depredict/sphere2dice

Sphere2dice converts Cortical Brain Morphometry data from FreeSurfer or Gifty type into a stack of 2D images and saves is as numpy array. This stack of images is easier to use as input to conventional CNNs, while avoiding projection issues other methods suffer from.

Sphere2Dice Readme

Sphere2Dice converts Cortical Brain Morphometry data from FreeSurfer (lh.thickness) or Gifty (lh-thickness.gii) type into a stack of 2D images and saves is as numpy array. This stack of images is easier to use as input to conventional CNNs, while avoiding projection issues other methods suffer from.

How to use?

Run python sphere2dice.py -h for help on the inputs.

To get started processing one single file of your own you only have to provide the --scalar_file parameter. python sphere2dice.py --scalar_files sub-01_Thickness.gii

This of course also works in series:

python sphere2dice.py --scalar_files sub-01_Thickness.gii sub-02_Thickness.gii

Alternatively, it is also possible to iterate over directories:

python sphere2dice.py --scalar_files lh.thickness --input_dirs sub-001/Surf sub-002/Surf-sub-003/Surf output_dirs sub-001 sub-002 sub-003

If we want to store our results in the same folder as we read them from, we can set output_dirs to None:

python sphere2dice.py --scalar_files lh.thickness --input_dirs sub-001/Surf sub-002/Surf-sub-003/Surf output_dirs None

How sphere2dice works

It uses fsaverage sphere coordinates (provided in separate file as lh.sphere) for fsaverage registered Gifty (.gii) files. Alternatively, for FreeSurfer Morphometry data, it is assumed that the scalar file read is accompanied by a corresponding ?h.sphere file.

The index of these sphere coordinates is linked to the index of scalar values (like thickness or curvature). These coordinates are projected to a 2D plane, and we cut a square out of the center of size size patch_size. Then we interpolate these points to a given resolution. We create multiple patches from all different sides but rotating the coordinates and repeating the process. If the code is run with the --verbose flag, extra figures are generated displaying this process.

Requirements

Compatible with Python 3. Modules used are basic so requirements are minimal. I have therefore not containerized the script. I did attach the requirements.txt for repeatability.

Terminology

• Patch: A 2D projection. You have two choices for the number of patches as of now: 6 or 18. This makes sense as 6 patches contain two opposite projections from three orthogonal axes, like folding a simple box or dice. An issue with this lies in the edges: As when folding a 6-sided die: there is no overlap and the anatomy at the edges lacks anatomical context. The solution is simple: Take an extra patch right in between every two patches, this comes down to 4 around the equator, 4 on the north pole and 4 on the south pole = 18. The code contains debugging modes that show this process (but might be a bit finicky to set up, so I’ll supply these as soon as I’m back at my machine.
• Dice: I use the word dice every so often (like in the name of the code) for folding a simple box, which is the most simple way to perceive this method (but it is not limited to it)
• Stack: A stack is a 3D volume of 2D patches.

Example code

Run the demo

Create additional figures using the verbose flag displaying the rotation process. See /output/angle-0*_overview.jpg for the results.

(ENIGMA) D:\...\sphere2dice-standalone>python sphere2dice.py --verbose
Computing patches for "example-thickness-file.gii".
100%|████████████████████████████████████████████████████████████████████████████| 6/6 [00:13<00:00,  2.28s/it]
Storing verbose figures to "out_name_prcfg".
100%|████████████████████████████████████████████████████████████████████████████| 6/6 [01:36<00:00, 16.02s/it]
Successfully created: "output\example-thickness-file.npy" from "input\lh.sphere"
sphere2dice.py complete succesfully

Real world example

There are two usefull ways to use the script:

1) Iterate over files in the same folder. Note the use of the is_right flag.

python shere2dice.py --scalar_files lh.thickness --input_dirs sub-001 sub-002 --output_dirs None

(ENIGMA) D:\...\sphere2dice-standalone>python sphere2dice.py --scalar_files 26567_space-fsaverage_hemi-R.thickness.gii 26677_space-fsaverage_hemi-R.thickness.gii --input_dir D:\...\AnatSurfRH\Thickness --is_right True True
Computing patches for "26567_space-fsaverage_hemi-R.thickness.gii".
100%|████████████████████████████████████████████████████████████████████████████| 6/6 [00:11<00:00,  1.88s/it]
0it [00:00, ?it/s]
Successfully created: "output\26567_space-fsaverage_hemi-R.thickness.npy" from "input\lh.sphere"
Computing patches for "26677_space-fsaverage_hemi-R.thickness.gii".
100%|████████████████████████████████████████████████████████████████████████████| 6/6 [00:11<00:00,  1.93s/it]
0it [00:00, ?it/s]
Successfully created: "output\26677_space-fsaverage_hemi-R.thickness.npy" from "input\lh.sphere"
sphere2dice.py complete succesfully

2) Iterate over folders: Same file (lh.thickness) in different input folders (sub-001, sub-002), with --output_dirs set to None output is stored in the respective input folders

python shere2dice.py --scalar_files lh.thickness --input_dirs sub-001 sub-002 --output_dirs None

Origin of accompanying files

Origin of the lh.sphere¹ file: CorticalParcellation_Yeo2011 - Free Surfer Wiki ² Then Yeo_JNeurophysiol11_FreeSurfer\fsaverage\surf\lh.sphere

Footnotes

1. ftp://surfer.nmr.mgh.harvard.edu/pub/data/Yeo_JNeurophysiol11_FreeSurfer.zip ↩

2. https://surfer.nmr.mgh.harvard.edu/fswiki/CorticalParcellation_Yeo2011 ↩