data normalization

[zoujing925]

thanks for the wonderful work and open-source code.

I have questions about data normalization. In the dataset.py file, the data normalization was set as 'std' with a data.normalize_coeff. but I didn't find inverse normalization in the test code, should I perform inverse normalization during evaluation?

Could you please tell me the range of the data that should be as input, [0, 1] or [-1,1], or others, the range of the photom data you provided is [-25, 82]

Many thanks for your help

[YuebyYue]

Thank you for your question and for pointing this out!

We do not perform inverse normalization during testing, as it does not affect the results of 2D reconstructions. However, we recommend applying inverse normalization after the reconstruction if you are planning to perform slice-by-slice reconstruction of 3D data.

Additionally, while we normalize the k-space data, there is no specific input data range requirement—it depends on the characteristics of your data.

[zoujing925]

Thanks a lot for your kind reply.

I train your model with fastMRI single coil knee dataset. But the sampling results have large noise, as shown in the figure.

The model was trained with the original hyperparameters provided in your code. Could you please give me some advice about training with the fastMRI knee dataset?

many thanks.

[zoujing925]

The sampling was testes with the checkpoint of the 275 epoch.

[YuebyYue]

There are some advice maybe helpful for you:

1. For the training, you can train it for a longer epoch. And execute reconstruction per xx epoch to check quality.
2. For the sampling, The "README" profile provides more details about setting proper hyperparameters especially the "snr" , "mse" and "corrector_mse". You can read it to fine tune these.