Age-associated changes in aging lungs: A study with auto-segmentation and radiomics based on CT images

Background: The senile lung undergoes physiologic changes that are well known but have not been investigated with computed tomographic (CT) texture analysis. The thin-section pulmonary computed tomographic (CT) microstructure features change in asymptomatic elderly individuals were not explored. Methods : We retrospectively selected 259 subjects under-going chest computed tomography (CT) between April 2018 and June 2019, as group A(consist of group1 and group2), without a history of smoking within the past 5 years, respiratory symptoms or any known chronic pulmonary disease. There were 118 patients in group 1(age ≥ 60 years-elderly) and 141 patients in group 2 (age≤50 years-young). Furthermore, 273 tests(PFTs) were included as group B, which divided into two cohorts, chronic obstructive pulmonary disease (COPD) (n=83) and non-COPD (healthy smoker [HS], n =90; healthy non-smoker [HNS], n =100) cohort. The radiomic features were extracted and selected from group A, trialed in group B, using the LASSO algorithm. Results : A total of 233 features were significant in group A. Among these features, 17 features exhibited distinct differences between COPD and non-COPD patients, 18 features exhibit distinct differences between HSs and HNSs. Meanwhile, five features were shared in group B. A negative correlation was determined between carbon monoxide diffusing capacity(DL CO ) and the two features: -0.63). Similarly, a positive correlation was found between FEV1/FVC and HighGreyLevelRunEmphasis_AllDirection_offset8_SD (ρ = 0.74). Conclusion : Radiomic features, which associated with the ages and significant in COPD patients and smokers, maybe reveal the microstructure changes of the aging lungs. Registered


Introduction
As it is known, aging has become a public health problem, which influences all aspects of human biology, such as aging lung. The aging lung undergoes anatomic changes, with a deterioration in terms of function. Furthermore, the aging of lung considered as a risk factor for respiratory diseases, and strongly correlates with the onset and progression of chronic respiratory diseases [1][2][3]. The structure of aging lung is characterized as alveolar space enlargement, alveolar wall thinning and bronchial basement membrane thickening. This usually presents with decreasing thoracic volume, reduced lung volumes and alterations in respiratory muscles [4]. However, the degree of lung aging is inconsistent and greatly varies among individuals [5,6]. Several factors, such as age, race, environmental exposure and living conditions, have been reported to have a potential effect on lung aging.
The lung aging eventually cause a decrease of pulmonary function [7]. However, prior to the function change, structural change perhaps occurs in the lungs [8]. It was previously reported that computed tomography (CT) could detect the smoking-related inflammation of small and/or distal bronchi (bronchiolitis), and even the changes in lung structure before airflow limitation [7,9,10]. In addition, CT can also reveal the structural change of the lung before PF starts to decrease. To the date, the role of imaging in lung aging remains unclear, and there is a lack of imaging biomarkers for lung aging.
In previous studies, Copley et al found in the older group, there is a high prevalence of a basal subpleural reticular pattern (60% of subjects), small cysts (25%), bronchial dilatation (60%), and wallthickening (55%) and confirmed interlobular septal thickening (18%), as described in previous studies [11]. A.M. Chiesa et al revealed the absence of A-lines in healthy elderly subjects [12]. Recently, radiomic signature has proven to be a significant biomarker for biological, clinical condition, or disease progression [13][14]. The aim of the present study was to explore the age-associated changes based on the radiomics texture analysis for lung aging and validate or trial this in another two cohorts: a cohort of COPD patients and a cohort of smoking patients.

Patients
The present retrospective study was approved by the Ethics healthy non-smoker (HNS) [16]. The demographic, radiology and clinical evaluation were obtained by reviewing the medical records.
Patients were excluded when they had any of the following conditions: missing clinical data and poor imaging quality.

Lung CT
The conditions for the lung CT examination were as follows: 120 kV, 10-40 mA, gantry rotation speed (0.5 seconds), and helical scan mode (pitch 1.2). All lung CT images were reconstructed at a slice thickness of 2 mm, with a spacing interval of 2 mm. All images were obtained at a lung window width of 1,500 HU and a window level of -400 HU.

PF
PF tests were evaluated using a flow spirometer (Vmax22; SensorMedics, Yorba Linda, CA), based on the American Thoracic Society guidelines [16] and European Community Lung Health Survey values [17]. Post-bronchodilator measurements were The degree of airflow limitation was evaluated using the GOLD definition and spirometric classification [15].

CT image feature extraction
For processing, all images were standardized according to the  [18].

Radiomic feature selection and model establishment
The significance of the feature related to aging lung was evaluated by paired t-test. Then, the significant features (P<0.05) were selected to distinguish between COPD and non-COPD patients, and HSs and HNSs. In order to compare the features between groups, feature normalization was performed. The values of each feature were normalized using z-score normalization [19]. Similarly, the normalization was performed in the validation groups (COPD or HS) using the corresponding data calculated in the training cohort.
A feature selection algorithm based on the least absolute shrinkage and selection operator (LASSO) method was adopted. This algorithm adopts the 10-fold cross validation method, and the features were selected by weighting the LASSO coefficients (Supplementary material). The radiomics signature used in the validation cohorts were calculated using the formula obtained from the training set.
The radiomics scores (rad-score) were calculated for each patient, and several methods, such as Logistic regression, support vector machine (SVM), Random Forest, Bayes, K-Near Neighbor (KNN), and Decision Tree [20], were used to establish the prediction models for grouping COPD and non-COPD subjects (Model 1) or HSs from HNSs (Model 2). Then, the discriminative accuracies were evaluated using the receiver operating characteristic (ROC) curve analysis.
The radiomic methods included the following steps ( Figure 1), and the LASSO method and dimensionality reduction framework presented in Figure S1 and S2.

Statistical analysis
The statistical analysis was performed using the R-project  Table 2).

Diagnostic performance of the established models
The rad-scores were calculated using the formula for the radiomics signature (Figure 2 In the validation cohort, the accuracy, sensitivity, specificity, PPV  Figures 4A and 4B).

The associations between features and PF tests
The associations between PF tests (six tests) and features (17 features) were analyzed. DLCO was found to be correlated with four coarse features:
FEV1/FVC was found to be correlated with three different radiomic features:

Discussion
In the present study, the lung radiomics features associated with aging were investigated. Radiomics signatures were constructed, and its value was verified in two tested populations. To our knowledge, this is the first report to investigate the radiomics features of lung aging. Several advantages of the study have been identified. First, many molecules have been studied in the aging process. For example, in the lungs during aging, TLR, p53/p66/p21 and reactive oxygen species (ROS) generation are increased [21][22].
In addition, the telomere becomes short [23]. Unfortunately, an invasive procedure is required to obtain a tissue sample, and investigate these levels during aging [24]. However, the method for lung aging in the present study can be easily accessed, has little additional cost, and favors the discrimination of aging subgroups.
These advantages make this method promising. Second, previously, clinicians have mainly focused on the analysis of imaging manifestations for lung aging. However, the present study revealed that the whole lung texture analysis could better reveal the microchanges of lung structures, which is helpful in identifying lung aging and earlier initiating lifestyle interventions, and this may slow or delay the onset of lung aging [25]. Third, the present study also evaluated the relationship between texture features and PF tests, and found that a close relationship exists between these. Fourth, as it is known, lung aging can be affected by many factors. However, in the present study, efforts were made to alleviate the interference: (1) CT images from the same patient at different time points were analyzed; (2) automatic segmentation was performed for feature analysis. Compared with manual segmentation, the automatic method demonstrates good effectiveness and reliability, and has less dependence on humans.
A total of 30 radiomics features selected from 233 textural parameters may have the capability to reveal the characteristics of aging lungs, including histograms and morphological features, which are low-and higher-order texture parameters (such as GLCM, RLM and form factor features). Among these 30 textural parameters, five parameters were selected, which included one histogram, one GLCM, and three RLM parameters. It is noteworthy that these features could detect the invisible micro-structural changes of the lung. For example: histogram parameters could be used to assess the attenuation of the lungs [26]; GLCM could reflect the degree of heterogeneity of the whole lung, and the micro-structural changes may give uneven signals in the CT scan, which is irreversible of any anatomy change in the lung [27]; form factor features could reflect the changes in the region of interest, in terms of volume, area and shape, and these may be displayed more visually, allowing the changes in lung volume and shape to be easily observed and compared [28]; in addition to those mentioned above, the study also revealed that the RLM features, which reflect the roughness and directionality, were also associated with the changes in pulmonary structure [29].
Two populations, COPD and HS patients, were the study subjects, and these subjects were included to verify the observed features.
The reasons for this were as follows: First, previous studies have shown that lung aging is closely correlated to the onset and development of chronic respiratory diseases [1]. Second, the risk factors of COPD included age, suggesting that COPD has an association with lung aging. Third, declining lung function is a significant feature for the development of COPD, which increases with age, especially in smoking individuals [29]. Furthermore, an injured lung function cannot recover after smoking cessation [1]. To the date, the mechanism of injury in lung function in COPD patients remains unclear. However, given the strong similarities between elderly and COPD lungs, the mechanism that involves aging has been investigated, and it has proven that accelerated lung aging occurs in COPD patients [1]. Similar to other organs, lung demonstrates physiological and structural changes associated with aging, resulting in a progressive decrease in lung function among healthy populations. In terms of smoking, this can accelerate the decline in lung function over time, suggesting that this has an association with lung aging [30]. This may be explained, as follows: (1) smoking could accelerate the aging of the airway epithelium [31], (2) changes occurred in the micro-structure of the lung between HSs and HNSs, and these are similar to that found in the COPD cohort.
Interestingly, in the present study, it was found that the discriminative performance of radiomic features between HSs and HNSs in the training cohort were slightly better than those between COPD and non-COPD patients. The possible reason for this is that the changes between COPD and non-COPD patients may be more obvious. Thus, radiomic features that reveal these micro-changes may be insensitive in distinguishing between these. In a word, aging would make an effect on the manifestations of lung disease and even lower the response to treatment, and the present findings may have an ancillary role in conventional diagnostic methods and clinical examinations.
Although the pathology of the lung is distinct, in terms of PF, all aging lungs can result in impairment of lung function, which can be measured using FEV1 and FVC [30,31]. spatial information may be more useful than intensity information in the identification. In addition to texture features, radiomic analysis often requires the analysis of a combination of intensity, morphology, fractal geometry and higher-order features [33]. This information is integrated, and may thereby provide novel insights and a better detail of the lung using CT images. Eventually, radiomic analysis may serve as a useful tool that could help to improve the clinical management of patients [34].
Although these present findings provide meaningful implications, the present study has several limitations. First, CT examinations are usually performed without spirometric control of lung volume.
However, all subjects in the present study received education on how to perform the respiratory maneuvers prior to the CT scan. Second, the present study has a relatively small sample size. Thus, the results may be biased. Hence, further analysis using a larger sample size is required. Fourth, the relationship between the severity of COPD and radiomics features was not investigated. Hence, another trial is needed to specifically address this issue.
In conclusion, some radiomics features that reveal the microchanges of lung structure were found to be associated with lung aging, and radiomic signatures, which are constructed using LASSO regression, could be used to identify a lung aging-related population of COPD patients and a population of smoking patients. However, a larger study is still needed to verify this finding, and further analysis are required to assess its prognostic value in an aging population.

Consent for publication
Not applicable.

Availability of data and materials
The datasets generated and/or analysed during the current study are not publicly available due [REASON WHY DATA ARE NOT PUBLIC] but are available from the corresponding author on reasonable request.

Competing interests
The authors declare that they have no competing interests XYG. Study design, manuscript writing and editing;