A question about undersampling

[Haozhoong]

In the demo of transforms: '\fastmri\data\transforms.py', UnetDataTransform applies udnersampling mask in k space and then crops input to correct size in image domain and VarNetDataTransform does the same thing. However MiniCoilTransform crops input to correct size in image domain and then applies udnersampling mask in k space.

There are zero-padding lines in phasing direction and they are recorded in h5py file at attrs["padding_left"] and attrs["padding_right"]. Therefore, I think the two undersampling approaches above have different effective acceleration factors. Which order is usd in the test dataset?

Thanks

[mmuckley]

Hello @Haozhoong, the 100% correct thing to do is what is done in UnetDataTransform and VarNetDataTransform. MiniCoilTransform was used in a sampling trajectory design paper to construct a miniature version of the full data for faster experimentation, but it doesn't do exactly what will occur in the scanner.

The test set applies the same undersampling as UnetDataTransform and VarNetDataTransform.

[Haozhoong]

OK, thanks!