Scholarlink

DICOM (Digital Imaging and Communications in Medicine) files contain sensitive Protected Health Information (PHI) embedded in both pixel-level image data and metadata headers. Effective deidentification is crucial for enabling medical imaging research, data sharing, and compliance with privacy regulations such as HIPAA and GDPR. This study presents a comprehensive methodology for DICOM de-identification using John Snow Labs Visual NLP, a specialized library designed for medical imaging that can understand both image content and associated text. A dual-level de-identification process is proposed, addressing both pixel-level PHI removal through text detection, extraction, and anonymization, and metadata-level PHI removal through tag reading, PHI detection, and anonymization. The methodology is evaluated on the MIDI-B dataset, demonstrating the effectiveness of Visual NLP pipelines for comprehensive DICOM de-identification. Experimental results show high accuracy across multiple validation metrics, including 100% success rates for tag retention, date shifting, and UID consistency, with overall text processing accuracy exceeding 99.9%. The approach maintains clinical utility for research purposes.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comprehensive DICOM De-identification Using Visual NLP: A Dual-Level Approach to Privacy-Preserving Medical Imaging

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nitin Kumar,
  • Alberto Andreotti,
  • Veysel Kocaman,
  • Yigit Gul,
  • Mehmet Butgul,
  • David Talby

摘要

DICOM (Digital Imaging and Communications in Medicine) files contain sensitive Protected Health Information (PHI) embedded in both pixel-level image data and metadata headers. Effective deidentification is crucial for enabling medical imaging research, data sharing, and compliance with privacy regulations such as HIPAA and GDPR. This study presents a comprehensive methodology for DICOM de-identification using John Snow Labs Visual NLP, a specialized library designed for medical imaging that can understand both image content and associated text. A dual-level de-identification process is proposed, addressing both pixel-level PHI removal through text detection, extraction, and anonymization, and metadata-level PHI removal through tag reading, PHI detection, and anonymization. The methodology is evaluated on the MIDI-B dataset, demonstrating the effectiveness of Visual NLP pipelines for comprehensive DICOM de-identification. Experimental results show high accuracy across multiple validation metrics, including 100% success rates for tag retention, date shifting, and UID consistency, with overall text processing accuracy exceeding 99.9%. The approach maintains clinical utility for research purposes.