Exploring AI-Based System Design for Pixel-Level Protected Health Information Detection in Medical Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Data Privacy

Background:

• De-identification of medical images is crucial for privacy in research and clinical data sharing.
• Detecting Protected Health Information (PHI) within image metadata and pixels is the primary challenge.
• Limited evaluation of existing AI solutions hinders the development of reliable de-identification tools.

Purpose of the Study:

• To present and evaluate an AI-based pipeline for detecting PHI in medical images.
• To benchmark different AI models for text detection, extraction, and analysis in this context.
• To determine the optimal configuration for a robust and efficient PHI detection system.

Main Methods:

• Developed an AI pipeline with three modules: text detection, text extraction, and text analysis.
• Benchmarked YOLOv11, EasyOCR, and GPT-4o across various configurations.
• Evaluated performance on two datasets with diverse imaging modalities and PHI types.

Main Results:

• The optimal setup utilizes dedicated vision and language models for each pipeline module.
• This configuration achieves a favorable balance between performance, latency, and the cost of using large language models (LLMs).
• LLMs proved effective not only for PHI identification but also for enhancing Optical Character Recognition (OCR) and enabling end-to-end detection.

Conclusions:

• An AI pipeline with specialized vision and language models offers an effective solution for PHI detection in medical images.
• The integration of LLMs significantly improves the overall performance and capabilities of the de-identification process.
• This approach demonstrates promising results for enhancing privacy in medical image data sharing.