Network and config files

Question

Network and config files

Closed this issue 3 years ago · 10 comments

Hi Andre,
Would it be possible to share the .mibDeep and .migCfg files for the H&E epithelium model?
Thank you!

Answer 1 · 2022-02-22T18:35:24.000Z

@ajr82 I have contacted the first author (@SahPet) and he will send me the files soon. I will let you know when they are ready to download :]

Answer 2 · 2022-02-23T18:08:20.000Z

@ajr82 seems like people are on winter holiday in Norway this week. Hence, I cannot access and upload the model files until next week.

Answer 3 · 2022-02-23T20:44:15.000Z

Hi ajr82!

Sorry about the late reply. Here are the best performing U-nets for the HE-dataset with input 256x256 and 512x512. They are additionally trained with hardmining so slightly better performance than the paper states. Here are the U-nets (*.mibDeep files) shared on my Google Drive: https://drive.google.com/drive/folders/1CgLdIpiM7Rn2v30V0NLor0tTtZmllD7Y?usp=sharing
There is a bit sparse augmentation during training, so they can probably be further improved (especially regarding transferability to WSIs from different labs/scanners) with optimization of augmentation settings.

Answer 4 · 2022-02-24T02:29:39.000Z

Hi @andreped and @SahPet, thank you very much!

Answer 5 · 2022-02-24T07:49:20.000Z

Hi again @ajr82! I was going to add a IBDColEpi HE trained deeplabV3resnet18 *.mibDeep-file, but need to figure out if it's going to be part of a publication first.

Answer 6 · 2022-02-24T07:53:04.000Z

@SahPet important that we post the models used for the paper here, at least firstly.

It's there now (https://drive.google.com/drive/folders/1CgLdIpiM7Rn2v30V0NLor0tTtZmllD7Y?usp=sharing). We can of course share additional network files if we want too though.

Answer 7 · 2022-02-24T11:05:14.000Z

@ajr82 the best performing models (512, separate models for HE and CD3) have been uploaded to the Google Drive:
https://drive.google.com/drive/folders/1eUVs1DA1UYayUYjr8_aY3O5xDgV1uLvH

Download the zip file "trained-models.zip".

I have also sent a request to DataverseNO to get the models added there as well.

However, we cannot upload the models to the github itself, as github is not meant to store these types of data. They simple are too large. I will update the README when this is done.

Note that you only need the .mibDeep files to use the two models in MIB. The .mibCfg files are not needed.

EDIT: README has now been edited. See here for information on how to access the trained models (both in the .onnx and .mibDeep formats relevant for FastPathology and MIB, respectively).

Answer 8 · 2022-02-24T22:10:26.000Z

@andreped and @SahPet Thank you very much!

Answer 9 · 2022-02-24T22:14:11.000Z

@SahPet important that we post the models used for the paper here, at least firstly.

It's there now (https://drive.google.com/drive/folders/1CgLdIpiM7Rn2v30V0NLor0tTtZmllD7Y?usp=sharing). We can of course share additional network files if we want too though.

Hi again @ajr82! I was going to add a IBDColEpi HE trained deeplabV3resnet18 *.mibDeep-file, but need to figure out if it's going to be part of a publication first.

It would be nice to see the outcome of this model too. We eventually plan on publishing giving due credit to you both. If the deeplabv3resnet18 model cannot be used to analyze data that will be published, we will not use it for the publication.

Answer 10 · 2022-02-24T22:16:23.000Z

The U-Net model should work quite well. There will always be a better model in the future. We will let you know when it is available.