How to de-identify medical images: HIPAA methods, then DICOM step by step

The HIPAA de-identification guidance HHS, 3 Feb 2025 (opens in a new tab) describes two methods, summarized in the table below. Safe Harbor removes the listed identifiers of the patient and of relatives, employers and household members, and the organization must have no actual knowledge that what remains could identify the person. Expert Determination relies on a person with appropriate statistical and scientific knowledge, who determines that the risk of identification by an anticipated recipient is very small and documents the methods and results.

Safe Harbor's list HHS, 3 Feb 2025 (opens in a new tab) includes names, geographic subdivisions smaller than a state, all elements of dates except the year, phone and fax numbers, email and web addresses, record, account, license and device numbers, biometric identifiers, full-face photographs and comparable images, and any other unique identifying number, characteristic or code. Image headers carry several of these, and so can the pixels. Full dates are outside Safe Harbor, so an archive that must keep them needs an Expert Determination.

Outside the United States, local law decides what counts as anonymous. Under the GDPR, health data is a special category of personal data Official Journal of the European Union, 4 May 2016 (opens in a new tab), so the clinic's counsel settles the legal basis before any file is shared.

General information, not legal advice.

The DICOM standard sets out the attribute rules in Annex E of PS3.15, the source of the actions in the table below NEMA, 27 Oct 2026 (opens in a new tab). Its basic confidentiality profile gives each attribute that can identify someone one action: remove it, replace it with a blank or dummy value, replace it with a non-identifying value, keep it, clean it, or replace a UID with a new one that stays consistent across the set. The profile also has the file say what was done: Patient Identity Removed is set to YES, and the method is recorded.

Options change the basic profile in both directions NEMA, 27 Oct 2026 (opens in a new tab). Clean Pixel Data, Clean Recognizable Visual Features and Clean Graphics remove more. Clean Structured Content and Clean Descriptors clean text instead of removing it. Retain options keep what the basic profile would remove: longitudinal temporal information, patient characteristics, device identity, institution identity, UIDs and safe private attributes. Record which options you used and why, with the tool and its version.

Annex E also names two risks NEMA, 27 Oct 2026 (opens in a new tab). A plain hash of a patient ID is weak, because anyone can hash a list of candidate IDs and compare. And anyone who already holds the original images could match the pixel data even after the UIDs are replaced, though the standard says the probability and impact of that may be judged negligible. Keep the key that links new IDs to old ones, and the originals, inside the clinic.

These free tools apply the Annex E actions on a workstation or server the clinic controls, so no file that still identifies a patient has to leave the clinic network.

The RSNA DICOM Anonymizer also has optional AI features that remove burned-in text and blur faces. They run on the workstation after a one-time model download, and images are not uploaded for processing Radiological Society of North America, 27 Oct 2026 (opens in a new tab).

Whatever the tool, test its output against the profile and options you chose: a tool's description of itself is not proof. Pages that carry no publication date, such as the current DICOM edition and the tool manuals, are dated here by the day they were read.

Cleaning the header does not touch the image. Annex E notes that CT images do not normally carry burned-in annotation while ultrasound images routinely do, and that the stored pixel values must be changed: an overlay or a shutter on top is not enough NEMA, 27 Oct 2026 (opens in a new tab). After cleaning, Burned In Annotation is set to NO.

Check every image type, not only the main series: secondary captures, scanned forms, screenshots and dose summary screens. The MIDI Task Group's best practices ask for every image type to be checked for burned-in text, not a sample of types Journal of Imaging Informatics in Medicine, 12 Jul 2024 (opens in a new tab).

High-resolution CT or MRI of the whole head and neck can be shown as a volume or surface rendering, and Annex E notes that such a rendering may be enough to identify a person under some circumstances. Its Clean Recognizable Visual Features option removes or distorts those features; it may need a person to approve the result and can make the images unusable NEMA, 27 Oct 2026 (opens in a new tab). The MIDI best practices ask for the face risk of head and neck images to be measured, and above a threshold for facial features to be removed or the images kept from public release Journal of Imaging Informatics in Medicine, 12 Jul 2024 (opens in a new tab).

Structured reports can hold identifying text in their content, such as the name of an observer. Annex E says a de-identifier without the Clean Structured Content option creates significant risk for them NEMA, 27 Oct 2026 (opens in a new tab). Remove the structured reports you do not need, or clean them with a tool that supports that option.

Radiology reports are free text, so names, dates and places can appear anywhere in them. Treat them as a separate job: drop the header block, replace names using the clinic's own lists of patients, relatives and staff, then review the result by hand.

Pilot steps

1. 01Pick a pilot batch that covers every scanner model, series type and object type in the archive, secondary captures and structured reports included.
2. 02Run the tool with the profile and options you chose, on a workstation inside the clinic.
3. 03Compare the headers before and after for every attribute the profile lists, and look at every image type for burned-in text.
4. 04Have a second person review the output. The MIDI best practices call for human quality control, with the share inspected set by a documented risk assessment, and describe a public archive that reviews every image during curation Journal of Imaging Informatics in Medicine, 12 Jul 2024 (opens in a new tab).
5. 05Keep the key that links new IDs to old ones, the date offsets and the originals inside the clinic, under its own access controls.
6. 06Record the method, the options, the tool and its version, the pilot results and the reviewer, and give that record to the expert or counsel who signs off.

De-identification decides what survives, and what survives decides what the archive can be used for. These choices keep its research value:

• Consistent pseudonyms: one replacement ID per patient across every study, so follow-up studies stay linked Radiological Society of North America, 27 Oct 2026 (opens in a new tab).
• Time between studies: the Retain Longitudinal Temporal Information options keep full dates, or dates modified so the time relationships between studies are kept as far as the use requires NEMA, 27 Oct 2026 (opens in a new tab).
• Patient characteristics: the Retain Patient Characteristics option keeps details such as age and sex NEMA, 27 Oct 2026 (opens in a new tab).
• Acquisition settings: settings such as slice thickness and reconstruction kernel are not on the basic profile list, so they stay unless a tool removes more NEMA, 27 Oct 2026 (opens in a new tab).
• Reports linked to their studies under the same pseudonym, with the text cleaned.
• A written record of all of it: the data card template and the data provenance log show what to write down.

How medical imaging licenses are usually structured is on the medical imaging data page. The health-data AI programs brief shows how to read a program's public evidence, and the health data hub collects the rest.

Stop conditions

• Stop if the HIPAA method, or the legal basis outside the United States, has not been chosen and written down.
• Stop if any attribute in the profile was skipped, or the options used are not recorded.
• Stop if a plain hash of the patient ID is the only pseudonym.
• Stop if any image type was not checked for burned-in text.
• Stop if head and neck series were not assessed for face risk.
• Stop if structured reports or free-text reports were neither cleaned nor removed.
• Stop if the key, the date offsets or the originals have left the clinic.
• Stop if no second person reviewed the pilot output.