Welcome to the official repository of HoverFast, a high-performance tool designed for efficient nuclear segmentation in Whole Slide Images (WSIs).

Overview

HoverFast utilizes advanced computational methods to facilitate rapid and accurate segmentation of nuclei within large histopathological images, supporting research and diagnostics in medical imaging. For more info on the inner workings of HoverFast, do not hesitate to go over our paper

Documentation

An overview of the documentation is provided in this repository, but for more details, please visit the full official documentation

Installation

Prerequisites

• Python 3.11.5
• CUDA installation for GPU support (version > 12.1.0)

Using Docker

We recommend using HoverFast within a Docker or Singularity (Apptainer) container for ease of setup and compatibility.

• Pull Docker Image docker pull petroslk/hoverfast:latest

Using Singularity

For systems that support Singularity (Apptainer), you can pull the HoverFast container as follows:

• Pull Singularity Container singularity pull docker://petroslk/hoverfast:latest

Local Installation with Conda

For local installations, especially for development purposes, follow these steps:

• Create and activate a Conda environment conda create -n HoverFast python=3.11 conda activate HoverFast

• Install HoverFast git clone https://github.com/choosehappy/HoverFast.git cd HoverFast pip install .

Usage

Command Line Interface

HoverFast offers a versatile CLI for processing WSIs, ROIs, and for model training.

For Whole Slide Images (WSI) Inference

• Basic Usage HoverFast infer_wsi --help

• Check Version HoverFast --version

• Example Command without binary masks HoverFast infer_wsi path/to/slides/*.svs -m hoverfast_crosstissue_best_model.pth -n 20 -o hoverfast_output

• Example Command with binary masks

Although HoverFast does have a simple threshold based tissue detection, we highly recommend the use of QC tools such as HistoQC for generating tissue masks to avoid computing on artefactual regions and reducing computation time. You can give the path to the directory where the masks are stored. HoverFast will search for a mask with the same name as the slide with a .png extension.

HoverFast infer_wsi path/to/slides/*.svs -b path/to/masks/ -m hoverfast_crosstissue_best_model.pth -n 20 -o hoverfast_output

• Example for IHC Nuclear DAB stain

If your IHC DAB stain is nuclear, you should use the ihc_dab flag to segment nuclei. If your IHC DAB stain is not nuclear, regular H&E segmentation might be a better option.

HoverFast infer_wsi path/to/slides/*.svs -b path/to/masks/ -m hoverfast_crosstissue_best_model.pth -n 20 -o hoverfast_output -st ihc_dab

For Region of Interest (ROI) Inference

• Example Command

HoverFast infer_roi path/to/rois/*png -m hoverfast_pretrained_pannuke.pth -o hoverfast_output

Using Containers

Containers simplify the deployment and execution of HoverFast across different systems. We highly recommend using them!

Docker

• Run Inference

docker run -it --gpus all -v /path/to/slides/:/app petroslk/hoverfast:latest HoverFast infer_wsi *svs -m /HoverFast/hoverfast_crosstissue_best_model.pth -o hoverfast_results

Singularity

• Run Inference

singularity exec --nv hoverfast_latest.sif HoverFast infer_wsi /path/to/wsis/*svs -m /HoverFast/hoverfast_crosstissue_best_model.pth -o hoverfast_results

Training

To train HoverFast on your data, you may need to generate a local dataset first using our provided container.

Generating Local Dataset

• Structure your data directory

└── dir config.ini └── slides/ ├── slide_1.svs ├── ... └── slide_n.svs

• Generate Dataset

docker run --gpus all -it -v /path/to/dir/:/HoverFastData petroslk/data_generation_hovernet:latest hoverfast_data_generation -c '/HoverFastData/config.ini'

This should generate two files in the directory called "datatrain.pytable" and "datatest.pytable". You can use these to train the model.

• Train model

You can use these to train the model as follows:

The training batch size can be adjusted based on available VRAM

HoverFast train data -o training_model -p /path/to/pytable_files/ -b 16 -n 20 -e 100

docker run -it --gpus all -v /path/to/pytables/:/app petroslk/hoverfast:latest HoverFast train data -o training_metrics -p /app -b 16 -n 20 -e 100

singularity exec --nv hoverfast_latest.sif HoverFast train data -o training_metrics -p /path/to/pytables/ -b 16 -n 20 -e 100

Testing

Since HoverFast utilizes GPU for almost all tasks, most tests have to be run locally using pytest.

First, install pytest:

pip install pytest

Then, you can just run the following command inside the HoverFast repo:

pytest -vv Note that the first time you run these, the infer_wsi test can take longer since the slide will be downloaded locally

For more detailed instructions, including setting up your environment and running specific tests, please refer to the testing documentation

How to Cite HoverFast

If you use HoverFast in your research, please cite our paper:

bibtex @article{Liakopoulos2024, doi = {10.21105/joss.07022}, url = {https://doi.org/10.21105/joss.07022}, year = {2024}, publisher = {The Open Journal}, volume = {9}, number = {101}, pages = {7022}, author = {Petros Liakopoulos and Julien Massonnet and Jonatan Bonjour and Medya Tekes Mizrakli and Simon Graham and Michel A. Cuendet and Amanda H. Seipel and Olivier Michielin and Doron Merkler and Andrew Janowczyk}, title = {HoverFast: an accurate, high-throughput, clinically deployable nuclear segmentation tool for brightfield digital pathology images}, journal = {Journal of Open Source Software} } By citing HoverFast, you help us to continue our research and development. Thank you for your support!

Authors

• Julien Massonnet - JulienMassonnet
• Petros Liakopoulos - petroslk
• Andrew Janowczyk - choosehappy