Benchmark of VAE models for Unsupervised Anomaly Detection
This repository aims to give all the steps and scripts to reproduce experiments in link to article.
For this work, we used the FDG PET available in the ADNI dataset. Data can be downloaded from the IDA plateform. We select the FDG PET that are co-registered, averaged, with standardized image and voxel size and with uniform resolution. Data are stored in the DATASET_DIRECTORY and the CSV files with the clinical data are stored in CLINICAL_DATA_DIRECTORY.
The data can then be converted to BIDS using Clinica software using the adni-to-bids pipeline. The BIDS_DIRECTORY is the output directory of the pipeline:
clinica convert adni-to-bids DATASET_DIRECTORY CLINICAL_DATA_DIRECTORY BIDS_DIRECTORY
PET images are then processed using the Clinica PET Linear pipeline. We use the cerebellum and pons as reference regions for the SUVR:
clinica run pet-linear BIDS_DIRECTORY CAPS_DIRECTORY 18FFDG cerebellumPons2
with CAPS_DIRECTORY being the output CAPS directory of the pipeline.
Once the data are preprocessed, we used ClinicaDL to extract Pytorch 3D tensors from Nifti files:
clinicadl extract CAPS_DIRECTORY image
As the registration algorithm is quite simple, we use the QC method in ClinicaDL to filter the images that are not well registered. QC T1 QC PET
We then use the get-labels command to select the cognitively normal subject with stable diagnosis and split function to generate our train, test and validation sets.
Install ClinicaDL with the correct branch. Install Pythae.
Launch the script.
Simulate hypometabolism using clinicadl
launch evaluation script