IGUANA is a graph neural network built for colon biopsy screening. IGUANA represents a whole-slide image (WSI) as a graph built with nodes on top of glands in the tissue, each node associated with a set of interpretable features.
For a full description, take a look at our preprint.
# create base conda environment
conda env create -f environment.yml
# activate environment
conda activate iguana
# install PyTorch with pip
pip install torch==1.10.1+cu102 torchvision==0.11.2+cu102 -f https://download.pytorch.org/whl/cu102/torch_stable.html
# install PyTorch Geometric and dependencies
pip install torch-scatter -f https://data.pyg.org/whl/torch-1.10.1+cu102.html
pip install torch-sparse -f https://data.pyg.org/whl/torch-1.10.1+cu102.html
pip install torch-geometric
doc: image files used for rendering the README - not necessary for running the code.dataloader: contains code for loading the data to the model.metrics: utility scripts and functions for computing metrics/statistics.misc: miscellaneous scripts and functions.models: scripts relating to defining the model, the hyperparameters and I/O configuration.run_utils: main engine and callbacks.
To see the full list of command line arguments for inference and explanation, run python run_infer.py -h and python run_explainer.py -h, respectively. We have also created two bash scripts to make it easier to run the code with the appropriate arguments. As an example, to run model inference enter:
python run_infer.py --gpu=<gpu_id> --model_path=<path> --data_dir=<path> --data_info=<path> --stats_dir=<path>
You will see above that the data_info csv file will need to be incorporated as an argument. This will determine the label and which images to process. By default, the code will process images with values in the fold column equal to 3. If considering a test set, there will be a single 'fold' column named test_info. The fold_nr and split_nr can be added as additional arguments if considering cross validation, which determines the subset of the data from the csv file.
We have made an interactive demo to help visualise the output of our model. Note, this is not optimised for mobile phones and tablets. The demo was built using the TIAToolbox tile server.
Check out the demo here.
In the demo, we provide multiple examples of WSI-level results. By default, glands are coloured by their node explanation score, indicating how much they contribute to the slide being predicted as abnormal. Glands can also be coloured by a specific feature using the drop-down menu on the right hand side.
As you zoom in, smaller objects such as lumen and nuclei will become visible. These are accordingly coloured by their predicted class. For example, epithelial cells are coloured green and lymphocytes red.
Each histological object can be toggled on/off by clicking the appropriate buton on the right hand side. Also, the colours and the opacity can be altered.
To see which histological features are contributing to glands being flagged as abnormal, hover over the corresponding node. To view these nodes, toggle the graph on at the bottom-right of the screen.
We have released a small portion of data to allow researchers to get the code running and see how our graph data is structured. We also include two 'data info' csv files - one as if the data is to be used for cross validation and the other as an external test set. Click here to download the sample dataset.
To download the IGUANA weights trained on each fold of the UHCW dataset, click here. To get the code running, you will also need the stats info used to standardise the input data. This includes the statistics (mean, mean, etc) of the features and the input node degree. Click here to download the stats info.
Code is under a GPL-3.0 license. See the LICENSE file for further details.
Model weights are licensed under Attribution-NonCommercial-ShareAlike 4.0 International. Please consider the implications of using the weights under this license.
@article{graham2022screening,
title={Screening of normal endoscopic large bowel biopsies with artificial intelligence: a retrospective study},
author={Graham, Simon and Minhas, Fayyaz and Bilal, Mohsin and Ali, Mahmoud and Tsang, Yee Wah and Eastwood, Mark and Wahab, Noorul and Jahanifar, Mostafa and Hero, Emily and Dodd, Katherine and others},
journal={medRxiv},
year={2022},
publisher={Cold Spring Harbor Laboratory Press}
}