search

AIM-Harvard / pyradiomics

Open-source python package for the extraction of Radiomics features from 2D and 3D images and binary masks. Support: https://discourse.slicer.org/c/community/radiomics
http://pyradiomics.readthedocs.io/
BSD 3-Clause "New" or "Revised" License
1.15k stars 497 forks source link

[FEAT EXTRACTION] Why are firstorder features influenced by gray level discretization? #666

Closed GiulioBen closed 3 years ago

GiulioBen commented 3 years ago

I changed the discretization between several values of binWidth (bw) and binCount (bc) and built a model with the radiomic features. I expected the model was the same, but it was not (I set the seed correctly)

I report below a table with some of the different configurations:

disc    filter          features    AUC
bw8 original+lbp+wavelet    firstorder  0.7515625  
bc128   original+lbp+wavelet    firstorder  0.7187500  
bc64    original+lbp+wavelet    firstorder  0.7171875  
bw32    original+lbp+wavelet    firstorder  0.7421875 
bc16    original+lbp+wavelet    firstorder  0.7140625 
bw16    original+lbp+wavelet    firstorder  0.7421875  
bw4 original+lbp+wavelet    firstorder  0.7406250  
bc256   original+lbp+wavelet    firstorder  0.7078125  
bc32    original+lbp+wavelet    firstorder  0.7218750  
bc8 original+lbp+wavelet    firstorder  0.7093750  
bw128   original+lbp+wavelet    firstorder  0.7390625  
bw64    original+lbp+wavelet    firstorder  0.7281250  
bw256   original+lbp+wavelet    firstorder  0.7171875

In the documentation I find that firstorder features are not influenced by discretization, and I can't find any mention of the fact that filters are influenced by the discretization too (they should be applied on the inizial image).

Can someone explain what I'm missing?

Thank you!

JoostJM commented 3 years ago

Both firstorder Uniformity and Entropy are affected by discretization, this is also detailed in the documentation.