Getting Rid of The Mohawk – Status Log for Correcting Falied coregistration in CPAC

Introduction

In the process of preprocessing ABIDE data for a release, we found that a lot of data failed the coregistration step in which the functional data were transformed into MNI space and coregistred to a template brain for across subject comparison. A typical such functional image which failed coregistration might show brain images being pulled out of grey matter (GM) boundaries around the mid-line of parietal lobe (See fig. 1). The parietal pulled-out often accompanied with an occipital brain-impressed-in. To get the most out of the ABIDE data, it's important find the cause of the problem and fix it before we calculate the resting-state fMRI measures.

On possible source of the error is the coregistraion program we are using. To test this hypothesis, we need to replace the program we are using. One candidate we now have is the Advanced Normalization Tools (ANTs) which is endorsed by Chris Gorgolewski.

We have noticed that among the subjects who are suffering the bad coregistration problem, some have T1 images failed skull stripping (Fig. 2), some have EPIs with bad brain coverage (e.g.: part of brain tissue got cut-off during data collection, Fig. 3), and some have the two combined. We also have to figure out what kind of role these factors are playing in coregistration and how much they are account for the problematic results.

In summary, to solve the bad registration problem, we need to

1. Try replacing Flirt/Fnirt with ANTS to see if coregistration will improve for images with bad coverage and failed skull stripping.
2. Try fixing bad skull stripping.

Fig 1. An example of failed registration showing EPI brain and surface boundaries (green line) of GM, white matter (WM) and cerebrospinal fluids (CSF) of a template brain in MNI space.

Fig 2. An example of failed Skull Stripping.

Fig 3. An example of EPI of bad brain coverage.

To-dos:

Coregistration Problem

1. Prepair 4 subject lists: subidbadSkullStrippingOnly subid_badCoverage_only subid_badCombined, subid_good [Done_, files are in /home2/data/Projects/ABIDE_Initiative/CPAC/solvingSkullStrippingProblem/scripts On Rocky]
2. Try ANTS on the 4 subset of subjects in the lists above [In progress]
3. Try rerun some subjects after fixing the Fnirt-not-using-config problem with CPAC to see if that helps.[code fixed, reprocdessing the subjects in 1.]

Skull Stripping Problem

1. Try skull stripping by FreeSurfer, FSL's BET and 3dSkullStrip with multiple parameters on the subjects in the lists above to see if the bad skull stripping problem can be fixed by those programs, or which program should be used. [In progress]
2. If none of the solutions works, then hand editing images to remove skull. Then try flirt/fnirt and ANTS, see what gives best results. [Pending]

Skull Stripping problem

I first tried to run the skull striping on 10 subjects who failed AFNI's skull stripping with FreeSurfer and FSL, and compare the results generated by the three packages. And we are using the default settings of the skull stripping function. It seems that FSL's BET does not have the leakage problem but at the same time have a problem of clipping out the GM. AFNI's 3dSkullStrip has the leakage. FreeSurfer gives results somewhere inbtween.

Below, I will be exploring the setting's for 3dSkullStrip to see what combination will give a optimal results.

The skull stripping problem we are having with most of our data is that we have leakage into meninges and skull. Here we are going to try to apply some parameters of AFNI's 3dSkullStrip to see if we can get better results. Before changing anything, we need to know what options were used for 3dSkullStrip in CPAC. In CPAC, the pipeline runs the skull striping step by calling the command below:

3dSkullStrip -input mprage_RPI.nii.gz -o_ply mprage_RPI_3dT.nii.gz

The command line above will take file “mprage_RPI.nii.gz” as input and remove the skull from it and output the brain image and a surface file (-o_ply option). All the other options will be using the program's default setting. Here I only list the parameters which might affect/fix the leakage problem. The whole set of parameters are detailed on the help page of 3dSkullStrip. (Type 3dSkullStrip -help to see)

• -niter N_ITER: Number of iterations. Default is 250. For denser meshes, you need more iterations N_ITER of 750 works for LD of 50.
• -ld LD: Parameter to control the density of the surface. Default is 20 if -no_use_edge is used, 30 with -use_edge. See “CreateIcosahedron -help” for details on this option.
• -shrink_fac SF: Parameter controlling the brain vs non-brainintensity threshold (tb). Default is 0.6.
• -var_shrink_fac: Vary the shrink factor with the number of iterations. This reduces the likelihood of a surface getting stuck on large pools of CSF before reaching the outer surface of the brain. (Default)
• -shrink_fac_bot_lim SFBL: Do not allow the varying SF to go below SFBL . Default 0.65, 0.4 when edge detection is used. This option helps reduce potential for leakage below the cerebellum. In certain cases where you have severe non-uniformity resulting in low signal towards the bottom of the brain, you will need to reduce this parameter.
• -pushout: Consider values above each node in addition to values below the node when deciding on expansion. (Default)
• -exp_frac FRAC: Speed of expansion (see BET paper). Default is 0.1.
• -avoid_vent: avoid ventricles. Default. -no_avoid_vent: Do not use -avoid_vent.
• -avoid_eyes: avoid eyes. Default.-no_avoid_eyes: Do not use -avoid_eyes.
• -use_edge: Use edge detection to reduce leakage into meninges and eyes. Default. no_use_edge: Do no use edges.
• -push_to_edge: Perform aggressive push to edge at the end. This option might cause leakage.
• -no_push_to_edge: (Default).
• -use_skull: Use outer skull to limit expansion of surface into the skull due to very strong shading artifacts. This option is buggy at the moment, use it only if you have leakage into skull.
• -no_use_skull: Do not use -use_skull (Default). -send_no_skull: Do not send the skull surface to SUMA if you are using -talk_suma
• -blur_fwhm FWHM: Blur dset after spatial normalization. Recommended when you have lots of CSF in brain and when you have protruding gyri (finger like) Recommended value is 2..4.

Possible thing to Try:

1. Use -shrink_fac higher than 0.6. start with 0.65 to 0.8.
2. Try using the -use_skull option.
3. Try use mesh with the density lower than default. should be less than 30.
4. Increase -shrink_fac_bot_lim from default value to deal with leakage in lower areas. Default is 0.4 with -use_edge option on.

Results:

1 ~ 4 does not rely solve the problem. The only thing that seems improving the result a bit is using the -use_skull option. Setting shrink factor to be 0.7 and using the -use_skull parameter helped a little bit but did not change things too much.

Please click here to see the result images for the tests I run.

Does failed skull stripping affect registration?

I re-processed several subjects with the newly updated version of CPAC (the fnirt problem fixed version).

In the figure below, I showed the functional and anatomical images before and after registration for one subject. It seems that the leakage into meninges and skull did not affect the registration between anatomical and MNI template too much. In Fig. 4E, the meninges in anatomical image aligned well with the meninges in the MNI template.

The registration between functional to anatomical images seem a little bit off, and there are still brain leaking (Mohawk) in functional image in MNI space even after the fix.

It seems that removeing skull manually does not improve registration, at least for subject 0050146. See Figure 4 and 5 below for output maps of registration.

Registration outputs 1	Registration outputs 2
Fig. 4: Registration maps. The anatomical image has some skull and meninges left after skull striping. Subid=0050146.	Fig. 5: Registration maps. The anatomical brain image was hand edited to remove skull and meninges left after skull striping. Subid=0050146.

Pending

Document created 07/09/2013 | Last updated on 13:05 07/17/2013