Deep-Learning-Based Automated Quantification of Ventilation Defects on 3d Isotropic Hyperpolarized 129xe Lung MRI

Semi-automated-segmentation quantifies non-isotropic 129Xe lung images in order to provide the ventilation-defect-percent (VDP), but this method is not ideally suited for 3D-isotropic-voxel analysis. 3D-isotropic-voxel 129Xe images can be used to calculate semi-automated and deep learning-based automated VDP values for lung disease assessment.  We aimed to develop a fully-automated deep learning-based (DL) segmentation algorithm for 3D 129Xe MRI analysis which generates VDP.  This is a prospective study which included 10 participants (COVID-19 Survivors) with ventilation heterogeneity.  A DL segmentation method was used to compute VDP and was compared to the semi-automated method serving as the benchmark reference.  Imaging was acquired from Fast-Gradient-Recalled-Echo in 16sec breath-hold.  Isotropic imaging was generated using a zero-filling approach.  A non-significant mean difference for semi-automated-segmentation and DL-based VDP values was observed. SNR values were above 5 (Rose criteria).  The two VDP estimates had an intercept of -0.03, a slope of 1.1 and r=0.89.  Bland Altman analysis indicated negligible bias and Sørensen-Dice (similarity) coefficients suggested a good match between the ground truth (semi-automated-segmentation) and DL segmentations.  This study establishes the use of a deep-learning-based algorithm for segmentation and suggests the proposed method can be an alternative for time-consuming and higher variability segmentation methods.