Normalization during training about covidprognosis

[mmuckley]

Closing due to inactivity and it seems the question is answered.

[mmuckley]

Hello @hoangtnm, to be honest it was just a convention. We wrote that transform pretty early in development and avoided altering our code afterwards so we could easily reuse models. If you're starting anew I think you can use the [0, 1.0] range with no issues if you prefer.

[hoangtnm]

Hi @mmuckley, your paper uses HistogramNormalize with CXR image, in my case, I want to use your paper's idea for chest CT images. Do you think HistogramNormalize would be necessary or just scale from [0, 255] from PNG images to [0, 1] is enough? Thank you

[mmuckley]

Hello @hoangtnm, hard to tell without knowing the process pipeline for the images. If your CT images are already PNG images then the difficult colorspace work might already be done - raw CT data has a huge dynamic range probably beyond what would normally go into a PNG. I would start with whatever is simplest for your setup.