Reproduce “Exploring AI-based System Design for Pixel-level Protected Health Information Detection in Medical Images” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tuan Truong, Ivo M. Baltruschat, Mark Klemens, Grit Werner, Matthias Lenga
A modular AI pipeline for detecting protected health information burned into medical-image pixels combines YOLOv11 localization, EasyOCR text extraction, and GPT-4o semantic classification. The best configuration achieved 98.69% instance-level precision and 98.08% recall on the synthetic RadPHI benchmark, while prompt refinement improved MIDI hold-out precision to 97.21% and recall to 97.81%, outperforming rule-based Presidio detection.
Key insights
- The strongest balance came from separating the tasks: YOLOv11 localized text, EasyOCR extracted it, and GPT-4o interpreted its meaning. GPT-4o could compensate for some OCR errors during classification, reducing the need to use it for every image crop.
- End-to-end multimodal analysis produced weaker detection and a 5.18% pipeline error rate, while EasyOCR-only localization was vulnerable to fragmented multi-line text and lost context.
- Prompt-based policy changes improved discrimination between patient-specific PHI and permissible study or image identifiers, showing that classification rules can be adapted without retraining the models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.