🧾 Intended Use

1. Clinical Context and Purpose

RadCloudAI is designed to assist board-certified radiologists and licensed medical professionals in the generation of structured radiology reports, impressions, and quality assurance (QA) feedback. It uses advanced large language models (LLMs) and vision models to provide AI-enhanced report drafts, diagnostic impressions, and guideline-linked recommendations.

RadCloudAI is not intended for autonomous diagnosis or decision-making. The tool is strictly used as a clinical decision support assistant, and final interpretation must be performed and approved by a qualified radiologist.

2. Acceptable Users

RadCloudAI is intended for use only by:

  • Licensed radiologists

  • Physician assistants or residents under supervision

  • Credentialed hospital staff with appropriate permissions

  • Internal QA teams reviewing reports for compliance and consistency

Use by laypersons, patients, or third-party vendors is explicitly prohibited.

3. Protected Health Information (PHI) Guidelines

Before entering any free-text report or uploading images:

✅ Users must make a reasonable effort to de-identify any patient health information (PHI), including names, dates of birth, MRNs, and other identifiers.

However, in recognition that inadvertent PHI may occasionally be entered:

PHI Scrubbing and Safeguards

RadCloudAI implements a dual-layer PHI defense system to protect sensitive data:

🔹 A. PHI Alert Layer (Pre-processing)

  • All uploaded images (e.g. screenshots or DICOM screenshots) are scanned using OCR-based detection via EasyOCR.

  • Text fields are scanned for likely PHI using structured rules (e.g., MRNs starting with D, 09-, 900, etc.).

  • If PHI is detected, the request is blocked immediately unless override permissions are present (e.g., for internal QA review).

  • PHI types (e.g. names, MRNs) are listed in the response for transparency.

🔹 B. PHI Scrubbing Layer (Fallback)

  • Even after passing the alert layer, all content passed to LLMs is auto-scrubbed using in-line PHI masking to catch missed or borderline cases.

  • PHI is replaced with placeholders (e.g., [PATIENT_NAME]) before being submitted to AI models.

These mechanisms operate across both:

  • Text-based workflows (e.g., full report generation, impression-only mode)

  • Image-based workflows (e.g., pasted screenshots, scanned PDFs)

4. Intended Input Sources

Users may paste or upload content from:

  • Radiology dictation software (e.g., Fluency, PowerScribe)

  • PACS viewers (image screenshots)

  • RIS output summaries (structured text or HL7-formatted snippets)

The platform supports both:

  • Text-only workflows for impressions and reports

  • Image + text workflows, where visual content is analyzed by vision models (e.g., Gemini 2.5)

5. Intended Outputs

RadCloudAI may generate:

  • Structured Findings and Impressions

  • Literature-linked Recommendations

  • Embedded Radiopaedia and PubMed references

  • Internal QA flags (actionable and review-level)

  • Custom routing based on exam type, CPT code, or body region

All output is customizable and can be reviewed, edited, and approved by the user prior to submission to RIS/PACS.

6. Security and Compliance Notes

  • All API traffic is secured using HTTPS with HSTS and strict Content Security Policy (CSP) headers.

  • MFA (Multi-Factor Authentication) is available for all user accounts.

  • All OpenAI, Gemini, and other LLM credentials are stored in encrypted format and scoped to user groups.

  • No PHI is persisted in long-term logs.

  • 7. Disclaimer

RadCloudAI is an assistive tool only. It does not replace professional clinical judgment and should not be relied upon as the sole basis for medical decision-making. Use of the system signifies acknowledgment of this limitation and agreement to follow institutional policies and guidelines regarding AI-assisted tools.