How are siblings similar? How similar are siblings? Large-scale imaging genetics using local image features

How are siblings similar? How similar are siblings? Large-scale imaging genetics using local image features

Toews, Matthew and Wells, William M.

Proceedings – International Symposium on Biomedical Imaging 2016

Abstract : A novel method is developed for analyzing large sets of 3D medical image data. The image manifold is approximated as a proximity graph between bags-of-features (BoFs), i.e. order-less sets of local invariant features extracted from images. The pairwise geodesic distance between images (or BoFs) is approximated locally by the Jaccard distance metric, which measures BoF overlap based on sets of nearest neighbor (NN) feature correspondences computed across a data set. An efficient computational framework allows the method to scale gracefully with the amount of data. Registration error associated with NN correspondences is shown to decrease with increasing numbers of images N. Experiments demonstrate strong links between manifold and genetic proximity, from T1-weighted head MRI of 511 twin and non-twin siblings. Siblings are automatically identified with high probability based on BoF distance. NN correspondences are more numerous in the frontal cortices of twin vs. non-twin siblings. An erroneous duplicate subject is identified based on manifold proximity.