Home / DICOM redaction software
DICOM Redaction Software
Patient identifiers in a DICOM study live in two places: burned into the pixels, and written into the header tags. Redactor handles both, as separate operations, across multi-frame studies.
Why DICOM Is Not Just Another Image
Most redaction tools treat a medical image as a picture. That misses half the problem and breaks the other half.
Identifiers are burned into the pixel data by the modality itself, often in a corner of every frame. They are also written into the DICOM header as textual tags. A tool that flattens a study to JPEG strips the study and series structure a PACS needs, and a tool that only cleans metadata leaves the patient name visible in the image.
A multi-frame study compounds it. Burned-in text does not necessarily appear on the first frame, so anything that samples one image and moves on will pass a study that is still identifiable.
How DICOM Redaction Works
Redactor treats DICOM as medical imaging rather than as a generic image, preserving study and series structure.
- Ingest —
.dcm,.dicand.dicomare recognized as a distinct medical-imaging content type, not converted to a general image format. - Process frame by frame — each frame of a multi-frame study is treated as one page by the processing pipeline, so burned-in text and objects are detected across the whole study rather than on the first image only.
- De-identify the header separately — extraction walks the DICOM tag set and returns the textual, PII-relevant tags, deliberately skipping elements that carry pixel data so imaging content is handled by image redaction rather than by metadata processing.
- Redact the pixels — pixel-burned identifiers are obscured frame by frame with a solid fill.
What DICOM Redaction Covers
| Extensions | .dcm, .dic, .dicom |
| Multi-frame studies | Handled a frame at a time, one frame per page |
| Pixel-burned identifiers | Detected and masked across every frame |
| Header tags | Textual PII-relevant tags extracted for de-identification; binary and pixel-bearing elements left to the imaging path |
| Masking | Solid fill |
| Clinical narrative | De-identified with a RoBERTa model trained on the i2b2 clinical corpus, running as a recognizer alongside the general PII pipeline |
| Deployment | Shared SaaS, dedicated SaaS, private cloud, on-premises, air-gapped |
Clinical text gets a clinical model
Where a study travels with a report, the narrative is de-identified using a model trained for clinical text rather than a general-purpose PII recognizer. The i2b2 de-identification model recognizes the identifiers that appear in medical records and the forms they take there, which is not the same as the forms they take in a contract or a police report.
Compared With Flattening to Images
| Flatten to JPEG or PNG | Redactor |
|---|---|
| Study and series structure lost; output no longer a DICOM study | Structure preserved, output remains DICOM |
| Header tags carried over or discarded wholesale | Textual tags de-identified as a separate, deliberate operation |
| First frame sampled, later frames unchecked | Every frame of a multi-frame study processed |
| Clinical narrative run through a general PII model | Clinical narrative run through a model trained on clinical text |
Where It Matters
- Medical and clinical documents — studies shared with sponsors, monitors, CROs and regulators
- HIPAA workflows — de-identification before a study leaves the organization
- Healthcare — imaging alongside records, recordings and consultation video
What DICOM Redaction Does Not Do
- Blur and pixelate are not available for DICOM. Masking is a solid fill. Blur and pixelate are explicitly deferred in the document masking work.
- The clinical de-identification model is one recognizer among those the PII pipeline runs; language coverage follows the pipeline rather than the model.
- Identifiers in the DICOM header and identifiers burned into the pixels are two operations. Running one does not cover the other.
How DICOM Redaction Is Evaluated
- DICOM is handled as a distinct medical-imaging content type, including multi-frame studies, rather than converted to a generic image.
- DICOM redaction covers multi-frame studies, processing frame by frame.
- Headers and pixel data are de-identified as separate operations, because the identifiers live in two different places.
- A dedicated clinical de-identification model handles PHI, running alongside the general recognizers.
- Masking for DICOM is a solid fill; blur and pixelate are not offered.
- Available on-premises and air-gapped, so studies need not leave the organization to be de-identified.
Redact a Study Without It Leaving Your Network
Send us a de-identified sample, or run it yourself on-premises. We will show you both surfaces: the pixels and the tags.