A novel approach to segment neonatal brain tissues and measure cortical folding in MRI with motion

Presenting Author Senior Author
Name: Hosung Kim Name: Duan Xu
Email: Email:
Presenting Author’s RIG/SRG: Informatics and Image Processing/Display MRI/MRS Pediatrics/Fetal Pediatric Radiology  
Presenting Author's Lab Location: Mission Bay   

Abstract Information
Imaging Modality: MR
Disease Application: preterm birth
Complete author list: Hosung Kim, Romir Maheshwary, A. James Barkovich and Duan Xu
Abstract highlights: Cortical folding is a well-established index of brain maturation. We proposed a brain tissue segmentation using probabilistic texture patches and measured cortical folding on WM boundary. Our segmentation performed reliably under various motion artifacts. In 157 neonates, we identified patterns of cortical folding increase over time, well-corresponding to cerebral morphogenesis
Cortical gyration becomes dramatically complex in the fetal brain during the 3rd trimester of gestation. This morphological characteristic is a well-established index of brain maturation. To quantify cortical folding, it is necessary to extract the interface between gray matter (GM) and white matter (WM) on MRI. This is challenging as neonatal brain MRI frequently present with a degree of motion (>30% affected; Figure B). Moreover, different degrees of myelination, neuronal proliferation, and cell migration among cortical regions, which are distinctive in developing brain, are manifested as regionally- and temporally-varying GM/WM contrast in structural MRI. We proposed a brain tissue segmentation method that addressed the aforementioned issues using probabilistic texture modeling (Figure A). We applied our approach to a large database of preterm neonates (n=157) and investigated its clinical utility.
MR images underwent intensity non-uniformity correction prior to tissue segmentation. We then masked out areas of non-brain tissues using a label fusion algorithm (Eskildsen et al., 2012). To segment GM/WM/cerebrospinal fluid (CSF), we created multi-individual texture patches. A patch was made of a 5×5×5 cube centered at voxel x. These patches were extracted within the brain area in 23 MRIs that were labeled manually into GM/WM/CSF. These MRIs were selected considering various ranges of age (22-40 weeks gestational age [GA]) and motion artifact. Based on texture similarity between the patches in the training-set and that at voxel x of the test image (Coupe et al., 2013), the patches within a searching area were ranked. Averaging the labels of high-ranked patches (n=10) created probability of GM/WM/CSF at voxel x (Wong et al., PAIM 2013). A level-set-based deformable model then evolved towards the WM boundary in the probability tissue map while constrained with a smoothing parameter. We finally triangulated the WM boundary using an icosahedral model, and surface-based registration, allowing for sampling 80,000 points and point-correspondence across subjects. We quantified the cortical folding as sulcal depth by computing the geodesic distance from the gyral crown to the sulcal fundus.
Our segmentation performed reliably under various motion amounts (Figure B; Dice similarity to manual segmentation: 93%). We analyzed cortical folding of 157 preterm neonates (GA at birth: 28±2.4). T1w MRI images were acquired near birth, re-scanned before discharge at late preterm age (34-38 weeks). An individual result is shown in Figure C. Our method identified increases in cortical folding over time in numerous cortical regions (mean sulcal depth: +0.01/wk) while folding unchanged in major sulci that are known to develop earlier (corrected p<0.05; Figure C). This pattern was nonlinear as 2nd order polynomial model fitted best. Comparing the folding pattern in the extremely preterm (EPT; <27 weeks GA) to the late preterm (LPT; >32 weeks), we found significant growth impairment in EPT, localized primarily in the regions developing postnatally (p<0.0001; Figure D).
The proposed method segmented the neonatal brain tissues reliably, even on images affected with a degree of motion artifact. Our approach successfully mapped cortical structural development, supporting current models of cerebral morphogenesis.