Useful Tools for AI Interpretability in Digital Pathology

Hi all!

I recently attended the Applied Machine Learning Days session on Interpretable AI for Digital Pathology and learned about a lot of new things thanks to these excellent speakers.

I wanted to share these resources here as I am sure they may be of use to someone else

The GitHub Repo of the Session, with a few notebooks that can be run in collab to see what was presented.

StainTools: A Python package specialized in H&E Stain Estimation, correction and augmentation (for training models)

HoVerNet: (on GitHub) Another nuclei segmentation model trained on H&E datasets that seems similar to the CellPose approach.

There are plenty of other references in the GitHub repository which all look fascinating, especially Graph Neural Networks (GNNs). I have not touched on the subject of the session which was to show methods to visualize and quantify what features (and where those features are) in an image a Deep Network uses as a discriminator when classifying images. Moreover it shows how one can map our own metrics (texture, object size, color, circularity, etc… ) to the results of a deep learning network and see how they correlate to the classification.

Hope others will find this useful. I personally realized that I was not up to date at all in this field, despite using DL already in our facility.



[edit] changed the title to be a bit less generic and buzz sounding


Thank you @oburri, these are very useful resources!

Best regards,