Validation of MRtrix tractography for clinical use

Abstract

Tractography is a technique which uses non-invasive diffusion magnetic resonance imaging (dMRI) to model the structural connections of the brain in vivo. Since it is not known to what extent the exact positions of white matter tracts vary between individuals, and how scan parameters affect the reconstructed location of tracts, the performance of tractography algorithms cannot be assessed by comparing them against a standardised model of white matter in the human brain. Until technology has advanced to a stage where these issues can be resolved, if this is indeed possible, the quality of tractograms must be assessed via alternative means if they are to be used in a clinical setting. MRtrix is a software package which offers a suite of tools for tractography. To investigate the clinical viability of tractography, the effect of adding artificial noise and patient movement to the input dMRI scans was examined. The performance of the following three MRtrix tractography algorithms was considered: the second-order integration over fibre orientation distributions (iFOD2) algorithm, the spherical deconvolution streamlines tractography (SD_STREAM) algorithm, and the probabilistic tractography (Tensor_Prob) algorithm. The accuracy of each tractogram and its robustness to added noise and motion was assessed through both quantitative and qualitative means. Streamline length histograms, polar plots, a resemblance metric, the RMS difference between tractmaps and polar plots, and false positive and negative rates were used to quantitatively measure tractogram quality. Each of these techniques provide measures of quality relative to an initial unmodified scan. Potentially, these techniques could be extended to provide a stand-alone metric of tractogram quality to eliminate operator dependence in diagnosing, treating and managing conditions associated with the integrity of white matter pathways in the brain. The tractography algorithm found to be most suitable for clinical applications was iFOD2, which provided the highest level of detail compared with the SD_STREAM and Tensor_Prob algorithms. iFOD2 was capable of consistently resolving tracts with added noise of up to about an added white Gaussian noise (WGN) of power 20, or a SNR of about 8 for the majority of patients (calculated using the single image method). Additionally, the iFOD2 algorithm successfully reconstructed tractograms with added translations up to 5 mm, and rotations up to 10 degrees, despite the reconstructed tracts being spatially shifted.