search

facebookresearch / fastMRI

A large-scale dataset of both raw MRI measurements and clinical MRI images.
https://fastmri.org
MIT License
1.3k stars 372 forks source link

Implementation of the cascade module in the code is different from the paper. #138

Closed akemisetti closed 3 years ago

akemisetti commented 3 years ago

The VarNetBlock is in line with the paper "End-to-End Variational Networks for Accelerated MRI Reconstruction" except for one part.

The equation for the cascade says image

But in the code return current_kspace - soft_dc - model_term

model_term is subtracted in the code but in the paper, the refinement module is added to the current_kspace in every iteration.
It looks like I am missing something in the implementation. Can you please explain.

mmuckley commented 3 years ago

Hello @akemisetti, practically, this shouldn't cause any difference. It just flips the sign of the weights. I checked and this is how the code was originally committed.

Since the published model weights that we have expect the current behavior, we won't be changing this.

mmuckley commented 3 years ago

Closing for now.