The tbpcxr Python module provides tools to detect outliers of chest x-ray (CXR) images. Machine learning or artificial intelligence (AI) methods are used to train a model representing "normal" CXR images, and then new images are classified as near the (normal) model, or significantly different from the model and therefore an "outlier".
The following steps are used in the algorithm to detect outliers:
- Normalize the image into a standard size and intensity range.
- Register the image to a CXR atlas.
- Transform the image into an approximations with a precomputed principle component basis.
- Predict if the image is an outlier determined by how well the image fits to the model.
The published Sphinx documentation is available here: https://niaid.github.io/tbp-cxr-outlier/
The master built Sphinx documentation is available for download from Github Actions under the build as "sphinx-docs".
The Python module is distributed as a wheel binary package which defines the package dependencies.
The package rap_sitkcore is an optional dependency used for reading DICOM files. It provides additional robust reading for difficult TBPortals DICOM files which may be addressed in some published datasets.
Option 1: Manually Download Wheels from Github Releases
The wheels can be downloaded from the Github Releases page and the optional dependency package rap_sitkcore can be downloaded separately from the Github rap_sitkcore Releases page. Then the two wheels should be installed together:
python -m pip tbpcxr-0.5.1-py3-none-any.whl rap_sitkcore-0.5.5-py3-none-any.whl
Option 2: Use Internal NIAID Artifactory
It can be specified when installing with the optional dependency package notation tbpcxr[sitkcore]. The rap_sitkcore package is hosted on the NIAID artifactory.
Internally the tbpcxr package is hosted on the NIAID Python Package Index (PyPI) hosted on artifactory and is installable with pip. This enables internal dependencies to be automatically download from the artifactory. The internal repository can be automatically used by setting an environment variable:
PIP_EXTRA_INDEX_URL
Then the tbpcxr package can be installed:
python -m pip install tbpcxr[sitkcore]
Wheels from the master branch can be download wheel from Github Actions in the "python-package" artifact.
Download the latest tagged release from the Github Releases page.
The wheel lists the package dependencies which are required for successful installation. This include internal NIAID packages. If the internal "artifactory" repository is not configured then these additional dependencies will need to be manually downloaded and installed before install tbpcxr. The downloaded wheels can be installed:
python -m pip install tbpcxr-0.5.1-py3-none-any.whl
Please use the GitHub Issues for support and code issues related to the tbp-cxr-outlier project.
Additionally, we can be emailed at: [email protected] Please include "tbp-cxr-outlier" in the subject line.