Abstract A053: Development and multicentric validation of a publicly-available head CT deidentification tool to facilitate inter-institutional stroke image sharing

Abstract

Purpose: Head CTs of stroke patients contain personal health information (PHI) in Digital Imaging and Communications in Medicine (DICOM) metadata and facial features in 3D reconstruction, raising privacy concerns and limiting multi-institutional data sharing. We developed, validated, and publicly shared a user-friendly tool for batch deidentification of head CTs to facilitate inter-institutional stroke image sharing. Methods: We created a head CT de-identification plug-in for 3D Slicer freeware. The tool batch-deidentifies head CTs by: (1) replacing folder names with new IDs; (2) removing 101 potentially identifying DICOM metadata elements; (3) detecting and redacting “burnt-in” text within images using optical character recognition (OCR); (4) stripping face/scalp tissue to blur 3D reconstructions and prevent facial recognition; and (5) purging dose report and reconstruction series, and retaining only primary axial acquisitions to reduce file size and further minimize the risk of retaining identifiable data. Feasibility was tested at three stroke centers and validated on a multicenter trial dataset. Results: The batch deidentification tool is available as the “SlicerHeadCTDeID” plug-in through Slicer Extension Manager, deployable in Windows, Max, or Linux OS (Fig 1), with source code available on GitHub. The inputs for this tool are a folder containing DICOM files, an Excel/CSV file mapping original folder names to new IDs, and a destination folder (Fig 2). The tool was tested by multiple users at three centers and validated in the Antihypertensive Treatment of Acute Cerebral Hemorrhage-II (ATACH-II) trial dataset. In total, 1,501 head CTs from 121 unique scanners were deidentified, with retention of axial image series only, and 35% reduction in average file size (52.59 MB to 34.20 MB). An average of 28.4 mL of superficial facial/scalp tissue was removed for facial deidentification (Fig 3). The automated process took 2–5 minutes per head CT, depending on local computer performance. Conclusion: We developed and validated a publicly available, user-friendly tool for automated batch deidentification of head CTs. This tool removes PHIs from DICOM metadata, strips superficial facial/scalp tissue to prevent 3D facial recognition, redacts burnt-in alphanumeric text on images, and purges non-axial head CT series to reduce storage memory requirement, facilitating research and AI model building that relies on secure, efficient multi-institutional stroke image sharing.

Ähnliche Arbeiten