BlueMatter
Last week at the 15th Annual Meeting of the Organization of the Human Brain Mapping Conference, Anthony Sherbondy presented a breakthrough result.
Title: BlueMatter: Optimal Estimation of Long Range White Matter Fascicle Networks using Diffusion Tensor Imaging
Authors: AJ Sherbondy, R Ananthanarayanan, RF Dougherty, BA Wandell, DS Modha
Abstract:
Introduction: Fiber tractography algorithms generally use greedy algorithms to estimate one fiber group at a time. Consequently, the full white matter network can have unlikely properties – for example, there can be empty regions or regions of implausibly high white matter density. We describe a new approach to white matter estimation (BlueMatter) that optimizes the full network solution, not just individual fiber groups. This algorithm makes significant use of the BlueGene/L architecture in order to find optimal solutions to a global scoring criterion from a database of 180 billion candidate connections. The network optimization (a) limits the relative fascicle density across the white matter volume, and (b) predicts the diffusion data from all fibers that pass through a voxel. We demonstrate an implementation of this method, and we show that the network solutions generated by BlueMatter have clear advantages over the greedy algorithms.
Methods: Diffusion-weighted data were acquired from a healthy male adult subject (age 35) using a single-shot spin-echo, echo planar imaging sequence (260mm FOV; 128x128 matrix size; 2mm thick slices (no skip); b=800 s/mm2; 12 diffusion directions). The BlueMatter technique comprises three parts: 1) Identification of potential gray matter voxels. The segmentation of neocortex and subcortical gray matter resulted in 50 thousand potential gray matter voxels. 2) Generation of candidate pathways that connect gray matter voxels: We generated 3.6 million potential pathways per voxel in 40 hours on the BlueGene supercomputer. The total database contained 180 billion candidate pathways. 3) Selection of a network of pathways subject to: (a) predicting the diffusion-weighted data, and (b) producing a uniform fiber group density within the white matter. The algorithms were run on a BlueGene supercomputer with 2048 processors each allocated 256 MB RAM. Optimal network solutions for each hemisphere were identified within 10 hours.
Results: We compared the resulting pathway network with a whole brain pathway network solution derived using the Streamlines Tracing Technique (STT) [1]. BlueMatter produced a network that was at least as accurate in predicting the acquired diffusion data at every measurement location. Within core white matter the pathway density variation for BlueMatter (50-150 fibers/voxel) was much more uniform compared to the STT solution (50-2000 fibers/voxel). The fiber group densities in a slice and along three typical lines within this slice are shown in Figure 1.
Conclusions: By leveraging software developed for the BlueGene architecture and a novel mathematical model for assigning a likelihood score to a white matter network, BlueMatter provides neuroscientists with a method for identifying fascicle networks that simultaneously optimizes the predictions of the diffusion-weighted measurements and normalizes pathway density to match biological variation.
References:
Basser, P. (2000), 'In vivo fiber tractography using DT-MRI data', Magnetic Resonance in Medicine, vol. 44, no. 4, pp. 625-632.
