Introduction:

Pulmonary hyperpolarized ³He and ¹²⁹Xe MRI provide a way to measure lung ventilation and parenchyma microstructure in patients with chronic obstructive pulmonary disease (COPD), including ventilation defects and terminal airspace enlargement or emphysema.^[1] X-ray computed tomography (CT) also provides a way to quantify extent of emphysema based on tissue density measurements that are susceptible to lung volume changes, with studies indicating that a change of 1.1% in extent of emphysema may be detected with 95% probability, measured as relative area of lung < -950 Hounsfield Units (RA₉₅₀).^[2] Unfortunately, and likely because of radiation dose considerations, CT follow-up of emphysema is rarely performed in patients, and hence its longitudinal progression is not well-understood.^[3] Here we evaluated a large database of volume-matched hyperpolarized ³He MRI and CT in patients with COPD and hypothesized that ³He MRI ventilation features may be used to predict those patients who will experience rapidly worsening emphysema, after 3 years.

Methods:

Participants and Data Acquisition:
We retrospectively evaluated 47 ex-smokers (>10 pack years, 40-85 years of age) with and without a diagnosis of COPD, who provided an informed consent (NCT02279329) to the TINCan cohort study which included spirometry, ³He MRI and thoracic CT scans.^[4] Long-term follow-up was prospectively planned for 24 ± 6 months after baseline visit.
Spirometry was acquired according to the American Thoracic Society/European Respiratory Society guidelines.^[5] COPD severity was defined according to the Global Initiative for Chronic Obstructive Lung Disease (GOLD) grades.^[6]
Thoracic CT scans were acquired while participants were in supine position using a 64-slice Lightspeed VCT scanner (General Electric Healthcare, Milwaukee, WI) during single breath hold of 1L bag of N₂ gas, with parameters 64×0.625mm, 120kVp, 100 effective mA, tube rotation time= 500ms, pitch= 1.0, slice thickness= 1.25mm, slices= 200-250. The mean total effective dose was 1.8mSv.
MRI was acquired using a whole-body 3.0 T Discovery MR750 (General Electric Healthcare, Milwaukee, WI), a whole-body radiofrequency coil and a fast-gradient-recalled echo (FGRE) sequence with a partial echo implementation to acquire the conventional ¹H images using the following parameters: total-acquisition time= 8s, repetition time[TR] / echo time[TE] / flip angle=4.7ms /1.2ms /30°, field of view[FOV]= 40×40cm², bandwidth[BW]= 24.4kHz, matrix= 128×80 (zero-padded to 128×128), partial echo percent= 62.5%, number of slices= 15-17, slice thickness= 15mm, 0-gap. Hyperpolarized ³He ventilation images were acquired using an FGRE sequence with a partial echo and following parameters: Acquisition time= 10s, TR/TE/flip angle= 3.8ms/1.0ms/7°, FOV= 40×40cm², matrix= 128×80 (zero-padded to 128×128), BW= 48.8kHz, partial echo percent= 62.5%, slices= 15-17, slice thickness= 15mm, 0-gap.

Image Processing and Statistics:
Participants were dichotomized based on a 1.1 %/year change in CT RA₉₅₀ ­measurement between baseline and follow-up visits, as previously described.^[7] Custom algorithm inputs were extracted from the 3-dimensional application of gray level run-length, gap-length, zone-size, dependence and co-occurrence matrices via PyRadiomics platform.^[8] Additional parameters used for training were: 1) CT RA₉₅₀ and 15^th percentile (HU₁₅), 2) MRI-ADC and ventilation defect percent (VDP).^[9] Feature selection was performed using Boruta analysis and a random forest classifier to independently identify CT and MRI features that significantly contributed to machine-learning model’s accuracy. The Shapiro-Wilk test was used to determine the normality of the data distribution. Non-parametric tests were performed for non-normally distributed data. Model performance was evaluated using area under the receiver operator curve (AUC), as well as sensitivity and specificity metrics.

Results:

We identified nine texture feature contributors, which resulted in four texture features that significantly contributed to machine-learning models that were used to predict participants with rapidly worsening CT emphysema at follow-up (2.5±0.6 years). The MRI-based model achieved 82% prediction accuracy with coarse decision tree algorithm and outperformed the standard clinical model (64%) as well as the clinical model with lung volumes included (70%). MRI-derived ADC and VDP were both outperformed by individual MRI-derived texture features, validated using the AUC metric from the generated ROC curves. The best performing clinical variable for this dataset was functional residual capacity or FRC (AUC=0.605), whereas GLCM-Idn was the overall best performing texture feature (AUC=0.767) in the MRI model. Additionally, all texture features except Elongation, FO-Skewness, GLCM-Cluster Shade and RLM-SRLGLE provided weak-to-moderate correlations (r=0.2-0.5, p<.05) with FEV₁, FEV₁/FVC, RA₉₅₀ and MRI-ADC measurements.

Conclusions:

We have identified hyperpolarized gas MRI texture features that independently and uniquely correlated and predicted abnormal changes in CT RA₉₅₀ nearly 3 years later. Future work will necessarily focus on detailed feature interpretation towards a deeper understanding of the physiological and pathological mechanisms of progressive emphysema worsening.

Acknowledgements

No acknowledgement found.