Query - different type of label and how to cover in training part

[kamlesh0606]

Hello ,

I read given Documents and provide good document details for understanding logic parts.

i have some different type of label Like 1) only cover Color 2) only cover Size/Shape
3) only cover Size/Shape + Color

For Example ,

• in 2D , Covid X-Ray have cover only Color of pixel point and then find out area. And also Color base provide level of Covid. Same way TB have multiple place as color parts. and also size/Share is not need.

And Covid and TB Cover inside lang segmentation So have you any way for segmentation as Parent and diseases as child parts. So child label only check inside Parent label.

i discussed With Dr. for ref. https://github.com/MIC-DKFZ/nnUNet/blob/master/documentation/region_based_training.md - Brain Tumor Segmentation. As Per Dr, Size/Shape is not important. Brain as Parent and Tumor as child way needs. and also Color important for compare it.

Please provide your input on it.

[seziegler]

Hi @kamlesh0606 , to me it seems like the region based training that you've referenced is exactly what you need. Can you please clarify what is not working in your use case with this approach?

[kamlesh0606]

@seziegler Thanks for provide update.

Currently we are planning for Brain Tumor finding. So Step 1 ) find Out Brain using Brain Segmentation Step 2) Find out Tumor using Color base detection due to Size/Shape is not fix.

So in Step 2 ) can we use region based training ? is it ok ? Due to U-Net Model Mostly consider Shape model training . So i asking this query.

[seziegler]

I'm not sure what you mean by color base detection. Do you have annotations for brain and tumor?

If yes, if you would want to use region based training you could do it like this: "labels": { "background": 0, "brain": [1, 2], "tumor": [2], }, "regions_class_order": [1, 2], ...

This means that besides background you have two classes: brain and tumor. Here, brain would be an overlapping area of class 1 and 2 (since a tumor in the brain is still part of the brain).

However, in your use case it seems like the region based training might not yield a large improvement. So you could simply use normal nnunet training: "labels": { "background": 0, "brain": 1, "tumor": 2, },

[kamlesh0606]

@seziegler Thanks For provide update.

color base detection Like tumor is not fix Size And Shape So if Size And Shape use detection Then tumor detection outcome is low.

So Color base Area finding and Give final result of tumor Area.

Logic of tumor area - https://www.mayoclinic.org/diseases-conditions/brain-tumor/symptoms-causes/syc-20350084