Oropharyngeal cancer is one of the most common head and neck cancers, with rising incidence and mortality rates. Transoral robotic surgery (TORS) is a minimally invasive treatment for oropharyngeal cancer, offering better functional outcomes compared with open surgery. However, TORS is challenging because surgeons must operate in a constrained environment without haptic feedback to assess tumor extents. Image guidance systems have the potential to enhance anatomy and tumor visualization, supporting intraoperative decision-making. Intraoperative ultrasound (US) can provide real-time and cost-effective internal images of patients during TORS. However, current US applications in TORS have focused on intra-oral US, which may interfere with robotic instruments. Alternative US scanning strategies need to be investigated. This thesis hypothesizes that transcervical US, which places the US transducer externally on the neck, can provide continuous imaging for TORS without obstructing the surgery. It proposes a novel transcervical US guidance workflow for TORS and evaluates its technical feasibility. It first assesses the usability of 3D transcervical US for oropharynx imaging and a semi-automatic US-MRI registration method. Building on this, it presents a novel US-guided augmented reality framework for TORS that can overlay MRI-derived anatomical models onto the endoscopic view. The thesis then introduces a novel point cloud registration method, SemICP, integrating semantic information and a novel biomechanical energy-based regularization term to improve the surface matching accuracy compared with traditional ICP. SemICP is then incorporated with automatic US segmentation to enhance the clinical feasibility of the MRI–US registration process for a virtual reality US scanning guidance system for TORS. Finally, the thesis develops novel methods for intraoperative US landmark tracking and view retrieval. It first proposes a self-supervised keypoint-tracking model for landmark tracking in intraoperative US sequences. It then presents a representation learning method that leverages view matching for markerless localization of the US probe in transcervical US. The methods introduced in this thesis have been evaluated on new datasets collected on patients and healthy volunteers. The thesis demonstrates that it is feasible to utilize transcervical US to provide image guidance for TORS and develops a novel workflow of transcervical US guidance in TORS.
Wanwen Chen (Thu,) studied this question.