Medical Image De-Identification Benchmark

A fundamental requirement for sharing medical images, particularly through public image repositories, is the deidentification (deID) of the protected health information (PHI), to comply with patient privacy regulations. However, there are very few guidelines for true compliance and best practices in image deID. Here, compliance refers to adherence to regulations for the protection of patient privacy (e.g., HIPAA regulation in the USA) and DICOM Standard, while best practice refers to the preservation of metadata values that may be important in downstream analysis of data by the research community.

We propose the Medical Image Deidentification Benchmark (MIDI-B), to help guide the assessment and further development of image deID toward the protection of patient privacy. In MIDI-B we use a large (n > 500) collection of clinical, multisite, and multimodality DICOM images, obtained from the Cancer Imaging Archive (TCIA), infused with synthetic protected health information (PHI), or patient identifying information (PII), into the DICOM header and/or the pixel (image) matrix. We invite participants to apply their image deID tools for a rule-based anonymization of medical images. A scoring script will assess performance accuracy of compliance with rules and guidelines for anonymization of PHI/PII, DICOM standard, and best practices, to be described in the Synapse challenge platform at Sage Bionetworks. If accepted, we plan on publication of a challenge report and keeping MIDI-B open to allow benchmarking of deID tools, following MICCAI 2023.

MIDI-B is the first challenge of its kind to help guide and support the community in development of image anonymization tools for data sharing, to further fuel advances in technical and clinical analytics data science, including imaging AI/ML.