Multi-site fixel-based analysis with dimension and parameter discrepancies

Hello everyone,

I am a novice Ph.D. student in the field of DTI, currently working on a fixel-based analysis project using data from a public database. While extracting some features from the available data, I observed variations in certain parameters between subjects. This discrepancy is likely attributed to changes in the data collection site. Specifically, when using the “mrinfo” command, I noticed variations in dimensions, voxel sizes, and phase encoding directions across subjects. Here are a few examples:


  • 116 x 116 x 81 x 127
  • 116 x 116 x 81 x 52
  • 116 x 116 x 81 x 6
  • 116 x 116 x 81 x 3

Voxel Size:

  • 2 x 2 x 2 x ?
  • 2 x 2 x 2.7 x ?
  • 2 x 2 x 3.5 x ?

Phase Encoding Direction:

  • j
  • j-

Shell B-values:

  • 0 500 1000 2000
  • 0 1000 2000

Shell sizes:

  • 13 6 48 60
  • 3 7 8

B 0,500,1000,2000 counts:

  • 13 b=0, 6 b=500, 48 b=1000, 60 b=2000
  • 3 b=0, 7 b=1000, 8 b=2000

I am seeking guidance on the best approach to address this issue. Is it feasible to work with subjects of different dimensions using the same pipeline? Alternatively, would it be more appropriate to process subjects with distinct pipelines? Your insights and suggestions would be highly appreciated. Thank you in advance for your assistance.

Hi @CalumaXs,

The fourth number is how many volumes is in an image. I doubt any subject’s full run should contain 6 or 3 volumes. Can you confirm if any specific dimension is specific to site?

During preprocessing, everything should be resampled to some isotropic resolution (you could do 2 mm or even try upsampling to 1.5 or 1.25 mm).

Is it possible one of these phase encoding directions are reverse phase-encoded fieldmaps used for susceptibility distortion correction.

I would hesitate to use the 3b0 7b1000 8b2000 data, that is very low angular resolution.

Other note:
In multi-site fixel based analysis, you should only average response functions within site, instead of applying a whole cohort average to everyone. Then, in your final statistical model, you can include a covariate for site or try out this recent implementation of COMBAT for fixel data: GitHub - remikamito/fixel_combat: This performs ComBat data harmonisation on fixel-based data.


Hi Steven,

Thank you for your response. Regarding the number of volumes, I’m not certain if it’s a site-specific characteristic or related to individual subjects, as this information is not specified (it’s a public database). I have an additional question to clarify: could you kindly confirm whether it’s necessary to run the population template separately for each data site and then create a single one, or if it’s possible to run a standard population template for all the data?

If separate templates for each site are required, could you please provide guidance on the process of creating individual templates for each site and subsequently merging them into a single template?

Thanks in advance for your response.

Hi @CalumaXs,

A single population template should be used, ideally with equal input from all sites such that the template is more generalizable.