Network Connectivity via Inference over Curvature-Regularizing Line Graphs

Diffusion Tensor Imaging (DTI) provides estimates of local directional information regarding paths of white matter tracts in the human brain. An important problem in DTI is to infer tract connectivity (and networks) from given image data. We propose a method that infers high-level network structures and connectivity information from Diffusion Tensor images. Our algorithm extends principles from perceptual contours to construct a weighted line-graph based on how well the tensors agree with a set of proposal curves (regularized by length and curvature). The problem of extracting high-level anatomical connectivity is then posed as an optimization problem over this curvature-regularizing graph - which gives subgraphs which comprise a representation of the tracts' network topology. We present experimental results and an open-source implementation of the algorithm.