Automation through Deep-Learning to Quantify Ventilation Defects in Lungs from High-Resolution Isotropic Hyperpolarized 129Xe Magnetic Resonance Imaging

Tuneesh Kaur Ranota, The University of Western Ontario

Abstract

Obstructive lung diseases are characterized by heterogenous ventilation. Hyperpolarized ¹²⁹Xe gas lung magnetic resonance imaging (MRI) can examine lung ventilation heterogeneity by acquiring high-resolution isotropic images. The current gold standard of semi-automated (SA) segmentation can be used to quantify non-isotropic ¹²⁹Xe lung images to generate ventilation defect percent (VDP), however, this method is not suitable for analysis of isotropic voxel ¹²⁹Xe images due to the large number of slices. Therefore, we used a fully automated deep learning-based (DL) lung algorithm to calculate VDP from isotropic images. SNR, SA and DL-based VDP were calculated, showing a strong positive linear correlation with a zero intercept and close to unity slope. This study demonstrates the feasibility of using DL-based segmentation methods to quantify ventilation defects, which has potential for clinical translation of ¹²⁹Xe MRI as a tool for treatment and monitoring for patients with pulmonary diseases.