Image preprocessing: Quantile normalization

[RobertoLorusso]

In the current version quantile normalisation is applied on the following axes: (0,1,2,3). Results within the first epochs aren't as good as I expected. Probably is due to the a wrong choice of the axes.

I load the images channel first, resulting in the following shape for each image: (4,240,240,240) Should try to train the model after normalisation on the following configurations:

• [x] axis = (0,1,2). As in the reference work. Not sure if they load images channel first.
• [x] axis = (1,2,3)

[dagshub[bot]]

Join the discussion on DagsHub!

[RobertoLorusso]

The quantiles must be obtained excluding the channel from the 'axis' parameter. I should try first with axis = (1,2,3) and see if there are any improvements.