Segmentation results

[swillek]

Hi Fabian,

After running HD-GLIO on data with paediatric patients with low grade gliomas, I get varying results. I know that the tool is tested on an adult population, but I wanted to know if it is applicable on a paediatric population as well.

After running and looking at my results, I have a question regarding tumours that have a cystic component. Some tumours have cystic components and sometimes only this portion is growing, which means that I need this portion in my segmentation mask to determine the entire tumour volume. However, as I have tested, this part of the tumour is not included in the non-enhancing tumour mask for some patients. The cystic component is hyperintens on T2, but is iso- or hypo-intens on FLAIR. Is this the reason that this component is not included in the tumour segmentation?

Thereby, I see varying results in the T2-FLAIR segmentation mask. In some cases, only half of the tumour is segmented, while that portion doesn't appear different on one of the scans compared to the included portion. Have you seen this before and do you know where and why this happens?

I figured that it may be useful to have contact via email. My email address is s.c.willekens@student.utwente.nl. I am looking forward to your reply.

Kind regards, Sanneke Willekens

[FabianIsensee]

Hi Sanneke,

as far as I know we did not have patients with cysts in our data set which would explain why these regions cannot be segmented properly. Even if we had cysts, these would probably have been annotated as background, resulting in a model that is again unable to segment them properly.

Could you provide examples on the T2/FLAIR segmentations sometimes being incomplete? It is difficult to say what might be going on without looking at the data first.

If you need a necrosis label you can also consider using our BraTS 2020 docker image (see here https://hub.docker.com/r/fabianisensee/isen2020, read this on how to use the docker images: https://github.com/BraTS/Instructions). The BraTS ground truth classes differ from the HD-GLIO dataset and may be more useful for your particular application. Again, BraTS is only adult patients so there is high probability that the segmentations are not perfect.

Best,

Fabian

[FabianIsensee]

Hi, I do not have access to the images you shared via google drive. When I click it it says I need to request access first. Best, Fabian

[FabianIsensee]

nope. You need to check an extra box for link sharing

[FabianIsensee]

Got it. So these cysts are rather atypical. I don't think we can fix that tbh. I ran some model that I trained on BraTS2020 on your data. The output is the attached file: 1.nii.gz The braTS model also struggles with the cyst. Best, Fabian

[swillek]

Your result does look better. Which model did you use? Is it open source? Maybe I can try them for the other patients that had similar issues as this one.

Best, Sanneke

Op ma 3 mei 2021 om 16:41 schreef Fabian Isensee @.***>:

Got it. So these cysts are rather atypical. I don't think we can fix that tbh. I ran some model that I trained on BraTS2020 on your data. The output is the attached file: 1.nii.gz https://github.com/NeuroAI-HD/HD-GLIO/files/6415528/1.nii.gz The braTS model also struggles with the cyst. Best, Fabian

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/NeuroAI-HD/HD-GLIO/issues/4#issuecomment-831304612, or unsubscribe https://github.com/notifications/unsubscribe-auth/ATHT5OOV4ZNJIXJYOXVACCTTL2YYVANCNFSM43SS2XRQ .

[FabianIsensee]

Install nnU-Net

use nnUNet_download_pretrained_model Task082_BraTS2020

use nnUNet_predict with any of the downloaded models, example:

nnUNet_predict -i INPUTFOLDER -o OUTPUTFOLDER -tr nnUNetTrainerV2BraTSRegions_DA4_BN -p nnUNetPlansv2.1_bs5

see https://github.com/MIC-DKFZ/nnUNet/blob/master/nnunet/dataset_conversion/Task082_BraTS_2020.py for more information on what models there are (see bottom of the page)

OR simply use our Docker image which reproduces our winning BraTS 2020 submission (see my post at the top ;-) )

[FabianIsensee]

(if you do down the nnU-Net route please read ALL the documentation before asking questions :-) )

[swillek]

Haha yeah I missed that one sentence the last time. I will try the other methods on my patient data, I hope those work better for my population. If not, then at least I have a starting point for my tumour segmentation.

Thanks for the help!

Op ma 3 mei 2021 om 17:11 schreef Fabian Isensee @.***>:

(if you do down the nnU-Net route please read ALL the documentation before asking questions :-) )

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/NeuroAI-HD/HD-GLIO/issues/4#issuecomment-831324596, or unsubscribe https://github.com/notifications/unsubscribe-auth/ATHT5OLG2XXQ7RK254S7M4LTL24LVANCNFSM43SS2XRQ .

[FabianIsensee]

let me know how well the docker performs for you