Abstract
BACKGROUND AND PURPOSE: Manual segmentation of white matter (WM) bundles requires extensive training and is prohibitively labor-intensive for large-scale studies. Automated segmentation methods are necessary in order to eliminate operator dependency and to enable reproducible studies. Significant changes in the WM landscape throughout childhood require flexible methods to capture the variance across the span of brain development. METHODS: Here, we describe a novel automated segmentation tool called Cortically Constrained Shape Recognition (CCSR), which combines two complementary approaches: (1) anatomical connectivity priors based on FreeSurfer-derived regions of interest and (2) shape priors based on 3-dimensional streamline bundle atlases applied using RecoBundles. We tested the performance and repeatability of this approach by comparing volume and diffusion metrics of the main language WM tracts that were both manually and automatically segmented in a pediatric cohort acquired at the UCSF Dyslexia Center (n = 59; 25 females; average age: 11 ± 2; range: 7-14). RESULTS: The CCSR approach showed high agreement with the expert manual segmentations: across all tracts, the spatial overlap between tract volumes showed an average Dice Similarity Coefficient (DSC) of 0.76, and the fractional anisotropy (FA) on average had a Lin's Concordance Correlation Coefficient (CCC) of 0.81. The CCSR's repeatability in a subset of this cohort achieved a DSC of 0.92 on average across all tracts. CONCLUSION: This novel automated segmentation approach is a promising tool for reproducible large-scale tractography analyses in pediatric populations and particularly for the quantitative assessment of structural connections underlying various clinical presentations in neurodevelopmental disorders.