Reproduce “DICOM De-Identification via Hybrid AI and Rule-Based Framework for Scalable, Uncertainty-Aware Redaction” 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
Kyle Naddeo, Nikolas Koutsoubis, Rahul Krish, Ghulam Rasool, Nidhal Bouaynaya, Tony OSullivan, Raj Krish
A hybrid DICOM de-identification pipeline combines deterministic tag-removal rules with fuzzy matching, LUKE-based entity recognition, Faster R-CNN text detection, OCR, and uncertainty-guided human review. Across reported DICOM, HIPAA, and TCIA tests, it processed 581,265 instances with a 99.88% overall pass rate, although performance varied substantially across repository-specific categories.
Key insights
- Combining interpretable rules with language and vision models broadens coverage beyond known DICOM fields and exact identifier matches while retaining a mechanism for institutional policy control.
- Uncertainty is used operationally rather than treating every detector output as equally reliable: confident detections can be automated, while higher-risk cases are quarantined for review.
- The reported results suggest strong aggregate compliance performance but also show that repository-specific criteria and real-world distribution shifts remain important sources of failure.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.