UNEST brain segmentation getting noise in the segmentation

[abhishekpatil32]

Hello everyone,

Thank you for the MONAI bundle feature—it's much easier to use!

I was trying out the bundle on different datasets and noticed that the model is producing segmentation errors, as shown in the figure. I tried keeping the largest component and filtering the intensities, but this didn't help.

Can someone please help me resolve this?

Thank you!

[ericspod]

CC @yiheng-wang-nv @tangy5 Hi both, I think this is your network, would you have any feedback here? Thanks!

[yiheng-wang-nv]

Hi @abhishekpatil32 , could you provide more details? Like give some sample code. I'm not familiar with this model, but usually in order to adapt to other dataset requires transfer learning.

[abhishekpatil32]

Hello @yiheng-wang-nv, I am using the monai-bundle and the model from the monai model zoo (https://proxy.monai.io/proxy/download/whole_brain_large_unest_segmentation/versions/0.2.5/files/whole_brain_large_unest_segmentation_v0.2.5.zip) using the instructions as provided.

The initial code is registration to MNI305 template using ants followed by the bundle code, similar to what is mentioned in the model zoo instructions.

I wonder if there are more pre-processing steps required for this?
I am not sure if its a problem of transfer learning since we can clearly see the segmented brain but the issue is the noise with it.

[yiheng-wang-nv]

Hi @abhishekpatil32 , since this bundle's training set is not public, I'm not sure if any extra pre-processing steps are required. Maybe @tangy5 (author of the bundle) can bring some information here.

If looking at the transform here: https://github.com/Project-MONAI/model-zoo/blob/dev/models/renalStructures_UNEST_segmentation/configs/train.json#L76
Only the range (-175, 250) of the input data are useful. I guess if your CT data has different value range, some informative part of the data will be ignored.

[abhishekpatil32]

Hello @yiheng-wang-nv, I think the model you have the link to is not what we are currently using and the model uses T1W MR images (not CT scans).

The correct model config is here: https://github.com/Project-MONAI/model-zoo/blob/dev/models/wholeBrainSeg_Large_UNEST_segmentation/configs/train.json

[yiheng-wang-nv]

I see. This bundle uses NormalizeIntensityd, thus there does not have normalization issue.