Evaluation of pseudo-healthy image reconstruction for anomaly detection in brain FDG PET

Overview

This repository provides the source code and data necessary to run the evaluation of a deep generative model trained to reconstruct pseudo-healthy images for unsupervised anomaly detection on FDG PET.

The model has been trained using the ClinicaDL open source software, and using the ADNI dataset.

The method is described in the following article Evaluation of pseudo-healthy image reconstruction for anomaly detection with deep generative models: Application to brain FDG PET (accepted).

If you use any ressources from this repository, please cite us:

@article{melba:2024:003:hassanaly,
    title = "Evaluation of pseudo-healthy image reconstruction for anomaly detection with deep generative models: Application to brain FDG PET",
    author = "Hassanaly, Ravi and Brianceau, Camille and Solal, Maëlys and Colliot, Olivier and Burgos, Ninon",
    journal = "Machine Learning for Biomedical Imaging",
    volume = "2",
    issue = "Special Issue for Generative Models",
    year = "2024",
    pages = "611--656",
    issn = "2766-905X",
    doi = "https://doi.org/10.59275/j.melba.2024-b87a",
    url = "https://melba-journal.org/2024:003"
}

Requirement

Environment

To come

Data

In order to improve reproducibility, the ADNI dataset have been converted to BIDS and preprocessed using Clinica open source software.

Once we obtain a BIDS of ADNI using the adni-to-bids command, we run the pet-linear command on the dataset:

clinica run pet-linear $BIDS_DIRECTORY $CAPS_DIRECTORY 18FFDG cerebellumPons2

The outputs are stored in a CAPS dataset at $CAPS_DIRECTORY.

All the details on the data selection are in the Appendix of the article.

Model training

The model has been trained using ClinicaDL and Pythae using the following command.

clinicadl train pythae $CAPS_DIRECTORY $PREPROCESSING_JSON $TSV_DIR $MAPS_DIRECTORY -c $CONFIG_FILE

See ClinicaDL documentation for more information on how to run this command. A script to run this command on a cluster is available, run it using the following command:

sbatch clinicadl_scripts/run_train.sh

after specifying in the file the different path to required files and folders.

A config file with the used parameters are available at in config/config_vae.toml

Outputs are stored in a MAPS directory at $MAPS_DIRECTORY.

Hypometabolism simulation

We then need to build the different simulated dataset used for our evaluation. For that we run following ClinicaDL command with the different parameters of $PATHOLOGY and $PERCENTAGE:

clinicadl generate hypometabolic $CAPS_DIRECTORY $GENERATED_CAPS_DIRECTORY --pathology $PATHOLOGY --anomaly_degree $PERCENTAGE

All the different test set build for the experiments are detailed in the article.

Run inference on the trained model

Finally, we need to store the latent vectors and model reconstruction in order to compute the different metrics and plot the graphs that we use for ou analysis. A script that compute this can be run through the following command:

sbatch clinicadl_scripts/run_predict.sh $MAPS_DIRECTORY $CAPS_ROOT_DIRECTORY $TSV_DIRECTORY

with $CAPS_ROOT_DIRECTORY the root directory where all the simulated test sets are stored, and $TSV_DIRECTORY the folder that stored the tsv files with the splits.

How to use

Reconstruction

python src/reconstruction_figure.py ${MAPS_DIRECTORY} --split ${SPLIT}

Comparison of simulation framework

python src/violin_plot.py ${MAPS_DIRECTORY} --split ${SPLIT}

Healthiness of reconstructed images

python src/healthiness.py ${MAPS_DIRECTORY} --split ${SPLIT}

Anomaly score

python src/anomaly_score.py ${MAPS_DIRECTORY} --split ${SPLIT}

Latent Space Analysis

python src/latent_space.py

python src/closest_participant.py

Linear mixed effect model

Some scripts needs to be run preliminary. To come.

python src/lmm.py

Run all analysis

python src/main.py  ${MAPS_DIRECTORY} --split ${SPLIT}