Mismatch issues between lesion mask of ground truth and input images for learning

Hello. According to the data released for learning, the lesion mask of ground truth expected to be generated based on MR-DWI and then co-registered to NCCT, CTA, and CTP images. And I think there were some locational errors to occur in the co-registration step. In particular, overlays were often observed around the skull area. Even if we adjust and perform learning ourselves during the learning stage, MR images will not be provided during the test stage, so if ground truth for test-set will generate based on the same manners, we cannot expect high performance. I wonder what the organizers thinks about this issue. Thank you.

Hello, can you please provide the case id?

That case number is 'sub-stroke0080.' It was just one example. There were many similar cases that were incorrectly co-registered, especially the boundaries between bone and tissue. I would like to believe that the test set does not contain cases like this.

Hi, We have just released a revised version of the dataset, with some improvements in image registration. The pointed case looks better. Note that there are multiple reasons for which registration might be sub-optimal (please take a look at this discussion https://grand-challenge.org/forums/forum/ischemic-stroke-lesion-segmentation-challenge-2024-722/topic/request-for-co-registration-code-2363/). In large infarcts, sometimes with edema, there might be tissue displacements, turning the task even harder. The case you have pointed out is indeed complex, as it consists on a massive MCA lesion. Registration imperfections might be even more pronounced in the lower/upper brain slices due interpolation and partial volume effects.

Last: performance metrics are computed within the NCCT brain cavity so that lesions falling outside will be excluded.

Hope this helps. Ezequiel.