Burned-in PHI

Unlike metadata tags that can be stripped programmatically, burned-in PHI is fused with the image pixels themselves. This text—such as a patient name or MRN overlaid on an ultrasound frame—becomes part of the rasterized image data. Removing it requires optical character recognition (OCR) followed by inpainting or black-box redaction, not simple header manipulation. The text has no machine-readable encoding; it exists only as a visual pattern of light and dark pixels.

Burned-in PHI is most prevalent in modalities that capture screen outputs or secondary captures:

• Ultrasound: Sonographers frequently annotate frames with patient demographics directly on the screen before capture.
• Endoscopy/PACS: Older Picture Archiving and Communication Systems often burned text overlays into exported secondary captures.
• Mobile Photography: Clinicians taking photos of monitor screens for consults inadvertently capture overlaid patient banners.
• Digitized Film: Scanned analog X-rays may have lead-letter annotations burned into the film emulsion.

Automated detection requires a specialized computer vision pipeline distinct from standard text-based PHI detection:

1. Region Proposal: Object detection models identify text-bearing regions within the image.
2. Text Recognition: OCR engines transcribe the pixel-based text into machine-readable strings.
3. PHI Classification: A downstream medical NER model classifies the transcribed strings against HIPAA Safe Harbor identifiers.
4. Redaction: The identified pixel region is permanently obscured using black bars or Gaussian blur.

A critical distinction in medical imaging de-identification exists between DICOM metadata and pixel data. Standard DICOM de-identification tools excel at clearing tags like (0010,0010) Patient Name from the file header. However, they are blind to the same information if it appears visually within the pixel array (7FE0,0010). A fully compliant pipeline must address both vectors, ensuring the image is 'clean' both structurally and visually.

The DICOM Secondary Capture Image Storage SOP Class is a high-risk source. These images are often screen grabs exported from a modality workstation, containing not just the medical image but the entire application chrome—including the patient banner, exam list, and status bar. De-identifying these requires cropping or redacting the non-diagnostic UI regions before release, a step frequently missed in automated workflows.

Verifying the absence of burned-in PHI cannot rely on metadata checks alone. A robust QA process includes:

• Visual Spot-Checking: Manual review of a random sample of de-identified images.
• Automated OCR Re-scan: Running the de-identified output through the detection pipeline again to confirm zero detections.
• False Negative Audits: Specifically testing on images known to contain burned-in text to ensure the pipeline's recall is 100% for visual identifiers.

The foundational technology for detecting burned-in PHI embedded in pixel data. Unlike metadata extraction, OCR processes the visual frame buffer of modalities like ultrasound, endoscopy, and nuclear medicine to locate and transcribe text rendered directly into the image. Modern clinical pipelines use deep learning-enhanced OCR to handle challenges unique to medical imagery:

• Low contrast text: White text on grayscale anatomical backgrounds
• Non-standard fonts: Proprietary typefaces from legacy modality workstations
• Multi-orientation text: Rotated annotations on lateral or oblique views
• Overlay interference: Text partially obscured by anatomical structures or measurement calipers

Post-OCR, detected text strings are fed into PHI detection models to classify whether the recognized characters constitute protected identifiers before redaction.

A critical distinction in medical image de-identification that determines the technical approach required:

Metadata PHI resides in structured DICOM header fields and is detectable through tag-level parsing. Examples include the PatientName, PatientID, and InstitutionName attributes. Removal is deterministic and rule-based.

Pixel Data PHI (burned-in) is visually rendered into the image itself during acquisition. Common sources:

• Ultrasound machines that overlay patient demographics on each frame
• Ophthalmology devices that burn identifiers into fundus images
• Endoscopy video with timestamp and patient ID overlays
• Secondary capture images that photograph original films containing labels

Detection requires computer vision pipelines operating on the raw pixel buffer, making it computationally more intensive than metadata scrubbing.

The proportion of actual PHI instances that a detection model incorrectly classifies as non-sensitive, representing a direct measure of privacy leakage in a de-identification system. For burned-in PHI, false negatives are particularly dangerous because:

• A single missed instance of a patient name in an ultrasound frame can constitute a HIPAA breach if the image is released for research
• Burned-in text often appears in unpredictable locations across different modality vendors
• Low-contrast overlays may be invisible to human reviewers but recoverable through image enhancement

Mitigation strategies:

• Dual-pass detection: Combining OCR-based text recognition with direct visual object detection of text regions
• Confidence thresholding: Routing low-confidence frames to human-in-the-loop review
• Adversarial testing: Deliberately crafting images with subtle burned-in PHI to stress-test detection pipelines

A system architecture that combines deterministic rule-based redaction with probabilistic machine learning models to maximize both precision and recall of PHI detection across all modalities.

For burned-in PHI specifically, the hybrid approach layers:

1. Heuristic pre-processing: Detecting known modality-specific overlay patterns (e.g., Siemens ultrasound footer format)
2. OCR text extraction: Running multiple OCR engines in parallel with voting mechanisms
3. NER-based PHI classification: Applying medical named entity recognition to distinguish patient names from anatomical labels
4. Visual region proposal: Using object detection models to locate text bounding boxes independently of OCR
5. Consensus scoring: Combining confidence scores across all detection layers before redaction or human review routing

This defense-in-depth strategy ensures that PHI missed by one layer is caught by another, minimizing residual risk.

The remaining probability that protected health information persists in a dataset after an automated de-identification pipeline has been executed. For burned-in PHI, residual risk assessment must account for:

• Vendor-specific overlay formats: Proprietary text placement that may not appear in training data
• Image degradation: Compression artifacts that degrade OCR accuracy in transmitted images
• Multi-frame instances: DICOM objects containing dozens or hundreds of frames where burned-in text may appear intermittently
• Annotation text: Physician-added measurements or labels that inadvertently include identifying information

Formal risk assessment methodologies:

• Stratified sampling across modality types and device manufacturers
• Red-team penetration testing using manual reviewers to hunt for missed identifiers
• Statistical extrapolation from false negative rates measured on ground-truth annotated test sets