Freehand 3D ultrasound imaging toward midfacial bone surface reconstruction for intraoperative image registration

Image-guided surgery is a critical technique in maxillofacial surgery. The foundation of image-guided surgery is image registration. Traditional image registration methods have limitations in terms of invasiveness, complexity, and unsatisfied accuracy. Freehand 3D ultrasound (US) imaging using a tracked 2D US probe may offer a non-invasive, real-time, and accurate alternative. Purpose This study aims to develop a novel freehand 3D US imaging framework for midfacial bone surface reconstruction and registration with preoperative 3D data (e.g., computed tomography), enabling accurate intraoperative surgical navigation in maxillofacial surgery.

First, a customized stereo camera is used to track the pose of a 2D US probe during the freehand US scanning toward the midfacial bone surface. Then, a short-term dense concatenate network (STDC) is employed to segment the bone surface from the US image. The segmented pixels with spatial information form a coarse 3D volume in real time. The 3D volume's voxels are then converted to a coarse point cloud. A template matching denoising technique is utilized to remove noisy and outlier points, followed by a self-supervised Freehand 3D Ultrasound Neural Surface Reconstruction network (FUNSR) to reconstruct the point cloud to a smooth surface mesh. Finally, the resulting fine bone surface is registered with preoperative 3D data for quantitative evaluation. A total of 1000 zygomatic ultrasound images (split into 700 training, 150 validation, and 150 test images) were used to train the segmentation network. The reconstruction network was trained with self-supervision. The reconstruction accuracy of the network was validated using surface registration error (SRE), and the registration accuracy was verified using target registration error (TRE). Method performance improvement was evaluated using t-tests and analysis of variance, with Tamhane's T2 test applied for multiple comparison correction to control the false discovery rate. Cohen's effect sizes were calculated to quantify performance differences.

In the phantom experiment, the average SRE was 0.387 $\pm$ 0.034 mm, and the average TRE was 0.802 $\pm$ 0.177 mm. Compared with registration using only voxel reconstruction results (SRE = 1.301 $\pm$ 0.133 mm, TRE = 1.155 $\pm$ 0.359 mm), the accuracy was improved (Cohen's d = 9.416 for SRE, Cohen's d = 1.247 for TRE, and 0.01 for both). Also, the accuracy remained uniform across various regions of the midface ( $p=$ 0.918). When using only local region reconstruction for registration, the decrease in overall accuracy is relatively minor ( $p=$ 0.025). In the volunteer trials, the average SRE was 0.445 $\pm$ 0.099 mm. Compared with the fundamental framework of our method (SRE = 0.955 $\pm$ 0.204 mm), the proposed template matching denoising and surface reconstruction components further enhance the registration accuracy ( 0.001, Cohen's d $>$ 2.0).

The proposed freehand 3D US imaging framework could offer a noninvasive, accurate, and quasi-real-time solution for midfacial bone surface reconstruction and image registration in maxillofacial surgery.