I am having issues with nilearn’s seed-based functional connectivity estimation functionality.
You can find a minimal self-contained example with all code and data, here.
The issue is that the z-scores are excessively small (ranging from min = -0.195 to max = 0.266) - although activation thresholded at +/-0.1 shows very low noise and a very clear distribution, consistent with the stimulation protocol (we stimulate in the midbrain, and past studies show that inhibitory transmission is to be expected particularly in the smatosensory and cingulate cortices. So basically we are seeing robust, high contrast-to-noise effects, but according to the statistic they are insignificant.
Not least of all, I find it particularly concerning that even in the seed ROI, correlation with the seed only goes up to 0.266. Any idea what might cause these very low statistic scores? The example in nilearn has somewhat higher scores, and it’s even a human brain, where no activity is introduced directly at the neuronal level, so you would expect a lower contrast-to-noise ratio.
One thing which improved statistics somewhat was shrinking the ROI, but no ROI I could chose even came close to the statistic scores in the example.