Scholarlink

Three-dimensional (3D) ultrasound (US) can enhance diagnostic and intraoperative imaging by providing volumetric insights in a non-invasive and cost-effective manner. However, most current methods rely on expensive tracking systems or 3D transducers, limiting the use in point-of-care settings. We propose a trackerless 3D US reconstruction method using a handheld US probe with an integrated inertial measurement unit (IMU) and a CNN-Transformer architecture. A dataset of 361 US sweeps from ex vivo specimens and phantoms was collected using a custom setup. The model uses Two-dimensional (2D) US images, optical flow, and IMU orientation data to predict local inter-frame and global transformations relative to the start frame. The final model achieved a mean Final Drift Ratio (FDR) of 11.63% and median FDR of 8.11% on unseen data. It accurately reconstructed 3D anatomy, enabling segmentation and visualization of structures, particularly along the x/y axes and x-axis rotation. This work demonstrates a feasible and competitive trackerless 3D US reconstruction approach, supporting future clinical use in diagnostic imaging, surgical guidance, and planning.

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

3D Ultrasound Volume Reconstruction Using a CNN-Transformer Model and IMU Data

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mark Wijkhuizen,
  • Chrissy A. Adriaans,
  • Lennard M. van Karnenbeek,
  • Tiziano Natali,
  • Theo Ruers,
  • Freija Geldof,
  • Behdad Dashtbozorg

摘要

Three-dimensional (3D) ultrasound (US) can enhance diagnostic and intraoperative imaging by providing volumetric insights in a non-invasive and cost-effective manner. However, most current methods rely on expensive tracking systems or 3D transducers, limiting the use in point-of-care settings. We propose a trackerless 3D US reconstruction method using a handheld US probe with an integrated inertial measurement unit (IMU) and a CNN-Transformer architecture. A dataset of 361 US sweeps from ex vivo specimens and phantoms was collected using a custom setup. The model uses Two-dimensional (2D) US images, optical flow, and IMU orientation data to predict local inter-frame and global transformations relative to the start frame. The final model achieved a mean Final Drift Ratio (FDR) of 11.63% and median FDR of 8.11% on unseen data. It accurately reconstructed 3D anatomy, enabling segmentation and visualization of structures, particularly along the x/y axes and x-axis rotation. This work demonstrates a feasible and competitive trackerless 3D US reconstruction approach, supporting future clinical use in diagnostic imaging, surgical guidance, and planning.