search

Ruyi-Zha / naf_cbct

"Neural Attenuation Fields for Sparse-View CBCT Reconstruction" (MICCAI 2022 Oral)
https://arxiv.org/abs/2209.14540
MIT License
119 stars 21 forks source link

Train on real projected data #25

Open springXIACJ opened 6 days ago

springXIACJ commented 6 days ago

When I used it on real projected data, the loss was very large and it never converged. But when I used the DRR generated by Tigre, there was no problem with training.

Ruyi-Zha commented 6 days ago

Hi, here are my suggested steps.

  1. Review the real projection data and rescale it if values are excessively large or small.
  2. Check the scanner geometry, ensuring parameters like DSD, DSO, dVoxel, and dDetector are consistent with the projection data’s unit.
  3. If feasible, normalize the scene to a unit cube to align with our dataset standards.
  4. Adjust hyperparameters, including the learning rate, and consider tuning bound, which should be larger than the half size of your region of interest.

It would be great if you could share more information about your real data so that I can try to figure out the problem.