Project to brain space (inverse_transform()) a statistic vector obtained from probabilistic atlas

, , ,
1

Hi,

I’m performing intersubject correlation (ISC) analyses using the BrainIAK API and have encountered a challenge with visualizing my results. Here’s my workflow:

  1. I used the probabilistic DiFuMo256 atlas (DiFuMo atlas - Nilearn) to extract time series with NiftiMapsMasker(), resulting in a 2D array of dimensions TRs x regions.
  2. The ISC output is a vector of n ROIs, containing the median pairwise correlation for each region.

The issue arises when I try to project these correlation values back into brain space for visualization:

  1. The probabilistic nature of the atlas seems to limit the tools available for displaying statistic values, compared to binary ROI-based atlases. I only saw tools to visualize the atlas, and not a statistic associated with each ROI.
  2. Transformation alters the scale of correlation values: Using NiftiMapsMasker.inverse_transform() to map correlation values back to brain space changes the value magnitudes within each ROI. For example, a correlation value of 0.30 for an ROI becomes 0.0004 at the voxel level. This makes it difficult to interpret and set thresholds, especially when trying to display regions that survive FDR correction. I have a separate vector of p values, which I can use to find the fdr threshold I need to apply to the correlation vector/brain map.

My goal is to retain the original correlation values when projecting them into brain space while applying a pre-computed threshold based on p-values. Ideally, I would display only significant cluster/regions, while the colorbar would indicate the correaltion values.

Nilearn offers helpful functions like threshold_img() and threshold_stats_img(), but:

  • threshold_img() requires inputting an explicit threshold, which becomes problematic when the voxel-wise values have already been transformed.
  • threshold_stats_img() requires a Z-map, which is not directly available from my current ISC outputs. I could r_to_z transform, but due the the inverse transform operation, I loose track of the threshold I need to apply.

Any suggestions on how to approach this problem, particularly with probabilistic atlases and preserving ISC values, would be greatly appreciated.

Thanks in advance for your help!

Best regards,
Dylan