Better template + better defaults for reorienting on brain damaged patients + mutual information coregistration (and automatic AC origin placement)

[lrq3000]

Hello,

This PR adds a T1 template that is the average of 10 healthy volunteers manually reoriented and realigned using CAT12. It also increases the smoothing kernel, as this provided better results on T1 for both healthy volunteers and brain damaged patients alike. The algorithm was changed to do non-linear coregistration to template (option 'mni') as recent studies found this is more efficient than rigid-body (but the applied transform is still rigid-body, we just don't use the scaling factor).

Also update the documentation according to the latest changes in the code and more logically order the readme.

This was tested on 576 T1 (450 different subjects) with 350 being brain damaged and lots being artefacted by movement, the auto reorienting always resulted in a better head orientation although not in the optimal AC-PC alignment for several brain damaged patients (I am not sure what is the culprit, since it works pretty well even with motion artefacts and more brain damaged patients, all I can think of is maybe a weird intensity scaling of the input image, so maybe next possible enhancement would be an intensity equalization before doing the coregistration on template?). Before the changes proposed in this PR, most auto reorientation (including on healthy volunteers) failed, with the changes the enhancement has been significant (all healthy volunteers are auto reoriented to AC-PC and most brain damaged patients too).

[lrq3000]

Update: I manually checked all the structural images, and in general the new script did a pretty good job.

I now added a new mode I did not yet test on the whole database but which shows a lot of promise: the new spm_coreg function in SPM12. The script can combine both the affine coregistration and then after the mutual information coregistration, this seems to give the best results.

A nice bonus is that spm_coreg also automatically place the origin on AC (or whatever the template has as origin).

Tested on a limited subset of brain damaged subjects, this shows very promising results.

[lrq3000]

Ironed out some issues and added possibility to do automatic coregistration between modalities, a helper script is provided in spm_auto_coreg.m. Here is the result on a moderately brain damaged subject and original EPI BOLD image way off:

This was achieved in a 3 steps process:

1. spm_auto_reorient the structural in MNI space, so that it is also matching other SPM templates
2. coreg the functional to SPM12 EPI template (this allows a correct translation matrix and a good basis for rotation)
3. coreg the functional onto the structural (this fine-tunes rotation to precisely match the subject's structural)

The last two steps are done in spm_auto_coreg with optimized defaults.

So far, this strategy seems to work very well. I'll now test on 450 subjects, I'll report back how good it works overall.

[lrq3000]

This PR is now final, I do not expect to change anything else, the results are quite satisfactory with more than 95% successful coregistration on a hundred subjects :-)

[lrq3000]

Tested on 434 sessions (about 410 different subjects), with most being brain damaged or artifacted, 28 only were mis-coregistered (11 really totally off, 17 slightly off but fixable with a few increments of forward/top parameters), thus 93.5% success rate on this dataset. In general, where it was slightly off, it was for healthy subjects.

Indeed, the algorithm seems to work particularly well on brain damaged subjects and artifacted images, with lesions and artifacts acting as landmarks for the algorithm (which was unexpected). For healthy volunteers, the landmarks are a lot smaller, hence some slight offset from correct coregistration (but always very close). I'll try to see if there is a way to enhance the results for these subjects.

[lrq3000]

Tested both reorientation and coregistration on 5 other datasets with healthy volunteers (the case that fails the most with this algorithm), and the success rate is similar, above 90% (I don't have an exact figure here since I did not track precisely these datasets contrary to the brain damaged subjects one). These datasets were of about 20 subjects each, and 4 acquisition sessions each.