Clario's AI Unlocks PHI/PII in DICOM with Bedrock

AI-Powered Medical Image De-identification

The article effectively showcases how Clario leverages Amazon Bedrock and Textract to automate PHI/PII detection in DICOM images, a critical but often manual and error-prone process in clinical trials. The chosen architecture, utilizing managed AWS services like EKS, RDS, and API Gateway, demonstrates a mature approach to scalability, security, and compliance. The separation of detection from redaction is a particularly noteworthy design decision, enhancing flexibility and auditability while maintaining human oversight for irreversible data modifications. The emphasis on addressing PHI/PII not only in metadata but also burned into image pixels highlights the comprehensive nature of their solution.

However, while the article touts the benefits of managed foundation models via Bedrock, it doesn't delve deeply into the specific challenges or nuances of fine-tuning or prompting Claude Sonnet 4.5 for this highly specialized medical imaging context. The article mentions the use of a 'configurable tolerance' for bounding box matching, which is crucial for accuracy, but the specifics of this tolerance and its impact on false positives/negatives could be elaborated upon. Furthermore, the 'limited window' for data retention, while good for compliance, might pose challenges for long-term research or retrospective analysis if not managed carefully. The article also assumes a certain level of technical understanding from its audience, particularly regarding DICOM standards and clinical trial workflows, which might limit its accessibility to a broader tech audience.

This solution is highly beneficial for organizations involved in clinical trials, particularly sponsors, CROs, and imaging vendors who must adhere to strict data privacy regulations like HIPAA and GDPR. The ability to automate a previously labor-intensive process leads to significant efficiency gains, reduced risk of human error, and faster data processing timelines. The technical implications are substantial for the healthcare AI sector, demonstrating a practical and scalable application of LLMs for sensitive data handling in a regulated environment. While comparisons to existing solutions are not explicitly detailed, the integrated approach of combining OCR with LLM analysis for both metadata and pixel data offers a more holistic solution than traditional rule-based or simpler OCR methods alone.

Key Points

• Clario automates PHI/PII detection in DICOM images using Amazon Bedrock and Amazon Textract.
• The solution handles sensitive data embedded in metadata (standard and custom) and directly burned into image pixels.
• Key AWS services used include Amazon Bedrock (Anthropic's Claude Sonnet 4.5), Amazon Textract, Amazon S3, Amazon EKS, Amazon RDS for PostgreSQL, and Amazon API Gateway.
• Design emphasizes scalability, security, compliance, and end-to-end observability.
• Separation of detection from redaction allows for flexibility, auditability, and human-in-the-loop validation.
• The solution supports both DICOM (.dcm) and PDF file formats.

---

📖 Source: How Clario technology detects PHI/PII in DICOM images using Amazon Bedrock