Chest MRI and CT Predictors of 10-Year All-Cause Mortality in COPD

Abstract

Pulmonary imaging measurements using magnetic resonance imaging (MRI) and computed tomography (CT) have the potential to deepen our understanding of chronic obstructive pulmonary disease (COPD) by measuring airway and parenchymal pathologic information that cannot be provided by spirometry. Currently, MRI and CT measurements are not included in mortality risk predictions, diagnosis, or COPD staging. We evaluated baseline pulmonary function, MRI and CT measurements alongside imaging texture-features to predict 10-year all-cause mortality in ex-smokers with (n = 93; 31 females; 70 ± 9years) and without (n = 69; 29 females, 69 ± 9years) COPD. CT airway and vessel measurements, helium-3 (³He) MRI ventilation defect percent (VDP) and apparent diffusion coefficients (ADC) were quantified. MRI and CT texture-features were extracted using PyRadiomics (version2.2.0). Associations between 10-year all-cause mortality and all clinical and imaging measurements were evaluated using multivariable regression model odds-ratios. Machine-learning predictive models for 10-year all-cause mortality were evaluated using area-under-receiver-operator-characteristic-curve (AUC), sensitivity and specificity analyses. DL_CO (%_pred) (HR = 0.955, 95%CI: 0.934-0.976, p < 0.001), MRI ADC (HR = 1.843, 95%CI: 1.260-2.871, p < 0.001), and CT informational-measure-of-correlation (HR = 3.546, 95% CI: 1.660-7.573, p = 0.001) were the strongest predictors of 10-year mortality. A machine-learning model trained on clinical, imaging, and imaging textures was the best predictive model (AUC = 0.82, sensitivity = 83%, specificity = 84%) and outperformed the solely clinical model (AUC = 0.76, sensitivity = 77%, specificity = 79%). In ex-smokers, regardless of COPD status, addition of CT and MR imaging texture measurements to clinical models provided unique prognostic information of mortality risk that can allow for better clinical management.Clinical Trial Registration: www.clinicaltrials.gov NCT02279329

Keywords:

• Ex-smokers
• mortality
• computed tomography
• hyperpolarized gas MRI
• machine-learning
• texture analysis

Introduction

Chronic obstructive pulmonary disease (COPD) is among the leading causes of global mortality [1]. Currently, mortality risk is predicted using a variety of clinical-based models [2–4], the most common being the BODE index [5] (Body-mass-index, Obstruction measured via forced-expiratory-volume-in-one-second (FEV₁), Dyspnea-score and Exercise limitation measured via six-minute-walk-distance (6MWD)). A meta-analysis comparing various prognostic clinical models showed that compared to BODE, models based on the Age-Dyspnea-score-FEV₁ (ADO) index [4] may be stronger, but these were not statistically significantly different at predicting three-year survival [6]. Decades after the predictions of the Fletcher-Peto model [7], the spirometry measurement of FEV₁ remains the clinical measurement of global lung function that helps diagnose and stratify COPD severity [8–10].

The small airways are considered the major site of airflow limitation in COPD [11] and spirometry measured at the mouth is not sensitive to small airway measurements. However, currently chest imaging measurements are not included in clinically-accepted mortality risk assessments, diagnosis, prognosis, nor staging of COPD. Chest x-ray computed tomography (CT) provides quantitative measurements of airway wall thinning [12], luminal narrowing and obliteration [13], and parenchymal measurements of gas-trapping and terminal airspace enlargement or emphysema [14].

Recent studies stemming from the COPD Genetic Epidemiology study (COPDGene) cohort showed that in ever-smokers with emphysema, emphysema progression over 5-years was associated with all-cause mortality [15]. Emphysema also worsened more quickly in COPD patients with preexisting emphysema who continued to smoke [16]. In addition, decreased number of distal vascular branches (i.e. pruning) on CT was associated with an increased mortality risk in both COPD and healthy adults [17]. CT measurements of ex vivo lung tissue cores with micro-CT also suggested that COPD may initiate in the small airways and that small airway abnormalities precede the development of emphysema and airflow obstruction [18–21].

Qualitative visual CT scoring by radiologists shows a stronger association with both pulmonary function and mortality than standard quantitative CT measurements [22,23]. This could, in part, be explained by visual estimates of emphysema describing both decreased tissue density and complexity of emphysema distribution (and the predominant emphysema type), which may only be identified by a highly-trained chest radiologist [24]. Importantly, CT images consist of embedded electron-density textural features, which can be exploited using texture analysis tools to map voxel intensity (attenuation) and spatial relationships [25] that are not easily identified by expert observers. In this regard, CT texture analysis and machine-learning tandems have been shown to predict COPD severity [26], progression [27], and showed stronger association with lung function compared to conventional densitometry measures [28,29]. Imaging textures have also been shown to differentiate emphysema types [30,31] and provided improved radiological finding assessments as a second-reader [32].

Pulmonary functional magnetic resonance imaging (MRI) using hyperpolarized helium (³He) and xenon (¹²⁹Xe) gases provides a way to measure pulmonary microstructure, ventilation, perfusion, and gas-exchange [33–35]. In patients with COPD, such MRI measurements of ventilation and parenchyma microstructure [36,37] are predictive of acute exacerbations [38], airway narrowing and remodeling [39], and symptoms and severity [40,41]; these measurements also correlate with longitudinal changes in quality-of-life [42]. Importantly, MRI measurements are sensitive to COPD disease-related changes in patients in whom CT and pulmonary function test results have not changed [43,44]. In addition, MRI ventilation texture features have been shown to predict longitudinal lung-function decline in ex-smokers with and without COPD [45]. However, to the best of our knowledge, ventilation and diffusion-weighted MRI measurements and radiomics-based CT/MRI textures have not been investigated for the prediction of 10-year all-cause mortality in ex-smokers with and without COPD.

Given all of this previous evidence, we hypothesized that, regardless of COPD status, incorporating chest MRI and CT measurements and image texture analysis in combination with machine-learning would provide unique prognostic information for mortality risk assessments in ex-smokers. Hence, here we evaluated MRI and CT measurements and employed imaging texture-analysis to predict all-cause mortality in ex-smokers with and without COPD after 10-years.

Materials and methods

Study participants

All participants provided written informed consent to a study protocol approved by a local research ethics board (Institutional Ethics Board #00000984) in compliance with the Health Canada approved and registered protocol (clinicaltrials.gov NCT02279329). All participants were recruited from a tertiary-care academic center and by advertisement in London, Canada between 2009 and 2012 as a convenience sample. These participants were followed for 10-years. Inclusion criteria were age of 50-85 years and a history of cigarette smoking >10 pack-years at baseline visit. Exclusion criteria included current smokers, claustrophobia and any contraindications for MRI or CT. Death dates were obtained from the health electronic record and survival time was calculated from the participant’s baseline visit date. Some longitudinal results from this study were previously reported [46–48]. In contrast to previous evaluations, this study quantified CT and MR imaging texture features at baseline and their association with 10-year all-cause mortality.

Pulmonary function tests and questionnaires

Spirometry, plethysmography and the diffusing capacity of the lungs for carbon monoxide (DL_CO) were measured according to the American Thoracic Society/European Respiratory Society standardization document [49] using a whole-body plethysmography system (MedGraphics Corporation, St Paul, MN, USA) and attached gas analyzer [46]. COPD was defined as post-bronchodilator spirometry according to the Global Initiative for Chronic Obstructive Lung Disease (GOLD) criteria [10]. Abnormal DL_CO was defined as DL_CO <75%_pred as previously reported [50]. The 6MWD [51] test and St. George’s Respiratory Questionnaire (SGRQ) [52] were administered under the supervision of study personnel.

CT Acquisition and analysis

Thoracic CT was acquired using a 64-slice Lightspeed VCT scanner (GE Healthcare, Milwaukee, WI, USA) (64 × 0.625 mm, 120 kVp, 100 effective mA, tube rotation time= 500 ms, pitch= 1.25, reconstructed using a standard convolution kernel to 1.25 mm slice thickness, slices= 200-250), as previously described [53]. Images were acquired in the supine position under breath-hold after inhalation of a 1 L bag of N₂ (from FRC lung volume) in order to match the lung volume for MRI. Total effective dose was estimated as 1.8 mSv using the ImPACT CT patient dosimetry calculator (Health Protection Agency [UK] NRBP-SR250). CT data were quantitatively evaluated using VIDAvision2.2 software (VIDA Diagnostics Inc., Coralville, IA, USA) to quantify lung density using the relative area of lung less than −950 Hounsfield units (RA₉₅₀), total lung volume (TLV) and create a binary lung mask. CT vessel measurements were automatically generated using Chest Imaging Platform (Brigham and Women’s Hospital, Boston MA) [54].

MRI Acquisition and analysis

Anatomic proton (¹H) and hyperpolarized ³He MR images were acquired using a whole-body 3.0 Tesla Discovery MR750 system (GE Healthcare, Milwaukee, WI, USA), a whole-body radiofrequency coil and a fast gradient recalled echo (FGRE) sequence with a partial echo implementation, with acquisition parameters as previously described [55]. Hyperpolarized ³He MRI was acquired using a linear bird-cage transmit/receive chest coil (RAPID Biomedical GmbH, Wuerzburg, Germany). A commercial system (HeliSpin™, Polarean Inc, Durham, NC, USA) was used to polarize ³He gas to 30–40% and doses (5 mL/kg body weight) diluted with N₂ were administered in 1 L Tedlar® bags (from FRC lung volume). Hyperpolarized ³He MRI diffusion-weighted imaging was performed using a 2D multi-slice fast gradient-echo method, as previously described [55], during breath-hold for acquisition of two interleaved images with and without additional diffusion sensitization with b = 1.6 sec/cm² (maximum gradient amplitude [G] = 1.94 G/cm, rise and fall-time = 0.5 ms, gradient duration = 0.46 ms, diffusion time = 1.46 ms) [48].

MR images were evaluated for the measurement of ventilation defect percent (VDP) using semi-automated custom-built software, as previously described [36]. Briefly, the anatomic ¹H and functional ³He images were first co-registered to segment and remove the large airways (trachea), then a k-means clustering approach was used to generate ventilation clusters, with the lowest cluster representing ventilation defects as previously described [36]. Ventilation abnormalities were quantified as the ventilation defect volume (VDV) and VDP was calculated as VDV normalized to the MRI-measured volume of the thoracic cavity [36]. Diffusion-weighted images were automatically processed to generate apparent diffusion coefficient (ADC) values and images, after the removal of the large airways, as previously described [46]. Abnormal ³He ADC was defined as ADC >0.25 cm²/s, as previously reported [56,57].

Texture feature extraction, selection and machine-learning

CT images were first pre-processed by extracting the lungs and removing the large airways using the binary lung mask generated from the segmented CT images. Signal normalization was applied to the MR images, while a threshold between −1000 Hounsfield Units (HU) and 0 HU was applied to the CT images. The binary mask was applied to the segmented CT images in order to create regions of interest (ROI) for feature extraction. Similarly, the ROI from MR images was generated using the binary lung mask created by co-registering the ¹H and ³He MRI acquisitions, as previously described [45]. Next, 110 unique, unfiltered texture features were extracted in a voxel-by-voxel manner from CT and MR images using the open-source PyRadiomics platform (version 2.2.0) [58]. For the best compromise between differentiation and resolution, a fixed bin number (FBN) discretization approach [59] was utilized to extract features between CT and MR modalities and ensure that textures are assessed against similar contrasts within the modality ROI, as previously described [60]. Histogram and shape, first-, and higher-order texture features from run-length, gap-length, size-zone, neighborhood-dependence and co-occurrence matrices were computed. In addition, 376 wavelet-based texture features were extracted using four high- and low-pass filter combinations applied to the original image in x- and y-directions for wavelet decomposition [61].

To maximize the model generalizability and avoid overfitting, a combination of principal component analysis (PCA) and Boruta analysis [62] were implemented for feature selection. Feature selection step is primarily required for removing any redundant features and/or misleading data for improved modeling accuracy and also used for dimensionality reduction of data, enabling more efficient computations. Feature selection included the generation of nine components from PCA, which explained >94% of the variance in the data. Components were generated using principal component scores for each participant using a Varimax rotation method with Kaiser-normalization that converged after 38 iterations.

All variables in the models, including the texture features and emergent components generated for every participant, were subjected to Boruta analysis for ranking. The Boruta algorithm generated shadow features for comparisons and used a two-step correction for multiple testing, with an optimizable random forest classifier for iterations (number of trees in the forest = 150, maximum iterations = 200, maximum tree depth = 10 [branches], percentage of shadow feature threshold = 95%, alpha-level = 0.05).

Once all the features and parameters were selected, four machine-learning models were generated using: 1) clinical measurements, 2) imaging measurements, 3) image texture measurements, and 4) a combination of all available measurements. Five-fold cross-validation was performed to avoid overfitting or selection bias of the machine-learning models during the training step. During cross-validation, all participant data (n = 162) were randomly and evenly divided into five groups (n = 32/33) and for each of the “folds”, one group was withheld for testing and remaining groups were used for training iteratively. Each fold utilized a different combination of testing and training groups in order to avoid data contamination and insure that no information was carried over from training to testing steps. Single (Naïve Bayes [63], Support Vector Machines [SVM] [64], Decision Trees [65], K-Nearest Neighbors [KNN] [65]) and ensemble (Bagged Trees [66], subspace Discriminant [67], subspace K-Nearest-Neighbors [67], and Random Under-Sampling Boosted Trees [RUSBoosted] [68]) machine-learning classifiers were implemented for predicting 10-year all-cause mortality in ex-smokers. Data were standardized and hyperparameter search was performed using MATLAB2021a (Classification Learner App) for every machine-learning model. Classification performance was evaluated using the mean of the respective 5-fold cross-validation area under the receiver-operator characteristic curve (AUC), as well as sensitivity and specificity calculated from model’s confusion matrix.

Statistical analysis

Statistical analysis was performed using SPSS Statistics v28.0 (IBM Statistics, Armonk, New York, USA). Predictors of 10-year all-cause mortality were evaluated using binary logistic regression to generate odds ratios (OR). Shapiro-Wilk tests were used to determine the normality of the data. The pvalue significance was determined using the Mann-Whitney U-test for non-parametric data, and a post-hoc analysis using Holm-Bonferroni correction was applied for multiple comparison tests for the selected texture features. Statistical significance was considered using a 5% Type-I error threshold (p < 0.05).

Results

Participant demographics and mortality

A CONSORT diagram provided in Figure 1 shows that 266 ex-smokers were enrolled and 99 were excluded from analysis due to enrollment in a sub-study using oscillatory positive expiratory pressure device (n = 33), due to cancelation or not completing all protocol tests (n = 61), and due to poor image quality (n = 5). In addition, five participants were excluded because 10-year follow-up was outside the Dec 2009-2022 window. As shown in Figure 2, the last participant enrolled on December 12^th 2012 and the mortality data window closed on December 13^th, 2022.

Figure 1. CONSORT Flow diagram. Of the 266 participants enrolled in the TINCan study, 33 were enrolled in a sub-study, 61 either canceled or did not complete all required tests during visit 1, and five had CT or MRI artifacts which precluded analysis. Of the 167 participants who completed Visit 1, five were not yet within their 10-year follow-up timeframe. At follow-up, there were 52 deceased participants, of whom 14 were ex-smokers and 38 were ex-smokers with COPD.

Figure 2. Participant enrollment and follow-up timeframe. Arrows showing the timeline of participant enrollment and follow-up period for mortality data collection within 10-years from the initial visit in all ex-smoker participants. Of the 162 ex-smokers analyzed, the first participant completed the baseline visit on December 1^st, 2009 and the last participant completed their baseline visit on December 12^th, 2012.

Demographic, clinical and imaging data for survivor and deceased ex-smoker subgroups are provided in Table 1 and summarized by COPD status in Table S1 (online supplement). In total, 162 ex-smokers were evaluated, including 93 ex-smokers with spirometry evidence of COPD and 69 ex-smokers without COPD. As shown in Table 1, 52/162 (32%) ex-smokers died within the 10-year window. Just over half of these (20/52, 53%) had CT evidence of emphysema, defined by the published threshold, RA₉₅₀ >6.8% [69]. There were no significant differences between survivors and deceased participants for sex, pack-years, total lung capacity, and the CT measurements of emphysema (lowest attenuating cluster) and the pulmonary vasculature (total blood volume, volume of blood in small and mid-sized vessels).

Table 1. Participant demographics, pulmonary function and imaging measurements in survivors and deceased ex-smokers.

Parameter mean(±SD)	All (n = 162)	Survivors (n = 110)	Deceased (n = 52)	p-value
Age [years]	71 (9)	69 (9)	73 (8)	.001
Females [n(%)]	60 (37)	40 (36)	20 (38)	.3
Pack Years	39 (25)	37 (23)	41 (27)	.4
BMI [kg/m^2]	28 (4)	28 (4)	26 (5)	.01
FEV1 [%pred]	80 (22)	81 (25)	65 (30)	.001
FEV1/FVC	60 (18)	66 (16)	53 (21)	<.001
TLC [L]	6.6 (1.3)	6.7 (1.2)	6.5 (1.5)	.3
RV/TLC	47 (11)	44 (12)	52 (11)	.001
DLCO [%pred]	63 (22)	70 (21)	49 (19)	<.001
6MWD [meters]	384 (97)	407 (81)	335 (104)	<.001
SGRQ [score]	33 (21)	29 (20)	44 (19)	<.001
Without COPD [n(%)]	69 (41)	55 (48)	14 (27)	.007
With COPD [n(%)]	98 (59)	60 (52)	38 (73)	.01
GOLD Severity
GOLD I [n(%)]	22 (14)	17 (15)	5 (9)	.3
GOLD II [n(%)]	42 (26)	29 (26)	13 (25)	.7
GOLD III [n(%)]	22 (14)	8 (7)	14 (27)	.004
GOLD IV [n(%)]	7 (4)	1 (1)	6 (11)	–
CT and MR Imaging Measurements
^3He VDP [%]	13 (10)	11 (9)	16 (11)	.002
^3He ADC [cm^2/s]	.35 (.11)	.32 (.09)	.41 (.12)	<.001
HU15^th Percentile	−914 (33)	−908 (29)	−927 (26)	.005
LAC	−1.9 (.3)	−1.9 (.3)	−1.8 (.3)	.1
TBV [ml]	265 (59)	264 (57)	272 (73)	.2
BV5 [ml]	115 (28)	117 (26)	109 (31)	.07
BV5-10 [ml]	60 (16)	59 (14)	62 (20)	.3
BV10 [ml]	89 (32)	86 (3)	101 (4)	.007
BV5/TBV [%]	44 (8)	45 (7)	40 (9)	.003
BV5-10/TBV [%]	22 (3)	22 (3)	22 (4)	.7
BV10/TBV [%]	33 (6)	32 (5)	36 (8)	<.001
RA950 [%]	6.8 (9.2)	4.6 (6.1)	11.7 (12.5)	<.001
Emphysema [n(%)]	50 (31)	30 (25)	20 (53)	.03

BMI: body mass index; FEV₁: forced expiratory volume in 1 s; %_pred: percent of predicted value; FVC: forced vital capacity; RV: residual volume; TLC: total lung capacity; DL_CO: diffusing capacity of lung for carbon-monoxide; 6MWD: six minute walk distance; SGRQ: St. George’s respiratory questionnaire; GOLD: Global Initiative for Chronic Obstructive Lung Disease; TBV: Total blood volume; BV5: vessel volume for vessels less than 5 mm²; BV5-10: vessel volume for vessels between 5-10 mm²; BV10: vessel volume for vessels greater than 10 mm²; LAC: lowest attenuating cluster; RA₉₅₀: relative area of lung less than -950 Hounsfield Units; ADC: apparent diffusion coefficient; VDP: ventilation defect percent.

p=uncorrected values showing significant differences between survivor and deceased ex-smokers.

Texture Extraction and selection

Table 2 shows the six MRI and seven CT texture features which were the highest performing machine-learning predictors of 10-year all-cause mortality. Texture feature predictors were grouped into fine and coarse textures based on their mathematical definitions and are described in Table S2. As shown in Table S3, most of these texture features were significantly correlated with clinically-relevant measures (BMI, FEV₁, 6MWD, and SGRQ). Figures 3 and 4 provide context for the CT and MRI textures respectively. Figure 3 shows coronal chest CT slices and in the inserts, representative ROI and example textures. In the survivor with COPD (top panel: S13), there was negligible CT evidence of emphysema (RA₉₅₀=10.8%) and the CT texture was visually homogeneous (e.g. GLCM-Imc2 = 0.77) similar to a survivor without COPD (top panel: S11, GLCM-Imc2 = 0.76). In contrast, for the deceased ex-smoker with COPD and quantitative evidence of emphysema (RA₉₅₀=24.9%) (bottom panel: S35, GLCM-Imc2 = 0.84) and deceased ex-smoker with no evidence of COPD (bottom panel: S50, GLCM-Imc2 = 0.82), CT textures appeared patchy and heterogeneously coarse.

Figure 3. Chest CT for representative surviving and deceased ex-smokers with and without COPD. Coronal center-slice of chest CT and the corresponding qualitative CT texture heterogeneity. Top panel: A 63 yo male ex-smoker with COPD: FEV₁=72% _pred, FEV₁/FVC = 50, BMI =27 kg/m², DL_CO=79%_pred, ADC = 0.38 cm²/s, VDP = 10%, RA₉₅₀=10.8%, GLCM-Imc2=.77, Wavelet-HH-GLDM-DV=.812; And a 66 yo female ex-smoker: FEV₁=80%_pred, FEV₁/FVC = 76, BMI =36 kg/m², DL_CO=80%_pred, ADC = 0.24 cm²/s, VDP = 5.4%, RA₉₅₀=2.3%, GLCM-Imc2=.76, Wavelet-HH-GLDM-DV=.766; Bottom panel: A 78 yo male ex-smoker with COPD that died: FEV₁=38%_pred, FEV₁/FVC = 39, BMI =20 kg/m², DL_CO=30%_pred, ADC = 0.55 cm²/s, VDP = 28%, RA₉₅₀=24.9%, GLCM-Imc2=.84, Wavelet-HH-GLDM-DV = 1.14; And a 64 yo female ex-smoker that died: FEV₁=111%_pred, FEV₁/FVC = 82, BMI = 36 kg/m², DL_CO=68%_pred, ADC = 0.26 cm²/s, VDP = 4.5%, RA₉₅₀=1.2%, GLCM-Imc2=.82, Wavelet-HH-GLDM-DV = 1.02.

Figure 4. Hyperpolarized gas MRI for representative surviving and deceased ex-smokers with and without COPD. Coronal center-slice of MRI ADC and ventilation with corresponding qualitative and quantitative MRI texture heterogeneity. Top panel: A 63 yo male ex-smoker with COPD: FEV₁=72%_pred, FEV₁/FVC = 50, BMI = 27 kg/m², DL_CO=79%_pred, ADC = 0.38 cm²/s, VDP = 10%, RA₉₅₀=10.8%, Shape-SVR=.43, Wavelet-LL-Skewness = 0.58; And a 66 yo female ex-smoker: FEV₁=80%_pred, FEV₁/FVC = 76, BMI = 36 kg/m², DL_CO=80%_pred, ADC = 0.24 cm²/s, VDP = 5.4%, RA₉₅₀=2.3%, Shape-SVR=.45, Wavelet-LL-Skewness=.81; Bottom panel: A 78 yo male ex-smoker with COPD that died: FEV₁=38%_pred, FEV₁/FVC = 39, BMI = 20 kg/m², DL_CO=30%_pred, ADC = 0.55 cm²/s, VDP = 28%, RA₉₅₀=24.9%, Shape-SVR=.62, Wavelet-LL-Skewness = 2.1; And a 64 yo female ex-smoker that died: FEV₁=111%_pred, FEV₁/FVC = 82, BMI = 36 kg/m², DL_CO=68%_pred, ADC = 0.26 cm²/s, VDP = 4.5%, RA₉₅₀=1.2%, Shape-SVR=.48, Wavelet-LL-Skewness = 1.25.

Table 2. MRI and CT texture features in survivors and deceased ex-smokers.

Parameter mean (±SD)	All (n = 162)	Survivors (n = 110)	Deceased (n = 52)	*p-value
Coarse textures
MRI FO-skewness	.77 (.46)	.63 (.38)	1.06 (.49)	<.001
MRI shape-SVR	.49 (.10)	.47 (.09)	.52 (.13)	.02
MRI W-LL-FO-kurtosis	3.68 (2.06)	3.25 (0.99)	4.58 (3.17)	.01
MRI W-LL-cluster shade	673 (765)	465 (539)	1110 (970)	<.001
MRI W-LL-FO-skewness	.75 (.47)	.61 (.38)	1.04 (.49)	<.001
MRI GLCM-Idn	.996 (.001)	.996 (.001)	.995 (.001)	<.001
CT Shape-sphericity	.455 (.069)	.447 (.060)	.473 (.084)	.052
CT RLM-run percentage	.944 (.009)	.948 (.010)	.940 (.013)	.001
Fine textures
CT GLCM-Imc1	−0.09 (.03)	−0.09 (.02)	−0.11 (.02)	<.001
CT GLCM-Imc2	.79 (.06)	.77 (.05)	.82 (.06)	<.001
CT W-HH-GLDM-DV	.80 (0.24)	.74 (0.21)	.92 (0.26)	<.001
CT W-HH-run variance	.12 (.05)	.11 (.04)	.15 (.05)	<.001
CT W-LL-GLCM-Imc1	−0.12 (.02)	−0.12 (.01)	−0.14 (.02)	<.001

VDP: ventilation defect percent; ADC: apparent diffusion coefficient; LAC: lowest attenuating cluster; TBV: Total blood volume; BV₅: vessel volume for vessels less than 5 mm²; BV_5-10: vessel volume for vessels between 5-10 mm²; BV₁₀: vessel volume for vessels greater than 10 mm²; RA₉₅₀: relative area of lung less than -950 Hounsfield Units; FO: first order features; SVR: surface to volume ratio; W: wavelet; GLCM: gray level co-occurrence matrix; Idn: inverse difference normalized; LL: low-low-pass filter; HH: high-high-pass filter; RLM: run length matrix; Imc2: informational measure of correlation; GLDM: gray level dependence matrix; DV: dependence variance; All selected texture feature abbreviations and descriptions can be found in Table S2.

*Holm-Bonferroni corrected pvalue.

Figure 4 shows coronal MRI ADC and ventilation slices and in the inserts, representative ROI and example MRI ventilation textures. In the representative survivor with COPD (top panel: S13, mean ADC = 0.38 cm²/s, VDP= 10%, Wavelet-LL-FO-Skewness = 0.58) and survivor without COPD (top panel: S11, mean ADC = 0.24 cm²/s, VDP = 5.4%, Wavelet-LL-FO-Skewness = 0.81), MRI ventilation textures were visibly homogenous. In contrast, in the deceased ex-smoker with COPD (bottom panel: S35, mean ADC = 0.55 cm²/s, VDP = 28%, Wavelet-LL-FO-Skewness = 2.1) and ex-smoker without COPD (bottom panel S50, mean ADC = 0.26 cm²/s, VDP = 4.5%, Wavelet-LL-FO-Skewness = 1.25), MRI ventilation textures were heterogeneous and patchy.

Predicting 10-year all-cause mortality

As shown in Table 3, machine-learning prediction models for 10-year all-cause mortality were generated using: 1) clinical measurements, 2) imaging measurements, 3) image texture measurements, and 4) a combination of clinical and imaging measurements. The best performing clinical model (77% accuracy) was based on age, pack years, BMI, FEV₁%_pred, TLC, DL_CO %_pred, 6MWD and SGRQ score. The best performing imaging-based model (77% accuracy), included MRI VDP and ADC as well as CT HU15^th percentile, TBV, BV₅/TBV, BV₁₀, BV₁₀/TBV, and %RA₉₅₀. The predictive model based exclusively on imaging texture features outperformed both clinical and imaging-based models (80% accuracy). Finally, the combined clinical-imaging model had the overall best performance (83% accuracy) which included: DL_CO, MRI ADC, as well as MRI and CT texture features.

Table 3. Machine-learning performance at predicting all-cause mortality after 10-years.

Best performing models	AUC	sensitivity (%)	specificity (%)	accuracy (%)
Clinical model†
Medium gaussian SVM	0.76	76.8	79.1	77.2
Imaging model‡
Cosine KNN	0.75	77.7	71.8	76.5
Texture-based model§
Subspace KNN	0.80	80.4	79.4	80.2
Combined model□
Subspace discriminant	0.82	82.8	84.3	83.3

AUC: area under the receiver-operating curve; KNN: K-nearest neighbors; SVM: support vector machine.

†Variables used for training included: Age, Pack years, BMI, %_pred FEV₁, TLC, %_pred DL_CO, 6MWD and SGRQ score.

‡Variables used for training included: ³He VDP, ³He ADC, HU15^th percentile, TBV, BV₅/TBV, BV₁₀, BV₁₀/TBV, and % RA₉₅₀.

§Features selected for training included: MRI wavelet-low-low-First-order-Skewness, MRI wavelet-low-low-kurtosis, CT gray level co-occurrence matrix (GLCM)-informational measure of correlation (Imc2), CT wavelet-high-high-gray-level dependence matrix (GLDM)-dependence variance, CT wavelet-high-high-Run variance, CT wavelet-low-low-first-order-skewness, CT wavelet-low-low-GLCM-Imc1.

□Combined model included: DL_CO, ADC, MRI Wavelet-LL-FO-Skewness, CT LL-GLCM-Imc1, CT Shape-Sphericity, CT Wavelet-HH-GLDM-Dependence variance and MRI Wavelet-LL-FO-Kurtosis.

In Figure 5, logistic regression models are shown in receiver operator characteristic (ROC) curves of all-cause mortality for individual clinical (top panel), imaging (middle panel) and imaging texture (lower panel) measurements. The best performing individual measurements included DL_CO (AUC = 0.736), MRI ADC (AUC = 0.738) and CT Wavelet-LL-GLCM-Imc1 (AUC = 0.787).

Figure 5. Receiver-operator characteristic curves of texture features and clinical variables. Top Panel: Logistic regression analysis of individual clinical variables at predicting 10-year all-cause mortality in ex-smoker participants. DL_CO had the best AUC=.736. Middle Panel: Logistic regression analysis of standard imaging measurements at predicting 10-year all-cause mortality in ex-smoker participants. ³He ADC had the best AUC=.738. Bottom Panel: Logistic regression analysis of imaging texture features at predicting 10-year all-cause mortality in ex-smoker participants. CT Wavelet-LL-GLCM-Imc1 had best AUC=.787.

Figure 6 shows a forest plot for associations adjusted for confounders (age, BMI, sex, and pack-years). The increased risk of 10-year all-cause mortality was strongly associated with the CT texture feature GLCM-Imc2 (OR = 3.546 [per 0.1 change], p = 0.001) and MRI ADC (OR = 1.843 [per 0.1 cm²/s change], p < 0.001).

Figure 6. Logistic regression models for associations between all-cause mortality and clinical, imaging and textural measurements. All-cause mortality assessment was conducted in 162 ex-smokers, of whom 52 deceased across the longitudinal study duration (10-years). Bolded values indicate categories where 95% CI did not include 1.0 (P < 0.05). *All odds ratios were adjusted for age, BMI, sex, and pack-years. GLCM = gray level co-occurrence matrix; 6MWD = six minute walk distance; ADC = apparent diffusion coefficients; DL_CO =diffusing capacity of the lung for carbon monoxide; GLDM = gray level dependence matrix; HH = high-high pass filter; LL = low-low pass filter; SVR = Surface volume ratio; DV = dependence variance; RV = run variance; FEV₁=forced expiratory volume in 1 second; FVC = forced vital capacity; SGRQ = St. George’s respiratory questionnaire; LAC = lowest attenuating cluster; RA₉₅₀=relative area of lung less than -950 Hounsfield Units; All texture feature abbreviations and descriptions can be found in supplementary Table S2.

Kaplan-Meier curves provided in Figure 7 show that ex-smokers with abnormal DL_CO, (log-rank χ²=11.95, p < 0.001), MRI ADC (log-rank χ²=6.38, p = 0.01) and an MRI texture (highest tertile wavelet-LL-FO-Skewness; χ²=21.81, p < 0.001) had a significantly greater risk of death.

Figure 7. Kaplan-Meier survival curves of 10-year all-cause mortality in ex-smokers. Orange: All-cause mortality analysis in ex-smokers with normal vs abnormal MRI ADC (ADC < 0.25cm²/s). Log-rank (Mantel-Cox) test χ²=6.38; P=.01. Black: All-cause mortality analysis in ex-smokers with normal vs abnormal DL_CO (DL_CO<75%_pred). Log-rank (Mantel-Cox) test χ²=11.95; P <.001. Green: All-cause mortality analysis in ex-smokers with tertiles of the MRI Wavelet-LL-FO-Skewness texture feature. Log-rank (Mantel-Cox) test across all tertiles: χ²=22.43; P <.001. Log-rank test between tertile-Low and tertile-Medium: χ²=7.86; P=.005. Log-rank test between tertile-Medium and tertile-High: χ²=4.99; P=.02. Log-rank test between tertile-Low and tertile-High: χ²=21.81; P <.001.

Table 4 summarizes the best performing measurements in descending order, across the key statistical tests and models generated. The common top performing measurements were DL_CO, MRI-ADC, MRI-Wavelet-LL-FO-Skewness and CT-GLCM-Imc2.

Table 4. Summary of top performing measurements for the key statistical analyses.

Statistical test	Measurement	Result
ROC analysis	CT-W-GLCM-Imc1	AUC = 0.787
	MRI-W-Skewness	AUC = 0.771
	CT-GLCM-Imc2	AUC = 0.744
	ADC	AUC = 0.738
	DLCO	AUC = 0.736
	RA950	AUC = 0.691
Logistic regression analysis	CT-GLCM-Imc2	OR = 3.55
	ADC	OR = 1.84
	MRI-W-Skewness	OR = 1.63
	MRI-SVR	OR = 1.60
	MRI-W-Kurtosis	OR = 1.59
	DLCO	OR = 0.955
Kaplan-Meier analysis	MRI-W-Skewness	χ^2=21.81 (1^st vs 3^rd tertile)
	CT-GLCM-Imc2	χ^2=17.67 (1^st vs 3^rd tertile)
	DLCO	χ^2=11.95
	ADC	χ^2=6.38
Combined ML model	MRI-W-Skewness	AUC = 0.82
	CT W-GLCM-Imc1	Sensitivity = 82.8%
	CT Shape-Sphericity	Specificity = 84.3%
	CT W-GLDM-Dependence variance	Accuracy = 83.3%
	MRI W-FO-Kurtosis ADC DLCO

ADC: apparent diffusion coefficient; DL_CO: diffusing capacity of lung for carbon-monoxide; RA₉₅₀: relative area of lung less than -950 Hounsfield Units; FO: first order features; SVR: surface to volume ratio; WL: wavelet; GLCM: gray level co-occurrence matrix; Idn: inverse difference normalized; Imc2: informational measure of correlation; ML: machine-learning; GLDM: gray level dependence matrix; DV: dependence variance; All selected texture feature abbreviations and descriptions can be found in Table S2.

Discussion

A recent investigation in patients with COPD [70] showed that unsupervised learning of chest CT measurements improved predictions of progression, exacerbation, and mortality risk. Another recent study [71] showed that machine learning models of clinical and CT imaging measurements outperformed the current standard (BODE and ADO indices) for predicting all-cause mortality.

Our investigation focused on predicting 10-year all-cause mortality in ex-smokers at risk of COPD and those with spirometry evidence of COPD and provides a number of key advantages relative to previous work including: 1) the addition of volume-matched MRI ventilation and ADC measurements, 2) MRI and CT texture features generated from images acquired within a few minutes of one another to capture structural and functional information, 3) inclusion of ex-smokers at-risk of COPD, and, 4) 10-year follow-up.

We applied machine-learning algorithms to generate models for predicting all-cause mortality using clinical, CT, MRI and imaging textures. We observed that a combined model consisting of DL_CO, MRI ADC and image textures, outperformed all other models. Surprisingly, none of the individual components of the BODE index were included in the “best performing” model after the feature selection step. Ensemble machine-learning models outperformed single machine-learning models, suggestive of complex, non-linear relationships between the individual imaging textures and mortality.

We were surprised to observe that the strongest individual imaging predictor was MRI ADC (AUC = 0.74), which outperformed all the individual components of the BODE index (except mMRC which was not measured) and all CT measurements. In fact, mortality risk increased by 84% for every 0.1 cm²/s increase in the ADC value, which underscores the high sensitivity of ADC measurements to terminal airspace enlargement due to air-trapping, emphysema or both [43,46]. This result also agrees with the COPDGene study results, whereby emphysema progression over 5-years was associated with mortality in ever-smokers with trace emphysema [15]. This result also agrees with the finding that the MRI ADC measurement is highly sensitive to terminal airspace enlargement as previously shown [43] and by comparison with histological measurements of the mean linear intercept [72].

We observed that one of the “fine” CT textures (GLCM-Imc2) had the strongest independent association with 10-year all-cause mortality, which after adjusting for confounders resulted in a three-fold increased mortality risk (OR = 3.55, p = 0.001). Such “fine” CT textures may be intuitively considered as reflecting tissue attenuation heterogeneity. To provide context, previous work demonstrated that the spatial arrangement of low attenuating voxels, or the size and arrangement of emphysematous clusters can differentiate patients with similar COPD disease severity [73]. However, unlike a more commonly used CT measurement of emphysema, RA₉₅₀, CT texture features reflect complex spatial heterogeneity, which may be argued is more sensitive to emphysema [31,74] and perhaps similar to what is often visually interpreted as emphysema by expert chest CT radiologists [24]. In contrast throughout this study, all MRI texture features selected were “coarse” and could be considered as reflecting the compactness and asymmetry of the ventilation distribution in the lungs.

Underscoring the power of MRI and CT measures as predictors of all-cause mortality, forest plots revealed significant associations with mortality for all selected MRI/CT textures, CT vascular measurements and MRI ADC. These results are consistent with those of recent studies reporting that CT emphysema [15,70,71] and vascular [17,54,75] measurements are associated with mortality and disease progression in COPD. Our findings contribute to the growing body of evidence about the utility of imaging for the management of patients at risk for, and with, a diagnosis of COPD.

In agreement with the ROC analysis and forest plots, Kaplan-Meier curves revealed that ex-smokers with abnormal DL_CO, MRI ADC and a specific MRI texture were at greatest risk of mortality. Of note, across all models, the measurements which remained significant and highly associated with 10-year mortality were DL_CO, MRI ADC, MRI Wavelet-LL-FO-Skewness and CT GLCM-Imc2. These consistent observations are indicative of the added-value of imaging measurements and textures for mortality risk assessments in ex-smokers with and without COPD.

We acknowledge a number of study limitations. First, direct comparison of our results with the BODE index should be undertaken with caution because we did not acquire the modified Medical Research Council dyspnea-score component of the BODE index. Second, our study included a relatively small sample size compared to other CT studies [76,77]. Fortunately, we employed methods to avoid overfitting so the machine-learning methods we used may be generalized with larger datasets and multicenter data in the future. We also acknowledge that lung imaging measurements are influenced by lung volume [78], and in this study, data were captured at FRC + 1L which for many ex-smokers is within 90% of TLC [78,79]. Finally, the availability of functional MRI for chest imaging is still limited to research sites. Thus, while such quantitative MR measurements are very helpful, CT measurements are certainly more readily available and generalizable to most clinical centers.

Inclusion of at risk ex-smokers may help identify individuals who are at high-risk of death and can potentially be included in clinical trials [80]. It is well established that asymptomatic ex-smokers with mildly abnormal DL_CO are at-risk of developing COPD within four years [50]. The abnormal MRI ADC values reported here have been shown to reflect mild air-trapping and/or subclinical emphysema [43]. Unfortunately, we did not make small airways disease measurements, so we are unable to comment on whether small airway abnormalities have already initiated in these participants with normal spirometry [18–21]. What we do know is that 14/69 (20%) of ex-smokers and 38/93 (41%) of COPD participants were deceased after 10 years and that similar MRI and CT measurements and textures indicative of emphysema helped explain risk. The results presented here expand on previous CT findings in COPD [15,17,54,70,71,75] and extend the application of such imaging measurements to ex-smokers without, but at risk of COPD.

Conclusion

In ex-smokers, regardless of COPD status, DL_CO, CT and MR imaging measurements and textures resulted in high accuracy models for predicting mortality risk. Texture measurements provide a way to reveal MRI and CT lung pathologies that are not visible to the human eye and may help predict 10-year all-cause mortality in ex-smokers.

Authors’ contributions

M.S. was responsible for image processing, statistical analyses and interpretation, as well as manuscript preparation and submission. P.V.W. and A.M.M. assisted with data analysis and interpretation. D.G.M. was responsible for recruitment of study participants, clinical input in the study design, and clinical interpretation of the data. G.P. was responsible for conception of the study, experimental design, data interpretation, and approval of the final manuscript, as well as being the guarantor of study data integrity. All coauthors had an opportunity to review and revise the manuscript and approved its final submitted version.

ABBREVIATIONS
COPD	=	chronic obstructive pulmonary disease
FEV1	=	forced expiratory volume in 1-second
6MWD	=	six minute walk distance
CT	=	computed tomography
COPDGene	=	COPD Genetic Epidemiology study
MRI	=	magnetic resonance imaging
DLCO	=	diffusing capacity of the lungs for carbon monoxide
GOLD	=	Global Initiative for Chronic Obstructive Lung Disease
SGRQ	=	St. George’s Respiratory Questionnaire
RA950	=	relative area of lung less than -950 Hounsfield units.
TLV	=	total lung volume
1H	=	proton
3He	=	helium-3 isotopic gas
FGRE	=	fast-gradient-recalled echo
VDP	=	ventilation defect percent
ADC	=	apparent diffusion coefficient
ROI	=	region of interest
PCA	=	principal component analysis
SVM	=	support vector machine
KNN	=	k-nearest neighbors
RUSBoosted	=	random under-sampling boosted trees
AUC	=	area under the receiver-operator characteristic curve
OR	=	odds ratio

Acknowledgments

We thank the participants who volunteered for this study.

Declaration of interest

The authors declare there is no Complete of Interest at this study.

Disclosure statement

M.S., P.V.W., A.M.M., D.G.M., and G.P. disclosed no potential conflicts of interest.

Data availability statement

Data generated or analyzed during the study are available from the corresponding author upon request.

Additional information

Funding

M.S. was supported by the Natural Sciences and Engineering Research Council (NSERC) of Canada Post-Graduate Doctoral Scholarship. A.M.M. is supported by the Natural Sciences and Engineering Research Council (NSERC) of Canada Canadian-Graduate Doctoral Scholarship. G.P. is supported by NSERC, CIHR, the Baran Family Foundation, and holds a Tier 1 Canada Research Chair.