Class Visualizations and Activation Atlases for Enhancing Interpretability in Deep Learning-Based Computational Pathology

This repository contains the code to reproduce the experiments in the paper.

Important

WIP! For model training, please refer to the modeling directory in this repository!

Setup

Use conda to setup this project's python environment:

conda env create -f conda_env.yml
conda activate activation-atlas

Usage

Check config/uni_nct.yaml for an example of how to set up your configuration file, and config/NCT-CRC-HE-100K.csv and config/CRC-VAL-HE-7K.csv for examples of the file(s) describing your dataset(s).

Extract activations and create activation atlases

python create.py --config path/to/config_file.yaml --vis_type atlas

Create class visualizations

python create.py --config path/to/config_file.yaml --vis_type class_vis

Interactively explore visualizations and metrics

Start the viewer via python view.py. Click on "Open new file", navigate to a previously created atlas or class_vis file (by default contained at <save_root>/<experiment name>/<atlas or class_vis>/<timestamp>/) and open it. You can now:

• Select your layer (for activation atlases) or class (for class visualizations) of interest from the list on the left.
• Pan across (hold left mouse button, move mouse) and zoom into (mouse wheel) the visualization.
• Enable overlays for attributions, ground truth data and metrics from the Overlay menu in the left sidebar.
• View detailed attribution, ground truth and metrics data for individual cells by hovering over them with your mouse.

External resources

• GUI icons (_viewer/_resources/*.svg, _annotator/_resources/*.svg) for the viewer and annotator components from fontawesome.com.
• Code in _creator/captum_fragments.py from Captum's optim-wip branch.

License

Unless otherwise noted, this project is licensed under the MIT License (see LICENSE).

This project includes third-party code derived from Captum (PyTorch), licensed under the BSD 3-Clause License. See LICENSES/BSD-3-Clause-Captum.txt.