cedricching / deepmri Goto Github PK

Hi~
You said in the paper that you use 2 datasets including fastmri.

How do you preprocess fastmri dataset? Especially, how to normalized the multi-channel data to have a maximum magnitude value of 1 as said in the paper?

I would highly appreciate it if you could provide the code to preprocess fastmri dataset and generate tfrecorders like train.tfrecords?

Hello,
I am trying to implement the code with our own dataset. I have a few questions regarding the training data:

1. I noticed the data was all reshaped to the defined dimension sequence. What is the sequence of direction? (like imgx, imgy, #of img, coils)?
2. Is it possible to provide the sample code of how to generate the train.tfrecords? Or could you provide the format of training data?
3. In the data.py and run.py, function "train(train_data,validate_data,test_data,mask) ", does this test_data denote the labeled data (reference data) instead of data used for training?

Thanks!

Best,
Cornelia W.

Hello,

I implemented this code on both brain and knee dataset. But in fastmri knee dataset, the reconstruction is worse than zero filled.

I don't know where the problem is or whether I skip steps or not. Is there anyone who take the satisfactory results in FastMRI knee dataset reconstruction?

Best,

你好，
`最近在用keras框架重现你们的工作。我们的数据是Calgaxy公开数据集，数据类型是12通道原始K空间数据。按照你们论文中归一化方式，我们对每张复数图像进行了除以最大模值进行归一化，代码如图。 不知道我这样归一化对不对？ 在用处理完的数据进行训练时，在第一个epoch中，我们发现训练损失从0.0016开始，后面就一直保持0.0015不变。这样正常吗？
希望能给一些指导，我们将万分感谢。

Hello, I am a graduate student, I would like to ask the version number of python3 and tensorflow you are using, thank you very much!