What are the exact preprocessing steps for the fluoroscopy images? about deepfluorolabeling-ipcai2020 HOT 4 CLOSED

I was able to reproduce the downsampling, thank you for the detailed instructions!

from deepfluorolabeling-ipcai2020.

Also the relationship between your transform and mine is not exactly linear, which is what I would expect if the difference were something small like adjusting for the linear attenuation coefficient.

from deepfluorolabeling-ipcai2020.

Hi @eigenvivek , thanks a lot for the great question and kind words. You can find the C++ code that was used for performing the log remapping here: https://github.com/rg2/xreg/blob/c2bdd99b5ddd91e4561ee5d8f1572274182ba2b8/lib/image/xregImageIntensLogTrans.cpp#L54

A brief comment on what it does is here: https://github.com/rg2/xreg/blob/c2bdd99b5ddd91e4561ee5d8f1572274182ba2b8/lib/image/xregImageIntensLogTrans.h#L34

Copying that here for convenience:
Assumes the basic imaging equation: I_f(i,j) = I_0 exp(-L(i,j)), where L(i,j) is the line integral that sums linear attenuation units and is equal to \int_0^1 V(l(i,j,t)) dt.
For each pixel: L(i,j) = -log(I_f(i,j) / I_0) for I_f(i,j) != 0.
For I_f(i,j) == 0, we set L(i,j) = L(i*,j*), where i*,j* = argmin_u,v I_f(u,v), such that I_f(u,v) > 0.

For this data I_0 is approximated by the largest I_f(i,j) (e.g. assumes that there is a pixel where there was no attenuating tissue). Due to noise, the image is first smoothed and then the max is found.

Take a look at the .cpp and let me know if I can help explain in more detail.

from deepfluorolabeling-ipcai2020.

I also used smoothing + a B-spline interpolator for resampling the image to smaller dimensions. That is probably different than what the torch resize function does. Looks it will do some smoothing followed by linear interpolation. That could explain some other differences.

from deepfluorolabeling-ipcai2020.