Realistic Ultrasound Synthesis Based on Diagnostic CT to Facilitate Ultrasound-Guided Robotic Spine Surgery

This paper aims to tackle the issues of unavailable or insufficient clinical ultrasound (US) data and meaningful annotations to enable bone segmentation and registration for US-guided spinal surgery. Although US-based spine surgery is being more widely utilized, the development of robot-assisted surgical navigation algorithms for this approach is still impeded by the limitation of insufficient training data. Moreover, due to the characteristics of US imaging, it is difficult to clearly annotate bone surfaces, which leads to a sub-optimal inference capacity of the trained model. Hence, we propose an in silico bone US simulation framework that synthesizes realistic US images from diagnostic computed tomography (CT) volumes. Afterward, using these simulated bone US images, we train a lightweight vision transformer model that can achieve accurate and on-the-fly bone segmentation for spinal sonography. In the validation experiments, a realistic US simulation is conducted by deriving from the diagnostic spinal CT volume to facilitate a radiation-free US-guided pedicle screw placement procedure. When the proposed approach is employed for training the bone segmentation task, the Chamfer distance reaches 0.599 mm; when it is applied to CT-US registration, the associated bone segmentation accuracy yields 0.93 Dice score, and the registration accuracy achieved based on the segmented point cloud reaches 0.13~3.37 mm in a complication-free manner. While bone US images exhibit strong echoes at the medium interface, the model may be unable to distinguish between thin interfaces and bone surfaces by simply relying on small neighborhood information. To overcome these shortcomings, we propose to utilize a long-range contrast learning module (LCLM) to fully explore the long-range differences between the candidates and their surrounding pixels. Comprehensive experiment results demonstrate that the proposed CT $\rightarrow $ US simulation dramatically increases the US segmentation performance without any labeled US bone samples, and the proposed LCLM is effective at precisely determining the position of the US region of interest (RoI) on bones.