A Method for Efficient De-identification of DICOM Metadata and Burned-in Pixel Text

Area of Science:

• Medical Imaging
• Data Security
• Radiology Informatics

Background:

• De-identification of DICOM images is crucial for medical research.
• Existing methods for removing protected health information (PHI) from DICOM metadata are established, but removing PHI
• burned-in
• to pixel data is often manual and lacks validated high-throughput automation.
• Optical character recognition (OCR) models can detect PHI in images but are computationally intensive for large datasets.

Purpose of the Study:

• To develop and validate an efficient, high-throughput method for de-identifying DICOM images, addressing the challenge of
• burned-in
• PHI.
• To combine automated metadata de-identification with a targeted OCR approach for images likely to contain PHI in pixel data.

Main Methods:

• A novel data processing method was implemented, performing metadata de-identification on all images.
• A targeted strategy was employed to apply OCR exclusively to images identified as having a high probability of containing burned-in PHI.
• The method was validated on a large dataset comprising 415,182 DICOM images across ten modalities.

Main Results:

• The validated method demonstrated high efficacy in de-identifying DICOM images.
• Out of 12,578 images with any burned-in text, only 10 instances were missed by the method.
• OCR was applied to only 6,050 images (1.5% of the total dataset), significantly improving processing efficiency.

Conclusions:

• The developed method provides an efficient and effective solution for de-identifying large volumes of medical images.
• This approach significantly reduces the computational burden associated with OCR for burned-in PHI in medical research.
• The validated method enhances the security and usability of medical image datasets for research purposes.