I'm preproccing data for a longitudinal VBM study and ran into a very weird problem. I was hoping someone has experience with this or can explain what's going wrong.
In short, preprocessing the same settings in different ways results in diffent results.
I'm using the VBM8 toolbox (revision 369) with the lastest SPM8 version and matlab 7.1 (x64) (on single node & on multinode system).
I created a batch in spm8 to process the VBM, tested the batch trough the GUI without errors and got good segmentation results.
Then I applied the batch to a 150 subjects using a multi-node cluster that uses exactly the same software as the above-mentioned.
The settings I used are as in the attached batch (spm_default normalization and only non-linear modulation of GM & WM )
Then when I compared the voxel values of the resulting images between the a subjects that I manually applied the batch to and the one I send into the multinode cluster, there was only a 77% match.
To the eye they seemed to match though. To check where this difference comes from I did a series of tests, all using the same settings of course.
- I ran everything twice on:
- a GUI-batch on a single node
- a matlabbatch on a single node
- the sh-batch of vbm8 toolbox on a single node
- the sh-batch of vbm8 toolbox on the multinode-system
- a matlabbatch on the multinode-system
These 5 different runs gave 4 different results, with the difference of every non-matching result having 13% of mismatching voxel values. The runs that did match each other was the GUI batch on single node with the sh-batch on the single node.
Every second run matched the first one, so it's not a randomisation difference.
Secondly, a difference between the methods used was that sometimes the VBM estimate&write gave a NU-correction error with some iterations to correct in the error log. This error only doesn't appear when using the GUI. But the non-error displaying GUI run matches the use of the sh-batch on the single node, so just the display or non-display of the error is not the determining factor in the differences .
I am completely out of options, does anybody have a suggestion?!!
Pieter van de Vijver