search

vqdang / hover_net

Simultaneous Nuclear Instance Segmentation and Classification in H&E Histology Images.
MIT License
537 stars 224 forks source link

eval resutls of stage1 and stage 2 #259

Open wustone1995 opened 1 year ago

wustone1995 commented 1 year ago

Hi, thanks for sharing your great work! I have a problem about reproducing the results on Consep dataset. For stage2, it's easy to obtain a result close to the one post in paper (like dice 0.83, original is 0.85). However, the result of stage 1 is very bad, like dice 0.3 or worse. But the log shows no big difference of these two stage (stage1 epoch50 dice=0.78, stage2 epoch50 dice=0.83). Is this a common scene? or something i make wrong?

simongraham commented 5 months ago

Hi,

I'm a little confused because you say that the Dice is 0.3 in stage 1, then you say it is 0.78 in the next sentence. Maybe you mean PQ is 0.3?

Anyway, I will put some notes here and hopefully this will help.

Despite the Dice being pretty good (binary segmentation performance) at stage 1, the HV outputs were not great and needed the second stage to get good results. Therefore, you will find that the instance segmentation results will not be good at stage 1 because the HV map is what is being used to split the nuclei.