Non integer labels in Colin27 2008 version

🐛Bug

Class labels in Colin27 dataset, version 2008, are not integer values.

To reproduce

from torchio.datasets import Colin27
colin = Colin27(2008)
label_values = colin['cls']['data'].unique()
print(label_values)

>>> tensor([ 0.0000,  0.9882,  2.0235,  3.0118,  4.0000,  4.9882,  6.0235,  7.0118,
         8.9882, 10.0235, 11.0118, 12.0000])

Expected behavior

Label encoding map should only contain integer values.

TorchIO version

0.17.7

[fepegar]

I removed the slope from the file

In [7]: with open('colin27_cls_tal_hires.nii', 'rb+') as f:
   ...:     f.seek(116)  # address of intercept, which I know is 0
   ...:     data = f.read(4)  # stored in float32 so 4 bytes
   ...:     f.seek(112)  # address of slope
   ...:     f.write(data)  # actually this should have been 1, but it looks like nibabel uses a 1 if it finds a 0
   ...:

and the values are now

In [14]: np.unique(lm)
Out[14]:
array([  0,  21,  43,  64,  85, 106, 128, 149, 191, 213, 234, 255],
      dtype=int16)

Indeed to get 12 from 255 you need to multiply by 12/255 = 0.047058823529411764, which is the intercept stored in the header. However, 234 * 0.047058823529411764 = 11.0118. It looks like, for some reason, someone decided to store the labels from 0 to 255 for some reason, and adding a slope. Maybe this worked with the software they use at the MNI because the labels are rounded after reading.

I think the best for now is just rounding the values ourselves when reading the CLS image.

[fepegar]

Thanks for reporting, @GFabien. This is weird, and I don’t think there’s much we can do about this. I tried reading the image with ITK and results are the same. I checked the intercept and slope values in the NIfTI header and I found nothing interesting:

scl_slope       : nan
scl_inter       : nan

I’ve been googling with no success. I think that the only solution for now is to round the data in the Colin27 class of the torchio.datasets.mni.colin module.