search

mahmoodlab / CLAM

Open source tools for computational pathology - Nature BME
http://clam.mahmoodlab.org
GNU General Public License v3.0
1.13k stars 366 forks source link

Inconsistent Tissue Segmentation #61

Closed artemi8 closed 3 years ago

artemi8 commented 3 years ago

Hello, Thank you for making this amazing project a public repo!

I used TCGA-BRCA WSI slides for experimenting with CLAM. When using the default segmentation parameters from the given tcga.csv file I observed some segmented WSIs had white background between the tissue region. I also observed that segmentation improved when I changed use_otsu parameter in tcga.csv to TRUE. I have attached two examples on both the cases for your reference.

Can you please point out the segmentation/thresholding parameters that need to be tuned to get more precise segmentation.

Using default parameters in tcga.csv with use_otsu = FALSE Tissue issue

Using default parameters in tcga.csv with use_otsu = TRUE Tissue issue_2

fedshyvana commented 3 years ago

Hi, you are welcome to refer to https://github.com/mahmoodlab/CLAM#custom-default-segmentation-parameters if you are interested in achieving really precise segmentation (additionally you can tune segmentation/filtering parameters for each slide as well in addition to using a global template). I think for most tasks even if you some how have a bunch of "empty" patches in your bags or imperfect segmentation, the classification model will learn just fine since it should learn to ignore those relevant regions and attend to diagnostically relevant regions.

artemi8 commented 3 years ago

Okay thank you for your response.