Acoustic 3D Shape Reconstruction of Zygomatic Bone for Non-Invasive Spatial Registration

Image-to-patient registration is an essential step in the overall image-guided surgery and determines the accuracy of precise surgery. Currently used registration methods, such as implanting titanium nails in the patient’s face, using occlusal splints or surface registration techniques based on soft tissues, etc., have their own problems such as invasiveness, high cost, or poor stability. Ultrasound imaging has the advantages of non-invasiveness and low cost, and facial bones are usually considered to have minimal deformation. Therefore, we hope to use ultrasound to non-invasively obtain facial bone surface features to complete the preoperative image-patient registration process. Therefore, we propose a technology for craniomaxillofacial surgery that segments and reconstructs 3D images from 2D ultrasound images obtained by ultrasound scanning of the zygomatic bone surface. In the segmentation module, we use nnU-Net with improved loss function and learning rate strategy to segment the zygomatic bone, achieving a Dice score of 0.8405. In the 3D reconstruction module, we propose a PNN algorithm with image downsampling that can quickly process data and a DSBezier algorithm that adds Bezier curves to fill voxels. The reconstruction results are registered with the head CBCT data, and the average residual can reach within 0.6 mm. The experimental results show that this system can accurately segment and reconstruct the zygomatic bone in three dimensions from ultrasound images.