Computed Tomography Radiomics Features Predict Change in Lung Density and Rate of Emphysema Progression

Abstract

Rationale

Emphysema progression is heterogeneous. Predicting temporal changes in lung density and detecting rapid progressors may facilitate the selection of individuals for targeted therapies.

Objectives

To test whether computed tomography (CT) radiomics can be used to predict changes in lung density and detect rapid progressors.

Methods

We extracted radiomics features from inspiratory chest CT in 4,575 subjects with and without airflow obstruction at enrollment, who completed a follow-up visit at approximately 5 years. We quantified emphysema using adjusted lung density (ALD) and estimated emphysema progression as the annualized change in ALD (ΔALD/yr) between visits. We categorized participants into rapid progressors (>1% ΔALD/yr) and stable disease (≤1% ΔALD/yr). A gradient boosting model was used 1) to predict ALD at 5 years and 2) to identify rapid progressors. Four models using demographics (base clinical model), CT density, radiomics, and combined features (clinical, radiomics, and CT density) were evaluated and tested.

Results

There were 1,773 (38.8%) rapid progressors. For predicting ALD at 5 years in the 20% held-out data, the base model explained 31% of the variance (adjusted R² = 0.31), whereas R² was 0.74 for the CT density model, 0.66 for the radiomics-only model, and 0.77 for the combined-features model. For detecting rapid progressors, the base model (area under the receiver operating characteristic curve [AUC], 0.57 [95% confidence interval (CI), 0.53–0.61]) was outperformed by the radiomics-only model (AUC, 0.73 [95% CI, 0.69–0.76]; Δ = 0.15; P < 0.001) and the combined model (AUC, 0.74 [95% CI, 0.71–0.77]; Δ = 0.17; P < 0.001).

Conclusions

Parenchymal and airway radiomics features derived from inspiratory scans can be used to predict temporal changes in lung density and help identify rapid progressors.

Chronic obstructive pulmonary disease (COPD) is a highly prevalent lung disease affecting more than 392 million adults and is the third leading cause of death globally (1). Although its course is variable, COPD is usually progressive. The predictors of lung function trajectories in COPD have been well described (2, 3). The structural changes underlying these trajectories, however, are not well understood. Emphysema, a major structural subset of COPD, is usually progressive, but the rate of progression is heterogeneous (4, 5). There are currently no specific therapies for emphysema, and the first step is to identify individuals who can be targeted in clinical trials of directed therapies to impede progression. The major causes of emphysema, such as cigarette smoking, biomass smoke, and occupational factors, with continued exposure, are also likely associated with its progression. Additional risk factors include advancing age and genetic conditions including alpha-1 antitrypsin deficiency (6).

There are now considerable data to suggest that several intrinsic factors related to alterations in the structure of the airways and lung parenchyma themselves may beget further disease (7, 8). Cross-sectional and longitudinal studies suggest that in some individuals, emphysema follows terminal small airway disease and may be a consequence of air trapping (4, 9). Other studies indicate that emphysema, once initiated, may itself impose mechanical stretch on adjacent normal lung regions and contribute to the further alveolar destruction (7, 8, 10). The location and distribution of emphysema may also affect its progression (11). In addition to lung density changes, which can be easily quantified, emphysema also alters lung tissue organization and texture. Recent advances in image analyses have enabled the quantification of lung density uniformity and contrast and thus an improvement in estimates of tissue texture. The remodeling of parenchyma in COPD does not happen in isolation and is almost always associated with remodeling of the small and medium-size airways, as either a cause or a consequence of the parenchymal changes. We hypothesized that a combination of parenchymal density and texture measures on computed tomography (CT) as well as features associated with airway shape and structure can help predict temporal changes in lung density and also aid the accurate detection of rapid emphysema progressors. We also aimed to extract important CT features predictive of progression, which may provide mechanistic insights.

Methods

Study Population

We analyzed data of participants enrolled in COPDGene (Genetic Epidemiology of COPD), which is a multicenter cohort study. The details of the study design have been described previously (12). Briefly, participants were current or former smokers with smoking histories of at least 10 pack-years between the ages of 45 and 80 years. Those with other lung diseases, such as interstitial lung diseases and bronchiectasis, or with prior lobar excision, active malignancy, and chest radiation were excluded. Spirometry and chest CT were acquired at enrollment and at a second visit approximately 5 years later. Participants underwent prebronchodilator and postbronchodilator spirometry with the administration of 180 μg of albuterol. COPD was defined by a postbronchodilator ratio of forced expiratory volume in 1 second to forced vital capacity of <0.70 (13, 14). The severity of COPD was determined as per the Global Initiative for Chronic Obstructive Lung Disease (GOLD) report (14).

CT Imaging

CT scans of the chest were obtained at full inspiration (total lung capacity) in the supine position. The images were reconstructed at B31f kernel with a slice thickness of 0.75 mm and a reconstruction interval of 0.5 mm. The lungs were segmented using a UNet-based deep learning framework (15), and airway trees were segmented using the imaging software LungQ (Thirona) (16).

Emphysema quantification

We calculated lung density on inspiratory CT scans as the 15th percentile Hounsfield units (HU) plus 1,000, to convert to grams per liter. To account for variance introduced by the volume of acquisition, we quantified emphysema using the adjusted lung density (ALD), which is the 15th percentile density of the lungs, with a correction factor of the actual total lung volume on CT divided by the predicted total lung volume using the MESA (Multi-Ethnic Study of Atherosclerosis) equations derived from a normative population, which account for age, sex, race, height, and body mass index (BMI) (5). Change in emphysema was quantified as the annualized change in ALD (ΔALD) between the two visits. Participants with ΔALD > 1% per year were categorized as rapid emphysema progressors, and those with ΔALD ≤ 1% per year were categorized as having stable disease (17). In sensitivity analyses, we also tested other thresholds of >0.5% and >1.5% annual ΔALD to define rapid progression.

Radiomics

We used the PyRadiomics (version 3.0.1) package, an open-source radiomics quantification platform, to compute features of 1) density, 2) texture, 3) lung shape, and 4) airway shape on the inspiratory chest CT scans (18). The lungs were segmented using a UNet-based algorithm (LungInsight) (15), and radiomics features were extracted from the segmented lungs. Density and texture features were computed for the whole lung, whereas shape features were computed for the individual lung separately. We applied the Laplacian of Gaussian filter to enhance the characterization of the texture features. This edge enhancement filter further highlights subtle textural variations in the lung parenchyma. We extracted the lung shape features from lung segmentation maps. A total of 116 density, texture, and shape features were obtained. The complete list of radiomics features and their definitions are provided in Table E1 in the data supplement.

Density features

The density-feature set included 18 features representing basic statistics of the inspiratory CT density histogram, such as mean, median, skewness, kurtosis, 10th percentile, and 90th percentile.

Texture features

The texture-based feature set included 56 features based on measurements derived from gray-level cooccurrence matrix, gray-level run length matrix, and gray-level size zone matrix such as inverse difference moment, gray-level nonuniformity (GLN), cluster prominence, and entropy.

Lung shape features

The lung shape–based feature set included 28 features derived from the right and left lungs separately. Lung shape features included elongation, flatness, major- and minor-axis lengths, maximum two- and three-dimensional diameters, and surface area.

Airway shape features

The airway shape–based feature set included 14 features derived from the airway tree and included the same shape features as were derived for the lungs.

Statistical analyses

We used a Light Gradient-Boosting Machine (LightGBM) regressor and classifier to point 1) predict the absolute ALD value at Year 5 and point 2) categorize participants into rapid progressors versus those with stable disease. LightGBM is a decision tree–based gradient boosting framework (19). LightGBM grows the decision tree leafwise, where the leaf yields maximum loss. This nature of growing the tree enables the achievement of global minimum loss at high speeds. LightGBM uses gradient-based one-side sampling (GOSS) and exclusive feature bundling techniques. GOSS retains every instance with a large gradient and randomly samples every instance with a small gradient. When calculating the information gain, GOSS inserts a constant multiplier for the data instances with small gradients to account for the effect of the data distribution. Without significantly altering the initial data distribution, GOSS increases its attention on the undertrained examples. Exclusive feature bundling is a near lossless technique for the identification of important features by bundling high-dimensional feature space into a low-dimensional space (19).

We tested four sets of predictors for both regression and classification analyses:

1. Base clinical model: The base clinical model included six features: age, race, sex, BMI, pack-years of smoking, and current smoking status.

2. CT density model: Baseline ALD and scanner type were included.

3. Radiomics model: All 116 radiomics features were included in this model.

4. Combined model: The base clinical model features, CT density model features, and radiomics features were combined (124 features).

Model Training and Evaluation

Regression

The LightGBM regressor was used to predict ALD at 5 years. The dataset was divided into training (80% of the sample) and held-out (20% of the sample) sets. The model was trained using 10-fold cross-validation on the training dataset and evaluated on the held-out sets separately on each of the four models (base, CT density, radiomics, and combined). Hyperparameter tuning was done on the following parameters of LightGBM: number of estimators, maximum depth of the tree, learning rate, minimum child weight, and number of leaves. We tuned hyperparameters for the four models separately using 10-fold cross-validation and extracted the best values for each of the estimators using tree of Parzen estimators (20, 21). The coefficient of determination (R²) was calculated for each model.

Classification

The LightGBM classifier was used to predict the classification of participants into rapid progressors and those with stable disease. We followed the same approach as in the regression analyses for splitting of the dataset into training and held-out sets for model training and testing. The area under the receiver operating characteristic curve (AUC) was calculated for each model. The DeLong test was used to calculate the confidence interval for AUCs of the training and held-out data. Differences in model discrimination were compared using the DeLong test (22).

Feature importance

We identified the top 10 features for both the regression and classification tasks for the radiomics-only and combined model using LightGBM’s built-in feature importance estimator that assigns a score for each feature on the basis of the feature’s contribution in the decision tree. The importance score reflects the frequency of the feature value applied toward prediction. We normalized all importance scores by the highest score.

Statistical analyses were performed in Python version 3.9 using PyRadiomics version 3.1.0 and LightGBM Python packages (18, 19). A two-sided α value of 0.05 was considered to indicate statistical significance for all analyses.

Results

Participants

Of 5,119 participants who completed 5-year follow-up visits, we included 4,575 subjects, excluding those with missing CT scans (n = 448), those with poor segmentation quality (n = 21), and lifetime nonsmokers (n = 75) (Figure 1). The mean (standard deviation [SD]) age of the cohort was 59.8 (8.6) years. There were 2,318 (49.8%) women, and 1,320 (28.4%) were of African American race. A total of 1,843 (40.3%) subjects had diagnoses of COPD. Of these, 409 (8.9%), 883 (19.3%), 451 (9.8%), and 100 (2.2%) had GOLD severity stages 1 through 4, respectively. There were 2,194 (48.0%) without airflow obstruction (GOLD stage 0), and 538 (11.8%) had preserved ratio impaired spirometry (Table 1) (23).

Figure 1.

Consolidated Standards of Reporting Trials diagram. COPDGene = Genetic Epidemiology of COPD; CT = computed tomography.

Table 1.

Characteristics of participants at baseline and after 5 years (n = 4,575)

Parameter	Visit 1	Visit 2
Age, yr	59.8 (8.6)	65.4 (8.6)
Female sex, n (%)	2,318 (49.8)	2,318 (49.8)
African American race, n (%)	1,320 (28.4)	1,320 (28.4)
Body mass index, kg/m2	29.2 (6.1)	29.1 (6.4)
Smoking, pack-years	42.5 (23.5)	44.0 (23.9)
Current smokers, n (%)	2,013 (46.3)	1626 (37.4)
FEV1, L	2.4 (0.8)	2.1 (0.8)
FEV1, pp	80.5 (22.7)	78.4 (24.5)
FVC, L	3.4 (1.0)	3.1 (0.9)
FVC, pp	89.4 (16.7)	87.2 (17.6)
FEV1:FVC ratio	0.7 (0.1)	0.7 (0.1)
GOLD severity, n (%)
PRISm	538 (11.8)	554 (12.1)
Stage 0	2,194 (48.0)	2,013 (44.0)
Stage 1	409 (8.9)	439 (9.6)
Stage 2	883 (19.3)	907 (19.8)
Stage 3	451 (9.8)	472 (10.3)
Stage 4	100 (2.2)	190 (4.2)
Pi10, mm	2.3 (0.6)	2.3 (0.6)
Adjusted lung density, g/L	86.3 (24.4)	85.7 (25.6)
Total lung capacity, L	5.5 (1.4)	5.5 (1.4)
Percentage emphysema	5.2 (8.0)	5.6 (9.2)

Definition of abbreviations: FEV₁ = forced expiratory volume in 1 second; FVC = forced vital capacity; GOLD = Global Initiative for Chronic Obstructive Lung Disease; Pi10 = square root of the wall area of a theoretical airway with 10-mm luminal perimeter; pp = percentage predicted, PRISm = preserved ratio impaired spirometry.

Data are expressed as mean (SD) unless otherwise specified.

At baseline, ALD ranged from 4.4 to 180.5 g/L, with a mean (SD) of 86.3 (24.4) g/L. The mean (SD) difference between the ALD at baseline and at 5 years was 0.6 (1.3) g/L. A total of 1,773 subjects (38.8%) were identified as rapid emphysema progressors, and 2,802 (61.2%) were identified as having stable disease. Table 2 shows the baseline characteristics of rapid progressors and those with stable disease.

Table 2.

Baseline characteristics for rapid progressors and stable disease (n = 4,575)

Parameter	Rapid	Stable
Age, yr	60.3 (8.5)	59.5 (8.7)
Female sex, n (%)	854 (48.2)	1428 (51.0)
African American race, n (%)	533 (30.1)	764 (27.3)
Body mass index, kg/m2	28.7 (6.2)	29.5 (6.0)
Smoking, pack-years	45.4 (23.3)	40.7 (23.4)
Current smokers, n (%)	937 (52.8)	1,251 (44.6)
FEV1, L	2.2 (0.9)	2.5 (0.8)
FEV1, pp	74.6 (24.2)	84.2 (20.8)
FVC, L	3.3 (1.0)	3.5 (1.0)
FVC, pp	87.7 (17.8)	90.5 (15.9)
FEV1:FVC ratio	0.6 (0.2)	0.7 (0.1)
GOLD severity, n (%)
PRISm	208 (11.7)	330 (11.8)
Stage 0	641 (36.2)	1,553 (55.4)
Stage 1	169 (9.5)	240 (8.5)
Stage 2	421 (23.7)	462 (16.5)
Stage 3	273 (15.5)	178 (6.4)
Stage 4	61 (3.4)	39 (1.4)
Pi10, mm	2.4 (0.6)	2.2 (0.6)
ALD, g/L	86.6 (27.1)	84.7 (22.0)
ΔALD per year, g/L	2.9 (1.6)	−1.8 (3.2)
Total lung capacity, L	5.6 (1.4)	5.5 (1.4)
Percentage emphysema	6.3 (9.5)	4.5 (6.8)

Definition of abbreviations: ALD = adjusted lung density; FEV₁ = forced expiratory volume in 1 second; FVC = forced vital capacity; GOLD = Global Initiative for Chronic Obstructive Lung Disease; Pi10 = square root of the wall area of a theoretical airway with 10-mm luminal perimeter; pp = percentage predicted, PRISm = preserved ratio impaired spirometry.

Data are expressed as mean (SD) unless otherwise specified.

Prediction of ALD at 5 Years

In the training set for the prediction of ALD at 5 years (80% of the cohort), the base clinical model explained 32% variance (adjusted R² = 0.32). R² was 0.75 for the CT density model and 0.66 for the radiomics-only model. When radiomics were combined with the base clinical and CT density model, R² was 0.78. In the held-out set (20% of the cohort), the base clinical model explained 31% of the variance (adjusted R² = 0.31, Pearson correlation = 0.56) whereas R² was 0.74 (Pearson correlation = 0.86) for the CT density model, 0.66 (Pearson correlation = 0.84) for the radiomics-only model, and 0.77 (Pearson correlation = 0.90) for the combined-features model. A scatterplot and Bland-Altman plot showing the relation between ground truth and predicted ALD at 5 years for the four models are shown in Figures E1–E4.

Figure 2 shows the top predictors of ALD at 5 years in each of the models. In the base clinical model, BMI was the best performing feature, followed by age and pack-years of smoking. In the radiomics model, the 10th percentile of CT density histogram was the best performing feature, followed by GLN of intensity values in the CT image and the ratio of surface area to volume of the airways. In the combined model, baseline ALD was the best performing feature, followed by 10th percentile of the CT density histogram and age.

Figure 2.

Top features for predicting adjusted lung density at 5 years. Results are shown for the (A) base clinical model, (B) radiomics model, and (C) combined model. The x-axis represents the importance score of each feature generated by Light Gradient-Boosting Machine, and the corresponding feature names are shown along the y-axis. 10th percentile = 10th percentile of computed tomography density histogram; Correlation = correlation of gray-level values in lung parenchyma; Maximum = maximum gray-level intensity; Run entropy = run entropy of gray-level runs of parenchymal texture; Sphericity = roundness of the shape of emphysematous regions.

Classification of Rapid Progressors versus Stable Disease

In the training set, for classifying rapid progression versus stable disease, the AUC was 0.58 (95% confidence interval [CI], 0.57–0.60) for the base clinical model, 0.61 (95% CI, 0.60–0.62) for the CT density model, 0.74 (95% CI, 0.73–0.75) for the radiomics model, and 0.75 (95% CI, 0.74–0.76) for the combined model. In the held-out dataset (20% of the cohort), 355 (39%) and 560 (61%) participants were identified as rapid progressors and subjects with stable disease, respectively. For the base clinical model, the AUC was 0.57 (95% CI, 0.53–0.61). The positive and negative likelihood ratios for the base clinical model were 1.97 and 0.98, respectively. For the CT density model, the AUC was 0.60 (95% CI, 0.56–0.64). The positive and negative likelihood ratios for the CT density model were 1.58 and 0.88. For the radiomics model, the AUC was 0.73 (95% CI, 0.69–0.76). The positive and negative likelihood ratios for the radiomics model were 2.81 and 0.63, respectively. For the combined model, the AUC was 0.74 (95% CI, 0.71–0.77). The positive and negative likelihood ratios for the combined model were 2.71 and 0.62, respectively. Compared with the base clinical model (AUC, 0.57 [95% CI, 0.53–0.61]), the discriminative accuracy was higher for the CT density (AUC, 0.60 [95% CI, 0.56–0.64]; Δ = 0.03; P = 0.27), radiomics-only (AUC, 0.73 [95% CI, 0.69–0.76]; Δ = 0.15; P < 0.001), and combined (AUC, 0.74 [95% CI, 0.71–0.77]; Δ = 0.17; P < 0.001) models.

The discriminative accuracy worsened with increasing disease severity in the base clinical model (AUCs of 0.56, 0.67, 0.62, and 0.53 for GOLD stages 0, 1, 2, and combined 3 and 4, respectively) and CT density model (AUCs of 0.62, 0.72, 0.52, and 0.50, respectively). The AUCs for the radiomics-only (0.70, 0.73, 0.73, and 0.78, respectively) and combined (0.71, 0.80, 0.72, and 0.74, respectively) models were consistently better than those of the base clinical and CT density models across all GOLD stages.

Figure 3 shows the top features of each model. In the base clinical model, pack-years of smoking was the best performing feature. In the radiomics model, the total energy from gray-level cooccurrence matrix–based texture features was the best performing feature, followed by cluster prominence, and sphericity of the left lung. In the combined model, baseline ALD was the best performing feature, followed by pack-years of smoking and cluster prominence. Figure 4 shows the receiver operating characteristic curves for the four models for the classification of rapid progressors versus stable disease.

Figure 3.

Top features for identifying rapid progressors versus stable disease. Results are shown for the (A) base clinical model, (B) radiomics model, and (C) combined model. The x-axis represents the importance score of each feature generated by Light Gradient-Boosting Machine, and the corresponding feature names are shown along the y-axis. Elongation = extent of deviation from being circular; Size zone non-uniformity = variability in the volumes of differently sized regions within lungs; Total energy = total energy of parenchymal texture.

Figure 4.

Receiver operating characteristic (ROC) curve showing the discriminative accuracy for the detection of rapid progressors versus subjects with stable disease. Results are shown for four models on the 20% held-out test data. (A) The blue line represents the ROC curve for the base clinical model, (B) the orange line represents the CT density model, (C) the green line shows the ROC curve for the radiomics model, and (D) the red line represents the ROC curve for the combined model. CT = computed tomography.

Sensitivity analysis

Radiomics features were extracted from CT images that were rotated 45° and 90° to evaluate model robustness with changes in CT acquisition protocol and patient positioning. Compared with the discriminative accuracy of the radiomics feature set derived from original images (AUC, 0.74), the AUC was 0.71 for radiomics features extracted from CT images rotated by 45° and 0.73 when images were rotated by 90°. Similarly, for predicting ALD at 5 years, an R² value of 0.66 was achieved using original CT image, and R² values of 0.70 and 0.71 were achieved after rotating CT image by 45° and 90° respectively. When rapid progression was defined by >0.5% and >1.5% per year ΔALD, the AUCs for the radiomics model were 0.73 (95% CI, 0.70–0.76) and 0.74 (95% CI, 0.70–0.77), respectively. Performance of all models using these thresholds for rapid progression is shown in Table E2.

Discussion

In a cohort of current and former smokers at high risk for disease progression, we demonstrated that radiomics from inspiratory scans alone can be used to predict the progression of emphysema over a 5-year period and to detect rapid progressors. The radiomics features alone had discriminative accuracy that was similar to that of a model that included additional demographics and CT density features. These findings have significant implications for the selection of patients for clinical trials targeting emphysema progression.

There are currently no known treatments that affect structural disease progression in COPD. Several large multicenter clinical trials of pharmacologic therapies have shown an improvement in lung function that is sustained over the period of the trials (24–26). There is no definitive evidence that existing medications decrease the rate of decline of lung function. Clinical trials of lung function decline are limited by the need for large sample sizes given the wide SD of spirometric measures. Spirometry is also nonspecific for the underlying structural processes of emphysema and airway remodeling. These factors have impeded the design and conduct of clinical trials targeting disease progression in COPD.

Recently, there has been more focus on targeting the structural disease underlying COPD using quantitative measures on chest imaging. These measures have less variance than spirometry, and hence studies targeting a change in these parameters are likely to need significantly lower sample sizes (27). This approach has been used in a subset of patients with COPD who have alpha-1 antitrypsin deficiency in which trials of augmentation with intravenous alpha-1 antitrypsin required a considerably smaller sample size to show meaningful effects of the medication (28, 29). However, there was considerable heterogeneity in the response to treatment in these trials. A recent placebo-controlled randomized trial of losartan, an angiotensin receptor blocker, did not result in slowing the progression of emphysema in those with mild to moderate disease (30). These data suggest that a more precision-based approach by targeting individuals at high risk for emphysema progression may result in the need for small sample sizes and a higher chance of impeding the rate of progression.

To enhance the likelihood of targeting individuals at high risk of progression of disease, risk predictors at baseline are needed. Data from the ECLIPSE (Evaluation of COPD Longitudinally to Identify Predictive Surrogate Endpoints) study suggest that blood concentrations of the biomarkers SP-D (surfactant protein D) and sRAGE (soluble receptor for advanced glycation end product) are both associated with emphysema progression, but no cutoffs were determined for prediction of disease progression (31). We improve on several prior models that have used a combination of clinical and imaging measures. Chen and colleagues used propensity-adjusted models using 27 clinical and imaging variables to define disease axes, which were then used to predict emphysema progression (32). These models explained approximately 42% of the variance in emphysema progression (32). Several variables used in this model, such as age of smoking initiation and 6-minute-walk distance, are not readily available in the electronic medical records and need to be prospectively collected. El Kaddouri and colleagues demonstrated that the pattern of emphysema on a semiquantitative scale on baseline scans (mild, moderate, and confluent centrilobular and paraseptal) is associated with emphysema progression but did not report whether these patterns can be used to discriminate between rapid progression and stable disease for a given individual (33). Furthermore, these patterns of emphysema often coexist, and hence these patterns may not accurately identify rapid progressors. We improve on these models by explaining a significantly larger proportion of the variance in emphysema progression and also by estimating the probability of rapid progression.

We identified several radiomic features on chest CT that were consistently associated with the degree of emphysema at 5 years and with the identification of rapid progressors, which add mechanistic insights into emphysema progression. The top features in the radiomics model were the 10th percentile of lung density histogram for the prediction of ALD at 5 years and total energy of parenchymal texture for the identification of rapid progressors. Tenth percentile represents the density (HU) value below which 10% of the lung voxels reside. The closer this value is to −1,000 HU (air), the higher the probability of emphysematous air trapping within the lung. An image’s energy is a metric for homogeneous patterns. A higher energy suggests that the image has more instances of intensity value pairings that are close to one another at higher frequencies. Total energy is calculated by multiplying the value of the energy feature by the volume of the voxel in cubic millimeters. A higher energy value of parenchyma represents more variations and heterogeneity in the density profile of a given region of interest. We, and others, have shown that the presence of emphysema begets more emphysema (7, 8). This is most likely by way of mechanical transduction. Periods of increased mechanical stress, such as during acute exacerbations, are associated with faster progression of emphysema (10). Other top features included aspects related to shape of the lungs that may be related to the remodeling associated with air trapping and hyperinflation noted with progressive emphysema. The ratio of airway surface area to volume was also noted to be an important feature, a measure we recently described as a useful metric for disease prognostication (34). In addition, GLN representing subtle texture changes associated with emphysema was also identified as a top feature in radiomics model. GLN reflects the distribution of intensities/gray levels in the region of interest. A higher GLN value represents how intensities in a texture deviate from a homogeneous or smooth textural pattern.

Strengths and Limitations

Our study has several strengths. The COPDGene study is a cohort of well-phenotyped participants that included both healthy control subjects and smokers with a wide spectrum of disease severity. Spirometry and CT images were gathered using a standard protocol and subject to stringent quality control. The robustness of the results was enhanced by running the models with 10-fold cross-validation and the validation of results in a separate held-out dataset. In sensitivity analyses, we rotated the input images by 45° and 90° and regenerated radiomics features. The performance of the models did not change significantly with rotation, supporting the robustness of the model to changes in model acquisition and patient positioning that may be encountered in clinical practice. We also note that in contrast to the clinical and CT density models, the performance of the radiomics models does not decrease with increasing disease severity, perhaps because of the inclusion of additional features of airway shape which may also be affected with increasing disease severity. The study also has a few limitations. The cohort was composed mostly of current and former smokers. This limits generalizability to a nonsmoking population, but smokers are at high risk for emphysema progression. We did not include any individuals with other coexisting lung diseases, which may alter the performance of the prediction models. More research is needed to validate these algorithms in nonsmokers as well as in the presence of coexisting chronic lung diseases. All parenchymal measures can be affected by the volume of acquisition. In COPDGene, breath-hold was not spirometry gated, but all participants were coached to inhale to total lung capacity, reflecting clinical practice. We used an arbitrary threshold of 1% ΔALD to define rapid progressors, but our models to predict ALD at 5 years are based on estimation of ALD as a continuous parameter. The 1% threshold has been used previously (17). Nonetheless, we conducted additional sensitivity analyses for >0.5% and >1.5% annual ΔALD, and the results were comparable with the primary results.

Conclusions

We demonstrate that lung radiomics can be used to predict lung density at approximately 5 years and to identify rapid progressors. Identifying rapid progressors is critical for the enrichment of future clinical trials targeting emphysema progression with those who are most likely to progress.