Scan pre-processing

#1
by JoHof - opened

Hi Pillar-0 team,

first of all, congratulations on the great work you have done with this model and the paper! Could you please clarify how the pre-processing of ChestCT volumes should look like for inference of this model? Specifically:

  • Should a 256^3 center crop be performed (should this be done on the original resolution or should be volume be downsampled?)
  • Should the volume be resampled to an isotropic resolution before the center crop or any other fixed spacing?

Could you please provide a code snipped that shows all the steps one should perform for pre-processing a CT scan before inference?

In addition, I would be interested on the decision to perform a 256^3 center crop as opposed to a larger matrix like you did on the abdominal scans with 384?

Thanks a lot!

Yala Lab org

Thank you!

We preprocess our images using this repository: https://github.com/YalaLab/rave, which accepts both DICOM directories and NIfTI files.

  • The volume is sampled to [1.25, 1.25, 1.25], then cropped or padded to 256x256x256. This is specified in our config for Chest CTs: https://github.com/YalaLab/rave/blob/main/configs/ct_chest.yaml.
  • We determined the pixel spacing and final dimensions based on our data distribution at UCSF. The abdomen/pelvis scans had a greater field of view than the chest scans.

Hope this helps, please let us know if you have more questions!

yala changed discussion status to closed

Sign up or log in to comment