Scholarlink

Single-view reconstruction (SVR) enables 3D reconstruction of vertebrae from a single radiograph and is particularly valuable in Minimally Invasive Spine Surgery (MISS), where intraoperative imaging is limited to 2D data obtained from C-arm fluoroscopy. SVR can reduce radiation exposure by avoiding multi-angle imaging. It supports various 3D representations, with mesh-based outputs offering greater memory efficiency and anatomical detail. However, SVR remains underexplored due to the lack of paired radiograph–mesh datasets. Datasets like VerSe, TotalSegmentator, and CTSpine1K offer Computed Tomography (CT) scans and segmentation labels, while others like MedShapeNet and VSD only provide surface models. No existing datasets offer paired radiographs and meshes needed for SVR. To bridge this gap, we present Rad-Surf, an automated pipeline that generates Digitally Reconstructed Radiographs (DRRs) and surface meshes from CT–segmentation labels pairs, generalizable across anatomies, with built-in post-processing for seamless integration into deep learning-based SVR workflows. Further, we improve the DRR quality in a Deep Image Prior (DIP) based super resolution pipeline. The generated DRRs were assessed based on Sign-to-Noise Ratio (SNR), Contrast-to-Noise Ratio (CNR), entropy and sharpness. In addition, a clinical analysis was performed on the generated mesh to study the measurements of four different clinical parameters of the vertebrae. The data generated through Rad-Surf for SVR of the lumbar vertebrae is made open source. It contains 475 unique DRR–mesh pairs and DRRs for each mesh are rendered from 24 diverse viewpoints, resulting in an augmented dataset containing 11,400 DRR-mesh pairs. The project page can be found in:

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

Rad-Surf: Automated Synthesis of Radiographs and Vertebral Surfaces for Single-View Reconstruction

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • R. Neeraja,
  • A. Anusha,
  • Vivek Maik,
  • Aparna Purayath,
  • Manojkumar Lakshmanan,
  • Mohanasankar Sivaprakasam

摘要

Single-view reconstruction (SVR) enables 3D reconstruction of vertebrae from a single radiograph and is particularly valuable in Minimally Invasive Spine Surgery (MISS), where intraoperative imaging is limited to 2D data obtained from C-arm fluoroscopy. SVR can reduce radiation exposure by avoiding multi-angle imaging. It supports various 3D representations, with mesh-based outputs offering greater memory efficiency and anatomical detail. However, SVR remains underexplored due to the lack of paired radiograph–mesh datasets. Datasets like VerSe, TotalSegmentator, and CTSpine1K offer Computed Tomography (CT) scans and segmentation labels, while others like MedShapeNet and VSD only provide surface models. No existing datasets offer paired radiographs and meshes needed for SVR. To bridge this gap, we present Rad-Surf, an automated pipeline that generates Digitally Reconstructed Radiographs (DRRs) and surface meshes from CT–segmentation labels pairs, generalizable across anatomies, with built-in post-processing for seamless integration into deep learning-based SVR workflows. Further, we improve the DRR quality in a Deep Image Prior (DIP) based super resolution pipeline. The generated DRRs were assessed based on Sign-to-Noise Ratio (SNR), Contrast-to-Noise Ratio (CNR), entropy and sharpness. In addition, a clinical analysis was performed on the generated mesh to study the measurements of four different clinical parameters of the vertebrae. The data generated through Rad-Surf for SVR of the lumbar vertebrae is made open source. It contains 475 unique DRR–mesh pairs and DRRs for each mesh are rendered from 24 diverse viewpoints, resulting in an augmented dataset containing 11,400 DRR-mesh pairs. The project page can be found in: