Repository for the paper Explainable diagnosis of oral cancer via deep learning and case-based reasoning
An in-browser demo is available here
Here you can find the generated explainable reports for the training and test datasets
| dataset | reports link |
|---|---|
| train | http://131.114.50.176/owncloud/s/BB8SuiBDUolvV9r/download |
| test | http://131.114.50.176/owncloud/s/zYMKHGHAe2NRTcA/download |
You can find the dataset here
git clone https://github.com/galatolofederico/oral-lesions-detection.git
conda create -n oral-lesions-detection python=3.7
conda activate oral-lesions-detection
conda install pytorch torchvision -c pytorch-lts -c nvidia
pip install sklearn shapely opencv-python
Check your CUDA version
nvcc --version
Install the corresponding detectron2 package
for CUDA 11.1
python -m pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu111/torch1.8/index.html
for CUDA 10.2
python -m pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu102/torch1.8/index.html
Download model and fitted PCA
mkdir ./models
wget http://131.114.50.176/owncloud/s/Tk6nl8Mjah2XeLW/download -O ./models/model.pth
wget http://131.114.50.176/owncloud/s/F2gPYU8HO7a4MoD/download -O ./models/pca.pkl
Download dataset
mkdir ./datasets
wget http://131.114.50.176/owncloud/s/zs7J18r37omcMmy/download -O ./datasets/lesions.zip
unzip ./datasets/lesions.zip -d ./datasets/
Download features
mkdir ./assets
wget http://131.114.50.176/owncloud/s/zqXO0hrOq2FVLPO/download -O ./assets/features.zip
unzip ./assets/features.zip -d ./assets
./explain.sh --model ./models/model.pth --pca-model ./models/pca.pkl --features-database ./assets/features/train-features.json --dataset-folder ./datasets/lesions --file datasets/lesions/images/a0b5f450f8603b13990805df60af640e.jpg
python predict-file.py --model ./models/model.pth --file datasets/lesions/images/a0b5f450f8603b13990805df60af640e.jpg
python predict-dataset.py --dataset-folder ./datasets/lesions --model ./models/model.pth
Simplify the dataset dataset into neoplastic, aphthous and traumatic
python -m scripts.simplify-dataset --folder ./datasets/lesions
Split dataset into train and test
python -m scripts.dataset-split --folder ./datasets/lesions
Run train
python train.py --dataset-folder ./datasets/lesions
To see the list of the hyperparameters run
python train.py --help
Simplify the dataset into 3 classes (neoplastic, aphthous, traumatic)
python -m scripts.simplify-dataset --folder ./datasets/lesions
Split the dataset into train and test
python -m scripts.dataset-split --folder ./datasets/lesions
View a dataset
python -m scripts.view-dataset --dataset-folder ./datasets/lesions
View an augmented dataset
python -m scripts.view-augmentation --dataset-folder ./datasets/lesions
Extract features for t-SNE plot
python -m scripts.extract-tsne-features --dataset-folder ./datasets/lesions/ --dataset ./datasets/lesions/dataset.json --model ./models/model.pth --output ./assets/tsne-features.pkl
Plot the t-SNE map
python -m scripts.tsne --features ./assets/tsne-features.pkl
Extract features from all the dataset images (used in explain)
python -m scripts.build-features-database --dataset-folder ./datasets/lesions --model ./models/model.pth --output ./assets/extracted-features
Fit the PCA used in explain
python -m scripts.fit-pca --features-database assets/features/features.json --output models/fitted_pca.pkl
Set the environment variable OPTUNA_STORAGE to your preferred optuna storage (available storages here)
export OPTUNA_STORAGE=sqlite:///optuna.sqlite
Install optuna
pip install optuna
Create study
python create-study.py --study-name <your_study_name>
Run trial
python run-trial.py --study-name <your_study_name>
The code is released as Free Software under the GNU/GPLv3 license. Copying, adapting and republishing it is not only allowed but also encouraged.
For any further question feel free to reach me at federico.galatolo@ing.unipi.it or on Telegram @galatolo