dddavid4real / HistGen

[MICCAI 2024] Official Repo of "HistGen: Histopathology Report Generation via Local-Global Feature Encoding and Cross-modal Context Interaction"
Apache License 2.0
33 stars 0 forks source link

Weights for DINOv2 ViT-L Feature Extractor #4

Closed BenPashley closed 3 months ago

BenPashley commented 3 months ago

Hi,

I'm really impressed with your work. Where can I download the weights for your feature extractor?

I can't see them on your downloads page.

Ben

dddavid4real commented 3 months ago

Thank you for your interest!

We are almost done with the organization of codes and checkpoints. The weight will be released within one week!

BenPashley commented 3 months ago

Wonderful. Thanks. As per the other logged request, will you be still making the WSI available from a single location? (to save having to download them individually using the IDs).

Ben

dddavid4real commented 3 months ago

We encourage you to download them from the TCGA data portal using the IDs we provided. The total file size is too large (~5T) and we haven't managed to find an online server to upload and save them efficiently.

Besides, for experimental reproduction, the .pt files (features of these WSIs) are available and could be directly used as the model inputs.

BenPashley commented 3 months ago

Thanks. I would like to replicate the full end to end process. Are you able to share the data-prepartion code for the patching and DinoV2 feature extraction and preparation? I assume this includes the local/global hierarchical pre-processing of the features?

dddavid4real commented 3 months ago

The preprocessing code will be uploaded along with the DINOv2 feature extractor model.

For the preprocessing, we utilized an accelerated version of CLAM. Note that there is no hierarchical pre-processing included in our method. The preprocessing is the same as CLAM. The local-global hierarchical encoding is after the feature extraction and can be seen as a MIL model.

dddavid4real commented 3 months ago

Hi,

We have updated the code and checkpoints for preprocessing and feature extraction. This issue will be closed.

Feel free to re-open it if you run into problems.