A Twin Study Using Brain Shape
Segmentation Algorithms

Abstract

In this pilot study, we validated a set of automated surface segmentation shape extraction algorithms to study genetic influences on the brain structure of normal twins. A set of manually delineated lateral ventricles was deformed, using a 3D Navier-Stokes fluid image registration algorithm, onto all the scans in the database of twin brain MRI images. The geometric transformations thus obtained were used to propagate the segmentation labels to all the other brain images. 3D radial distance maps were derived to encode anatomical shape differences. The proportion of shape variance attributable to genetic factors, known as the heritability, was estimated from the shape models using a restricted maximum likelihood formula to increase statistical power. Segmentation errors associated with the projection of labels onto new images were greatly reduced through multi-atlas averaging. The resulting algorithms provide a convenient and sensitive tool to recover and analyze small intra-pair image differences. In summary, here we show how computer vision approaches based on fluidly deformable parametric surfaces can be applied to automatically delineate and parameterize brain structures in an image database, and detect genetic influences on brain shape.