cOOpD

Reformulating COPD classification on chest CT scans as anomaly detection using contrastive representations

Official Code Repository Going alongside the Paper

Please cite the following work if you find this model useful for your research:

Silvia D. Almeida, Carsten T. Lüth, Tobias Norajitra, Tassilo Wald, Marco Nolden, Paul F. Jaeger, Claus P. Heussel, Jürgen Biederer, Oliver Weinheimer, Klaus Maier-Hein (2023). 
cOOpD: Reformulating COPD classification on chest CT scans as anomaly detection using contrastive representations. arXiv preprint  arXiv:2307.07254

cOOpD

General Information

Official PyTorch implementation for paper cOOpD: Reformulating COPD classification on chest CT scans as anomaly detection using contrastive representations , accepted in MICCAI 2023.

Abstract

Classification of heterogeneous diseases is challenging due to their complexity, variability of symptoms and imaging findings. Chronic Obstructive Pulmonary Disease (COPD) is a prime example, being underdiagnosed despite being the third leading cause of death. Its sparse, diffuse and heterogeneous appearance on computed tomography challenges supervised binary classification. We reformulate COPD binary classification as an anomaly detection task, proposing cOOpD: heterogeneous pathological regions are detected as Out-of-Distribution (OOD) from normal homogeneous lung regions. To this end, we learn representations of unlabeled lung regions employing a self-supervised contrastive pretext model, potentially capturing specific characteristics of diseased and healthy unlabeled regions. A generative model then learns the distribution of healthy representations and identifies abnormalities (stemming from COPD) as deviations. Patient-level scores are obtained by aggregating region OOD scores. We show that cOOpD achieves the best performance on two public datasets, with an increase of 8.2% and 7.7% in terms of AUROC compared to the previous supervised state-of-the-art. Additionally, cOOpD yields well-interpretable spatial anomaly maps and patient-level scores which we show to be of additional value in identifying individuals in the early stage of progression. Experiments in artificially designed real-world prevalence settings further support that anomaly detection is a powerful way of tackling COPD classification.

Technologies Used

Python - version 3.7+
See environment.yml
FrEIA Framework

Features

Contrastive Pretext Training on Medical Data

Setup

Set this up with conda:

$ conda env create -f environment.yml

Pre-processing

Follow the instructions here

Steps

Set all the paths to the data and logs in:

config/global_config
config/datasets/lung.py
create own datamodule carrying your data
- see --> datamodules/lung_module.py \& datamodules/lung.py

Verify that everything is working

$ python train_pretext.py --fast_dev_run True

See in the logs folder, whether a log has been created, access with tensorboard via:

$ tensorboard --path {path_to_logs}

Fitting of the generative models

$ python train_latent_gen.py -p {path_to_experiment}

Evaluation of the generative models

$ python eval.py -p {path_to_experiment}

Running with new datasets

To run experiments with new datasets:

Add a new datamodule
Add a new option for get_data_loaders(name_exp) in eval.py

Changing the generative models on the latent space

To do this, add a new entry to the dictionary in the file config/latent_model.py.

Usage

Anomaly Detection with the cOOpD framework on 3D data.

Project Status

Project is in progress

Acknowledgements

CRADL: Contrastive Representations for Unsupervised Anomaly Detection and Localization [link]

MIC-DKFZ / cOOpD

readme