Clinical imaging data, often stored as DICOM files, is the fuel needed to train new machine learning models. But before clinical images can be used, any identifying information must be removed. This process is called DICOM de-identification. It is a challenging problem with many custom needs per institution.
We designed, developed, and have since maintained a custom DICOM de-identification pipeline for the institution.
To provide their researchers easier access to clinical images, UAB’s Department of Radiology wanted to streamline its DICOM de-identification process. Their initial process involved manual steps that could be automated—speeding up the process while also avoiding human error.
UAB hired Innolitics to provide a DICOM de-identification solution.
We worked with the Department of Radiology’s vice chair of clinical research to understand UAB’s requirements. They needed a solution that would:
Allow researchers to specify DICOM UID mappings with simple excel files.
Provide a database that allows DICOM files to be re-identified.
Export de-identified files to a research PACS or the filesystem.
Communicate with the Philips iSite PACS.
Throttle requests to the clinical PACS.
Support multiple simultaneous research projects.
Support scheduling de-identification tasks during off-hours.
Not require outside network access.
Be straightforward for IT to install (we used Docker Images).