How will lesion voxels outside the field of view of the CTP imaging be handled?

For a number of patients, the field of view captured by the CTP is smaller than that of the DWI and, as a result, some of the tissue that is visible on DWI simply was not imaged via CTP. In many cases, some of the tissue that was imaged via DWI but not CTP (these voxels are completely empty in the derived CTP images) are segmented as part of the lesion. For patient 0019, for example, there are 1685 voxels (occurring on slices 58, 59 and 60) that were segmented as lesion on DWI, but are outside the image extents of the CTP data. Naturally, it is impossible to predict this part of the lesion using information from the CTP data since the CTP image is completely blank on the slices where this part of the lesion occurs.

It is simple to mask out any lesion voxels that fall outside the CTP for training, but considering that the lesion segmentations cannot be manipulated at test time, how will this be handled during testing?

Greyscale image is derivatives/sub-stroke0019/ses-01/sub-stroke0019_ses-01_space-ncct_ctp.nii.gz; red overlay is derivatives/sub-stroke0019/ses-02/sub-stroke0019_ses-02_lesion-msk.nii.gz

Hi Anthony,

We are not masking out voxels outside the field of view, as the lesions do exist in the considered brain, and the reduced field of view is just a technical limitation to tackle a clinical problem. This is one of the reasons for which we are also including all image data available at acute time, from which CTA and NCCT do provide full head coverage. In the ideal scenario, performance could be improved with these other modalities.

Ezequiel

Thanks @ezequieldlrosa. I am mainly asking to confirm whether a method that uses only the CTP data inherently faces an upper ceiling on the accuracy that it can achieve, which seems to be the case. As you suggest, our team will consider integrating additional imaging modalities with our model.

I am also familiar with and sympathetic to the fact that technical limitations come with using real-world data to tackle a clinical problem. If I were to be a devil's advocate, though, I might argue that the patient's full brain tissue also does exist in the considered brain, albeit outside the CTP field of view, and so some of what is 'truly there' has already been lost from the CTP images. The consequence of losing this information only from the CTP and not from the lesion segmentation, NCCT, or CTA seems to me to be that the CTP becomes significantly less informative than the other imaging modalities for rather unintuitive reasons. I have not seen any of the literature on CTP-based stroke outcome prediction, for example, report on the volume of lesion tissue that was outside the CTP FOV and count that tissue against the model's accuracy. This is not meant to be a criticism of the challenge or its data, especially considering that even CTP from our own institution typically does not cover the whole brain, but I think that it is important to note for the challenge participants that the decision not to harmonize the FOV of the baseline imaging and ground truth puts CTP-only models at a potentially-significant disadvantage compared to multi-modal models. That contextual information might also be important for drawing scientific conclusions from the results of the challenge, even though it accurately reflects the technical limitations of a real-world clinical scenario.

Regardless, thank you for the information. We will plan accordingly!

Hi Anthony,

Glad to hear your inputs!

We are aware of the CTP limitations w.r.t to the field of view. Our decision - after discussing and agreeing with the challenge organizers- came from the fact that this is a lesion segmentation challenge (not a sole CTP-centered lesion segmentation challenge). In that case, we agree, and your suggestion would have made 100% of sense. Being a multimodality imaging challenge, then CTP is what it is, with its benefits and drawbacks (the same happens with the other modalities). My personal thoughts are that 1) most of the time CTP would cover the majority (if not the entire) lesion and 2) it might still be the most meaningful modality (being the current clinical reference). Then, choosing which modalities to be included depends on each team's strategy.

I have not seen any of the literature on CTP-based stroke outcome prediction, for example, report on the volume of lesion tissue that was outside the CTP FOV and count that tissue against the model's accuracy. True. But as you clearly mentioned, for CTP-based outcome prediction, limiting the lesions to the FOV makes total sense. However, this is not the case for ISLES'24.

That contextual information might also be important for drawing scientific conclusions from the results of the challenge, even though it accurately reflects the technical limitations of a real-world clinical scenario. Agree!

Please let me know if you have further questions.

Best, Ezequiel