Incorporate DTI-TK into our evaluation of registration methods

[ebrahimebrahim]

https://dti-tk.sourceforge.net/pmwiki/pmwiki.php

[ebrahimebrahim]

Still making sure we can save the warps. It will also be important to have the warps in a format that mrtrix can accept.

[ebrahimebrahim]

Update: It seems that figuring out exactly how DTI-TK is storing and using warps is a massive pain. Some options:

• Keep banging head on wall till we figure it out. Submit request for source code possibly.
• Talk to Gary?
• Talk to Martin about whether they have used the warps in their group. They use DTI-TK for sure, but IDK if they have used the warps directly.
• Give up because if it's that hard to understand the warp format then this is not a good option anyway and doesn't need to be part of our evaluation and recommendations.

[ebrahimebrahim]

(didn't mean to close this yet)

[ebrahimebrahim]

@jamesfishbaugh When you get a chance please include your notes on this tricky dti-tk thing

[jamesfishbaugh]

Notes about dti-tk:

Preprocessing

• Requires source and target (and target mask) as power of 2. We found that a center crop of 128x128x128 worked for all the cases in the evaluation set. If anything changes with the eval set this will need to be re-verified.
• We have to scale the diffusivity values to make the units compatible with dti-tk (mean diffusivity for CSF around 3). For now I have checked a couple of subjects and found MD to be around 0.002 --> so I set the scale to 1500 as default. Uses: TVtool
• Set origin to [0,0,0] as recommended by dti-tk documentation. Uses: TVAdjustVoxelspace

Registration

• Run rigid, affine, then diffeo in sequence: dti_rigid_reg, dti_affine_reg, dti_diffeomorphic_reg.
  • Outputs an affine portion and a diffeo portion. They need to be combined using dfRightComposeAffine

Hack the transformation to the original coordinate space

The resulting displacement field is round(128*1.7)^3 with 1^3 spacing and (0,0,0) origin. We will maneuver it back to the coordinate space of the ABCD input.

• Resample to the crop size. Uses: c3d (convert3D)
  • Does the resampling for each channel (i.e. doesn't scale the vector field)
• Uncrop the displacement field back to it's original size (140x140x140) as shape (140, 140, 140, 1, 3). This appears to be a nifti specification for vector images.
• The final step is to move to original origin, and also reorient to RPI. Uses: c3d

Evaluation

At this point, the registration is complete and the output is in the format that the evaluation framework anticipates.