search

TissueImageAnalytics / cerberus

One Model is All You Need: Multi-Task Learning Enables Simultaneous Histology Image Segmentation and Classification
GNU General Public License v3.0
72 stars 12 forks source link

Human-in-the-loop Segmentation #5

Closed FabianHoerst closed 1 year ago

FabianHoerst commented 1 year ago

Hello!

First of all, thank you very much for providing the code. Your paper is very interesting and covers a tremendous range of topics. However, I have a question about section 3.4 (Collecting large datasets for segmentation): Which Tool did you use for performing active learning? Did you implemented it on your own specifically for WSIs or did you used software designed for any type of images, since you are working patch-wise?

Thanks in advance!

simongraham commented 1 year ago

Hi @FabianHoerst

Thanks for the kind words! For the pathologist-in-the-loop annotation process, we generated the segmentation results after each step and refined them using GIMP. Then, we fed re-extracted patches from the refined annotations and retrained each time. Using an integrated interface for this would have been ideal, but we didn't have one in mind.