search

TIO-IKIM / CellViT

CellViT: Vision Transformers for Precise Cell Segmentation and Classification
https://doi.org/10.1016/j.media.2024.103143
Other
189 stars 27 forks source link

Question about the bPQ metric. #25

Closed windygoo closed 9 months ago

windygoo commented 9 months ago

From your provided logs in 'logs_paper/PanNuke/CellViTHV/SAM-H', the bPQ for three folds is 0.6633524845383327, 0.6624584712676017 and 0.6711352126330197, respectively. And thus, the average value is 0.6656487228129847, which I find is different from 0.6793 reported in your paper. How to explain it?

FabianHoerst commented 9 months ago

Could you be more specific? Which bPQ do you mean? Please refer to the respective table.

FabianHoerst commented 9 months ago

I hope i get you right. The first values reported in the inference.log files are the global averaged results from the dataset (averaged within one fold). The bPQ in table 3, on the other hand, is calculated differently. For this, you need to average the bPQ values of each individual tissue class over all three folds, and then calculate the average over this to get the bPQ value. This averaging is different to the fold wise average. For further information, please refer to this script which was used to build the inference pipeline: https://github.com/TissueImageAnalytics/PanNuke-metrics

windygoo commented 9 months ago

I hope i get you right. The first values reported in the inference.log files are the global averaged results from the dataset (averaged within one fold). The bPQ in table 3, on the other hand, is calculated differently. For this, you need to average the bPQ values of each individual tissue class over all three folds, and then calculate the average over this to get the bPQ value. This averaging is different to the fold wise average.

Thanks very much! I have got it.