By: JeroenBertels on July 22, 2022, 9:20 a.m.
Dear challenge organisers,
Just to make sure: the FLAIR has always a different affine compared to the other three modalities (e.g. larger FOV and smaller voxel sizes). Is that the case also for the test images? Meaning that if participants want to include registered FLAIRs in their algorithm, they also need to build this into their submitted docker algorithms?
Thank you in advance!
Kind regards,
Jeroen
By: JeroenBertels on July 22, 2022, 1:56 p.m.
Furthermore, not all affines of the masks are identical to the affines of the dwi/adc... For example sub-strokecase0073 and sub-strokecase0086.
By: ezequieldlrosa on Aug. 11, 2022, 7 a.m.
Dear Jeroen,
Apologies for the very late reply. I just found this message in the long notifications list received from GC. We hope this has not hampered your algorithm development.
Methods using the FLAIR image are indeed expected to handle registration/whatever transformation needed in order to match the other modalities. This is up to each participant strategy (we aimed to leave full freedom to participants to design their solutions).
As for the affine issues, we will re-check the dataset. Thanks a lot for finding this out.... (Just FYI, for evaluating the challenge results we do not take the affines into account, but only the binary map arrays. Anyway we'll fix this!)
Best, EDLR
By: JeroenBertels on Aug. 17, 2022, 8:21 a.m.
Dear Ezequiel,
No problem! We can imagine that it must have been quite a hassle to organize everything, so thanks alot you for doing that!
Okay, we indeed understood the FLAIR was in a different coordinate system and had included a registration step to the DWI space :-)
Regarding "we only use the binary map arrays for evaluation": I had assumed these needed to be provided in the space of the DWI. However, in the training set not all affines of the segmentations exactly match the affine of the DWI. Anyway, the differences, if any, were mostly insignificant.
Kind regards,
Jeroen (team MIRC)
