Vision-based context-aware assistance for minimally invasive surgery.

Conference: Bournemouth University, Faculty of Media and Communication

Abstract:

This thesis aims to solve the challenges of surgical activities analysis and provide context-aware assistance for MIS. In our study, we consider the surgical visual data as the only input. Because videos and images own high-dimensional and representative features, and it is much easier to access than other data format, for example, kinematic information or motion trajectory.

Following the granularity of surgical activity in a top-down manner, we first propose an attention-based multi-task framework to assess the expertise level and evaluate six standards for surgeons with different skill level in three fundamental surgical robotic tasks, namely suturing, knot tying and needle passing. Second, we present a symmetric dilated convolution structure embedded with self-attention kernel to jointly detect and segment fine-grained surgical gestures for surgical videos. In addition, we use the transformer encoder-decoder architecture with reinforcement learning to generate surgical instructions based on images.

Overall, this thesis develops a series of novel deep learning frame- works to extract high-level semantic information from surgical video and image content to assist MIS, pushing the boundaries towards integrated context-aware system through the patient care pathway.

https://eprints.bournemouth.ac.uk/35389/

Source: Manual

Vision-based context-aware assistance for minimally invasive surgery.

Authors: Zhang, J.

Conference: Bournemouth University

Abstract:

Context-aware surgical system is a system that can collect surgical data and analyze the operating environment to guide responses for surgeons at any given time, which improves the efficiency, augment the performance and lowers the risk of minimally invasive surgery (MIS). It allows various applications through the whole patient care pathway, such as medical resources scheduling and report generation. Automatic surgical activities understanding is an essential component for building context-aware surgical system. However, analyzing surgical activities is a challenging task, because the operating environment is considerably complicated. Previous methods either require the additional devices or have limited ability to capture discriminating features from surgical data. This thesis aims to solve the challenges of surgical activities analysis and provide context-aware assistance for MIS. In our study, we consider the surgical visual data as the only input. Because videos and images own high-dimensional and representative features, and it is much easier to access than other data format, for example, kinematic information or motion trajectory. Following the granularity of surgical activity in a top-down manner, we first propose an attention-based multi-task framework to assess the expertise level and evaluate six standards for surgeons with different skill level in three fundamental surgical robotic tasks, namely suturing, knot tying and needle passing. Second, we present a symmetric dilated convolution structure embedded with self-attention kernel to jointly detect and segment fine-grained surgical gestures for surgical videos. In addition, we use the transformer encoder-decoder architecture with reinforcement learning to generate surgical instructions based on images. Overall, this thesis develops a series of novel deep learning frame- works to extract high-level semantic information from surgical video and image content to assist MIS, pushing the boundaries towards integrated context-aware system through the patient care pathway.

https://eprints.bournemouth.ac.uk/35389/

Source: BURO EPrints