Real-Time and Precise 3-D Lesion Reconstruction With Boundary Point Clouds for Robotic Ultrasound Scanning.

In robot-assisted breast ultrasound scanning, conventional 2-D imaging often fails to fully capture the spatial morphology of lesions, limiting clinicians' ability to obtain intuitive structural information. In addition, variations in patient body type and examination posture lead to significant scanning range fluctuations. Conventional voxel-based 3-D reconstruction methods rely heavily on pre-defined voxel grid size and position but inappropriate settings can adversely affect lesion feature data acquisition and computation, potentially resulting in inaccurate 3-D lesion reconstruction. To address these issues, we proposed a lesion boundary point cloud-based 3-D reconstruction method. The method first employed a neural network to extract lesion boundary features from ultrasound images in real time, generating a 2-D contour point set. These points were then projected into 3-D space to form a lesion point cloud, which was subsequently tetrahedrally meshed to generate a complete 3-D lesion mesh. Comparative experiments on two models demonstrated that our method significantly improved reconstruction accuracy, reducing the average error by approximately 47%. Visualization results further validated the structural reliability of the reconstructed models. Moreover, our method increased the data-processing speed by 130% and reduced the mesh construction time by 93%, achieving higher efficiency compared with voxel-based reconstruction. The real-time 3-D lesion models provide clinicians with intuitive structural information and have the potential to support intraoperative navigation and interventional procedures in robot-assisted surgery. Breast lesion, 3-D reconstruction, Point cloud, Visualization, Robot-assisted ultrasound.