Working With DTI Brain Scan Art Isn't What You Think It Looks Like
I've spent years processing diffusion tensor imaging data and creating visual outputs from it. People see the swirling colorful tractography images online and assume it's straightforward. It isn't. The gap between what these scans actually contain and what gets rendered is enormous. Diffusion tensor imaging measures the directionality of water molecule movement in tissue. In white matter tracts, water moves preferentially along the axis of axons rather than across them. That anisotropic diffusion is what you're visualizing when you see those rainbow fiber bundles. The colors correspond to direction: red for left-right, green for anterior-posterior, blue for superior-inferior. That's the basic mapping. Everything after that is a series of choices that dramatically change the output. The first practical step is acquiring good data. Most public datasets use a 3t scanner with around 64 diffusion gradient directions. I've run pipelines on 1.5t data too, and the tractography quality drops noticeably below 30 directions. If you're working with clinical data where the protocol was 20 directions or less, you're not going to get clean results. Accept that upfront and adjust your expectations.
Processing typically starts with eddy current and motion correction. FSL's eddy tool handles both, and I recommend passing the --repol flag to replace outlier slices. Without that, you'll get artifacts in the tensor estimation that propagate through to the final visualization. Next comes tensor fitting using dtifit, which outputs fractional anisotropy and eigenvector maps. The FA map is what you'll use for seeding regions in tracking. For probabilistic tractography, I use MRtrix3's commit2 algorithm rather than the older CSD methods. It constrains the response function iteratively and reduces false-positive streamlines significantly. The default parameters work for most brain data, but if you're tracking through the optic radiations or uncinate fasciculus, you'll want to lower the --step value to 0.5mm and increase the number of samples per voxel to at least 1000. These tracts are where partial volume effects cause the most problems. Here's something most tutorials don't mention: the seed density matters far more than people realize. I once spent three days debugging what I thought was a pipeline error, only to discover my seeding mask had a resolution mismatch of 0.5mm between the FA image and the segmentation. The result was a handful of streamlines instead of the expected bundles. The workaround was to resample the seed mask to the native diffusion space using flirt with nearest-neighbor interpolation instead of Trilinear, which smeared the boundaries and excluded entire regions.
For the surreal visualization aspect, the raw streamlines are just the beginning. I export them as .tck files and then use tools like SlicerDMRI or TractVis to render them. The color mapping is where the artistic choices happen. Standard RGB direction encoding looks clinical. I've found that applying a Gaussian blur to the streamline density before coloring it creates much more visually interesting outputs. A sigma of 2-3mm on the voxel grid smooths out the noise without obliterating anatomical detail. Another technique that works well is thresholding by FA values before rendering. Setting a minimum FA of 0.15 or 0.2 removes the noisy streamlines in gray matter and CSF spaces while preserving the coherent white matter pathways. This alone makes the visualization look cleaner without any post-processing in external software. If you want to push further into surreal territory, try combining DTI tractography with structural MRI contrast. Superimposing the fiber bundles on a T1-weighted image with adjusted opacity reveals how tracts relate to cortical geography. I've also experimented with taking the eigenvector field and using it as a displacement map for mesh deformation, which creates warped anatomical views that look nothing like standard radiology displays.
Get the Full Details
There are real limitations you need to account for. DTI can't resolve crossing fibers well. In regions like the centrum semiovale where multiple fiber populations intersect, the single-tensor model breaks down and you'll see distorted or terminated tracts. Constrained spherical deconvolution helps but doesn't eliminate the problem entirely. If your project depends on accurate visualization of the corona radiata or the superior longitudinal fasciculus near the parietal lobe, be aware that the data simply doesn't support perfect reconstruction at clinical field strengths. The other bottleneck is computation time. A full probabilistic tractography run on 100 million streamlines from a standard 2mm resolution dataset takes roughly 45 minutes to 2 hours on a modern CPU, depending on the number of directions and the tracking algorithm. GPU-accelerated versions in MRtrix3 can cut that to 10-15 minutes, but they require CUDA-compatible hardware. If you're processing a batch of subjects, this adds up quickly. For download and processing tools, the main packages are FSL, MRtrix3, and 3D Slicer with the DMRI extension. All three are free and open source. MRtrix3 is the most actively maintained for tractography work. FSL remains the standard for preprocessing. SlicerDMRI is useful if you want an interactive GUI for visualization rather than command-line processing.
The workflow I consistently come back to is: preprocess with FSL eddy and dtifit, compute CSD with MRtrix3, run probabilistic tracking with COMMIT2 filtering, export to Slicer for visualization, and apply custom shaders or density-based coloring for the final renders. It takes about 90 minutes total for a single subject on decent hardware, and the results are as reliable as the input data allows.
