CTFSH

This repository contains the code for the paper CTFSH: Full Head CT Anomaly Detection with Unsupervised Learning.

Requirements

All the necessary package requirements are listed in file requirements.txt and can be installed by running pip install -r requirements.txt:

matplotlib==3.3.4  
monai==0.7.0  
nibabel==3.2.1  
numpy==1.19.5  
pandas==1.1.5  
pytorch_lightning==1.5.2  
pytorch_msssim==0.2.1  
scikit_image==0.17.2  
scikit_learn==1.0.2
scipy==1.5.4
seaborn==0.11.2
torch==1.11.0.dev20211118+cu102
torchvision==0.12.0.dev20211119+cu102

Dataset

The CQ500 dataset can be downloaded from here. It should then be processed accordingly and divided with the following structure (in case you wanted to use the provided dataset class):

dataset
|
|--- TRAIN
|    |--- head_volumes
|    |--- stripped_volumes
|
|--- VAL 
|    |--- head_volumes
|    |--- stripped_volumes
|
|--- TEST
     |
     |--- ATY
     |    |--- head_volumes
     |    |--- stripped_volumes
     |
     |--- FRAC      
     |    |--- head_volumes
     |    |--- stripped_volumes
     |        
    ...

Running!

In order to run training, you should run the following command with the chosen parameters (here we chose toy parameters as example):

python src/train_lightning.py --gpu 0 --architecture AE --dataset_dir ./data --stripped False --output_dir ./runs

If you then would like to run inference/testing, you should memorize the path to your checkpoint and pass it to the following command:

python src/test_lightning.py --gpu 0 --architecture AE --checkpoint ./runs/AE-stripped/checkpoints/last.ckpt --stripped True --dataset_dir ./data

Checking the results

All the testing data will be saved in folder ./<OUTPUT_DIR>/<ARCHITECTURE_NAME>/testing. You can run Tensorboard to visualize losses, reconstructions, residuals, PR curves, confusion matrices and violin plots:

tensorboard --logdir ./runs/AE-stripped/testing/

Metrics are also saved inside this folder in Excel file format. Good luck!