MRtrix 3.0_RC3 -> 3.0.0 changelog

2

1. Changes for priority attention

dwidenoise

  • Noise level estimator: Fix denoising efficacy when the number of DWI volumes closely matches the number of volumes in the patch. (Thanks for reports @Kar, @uclpz, @Shawn_Yeh, @Forrest, @DorianP)

    • The noise level estimator as initially defined in MP-PCA is biased when the number of DWI volumes closely matches the number of voxels in the patch. Crucially, this bias manifested itself differently in cases M < N and M > N, M being the no. dMRI images and N being the patch size, leading to inconsistent results in some cases. This issue has been fixed in a new noise level estimator that reduces the bias and has symmetric behaviour across patch size. A detailed comparison is described in Cordero-Grande et al. (2019). This new estimator is now the default.

    • Users can revert to the legacy behaviour (not recommended) with the new command option -estimator. However, this will match the previous release candidate if and only if the patch size exceeds the number of DWI volumes in the dataset. The opposite case is handled symmetrically by using the matrix transpose.

dwifslpreproc (previously dwipreproc)

  • Brain mask: The brain mask generated within the script to constrain the optimisation within FSL eddy was erroneous if the input DWI was not in native RAS orientation. This is expected to not have been severely problematic in many cases, since it is only used to control which voxels contribute to the estimation of motion / eddy current distortion / outlier classification parameters, but users may wish to investigate the presence and magnitude of this issue in their own data. (Thanks for report @HelenaV)

fixelcfestats

  • Output parameter magnitudes: Specifically the beta coefficients, absolute effect size and standard deviation images were affected by a bug in the smoothing of subject fixel data; all other outputs (e.g. p-values, standardized effect size, t-values) were unaffected. In each fixel, the magnitudes of these values were erroneously scaled by some factor, where that factor was different for each fixel.

    The effect was self-cancelling for internally self-consistent calculations; e.g. the standardized effect size, which is the ratio of the absolute effect size and the standard deviation, was unaffected. Direct interpretation of e.g. beta coefficient magnitudes was however affected.

tckgen

  • SH precomputer: Fix quantification of FOD amplitudes specifically during FOD-based tractography, where pre-computed versions of complex mathematical functions are used to speed up execution. Effect was that FODs were interpreted as being 30-40% smaller than reality, and therefore comparable to the threshold being 30-40% larger than intended. Orientation dependency was minimal; i.e. the interpreted FOD shapes were not particularly affected.

  • Tracking parameter updates:

    • Changed default tracking angles for some algorithms, and made these angles independent of step size for all algorithms. Previously the maximum angle per step was calculated based on a minimal radius of curvature and the tracking step size; they are now instead hard-coded to a fixed value for each algorithm, and the value utilised will not change if the user manually sets only the step size.

    • Changed default step size to 1/4 voxel size when using 4th-order Runge-Kutta.

    • Changed default FOD amplitude cutoff for iFOD1 & iFOD2. In 3.0_RC3, these were changed from 0.10 to 0.05, due to a combination of fixing of an SH amplitude calculation bug and increased use of multi-tissue CSD. Based on the updated SH amplitude calculation fix and alterations to the recommended WM response function estimator, these have been reverted to 0.10. A lower value may however be preferable in some cases, particularly if using multi-tissue CSD; precise tuning of this parameter cannot currently be performed using automated heuristics and so is deferred to the user.

    • Default tracking cutoff threshold is halved for all algorithms if ACT is employed.