NLL_anomaly_detection

A simple anomaly detection algorithm for medical imaging based on multi-atlas image registration and negative log likelihood (NLL).

1. Paper

MICCAI-2021: Improved Brain Lesion Segmentation with Anatomical Priors from Healthy Subjects.

Chenghao Liu, Xiangzhu Zeng, Kongming Liang, Yizhou Yu, Chuyang Ye

2. Prerequisites

You need to install the following softwares for running the code.

ANTs-2.3 (older versions should also work) and FSL-6.0

If you already have ANTs and FSL installed before, just skip the following steps.

2.1. ANTs (Advanced Normalization Tools)

Link: https://github.com/ANTsX/ANTs

How to install: compile from source code (recommended) or pre-built binaries

Verify your install: to see whether ANTs is installed correctly on your system, after the installation you need to type in

antsRegistration --version

and

N4BiasFieldCorrection --version

in your console. It should produce output such as:

ANTs Version: 3.0.0.0.dev13-ga16cc
Compiled: Jan 22 2019 00:23:29

Then test if antsApplyTransforms can work:

antsApplyTransforms

if no error shows, then ANTs is successfully installed. We need to use antsRegistration to reigster medical images, antsApplyTransforms to apply transforms to images, and N4BiasFieldCorrection to correct image intensity bias.

2.2. FSL (FMRIB Software Library)

Link: https://fsl.fmrib.ox.ac.uk/fsl/fslwiki

How to install: FSL is installed using the fsl_installer.py downloaded from here. You need to register your personal information to the FSL site before download. After download you need to run the installer script and wait for the installation to finish.

Verify your install: when the installation finished, type in

bet -h

in your console. If no error occurs then everything is OK! :). We use bet (brain extraction tool) to calculate brain mask for each image.

3. Install this script

After the installation of ANTs and FSL, you need to do the following steps to install this script:

1. Create a new Python3 venv (link).

2. Activate the newly created virtual environment, download and unzip the codes, then cd into the directory where setup.py is located, and type:

pip install -U setuptools
pip install -e .

4. Prepare your data

In this example we used two datasets, HCP (Human Connectome Project) and ATLAS (Anatomical Tracings of Lesions After Stroke). They are all publicly avaliable. HCP dataset provides healthy MRI T1-weighted scans and ATLAS dataset provides T1-weighted scans of chronic stroke patients. Here is a screenshot of the downloaded images, you need to check if the appearances of the downloaded images are similar with the following example:

Note that the image appearances from HCP may vary in different releases. In this demo the HCP image release is not bias field corrected (you can see the center brain darker than the outside), which is a more generalized case. If you use your own dataset instad of HCP, and the images are bias field corrected, you can skip "N4 correction". For more details, please refer to "Sect. 5: Quick start".

5. Quick start

Once you installed the scripts, you can use the following command to calculate anomaly scores for a target image:

NLL_anomaly_detection \
--source-images /path/to/source/image1.nii.gz /path/to/source/image2.nii.gz /path/to/source/image3.nii.gz ... /path/to/source/image10.nii.gz \
--target-image /path/to/target/image.nii.gz \
--output-dir /path/to/output/dir/

In the above example, "--source-images" tells the script how many reference images there are and their locations (at least two images are needed), "--target-image" specifies the image we want to analyze. All the outputs will be saved to a directory, which is determined by "--output-dir" parameter.

5.1. Skipping bias field correction

More advanced options are available. For example, if the datasets you used are not affected by intensity bias fields, then you can skip intensity bias field correction. This is achieved by passing an additional "--do-n4-correction none" to the command line. The default setting for this parameter is "--do-n4-correction source", which means the source dataset (HCP dataset in this example) needs to do intensity bias field correction, while the target dataset (ATLAS dataset in this example) do not need to. The available options are: "none", "both", "source" and "target".

Implementation details will be provided in "Sect. 6".

6. Implementation details

6.1. Pre-processing

1. N4 bias field correction

...

2. Skull-stripping

...

6.2. Image registration using ANTs

6.3. NLL anomaly score calculation