Scholarlink

Retained foreign objects (RFOs) are rare but serious medical errors that can lead to significant patient complications and medicolegal consequences. While deep learning-based detection methods have shown promise in improving RFO recognition, their development is limited by the lack of large-scale, publicly available datasets. Additionally, most existing datasets include only non-critical RFOs, which pose minimal risk, while critical RFOs, which can be retained internally and cause severe complications, are largely absent due to their rarity. To address these limitations, we propose RFO-DeepDRR, a novel physics-based pipeline for generating synthetic radiographs with realistic renderings of critical RFOs. Our approach leverages high-fidelity CT segmentation, accurate 3D rendering of RFOs, and physics-based simulation to generate high-quality, large-scale synthetic datasets. We demonstrate that incorporating synthetic RFO images significantly enhances the performance of deep learning models in localizing these RFOs, thereby improving their ability to accurately identify and classify potential RFO cases across diverse clinical environments. Our code and datasets, which contain over 4,000 synthetic images, are made publicly available at GitHub .

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

A Dataset and Benchmark for Enhancing Retained Foreign Object Detection Through Physics-Based Image Synthesis

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Liwei Zhou,
  • Yuli Wang,
  • Victoria Shi,
  • Jonathan Feng,
  • Lin-mei Zhao,
  • Norman Atagu,
  • Praneeth Madhu,
  • Tej Mehta,
  • Premal Trivedi,
  • Cheng Ting Lin,
  • Web Stayman,
  • John Eng,
  • Pamela Johnson,
  • Elliott Haut,
  • Harrison Bai

摘要

Retained foreign objects (RFOs) are rare but serious medical errors that can lead to significant patient complications and medicolegal consequences. While deep learning-based detection methods have shown promise in improving RFO recognition, their development is limited by the lack of large-scale, publicly available datasets. Additionally, most existing datasets include only non-critical RFOs, which pose minimal risk, while critical RFOs, which can be retained internally and cause severe complications, are largely absent due to their rarity. To address these limitations, we propose RFO-DeepDRR, a novel physics-based pipeline for generating synthetic radiographs with realistic renderings of critical RFOs. Our approach leverages high-fidelity CT segmentation, accurate 3D rendering of RFOs, and physics-based simulation to generate high-quality, large-scale synthetic datasets. We demonstrate that incorporating synthetic RFO images significantly enhances the performance of deep learning models in localizing these RFOs, thereby improving their ability to accurately identify and classify potential RFO cases across diverse clinical environments. Our code and datasets, which contain over 4,000 synthetic images, are made publicly available at GitHub .