If you use this implementation, please cite paper:
@article{marimont2021implicit,
title={Implicit field learning for unsupervised anomaly detection in medical images},
author={Naval Marimont, Sergio and Tarroni, Giacomo},
journal={arXiv preprint arXiv:2106.05214},
year={2021}
}
Given that no augmentations were used, we extracted and stored the images already pre-processed. Pre-processing includes padding, clip intensities, normalize, downsample (optional), discretize using pixel intensity clusters, mode-pooling filter. Script provided below does this for HCP and BraTS'18 datasets:
python extract_and_preproc_brains.py --dataset <HCP or BraTS> --source <root directory of Brats18 / HCP with the original nii.gz> --target <directory to store pre-processed dataset> --res <160 or full> --cluster_centers <pix_cluster_centers_brain.cp (file with the k-Means cluster centers)> --mode_pool <2 or 3 or not specified if not used>
Pre-processed files are pickle files with img_extended = namedtuple( 'img_extended', ('img', 'seg', 'cid') ), containing the pre-processed Image, Segmentation (None in HCP dataset) and the identifier of the case (original file name).
The network architecture and implementation is based on DeepSDF repository. To train an experiment run the following:
python train_network.py --experiment <experiment directory> --continue <latest or epoch number>
Each experiment is organized in an "experiment directory", which collects all of the data relevant to a particular experiment. The structure is as follows:
<experiment_name>/
specs.json
Logs.pth
LatentCodes/
<Epoch>.pth
ModelParameters/
<Epoch>.pth
OptimizerParameters/
<Epoch>.pth
The only file that is required to begin an experiment is 'specs.json', which sets the parameters, network architecture, and data to be used for the experiment. Examples of spec.json are provided.
Note that the first time that a data loader is instantiated on the training set, a file called "dict_cid.k" is created in the training set directory. This file contains a dictionary assigning a unique sequential Integer to each case id.
Evaluation scripts are implemented in Implicit Fields Evaluation jupyter notebook
Sergio Naval Marimont, Giacomo Tarroni
Implicit field learning for unsupervised anomaly detection in medical images is relased under the MIT License. We use a substantial portion of DeepSDF which is relased under the MIT License. See the LICENSE file for more details.