Hyperpolarized <sup>129</sup>Xe MRI Features Associated with Interstitial Lung Disease Identified Using an Interpretable Diagnostic Algorithm.
Authors
Affiliations (6)
Affiliations (6)
- Department of Biomedical Engineering, Duke University, Durham, North Carolina (F.L., S.L., B.D.).
- Medical Physics Graduate Program, Duke University, Durham, North Carolina (A.C., S.L., D.M., B.D.).
- Department of Radiology, Duke University, Durham, North Carolina (H.Q., B.O., D.M., B.D.).
- Department of Medicine, Division of Pulmonary and Critical Care Medicine, Duke University Medical Center, Durham, North Carolina (A.S.).
- Medical Physics Graduate Program, Duke University, Durham, North Carolina (A.C., S.L., D.M., B.D.); Department of Radiology, Duke University, Durham, North Carolina (H.Q., B.O., D.M., B.D.).
- Department of Biomedical Engineering, Duke University, Durham, North Carolina (F.L., S.L., B.D.); Medical Physics Graduate Program, Duke University, Durham, North Carolina (A.C., S.L., D.M., B.D.); Department of Radiology, Duke University, Durham, North Carolina (H.Q., B.O., D.M., B.D.). Electronic address: [email protected].
Abstract
Hyperpolarized <sup>129</sup>Xe magnetic resonance imaging and spectroscopy (MRI/MRS) have been used to identify numerous imaging and spectroscopic features that distinguish interstitial lung disease (ILD) from chronic obstructive pulmonary disease (COPD) and healthy controls. However, it remains unclear which specific <sup>129</sup>Xe MRI/MRS features, and at what thresholds, are most strongly associated with ILD, relative to COPD and health. <sup>129</sup>Xe MRI/MRS from 155 participants (age 56.6 ± 17.8 years; 80 females), including 84 with ILD, 21 with COPD, and 50 healthy controls, were analyzed to derive eight quantitative metrics. The most informative of these metrics were selected using L1-regularized logistic regression with SHapley Additive exPlanations (SHAP) analysis. These selected metrics then informed a threshold-based decision tree that was evaluated for comprehensive diagnostic performance. Its stability was assessed with bootstrap resampling. The decision tree utilized the following <sup>129</sup>Xe metrics: high membrane uptake (Mem<sub>high</sub>), ventilation defect percentage (VDP), red blood cell (RBC) transfer defect gradient (ΔRDP<sub>BA</sub>), and RBC chemical shift, achieving an overall classification accuracy of 93.5%. Class-specific sensitivity/specificity for healthy, COPD and ILD were 90.0%/95.2%, 100%/ 100% and 94.1%/92.9% respectively, with the stability of features and thresholds confirmed by bootstrapping. This interpretable approach highlights four physiologically meaningful <sup>129</sup>Xe MRI/MRS metrics strongly associated with ILD, supporting their further development for early detection and individualized physiological assessment.