Global tractography is the process of finding the full track configuration that best explains the measured DWI data. As opposed to streamline tracking, global tractography is less sensitive to noise, and the density of the resulting tractogram is directly related to the data at hand.
As of version 3.0, MRtrix supports global tractography using a multi-tissue spherical convolution model, as introduced in Christiaens et al. (2015). This method extends the method of Reisert et al. (2011) to multi-shell response functions, estimated from the data, and adopts the multi-tissue model presented in Jeurissen et al. (2014) to account for partial voluming.
For multi-shell DWI data, the most common use will be:
tckglobal dwi.mif wmr.txt -riso csfr.txt -riso gmr.txt -mask mask.mif -niter 1e9 -fod fod.mif -fiso fiso.mif tracks.tck
In this example,
dwi.mif is the input dataset, including the
gradient table, and
tracks.tck is the output tractogram.
csf_response.txt are tissue response functions (cf. next
section). Optional output images fod.mif and fiso.mif contain the
predicted WM fODF and isotropic tissue fractions of CSF and GM
respectively, estimated as part of the global optimization and thus
affected by spatial regularization.
Per tissue response function estimation¶
Input response functions for single-fibre WM, GM and CSF can be estimated directly from the data.
The most convenient way of doing so, is via the
dwi2response dhollander algorithm
(Dhollander et al. (2016)):
dwi2response dhollander dwi.mif wm_response.txt gm_response.txt csf_response.txt
dwi.mifis the same dwi data set as used above (input)
<tissue>_response.txtis the tissue-specific response function as used above (output)
Note that the order of the tissue responses output for this algorithm is always: WM, GM, CSF.
Other methods exist, notably
dwi2response msmt_5tt, but this requires a co-registered T1 volume
and very accurate correction of EPI geometric distortions (both up to sub-voxel accuracy), as well as
accurate segmentation of the T1 volume.
Even then, still,
dwi2response msmt_5tt may be less accurate than
in a range of scenarios (Dhollander et al. (2016)).
-niter: The number of iterations in the optimization. Although the
default value is deliberately kept low, a full brain reconstruction will
require at least 100 million iterations.
-lmax: Maximal order of the spherical harmonics basis.
-length: Length of each track segment (particle), which determines
the resolution of the reconstruction.
-weight: Weight of each particle. Decreasing its value by a factor
of two will roughly double the number of reconstructed tracks, albeit at
increased computation time.
-ppot: The particle potential essentially
associates a cost to each particle, relative to its weight. As such,
we are in fact trying to reconstruct the data as well as possible, with
as few particles as needed. This ensures that there is sufficient
proof for each individual particle, and hence avoids that a bit of
noise in the data spurs generation of new (random) particles. Think of
it as a parameter that balances sensitivity versus specificity. A higher
particle potential requires more proof in the data and therefore leads
to higher specificity; a smaller value increases sensitivity.
-cpot: The connection potential is the driving
force for connecting segments and hence building tracks. Higher values
increase connectivity, at the cost of increased invalid connections.
-fod: Outputs the predicted fibre orientation distribution function
(fODF) as an image of spherical harmonics coefficients.
This fODF is estimated as part of the global track optimization, and
therefore incorporates the spatial regularization that it imposes.
Internally, the fODF is represented as a discrete sum of apodized point
spread functions (aPSF) oriented along the directions of all particles in
the voxel, akin to track orientation distribution imaging (TODI,
Dhollander et al., 2014). This internal representation
is used to predict the DWI signal upon every change to the particle
-fiso: Outputs the estimated density of all isotropic tissue
components, as multiple volumes in one 4-D image in the same order as
-riso kernels were provided.
-eext: Outputs the residual data energy image, including the
L1-penalty imposed by the particle potential.
- D. Christiaens, M. Reisert, T. Dhollander, S. Sunaert, P. Suetens, and F. Maes. Global tractography of multi-shell diffusion-weighted imaging data using a multi-tissue model. NeuroImage, 123 (2015) pp. 89–101 [SD link]
- M. Reisert, I. Mader, C. Anastasopoulos, M. Weigel, S. Schnell, and V. Kiselev. Global fiber reconstruction becomes practical. NeuroImage, 54 (2011) pp. 955–962 [SD link]
- B. Jeurissen, J.D. Tournier, T. Dhollander, A. Connelly, and J. Sijbers. Multi-tissue constrained spherical deconvolution for improved analysis of multi-shell diffusion MRI data. NeuroImage, 103 (2014), pp. 411–426 [SD link]
- T. Dhollander, L. Emsell, W. Van Hecke, F. Maes, S. Sunaert, and P. Suetens. Track Orientation Density Imaging (TODI) and Track Orientation Distribution (TOD) based tractography. NeuroImage, 94 (2014), pp. 312–336 [SD link]
- T. Dhollander, D. Raffelt, and A. Connelly. Unsupervised 3-tissue response function estimation from single-shell or multi-shell diffusion MR data without a co-registered T1 image. ISMRM Workshop on Breaking the Barriers of Diffusion MRI (2016), pp. 5 [full text link]