Replies: 3 comments 8 replies
-
Hello @arp95, there are quite a few issues with extending to other MRI datasets. The data could have different preprocessing that gives different normalization values. The data could be from different scanners. The data could be with different pulse sequences. All of these could cause generalization issues. It's not really possible to specify a procedure that will work in general. The best thing would be to work with an MRI physicist that understands how the systems work, and maybe be at an institution where you could acquire a small amount of data for validation and/or fine-tuning. Another option might be to ask the other people that released the datasets you're interested in how Siemens 2D imaging is related to their data. Lastly, if you post the datasets you're interested in here, maybe someone could give some advice that has experience with it. |
Beta Was this translation helpful? Give feedback.
-
So the fastMRI scans were acquired using Siemens scanners and if any dataset i am planning to validate this on is from Siemens should have same values? |
Beta Was this translation helpful? Give feedback.
-
Just one last question to clear my understanding over here @mmuckley |
Beta Was this translation helpful? Give feedback.
-
Hi,
I was currently working on training deep learning models mapping low resolution to high resolution images. First of all, thanks for this great dataset, it is an amazing resource.
I trained the dataset on fastMRI train and val datasets. I used the tutorial of fastMRI to get the images from raw kspace to images where the notebook applied a series of operations inverse fourier transform, complex_abs to get the image. I trained the images using the fastMRI as i mentioned and normlized using the mean and std from the train dataset.
The question becomes I want to validate this model on images outside of fastMRI. The first issue is that once fastMRI applies complex_abs we get image pixels in floating point values ranging between -1 to 1 (can be more than 1 or -1 but the intensities are not in 100's as we usually expect image intensity to be). When i am using other mri datasets for validation they have pixel range between 0 and 600. how do i convert them to fastMRI data distribution range so that my trained model can be tested? I cant use my mean and std on train dataset as that was 0.000326 and 0.000557 respectively due to the range of values I got from fastMRI train dataset. Am i doing something wrong?
Beta Was this translation helpful? Give feedback.
All reactions