47 fromData vs Model in Medical Robotics Workshop - IROS 2023 Computer Vision News One primary focus of my current research for my postdoc at UCSD is nonrigid registration and 3D reconstruction of surgical scenes, with an emphasis on handling challenging conditions such as large deformations and achieving accurate performance. Endoscopic videos are an important type of sensory data that I am working with. They provide real-time information of the surgery and are widely captured in modern surgical procedures, thereby eliminating the need for substantial surgical workflow modifications. We have developed a comprehensive surgical perception framework, Semantic-SuPer, which integrates geometric and semantic information extracted from endoscopic videos to achieve more accurate data association and lead to robust tissue deformation tracking and reconstruction. Currently, we are working on further improving the capability of this framework to handle larger deformations, as well as enhance robustness to noisy input data and reduce error accumulation during longer manipulations. In addition to endoscopes, I am also investigating non-rigid registration techniques for other sensor sources. I mentored a project focused on developing a recursive registration network to track respiratory motions in lung 4DCTs, i.e., sequences of 3D CT scans. Furthermore, I am extending the endoscopic video-based registration and reconstruction results to guide the manipulation of deformable objects. The proposed approach could potentially serve as compensation or an alternative for the methods that require accurate simulations of heterogeneous surgical scenes, which are hard to achieve.
RkJQdWJsaXNoZXIy NTc3NzU=