How to recreate First Level Model output from FSL FEAT in Nilearn

Software Support
, , , ,
1

I am currently trying to replicate a first level GLM in Nilearn; the analysis was originally done in FSL FEAT. However, I am getting significantly different results for the same, simple contrast (e.g., 1 0 0) between the two tools. The cope values from FSL FEAT are in the hundreds, while the beta maps (output_type=effect_size) from Nilearn ranges from ~2 to 2.

Additionally, when plotting the data across runs for the same contrast, the trajectory of the time courses are also extremely different. For instance, while FSL might indicate an increase in value from Run 1 to 2, Nilearn indicates a decrease.

The parameters for my first level model in Nilearn are as follows:
FirstLevelModel(t_r=2,
noise_model=‘ar1’,
hrf_model=‘spm’,
drift_model=‘cosine’,
high_pass=1./128,
signal_scaling=False,
minimize_memory=False)

2

FSL is doing a grand mean scaling (median value = 10000). Therefore, your beta values are different. Without scaling the grey values are ‘arbitrary’ and can not compared between runs. You need a scaling like percent scaling to compare runs.