MITTENS

MITTENS is a python library that performs analytical tractography on reconstructed diffusion-weighted MRI data. Currently the output from DSI Studio is supported.

Table of Contents

1. Installation

• Adding Geometric Constraints

2. Preparing your data
3. Calculating intervoxel fiber transition probabilities
4. VoxelGraph Construction

• Finding and visualizing connections between regions
• Building an AtlasGraph

Installation

NOTE: On Mac OS you may need to specify the path to a recent c++ compiler (download gcc-8 from homebrew) E.g. export CXX=/usr/local/bin/g++-8. You might also need to edit the setup.py script to include

# initialize Extension module with the appropriate source file
modules = [Extension("_NetworKit",
        src,
        language = "c++",
        extra_compile_args=["-fopenmp", "-std={}".format(stdflag), "-O3", "-DNOGTEST"],
        extra_link_args=['-fopenmp', "-std={}".format(stdflag),'-lgomp', '-Wl,-rpath,/usr/local/opt/gcc/lib/gcc/8/'], # <- NEW!
        libraries=["NetworKit-Core-{0}".format(optimize)],
        library_dirs=["./"])]

While this software can be installed like any other python package, it is possible to add custom compiled functions before installation. Analytic tractography requires the specification of a set of geometric constraints for inter-voxel tract transition expectations can be solved. MITTENS comes with the functions used to perform the analyses described in [our paper], but nothing more.

You can add as many sets of geometric constraints as you like, but be warned that it can take awhile to compile them. Remember you only need to compile a solution once. Also beware that some combinations of step size and turning angle maximum result in infinite sets of turning angle sequences. Python will crash with a recursion error in this case.

Adding Geometric Constraints (optional)

Download the current version of MITTENS and enter its source tree:

$ git clone https://github.com/mattcieslak/MITTENS.git
$ cd MITTENS

Now, launch a python session in the mittens directory

>>> from mittens.write_solution import write_solution
>>> write_solution("odf8", 35, 1.0)

This will write out two fortran files in the current directory. They are named after the parameters you chose. This particular example will read data reconstructed with 8-fold ODF tesselation, a turning angle maximum of 35 degrees and a step size of 1.0 voxel units (voxels are assumed to be isotropic).

After some calculations you will find doubleODF_odf8_ss1_00_am35.f90 and singleODF_odf8_ss1_00_am35.f90 in the current directory. Move these files into the src/ directory and install the package with pip. We recommend using an editable install, which will keep all these files in their current directory:

mv *.f90 src/
pip install -e .

This will compile everything in the src/ directory into a python module. You can now use these geometric constraints to calculate transition expectations.

Preparing your data

DSI Studio can be used to reconstruct diffusion MRI using a variety of methods. You can use any method you like except for DTI. If you acquired DTI data, choose GQI reconstruction instead of DTI reconstruction.

DSI, GQI and QBI can be selected from this menu. There are additional options not listed on this GUI but are accessible through the commandline. For example, ODF deconvolution and decomposition can be used. All options in purple boxes can be changed, but it is required that ODF data is written in the output file. Critically, the ODF Tesselation option determines the angular resolution of your output ODFs. The choice here will result in "odf8" being the appropriate choice for the call to write_solutions() above.

Data from other diffusion MRI packages such as FSL and MRTRIX can be loaded (theoretically) after being .

Calculating intervoxel fiber transition probabilities

DSI Studio files are read directly by MITTENS:

>>> from mittens import MITTENS
>>> mitns = MITTENS(reconstruction="HCP.src.gz.odf8.f5rec.fy.gqi.1.25.fib.gz")

From here you can estimate none-ahead or one-ahead, where NIfTI-1 files are saved for each neighbor direction:

>>> mitns.calculate_transition_probabilities(output_prefix="hcp")

You will find the output in the current working directory (unless you specified an absolute path as the argument). There is a single 3D file for each neighbor voxel named hcp_doubleODF_r_prob.nii.gz, hcp_singleODF_r_prob.nii.gz, hcp_doubleODF_lpi_prob.nii.gz, etc. There will also be CoDI and CoAsy output. Instead of re-running the estimations again, you can create a MITTENS object by specifying the prefix of the NIfTI files written out during estimation.

>>> nifti_mitns = MITTENS(nifti_prefix="hcp")

This is very fast.

VoxelGraph Construction

MITTENS uses networkit to construct directed graphs of voxels, where edges are weighted by the tract transition expectation from one voxel to another. A graph can build directly from MITTENS object.

>>> voxel_graph = mitns.build_graph(doubleODF=True, weighting_scheme="negative_log_p")
>>> voxel_graph.save("hcp_voxel_graph.mat")

There are different schemes available for weighting edges in the graph.

Option	Transformation	Reference
"negative_log_p"	negative log of transition probability	Zalesky (2008)
"minus_iso_negative_log_p"	max(0, transition prob - isotropic transition prob)	None yet
"minus_iso_scaled_negative_log_p"	Same as above but scaled to sum to 1 across all neighbors of the voxel	None yet
"transition_probability"	The transition probability. Not useful for shortest paths	None yet

You can also build a matching voxel graph using probabilities from isotropic ODFs

>>> null_graph = mitns.build_null_graph(doubleODF=True, purpose="shortest paths")
>>> null_graph.save("hcp_null_voxel_graph.mat")

Both voxel_graph and null_graph will be of class mittens.VoxelGraph, which contains a networkit.Graph and spatial mapping information from the original dMRI data. You can load these directly from disk using

>>> from mittens import VoxelGraph
>>> loaded_voxel_graph = VoxelGraph("hcp_voxel_graph.mat")
>>> loaded_null_graph = VoxelGraph("hcp_null_voxel_graph.mat")

Finding and visualizing connections between regions

Assuming the previous steps have been run, we can add an atlas to the graph and query for connections between regions. Here we have two cortical regions we'd like to connect. These regions are spheres around peak coordinates in primary motor cortex (M1) and supplementary motor area (SMA). Here we find the shortest path from each voxel in the M1 sphere to any voxel in the SMA sphere. The paths are saved to be loaded into DSI Studio as streamlines, optionally with their corresponding probabilities.

>>> paths, probs = voxel_graph.region_voxels_to_region_query("lm1_sphere.nii.gz",
                      "sma_sphere.nii.gz", write_trk="test", write_prob="test")

This will write out two files. One is lm1_sphere.nii.gz_to_sma_sphere.nii.gz_test.trk.gz and the other is lm1_sphere.nii.gz_to_sma_sphere.nii.gz_test.txt. To load in DSI Studio, open the fib file used earlier via the "Step 3: Fiber Tracking" button. Load the streamlines through the context menu and color-code them based on their probabilities:

the color-coded streamlines are more informative:

Building an AtlasGraph

Most network approaches to date use brain regions as nodes. You can add brain regions to your VoxelGraph like so:

voxel_graph.add_atlas("aligned_AAL.nii.gz")

This adds a region label to voxels contained within each region. The node labels can be examined in the VoxelGraph.atlas_labels. No additional actions (such as adding abstract "region nodes" or adding edges) are taken initially. An AtlasGraph, a graph where nodes are brain regions and edges are based on shortest paths through the VoxelGraph that connect them, can be build like so:

asym_raw_prob_graph, asym_mean_prob_graph, asym_path_length_graph, \
conj_raw_prob_graph, conj_mean_prob_graph = voxel_graph.build_atlas_graph()

Credits

This source code was sponsored by a grant from the GE/NFL head health challenge. The content of the information does not necessarily reflect the position or the policy of these companies, and no official endorsement should be inferred.

Authors

• Matt Cieslak
• Tegan Brennan
• Clint Greene

Logo by Allison Shapiro