histoencoder 0.1.0

Foundation models for digital pathology.

HistoEncoder

Foundation models for digital pathology.

Description • Why? • Installation • Usage • API Documentation • Citation

Description

HistoEncoder CLI interface allows users to extract and cluster useful features for histological slide images. The histoencoder python package also exposes some useful functions for using the encoder models, which are described in the API docs.

Why?

The models provided in this package produce similar features for tile images with similar histological patterns. This means that when we cluster the tile images based on their features, each cluster contains tile images with similar histological patterns.

Thus, visualising the clusters allows us to automatically annotate whole datasets! Additionally, calculating cluster percentages for a given patient would give us the distribution of histological patterns for the patient. This information could then be combined with other data modalities.

Installation

pip install histoencoder

Usage

1. Cut histological slide images into small tile images with HistoPrep.

HistoPrep --input './slide_images/*.tiff' --output ./tile_images --width 512 --overlap 0.5 --max-background 0.5

2. Extract features for each tile image.

HistoEncoder extract --input ./tile_images --model-name prostate-small

3. Cluster extracted features.

HistoEncoder cluster --input ./tile_images

Now train_tiles contains a directory for each slide with the following contents.

train_tiles
└── slide_image
    ├── clusters.parquet # Clusters for each tile image.
    ├── features.parquet # Extracted features for each tile.
    ├── metadata.parquet # Everything else is generated by HistoPrep.
    ├── properties.json
    ├── thumbnail.jpeg
    ├── thumbnail_tiles.jpeg
    ├── thumbnail_tissue.jpeg
    └── tiles  [52473 entries exceeds filelimit, not opening dir]

Citation

If you use HistoEncoder models or pipelines in your publication, please cite the github repository.

@misc{histoencoder,
  author = {Pohjonen, Joona},
  title = {HistoEncoder: Foundation models for digital pathology},
  year = {2023},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {https://github.com/jopo666/HistoEncoder},
}