fastMRI

Brain data crop size

Hi,
1.
I want to know exact image crop sizes(=reconstruction size) of challenge test sets which will be released in October 1st.
Are they encoding.reconSpace.matrixSize in ismrmrd_header or
smallest_width = min(resolution, image.shape[-2])
smallest_height = min(resolution, image.shape[-3]) ?
2.
In evaluate function in evaluate.py,
target = transforms.center_crop(target, (target.shape[-1], target.shape[-1]))
recons = transforms.center_crop(recons, (target.shape[-1], target.shape[-1])).
Is this function used to calculate challenge metric ?

Thanks in advance

Hello - (1) is more or less correct. I just rolled out an update today for the zero-filled solution that should be working correctly for the brain data. I’m going to update the U-Net and VarNet models as well when I get time.

Thanks.

I have one more question.
In evaluate function in evaluate.py, we can see that the following code is performed.

transforms.center_crop(recons, (target.shape[-1], target.shape[-1]))

This means that a ‘square’ crop is performed with length of the y dimension.

For example, if enc.reconSpace.matrixSize.x=320 and enc.reconSpace.matrixSize.y=260, then it seems to me that only the 260 by 260 portion of the reconstructed image will evaluated in challenge competition. Is this true?

Thank again.

Hello - this is true in this case.

Hi, thanks for answering.

I am still wondering about the ‘square crop’ in evaluate function because this operation can remove not only the background but brain structure also.

In case of file_brain_AXT1_202_6000282 file in test set, shape of ksapce is (16, 16, 640, 272) and enc.reconSpace.matrixSize.y=270. If we crop combined image to 270 by 270 matrix, it becomes like the attached image. It is awkward for me to measure metrics and evaluate perceptual quality with such images.

Despite these cases, will 'square crop` be used to evaluate reconstruction in the challenge?

Thank you.

Let me double-check with everyone on this. We might be able to tweak how the challenge evaluations are. But for the current public leaderboard they will continue with square cropping.

Is there any update on this issue?

My apologies. I conferred with some other organizers on the challenge and we’ve decided to keep the existing behavior with square cropping.

Thanks for the reply, I have just a few more questions.
Will the image from the square crop be used not only for the metric measurement stage, but also for the perceptual quality stage?
If the square-cropped image will be used for the perceptual quality stage, is it possible for us to submit the square-cropped image? When I checked the updated code, I found that run_zero_filled.py‘s save_zero_filled function doesn’t square crop but rather crops using the sizes (enc.reconSpace.matrixSize.x, `enc.reconSpace.matrixSize.y’).
Thanks again for replying to such specific questions.

For the perceptual quality stage we’ll send the full images. So don’t try to hide any artifacts there :slight_smile:.

I’m sorry for yet another question. When you say ‘full image’, do you mean images of size ( enc.reconSpace.matrixSize.x , enc.reconSpace.matrixSize.y )? So ultimately, should I submit images of size ( enc.reconSpace.matrixSize.x , enc.reconSpace.matrixSize.y )?Thanks again!

Yes that’s correct. The only exception is the FLAIR_203 files, because for those files enc.reconSpace.matrixSize is incorrect.

In the challenge test set, will the FLAIR_203 files appear? (I remember in a previous Q&A that the FLAIR 203 files will be omitted) If they are to be included, what size should they be for submission?

You can omit them. I also think if you submit to the leaderboard without them then your submission should process fine - they should be deleted from the evaluation set.