search

vqdang / hover_net

Simultaneous Nuclear Instance Segmentation and Classification in H&E Histology Images.
MIT License
523 stars 221 forks source link

about the datasets Kumar #86

Closed Kaiseem closed 3 years ago

Kaiseem commented 3 years ago

Hi, i found that the kumar datasets provided in this project is different from the original version "A Dataset and a Technique for Generalized Nuclear Segmentation for Computational Pathology" on webset url: http://nucleisegmentationbenchmark.weebly.com/

To be more specific, the following files are different, although they have the same name: train: TCGA-A7-A13E-01Z-00-DX1 TCGA-B0-5711-01Z-00-DX1 test: TCGA-21-5784-01Z-00-DX1 TCGA-AY-A8YK-01A-01-TS1 TCGA-B0-5698-01Z-00-DX1 TCGA-CH-5767-01Z-00-DX1

which is the correct one?

simongraham commented 3 years ago

I am confident that we are using the correct data. I will dig further into this issue and report back directly on here.

Kaiseem commented 3 years ago

Okay but i check Kumar (http://nucleisegmentationbenchmark.weebly.com/) and MoNuSeg (https://monuseg.grand-challenge.org/Data/) datasets, their images are the same, but some of the images in the provided kumar datasets are different from the above datasets, how to explain it?

simongraham commented 3 years ago

Email reply from the MoNuSeg organisers:

..................... Thanks for pointing this out - we wanted to add some additional images for which we had nuclear annotations lying around the names were messed while adding those to the original 2017 dataset.

Now, we have fixed this and you will find 30 original +7 bonus images with nuclear annotations on the Nuclei Segmentation Benchmark website.

However, for MoNuSeg training data we have kept only the original 30 images on the monuseg website data page. .....................

So, in short the data that we used in this repository is correct. The data from http://nucleisegmentationbenchmark.weebly.com/ was recently updated and a mistake was made. This has now been fixed.

1] 30 original images used found here 2] 30 original images + 7 extra images (added recently) found here

The 2017 IEEE TMI publication uses 1] as well as other subsequent publications such as HoVer-Net, deep distance map regression (DIST), CIA-Net etc.

Please let me know if you have any further questions.

Kaiseem commented 3 years ago

thank you for your checking, it confused me for a long time