Brain data reconstruction size

#1

Hi,

I am starting to look at the brain data, and I noticed that the reconstruction size is not as formatted as for the knee data, it’s not always 320 x 320.

I thought that I could obtain the information of the expected reconstruction size using the ismrmrd header, more specifically the encoding.reconSpace.matrixSize, but it doesn’t seem to match the actual reconstruction size found in 'reconstruction_rss'. An example of this mismatch is for example the validation file file_brain_AXFLAIR_203_6000881.h5.

In this file, the kspace has a shape of (16, 4, 512, 213), the reconstruction has a shape of (16, 213, 213) and the encoding reconstruction space has a shape of (512, 408).

For the train and validation datasets it’s of course not a problem because you can always get the reconstruction size from the actual reconstruction but for the test dataset (and probably the challenge dataset), I don’t know how to proceed.

How can I know the expected reconstruction size from a test file?

#2

I’m looking into it. I think there were some issues with the FLAIR scans due to parital Fourier encoding.

#3

Okay, just heard back. These scans have the wrong values for “matrixSize” and as a workaround we just use the minimum encoding size - (213, 213). These cases should be excluded from the test set.

#4

Thanks for asking.

Yes anyway for the validation and training I was planning to use the target size anyway.

One last question, for even to odd cropping, how are we supposed to crop? I guess I will see when experimenting but if you have the answer right now it could be simpler for me.

#5

I am new for the brain data so I am learning with you. :slight_smile:

#6

Hahaha ok !

I think that the cropping for odd output sizes is the following (in pseudo-code):

start_crop = orig_image.shape[0]//2 - (output_size[0]//2) - 1
orig_image[start_crop:start_crop + output_size[0]]
#7

I think the only case where we have an odd number is for one of the FLAIR_203 images, correct? These should be excluded from leaderboard calculations.