline provides similar reproducibility of LA and WAP as previous methods on similar datasets [
11], but it also gives better reproducibility of Pi10 [
28]. Additionally, it offers fully automatic bronchial parameter measurement using low-dose noisy scans.
Various methods can be used as an initial step for lumen segmentation. We used Bronchinet due to its state-of-the-art performance [
3], speed, and open-source availability which enabled retraining on the low-dose scans in this study. Fully automated bronchial parameter calculation has been previously proposed using tools trained on manually traced borders alongside older algorithms such as FWHM, intensity-based, and phase congruency [
29,
30]. However, previous research shows that manual and FWHM measurement overestimates the airway wall [
31], which is also evident when used to measure the COPDGene phantom (Table
S2). Compared to these approaches the advantage of our method is that Opfront was optimised on a phantom with precise physical measurements, eliminating the bias in wall measurements that comes with the previously mentioned approaches. The pipeline output is a ready-to-use 3D model of the airways, which has potential applications in tasks such as virtual bronchoscopy, airflow simulation, and 3D printing. Deploying the pipeline in a docker image provides the method as ready-to-use and implementable in clinical practice. For lumen segmentation, good results could be readily achieved by using the publicly available trained model bundled with Bronchinet [
3], which uses airway segmentations for training from the Danish Lung Cancer Screening Trial [
32] in combination with an Erasmus-MC Sophia (cystic fibrosis) dataset [
33]. The ImaLife scan protocol has a lower radiation dose with a total DLP of < 100 mGycm, and more noise in the scans; retraining the tools resulted in better performance [
13]. For maximum performance on different datasets, optimising the pipeline for the target CT protocol may be necessary. This was achieved by re-training the Bronchinet with efficiently generated ground truths, and tuning Opfront using a physical phantom.
A limitation of this study is the lack of severe airway disease in the cohort as the ImaLife study comprises a general population. Evaluation of severe cases is important prior to adoption in a clinical setting, where scan protocol may also change. For the analysis, we assumed that there are no short-term differences in bronchial parameters between the scans. However, factors such as illness or smoking before the scan could have an impact on the bronchial parameter results. This would tend to increase variability between scans, which could mean that the actual scan-rescan repeatability may be better than we currently report. The methods used do not perform anatomical airway labelling, and so we could not compare the repeat measurements of specific airway branches directly. Instead, we focused on average values per generation for participants. Lastly, Bronchinet does not guarantee a fully connected airway segmentation, some peripheral branches may be discarded during measurement. For cases with an occluded lumen, this could result in the exclusion of segmented airways beyond the blockage.
In conclusion, we demonstrate a comprehensive and fully automatic pipeline for bronchial parameter measurement on low-dose CT using open-source tools. Based on the results of short-term repeat CT scanning, the pipeline provides reliable bronchial parameters down to the 6th generation. Overall, these methods enable the exploration of bronchial parameters in large low-dose CT datasets after an initial investment in the training and optimisation of deep learning and optimal-surface graph-cut methods.
Publisher's note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.