Surgical data science for intelligent guidance and control in image-guided and robotic interventions

Research in surgical data science aims at improving the quality of interventional healthcare by leveraging information from heterogeneous data sources (e.g., medical imaging, sensors). By analysing these datasets, we will be able to model optimal surgical task execution and design skill-based controllers that will increase the autonomy of surgical robots resulting in more efficient surgeries with minimum invasiveness and trauma caused.

This project will investigate hybrid control approaches incorporating both model-based and learning-based ones to provide increased manoeuvrability and dexterity in surgical robots. The model-based ones will ensure that safety constraints are satisfied, while the learning-based ones will navigate the various uncertainties during the procedure and increase the performance of the surgical task execution.

Vision methods based on Deep Learning, would also be explored for localization and mapping of the environment using surgical cameras, offering intra-operative image guidance and enhancing clinical decision-making.

The specific objectives are to:

Model optimal surgical execution by fusing heterogeneous datasets
Translate these models into control policies for surgical robots leading to increased automation
Combine model-based and learning-based approaches to ensure that safety constraints are satisfied
Implement and evaluate intra-operative guidance/assistance systems to enhance surgical performance

Analysing and building models from multimodal data (e.g., images, surgical videos, haptic and position sensors) can enhance the efficiency of surgical and robotic interventions. To model optimal surgical task execution, first, segmentation of the different actions during the procedure needs to be done, using tools from unsupervised learning. Then, each action can be modelled by fusing the aforementioned heterogeneous datasets. The resulting models can be used to develop novel solutions in surgical training and simulation, and also to increase robot autonomy by translating them into control policies (e.g., skill learning). However, the ability of the robot to perform the same skill effectively depends both on the quality of the data collected from skill demonstration, and the modelling approach itself. To improve the autonomous learning and imitation ability of robots, Reinforcement and Deep Learning are proposed, instead of traditional modelling methods. However, these powerful learning-based approaches need be combined with model-based control to ensure safe operation and high performance at the same time.

Key EPSRC's research areas such as Medical Imaging, Artificial Intelligence Technologies, Robotics and Control Engineering are the project's core technical areas. The project is well-aligned with the current portfolio of EPSRC's research themes and specifically with Artificial Intelligence and Robotics and Healthcare Technologies. It also addresses two EPSRC Healthcare Technologies Grand Challenges in "Frontiers of Physical Intervention" and "Optimising Treatment" by focusing on data-driven methods and real-time analytics to realise new capabilities (autonomy, performance-based guidance) in surgical robotics. Project outcomes present opportunities for broad healthcare impact towards improving interventional outcomes with more accurate and safe procedures, increased access to surgery and personalising treatment, while reducing healthcare costs and patient recovery times.

Opportunity for integration and further development of project outcomes in the surgical robotics industry is expected to arise during the project duration. Beneficiaries include manufacturers of commercial surgical robotics systems as well as developers of novel interventional robotic platforms.