This repository contains the code for the preprocessing pipeline used in the Deep Unsupervised Anomaly Detection in Brain Imaging: Large-Scale Benchmarking and Bias Analysis publication. Please cite the mentioned publication when using this repository for your research.

The pipeline implements preprocessing for structural MRI and will generate skullstripped images that are coregistred (between individual modalities) and rigidly registered to the SRI24 atlas. The skullstripping follows the UKB pipeline. Additionally custom QC is implemented that follows fsqc. Note that for this step Freesurfer aseg files are needed that have to be located in the /anat directory and transformed to the original/individual coordinate system and the .nii.gz file format for every individual. Anomaly masks (binary segmentations indicating anomaly/normal) are automatically handled correctly if they are named sub-ID_seg-lesion_mask.nii.gz.

To use this repository your datasets have to be in the BIDS format:

 ├── Dataset
    │   ├── sub-ID
    │   │   ├── anat
    │   │   │   ├── sub-ID_T1w.nii
    │   │   │   ├── sub-ID_T2w.nii
    │   │   │   ├── aseg.nii
    │   │   │   └── ...

Additionally a .csv file is needed that specifies which individuals to preprocess. The path to the subjects directory is needed.

/home/user/dataset/sub-ID/
/home/user/dataset/sub-ID2/
...

I apologize that this repository is not well documented. This pipeline is setup to run on a HPC cluster that has Python and FSL installed and uses the job scheduler and workload manager Slurm. Before using the pipeline check the init_vars.sh script and set all the needed (SRI24, MNI_mask, FSL) paths appropriately. Additionally you need a T1w image for every subject or you will need to change the behaviour of the pipeline in pipeline_struct.py. The pipeline will NOT handle multiple T1w images! To run the pipeline use:

bash init_vars.sh path_to_csv.csv

This will submit a single job via Slurm for every individual with the start_pipeline.sh script. The pipeline needs roughly 30 minutes to finish one individual. They will be stored in a new directory named dataset_processed that contains a data and logs directory.