Sampler w/ resample option -- exclude midline

[kristianeschenburg]

Hi Dr. Burt

Thanks for this software, it's been super helpful to me so far. I have two questions:

1. I have some multivariate data that I'm interested in generating random samples from. I believe that answer is no, but does your code preserve the covariance between features in the multivariate case?

2. I'm having a bit of difficult using the resample option in the Sampler class. I have need to maintain the same scaling as the original data, so sampling from the original data is important. When computing the geodesic distance matrices, I can exclude the medial wall in the distance calculations. However, when using the resample option in the Sampler, the Sampler pulls NaNs / Infs / etc. from the medial wall. Is there a mask option that I can provide to simply exclude sampling from certain indices?

Thanks for your help!

k

[jbburt]

Hi Kristian,

I apologize for the delay in getting back to you. Regarding your questions:

1. I'm not entirely sure what your question is. By multivariate data, do you mean you have multiple (empirical) brain maps? In such a case, no, the code will not preserve the statistical relationships among the brain maps.

2. I'm having trouble reproducing your issue. When the empirical brain map contains NaNs (e.g., for medial wall voxels), then the resampled surrogate maps will also contain NaNs for those voxels. In other words, the "location" of the missing values should not change. In the test that I ran, this was indeed the case. Perhaps I am missing something? I've attached the code that I used to test this below. Could you take a look at it and see whether similar conditions are satisfied for your use-case?

from brainsmash.workbench.geo import cortex
from brainsmash.mapgen.memmap import txt2memmap
from brainsmash.mapgen.sampled import Sampled
from brainsmash.utils.dataio import dataio
import numpy as np
import os

brain_map_file = '/Users/jbb/Desktop/LeftDenseMyelin.txt'
surface = "/Users/jbb/Desktop/S1200.L.midthickness_MSMAll.32k_fs_LR.surf.gii"
medial_LR = "/Users/jbb/Desktop/Human.MedialWall_Conte69.32k_fs_LR.dlabel.nii"
distmat = "/Users/jbb/Desktop/LeftDenseGeodesicDistmat.txt"

# isolate left hemisphere of medial wall mask file
medial_L = "/Users/jbb/Desktop/Human.MedialWall_Conte69.32k_fs_L.dlabel.nii"
os.system(f"wb_command -cifti-separate {medial_LR} COLUMN -label CORTEX_LEFT /Users/jbb/Desktop/Human.MedialWall_Conte69.32k_fs_L.label.gii")
os.system(f"wb_command -cifti-create-label {medial_L} -left-label /Users/jbb/Desktop/Human.MedialWall_Conte69.32k_fs_L.label.gii")

# compute geodesic distance matrix for voxels in left hemisphere
cortex(
	surface=surface, 
	outfile=distmat, 
	medial=medial_L, 
	euclid=False, 
	use_wb=False, 
	n_jobs=-1
)

# memory-map the distance matrix
output_files = txt2memmap(
	dist_file=distmat, 
	output_dir='/Users/jbb/Desktop/', 
	delimiter=' '
)

# load brain map values from text file
myelin = np.loadtxt(brain_map_file)

# load medial wall mask file, which I will use 
# to insert NaNs into the myelin map
mask = dataio(medial_L)
myelin[np.where(mask)] = np.NaN
assert np.isnan(myelin).sum() == 2796

# Generate five surrogate maps
gen = Sampled(
	x=myelin,
	D="/Users/jbb/Desktop/distmat.npy",
	index="/Users/jbb/Desktop/index.npy",
	resample=True
)
surrogates = gen(n=5)

# Ensure that the surrogate maps contain the 
# same number of NaNs as the empirical map
assert all([np.isnan(x).sum() == 2796 for x in surrogates])

# Ensure that NaNs in each surrogate map correspond to the
# appropriate medial wall voxels
assert all([np.isnan(x[np.where(mask)]).all() for x in surrogates])

[christineecker]

Dear Joshua,

Thank you for a brilliant toolbox. Amazing work!

I also have trouble with the medial wall label. I use FreeSurfer data and read in the medial wall label as a long vector of zeros (outside label) and ones (inside label). However, the distance matrix contains a lot of 'inf' values, and the surrogates don't compute. Is there anything I am doing wrong? Your help would be much appreciated. Thank you!