Introduction

The forced-oscillation-technique (FOT) provides a way to directly measure central and peripheral airway resistance (R), which in participants with asthma, points to airway abnormalities that may be responsible for ventilation heterogeneity.¹ Hyperpolarized noble gas magnetic-resonance-imaging (MRI) provides a way to spatially locate and quantify inhaled gas abnormalities related to airway dysfunction, inflammation and remodeling.^2-4 In previous work, MRI ventilation-defect-percent (VDP)⁵ was shown to be related to peripheral airways resistance measured using FOT in patients with asthma and chronic-obstructive-pulmonary-disease.^6,7

VDP is a binary measurement of ventilation-defects which disregards signal intensity differences and assumes all ventilated regions contribute equally to global lung function. These signal intensity differences, or ventilation heterogeneity, may be quantified as texture features extracted from gray-level run-length (GLRLM) and co-occurrence matrices (GLCM).⁸ Hyperpolarized gas MRI-VDP and MRI texture analysis have been previously used to measure post-bronchodilator improvements^9,10 in airway function and ventilation and to predict methacholine and biologic therapy responsive participants with asthma. Unfortunately, the physiological interpretation of texture features in the context of pulmonary ventilation MRI remains uncharacterized.

Here we hypothesized that MRI-VDP in combination with shape-based, first-order and higher-order texture features will independently describe central (R_19Hz) and peripheral (R_5-19Hz) airways resistance measured using FOT. Therefore, the two objectives of this work were to: 1) extract texture features from ¹²⁹Xe MRI and use a machine-learning regression approach to identify features that explain respiratory system resistance and 2) perform principal-component-analysis (PCA) on significant features to relate these features to oscillometric biomechanics.

Methods

Participants and Data Acquisition:
We retrospectively analyzed 71 participants with eosinophilic-asthma (n=24) or post-acute COVID-19 syndrome (PACS) (n=47) and eight healthy volunteers. Data were randomly split into training (n=60) and testing (n=19) sets. Anatomic ¹H and ¹²⁹Xe static ventilation MRI were acquired using a 3.0 Tesla scanner as previously described.¹¹ Anatomic ¹H MRI was acquired using a fast-spoiled gradient-recalled-echo sequence (partial-echo acquisition; total acquisition time=8s; repetition-time msec/echo time msec=4.7/1.2; flip-angle=30°; field-of-view=40×40cm²; bandwidth=24.4kHz; 128×80 matrix, zero-padded to 128×128; partial-echo percent=62.5%; 15-17×15mm slices). ¹²⁹Xe MRI was acquired using a three-dimensional fast-spoiled gradient-recalled echo sequence (total acquisition time=14s; repetition-time msec/echo time msec=6.7/1.5; variable flip-angle; field-of-view=40×40cm²; bandwidth=15.63kHz; 128×128 matrix; 14×15mm slices). Supine participants were coached to inhale a 1.0L bag (400mL ¹²⁹Xe + 600mL ⁴He for ¹²⁹Xe MRI; 1.0L N₂ for ¹H MRI) from the bottom of a tidal breath with acquisition under breath-hold conditions. Participants performed spirometry¹² according to American Thoracic Society guidelines and FOT¹³ according to European Respiratory Society guidelines using a tremoFlo C-100 Airwave Oscillometry System to measure R between 5 and 37 Hz.

Image Processing and Statistics:
Quantitative MRI analysis was performed using a semi-automated segmentation algorithm, as previously described¹⁴ using Matlab 2021b. Texture features were extracted from the 3D-application of GLRLM and GLCM using the PyRadiomics platform.¹⁵ Feature selection was performed using f-test statistics to independently rank and identify extracted texture features, including MRI-VDP, which significantly contributed to each machine-learning model’s performance. Two machine-learning models were developed to explain R_19Hz and R_5-19Hz using Matlab 2021b. Five-fold cross-validation was implemented to prevent over-fitting. Model performance was evaluated using root-mean-square-error (RMSE). Selected features from both models underwent PCA to describe the physiology responsible for MRI ventilation texture features.

Results

Table 1 provides demographic characteristics for all 79 participants. Figure 1 provides ¹²⁹Xe MRI center slice ventilation images for representative participants with normal or abnormal central or peripheral airways resistance.¹⁶

As shown in Figure 2, we identified six features per model (four unique features per model, two common features) that were used to explain R_19Hz and R_5-19Hz; VDP was not identified as a significant feature (rank: R_19Hz-48/72, R_5-19Hz-65/72) in either model. Table 3 shows the five highest performing models for R_19Hz and R_5-19Hz. Highest performance for R_19Hz and R_5-19Hz was achieved using coarse regression trees (RMSE=0.97) and medium regression trees (RMSE=0.71), respectively.

The ten selected features for both models underwent PCA, which revealed that seven features described ventilation heterogeneity (Shape: Maximum 2D Diameter Slice; First Order: Robust Mean Absolute Deviation; GLCM: Difference Average, Difference Entropy, Inverse Variance, Joint Energy; GLRLM: Gray-Level Non-Uniformity), two features described whole-lung ventilation (First Order: Root Mean Squared, Interquartile Range) and one feature described the shape of the ventilation image (Shape: Flatness).

Discussion

We developed two machine-learning regression models to explain oscillometry-measured central and peripheral airways resistances using ¹²⁹Xe MRI texture features. Central and peripheral airway models each contained four unique and two common features. Both common features were revealed through PCA to describe ventilation heterogeneity, while the central airway feature First Order Root Mean Squared describes highly ventilated pixels while the peripheral airway feature First Order Interquartile Range describes the number of normally ventilated pixels. This suggests that highly ventilated pixels are related to the central airways while the number of normally ventilated pixels are related to the peripheral airways.

Conclusion

Hyperpolarized ¹²⁹Xe MRI texture features that explained central and peripheral airways resistance in patients with eosinophilic-asthma and PACS. We showed for the first time that central and peripheral airway resistance measured using FOT may be explained using hyperpolarized ¹²⁹Xe MRI texture features. These promising results suggest that ¹²⁹Xe ventilation texture analysis may reveal hidden anatomic-physiologic measurements that lead to ventilation heterogeneity.

Acknowledgements

No acknowledgement found.