iolag / upd_study

This repository contains code from our comparative study on state of the art unsupervised pathology detection and segmentation methods.

Official Repository of: "Unsupervised Pathology Detection: A Deep Dive Into the State of the Art"

This repository contains code to reproduce experiments from the paper "Unsupervised Pathology Detection: A Deep Dive Into the State of the Art" (arXiv preprint).

In this work, we perform a comprehensive evaluation of the state of the art in unsupervised pathology detection. We find that recent feature-modeling methods achieve increased performance compared to previous work and are capable of benefiting from recently developed self-supervised pre-training algorithms, further increasing their performance.

Citation

If you find our work useful for your research, please consider citing:

@ARTICLE{10197302,
  author={Lagogiannis, Ioannis and Meissen, Felix and Kaissis, Georgios and Rueckert, Daniel},
  journal={IEEE Transactions on Medical Imaging}, 
  title={Unsupervised Pathology Detection: A Deep Dive Into the State of the Art}, 
  year={2024},
  volume={43},
  number={1},
  pages={241-252},
  doi={10.1109/TMI.2023.3298093}}

Usage

Download this repository by running

git clone https://github.com/iolag/UPD_study/

in your terminal.

Environment

Create and activate the Anaconda environment:

conda env create -f environment.yml
conda activate upd

Additionally, you need to install the repository as a package:

python3 -m pip install --editable .

To be able to use Weights & Biases for logging follow the instructions at https://docs.wandb.ai/quickstart.

Data

CheXpert

To download CheXpert you must first register at https://stanfordmlgroup.github.io/competitions/chexpert/. After you receive the subscription confirmation e-mail, download the downsampled version (11G) and extract the CheXpert-v1.0-small folder in data/datasets/CXR. No other steps are required and all splits are provided.

Update 12/12/23

It seems that the small version of the dataset isn't available from the official source anymore. You can find it in Kaggle.

DDR

To download and prepare the DDR dataset, run:

bash UPD_study/data/data_preprocessing/prepare_DDR.sh

MRI: CamCAN, ATLAS, BraTS

To download and preprocess ATLAS and BraTS, first download ROBEX from https://www.nitrc.org/projects/robex and extract it in data/data_preprocessing/ROBEX. Then run:

bash UPD_study/data/data_preprocessing/prepare_ATLAS.sh
bash UPD_study/data/data_preprocessing/prepare_BraTS.sh

For ATLAS you need to apply for the data at https://fcon_1000.projects.nitrc.org/indi/retro/atlas.html and receive the encryption password. During the run of prepare_ATLAS.sh you will be prompted to input the password.

For BraTS, Kaggle's API will be used to download the data. To be able to interact with the API, follow the instructions at https://www.kaggle.com/docs/api.

To download the CamCAN data, you need to apply for it at https://camcan-archive.mrc-cbu.cam.ac.uk/dataaccess/index.php. After you download them, put them in data/datasets/MRI/CamCAN and run:

python UPD_study/data/data_preprocessing/prepare_data.py --dataset CamCAN

Experiments

To generate the "Main Results" from Tables 1 and 3 over all three seeds run:

bash UPD_study/experiments/main.sh 

Alternatively, for a single seed run:

bash UPD_study/experiments/main_seed10.sh 

To generate the "Self-Supervised Pre-training" results from Tables 2 and 4 over all three seeds run:

bash UPD_study/experiments/pretrained.sh

Alternatively, for a single seed run:

bash UPD_study/experiments/pretrained_seed10.sh      

To generate the "Complexity Analysis" results from Table 5 run:

bash UPD_study/experiments/benchmarks.sh

To generate "The Effects of Limited Training Data" results from Fig. 3 run:

bash UPD_study/experiments/percentage.sh

The repository contains PyTorch implementations for VAE, r-VAE, f-AnoGAN, H-TAE-S, FAE, PaDiM, CFLOW-AD, RD, ExpVAE, AMCons, PII, DAE and CutPaste.

If you face any issues or have any suggestions do not hesitate to contact me or open an issue.

upd_study's Issues

Questions about computing dice

Thanks for your great work. I have a question about computing dice in this code.

In utilies/metrics.py, function compute_best_dice() concatenates all the testing samples together to compute the dice score, instead of computing the dice of each sample one by one.

I am wondering whether this operation is reasonable as the dice score is usually caculated for each case seperately. Thanks.

Path to dataset

Hi,

This is a great work! Could you please tell me the direct path to the cheXPert dataset? I cannot find the downsampled one in the size about 11G, but only the fullsize one (400G+) here https://stanfordaimi.azurewebsites.net/datasets/8cbd9ed4-2eb9-4565-affc-111cf4f7ebe2.

Thank you very much for your help!

Best,
Cathy

Performance on the MVTec dataset

Hello,

Have you evaluated these methods using the MVTec dataset? I attempted to use your code on the MVTec dataset, but the outcomes were significantly inferior compared to those documented in the original paper. Specifically, the AUC for most methods was about 0.5-0.8 after 2000 training steps, except the method RD and DFR.

I used the following command. I changed the small testing set in your code to the complete testing set and observed the anom_val/sample-auroc in the wandb.
UPD_study/models/H-TAE-S/HTAEStrainer.py --shuffle t -ev t -mod MyData --percentage 100 --eval False --seed 10 --val_frequency 200 --log_frequency 200 --anom_val_frequency 200

Is there any problem with my usage? Look forward to your reply.

Thank you in advance!