User:Jadrian Miles/vis2011 tractography statistics paper

Idea: perform tractography on a suite of small computational phantoms and measure how tractography quality degrades with increasing noise in the input images. May also be worth considering other factors: interpolation scheme (tensors vs. eigenvectors), QBI vs. DTI, differential measures of the true vector field, number of b-vectors, number of b-values, blurring of the underlying true DWI signal.

Measures of tractography errors (two curves advected from the same seed point, one in a "gold standard" noise-free field, the other in some degraded field): length difference, absolute distance between corresponding points, rate of distance error accumulation, angle difference between corresponding segments.

Contributions:

1. Prior distributions for "normal" behavior of tractography in a uniform structure given noise level.
2. A new measure for the "plausibility" of a clustering: reconstruct the bounding volume, and reconstruct curves from this, compare reconstructed curves to input curves, and compute the goodness of fit given the prior distribution of tractography errors. In this case the reconstructed curves are actually considered "gold" while the input curves are noisy; the likelihood that the noisy curves actually came from the gold curves, given the known noise level and the corresponding error priors, tells us the likelihood of the clustering itself.
3. A new locality prior for region-finding MRFs: since continuity of the vector field guarantees that we can create two distinct but arbitrarily similar curves by starting from seed points sufficiently close to each other, the priors on error for degraded-vs-gold for a single curve also apply to degraded-vs-degraded on nearby curves.