Hyperpolarized 3He Magnetic Resonance Functional Imaging Semiautomated Segmentation

Abstract

To improve intra- and interobserver variability and enable the use of functional magnetic resonance imaging (MRI) for multicenter, multiobserver studies, we generated a semiautomated segmentation method for hyperpolarized helium-3 ((3)He) MRI. Therefore the objective of this study was to compare the reproducibility and spatial agreement of manual and semiautomated segmentation of (3)He MRI ventilation defect volume (VDV) and ventilation volume (VV) in subjects with asthma, chronic obstructive pulmonary disease (COPD), and cystic fibrosis (CF). The multistep semiautomated segmentation method we developed employed hierarchical K-means clustering to classify (3)He MRI pixel intensity values into five user-determined clusters ranging from signal void to hyperintense. A seeded region-growing algorithm was also used to segment the (1)H MRI thoracic cavity for coregistration to the (3)He cluster-map, generating VDV and VV. We compared manual segmentation performed by an expert observer and semiautomated measurements of (3)He MRI VDV and observed strong significant correlations between the volumes generated using each method (asthma, n = 5, r = 0.89, P < .0001; COPD, n = 5, r = 0.84, P < .0001; CF, n = 5, r = 0.89, P < .0001). Semiautomated VDV had high interobserver reproducibility (coefficient of variation [CV] = 7%, intraclass correlation coefficient [ICC] = 0.96); intraobserver reproducibility was significantly higher for semiautomated (CV = 5%, ICC = 1.00) compared to manual VDV (CV = 12%, ICC = 0.98). Spatial agreement for VV determined using the Dice coefficient (D) was also high for all disease states (asthma, D = 0.95; COPD, D = 0.88; CF, D = 0.90). Semiautomated segmentation (3)He MRI provides excellent inter- and intraobserver precision with high spatial and quantitative agreement with manual measurements enabling its use in longitudinal studies.