search

Bingchao-Zhao / Triple_U-net

The code of Triple U-net: Hematoxylin-aware Nuclei Segmentation with Progressive Dense Feature Aggregation.
18 stars 9 forks source link

The mistake about getting AJI? #1

Open Lewislou opened 3 years ago

Lewislou commented 3 years ago

Hi, I am curious about the AJI calculation of prediction images and the ground truth. From the code, looks like you are using binary predictions with binary masks to calculate the AJI. Would this kind of calculation wrong because AJI calculation needs instance labels?

caodoanh2001 commented 1 year ago

Agree. Seems like the authors using only nuclei binary masks to compute AJI. It is not correct.

caodoanh2001 commented 1 year ago

Hi, I figured it out.

Actually the authors did not use the binary nuclei map to compare with binary GT masks. But, the predicted nuclei map was binary ones. They were then performed instance segmentation from the binary maps by the function label from

https://github.com/Bingchao-Zhao/Triple_U-net/blob/44a50ab74acc78e69d8b8b59d4d91533e1d7d21d/source/metrics.py#L5

This function was performed in get_fast_aji function to compare them to GT instance map.

https://github.com/Bingchao-Zhao/Triple_U-net/blob/44a50ab74acc78e69d8b8b59d4d91533e1d7d21d/source/metrics.py#L26

I think it is clear now.