A growing number of tools and procedures claim to properly de-identify image data. Based on our decade of experience managing the Cancer Imaging Archive (TCIA) on behalf of NCI, we developed a DICOM dataset that can be used to evaluate the performance of de-identification algorithms. DICOM objects were selected from datasets published in TCIA.  Synthetic Protected Health Information (PHI) was generated and inserted into selected DICOM Attributes to mimic typical clinical imaging exams.  The DICOM Standard and TCIA curation audit logs guided the insertion of synthetic PHI into standard and non-standard DICOM data elements. An answer key was created based on our knowledge of the placement of synthetic data and the DICOM standard’s guidelines for what actions should be taken in regard to the synthetic PHI.  A TCIA curation team tested the utility of the evaluation dataset and answer key.

Patient privacy rules require the removal of Protected Health Information (PHI) before sharing images publicly. Manual de-identification is no longer scalable due to the rapid increase in imaging data volume. Our goal was to configure and test the efficacy of a cloud service for automated medical image de-identification (MIDI).  One such service for DICOM images is Google Cloud Platform’s Healthcare API. Training and test datasets for validation of image de-identification, specifically prepared by the placement of synthetic PHI in DICOM headers and image pixel data, were obtained from The Cancer Imaging Archive. The customized MIDI pipeline correctly performed 99.8% of expected actions on DICOM header data elements. For image pixel data, one false-positive case was noted, while all sensitive information was correctly removed from image pixel data. Throughput averaged at 58.4 images per second. This implementation of the MIDI pipeline holds promise for automated de-identification at scale. However, verification by a human expert is still currently recommended.