kaczmarj/camelyon16-uni

CAMELYON16 patch embeddings made with UNI

This repository contains patch embeddings for CAMELYON16 made with the UNI foundation model.

Patches are 128x128 micrometers. Tissue segmentation and patching was done with a modified version of the CLAM toolkit. The toolkit was modified to extract constant physical size patches.

The patches directory contains HDF5 files with patch coordinates. The attribute patch_size on the /coords dataset contains the patch size in pixels. This is equivalent to the physical size of 128 micrometers.

The embeddings directory contains PyTorch files with embeddings. Each specimen in CAMELYON16 is in a separate file, and each file contains a 2D tensor of shape (n_patches, n_features). The value of n_patches can differ across specimens.

The order of the patches is the same between the patch HDF5 file and the features PyTorch file.

Intended use cases

This dataset is intended for training weakly-supervised neural networks on CAMELYON16. It is also intended to help others reproduce the experiments in the HIPPO manuscript.

Links

• CAMELYON16: https://camelyon16.grand-challenge.org/
• CLAM: https://github.com/mahmoodlab/CLAM