De-identifying medical images: why removing the DICOM header isn't enough

Ask most teams how they de-identify a scan and they'll describe one step: strip the patient's name and ID from the DICOM header. It's the right first step. It's also where many datasets stop, and where identifying information gets through.

A medical image carries identifying information in at least four places, and each needs its own method.

1. The DICOM header

Every DICOM file has a header of tagged fields: patient name, patient ID, birth date, accession number, study dates, the hospital, the referring doctor, the scanner's serial number and more. Manufacturers also add private tags, whose contents aren't standardised and can hold anything, including names.

The DICOM standard sets out what to remove. Part 15, Annex E defines a Basic Application Level Confidentiality Profile: a table of which attributes to remove, blank or replace, plus options for when some information must be kept or cleaned more thoroughly. Applying a recognised profile, and recording which options were used, is far better than a home-grown list of tags.

Two details matter for research:

• Dates. Removing every date destroys the timeline between scans. A common approach is to shift all of a patient's dates by the same random offset, so intervals survive but real dates don't. Annex E has an option for keeping longitudinal temporal information in a modified form.
• Identifiers that link studies. Patient and study identifiers should be replaced consistently, so one patient's studies still link to each other without linking back to the hospital's records.

2. The pixels

This is the step most often missed. On many images, identifying text is part of the picture itself: burned into the pixels as a label in a corner of the screen. Cleaning the header doesn't touch it.

It's common in:

• ultrasound and echocardiography, where the machine prints patient details and the date on the screen it captures
• secondary captures and screenshots, such as 3D reconstructions saved as images
• scanned paper documents stored as DICOM
• some radiography and fluoroscopy, depending on the device and the site's settings

DICOM has a Burned In Annotation attribute meant to flag these images, but it isn't reliably filled in, so it can't be trusted on its own. Practical approaches combine masking the known text regions for each device model, text detection (OCR) to find what's left, and a human review of a sample before release. Annex E's Clean Pixel Data option covers this step.

3. The face

A head CT or MRI can be rendered in 3D into a recognisable face. For datasets that include head imaging, defacing (removing facial features from the image volume) is often needed. Annex E calls this the Clean Recognizable Visual Features option, and HIPAA's Safe Harbor list includes full-face photographs and comparable images.

4. The report

Radiology and cardiology reports are free text. They can name the patient, the referring doctor, the hospital, a date, a relative or a place, and header tools don't touch them. Reports need their own de-identification pass, usually software that finds names, dates and numbers followed by human checks, with dates shifted the same way as in the images so the report still matches the scan.

Don't forget the files themselves: folder and file names often contain patient IDs or accession numbers.

What the law expects

In the United States, HIPAA offers two routes to de-identified data. Safe Harbor removes 18 types of identifiers; Expert Determination has a qualified expert confirm that the risk of re-identification is very small. Either way, the standard applies to everything in the dataset: headers, pixels and reports.

In India, the Digital Personal Data Protection Act 2023 governs personal data, and what counts as adequately anonymised data is still being settled. Other countries have their own rules. In every case, the contract should state which method was used and who checked it.

Questions to ask a data supplier

1. Which de-identification profile did you apply to the DICOM headers, and with which options?
2. How do you find and remove text burned into the pixels: by device, by OCR, by human review?
3. Are head studies defaced?
4. How are reports de-identified, and are dates shifted consistently with the images?
5. Was the result checked by a person or an independent expert, and on what sample?
6. Which legal route applies (for example HIPAA Safe Harbor or Expert Determination), and can you show the documentation?

If a supplier can answer the first question but not the second, the dataset probably still contains names.

We ask these questions of every data source before we work with it. If you need de-identified imaging data with reports for research or AI development, see medical imaging datasets or tell us what you need.