Machine Learning for Identification of Human Movements in Usage of a Minimally Invasive Surgery Instrument

Minimally Invasive Surgery (MIS) operations in medicine, such as laparoscopy, require high skills that can be attained only by special training and years of experience. Identifying skilled and unskilled manipulations through analysis of visual cues or robot recorded position of tool movement could significantly improve the skill assessment during lengthy and expensive MIS training procedures. Such identification requires the capability to identify the tool tip movements in real-time. This project will be an integral part of an EPSRC funded ongoing project for MIS skill identification. In this project we will use a co-manipulated robot arm for monitoring and recording the hand movements of a user during experimental MIS suturing. We will also use a camera based motion tracking algorithm for position recording. The recorded position data will be analysed to identify the different and significant parts of hand movement. For example, the developed software will be capable of identifying and distinguishing between the movements of pushing the needle, pulling it from the other side, and translating the needle from one side to another in a suturing operation. Techniques such as or similar to linear discriminant analyser, hidden Markov models, motion primitives, and non-negative matrix factorization will be used. A co-manipulated robotic setup with a laparoscopy training platform is available for experimental MIS operations to be used in this project.

In this project the student will perform the following:
1) Learn about the existing robotic setup and existing visual tracking algorithm in order to gain an understanding of how co-manipulation is achieved and how to record the position/orientation data of the instrument.
2) Develop a machine learning algorithm to identify and distinguish between different movements of the tool tip (different phases of operation) using recorded position/orientation data.
3) Test and verify the system with the robotic-MIS training setup.
4) (Optional) Use the same machine learning algorithm to find out some characteristic differences between the tool tip trajectories of novice versus professional subjects and/or in pre-training versus post-training experiments.

Supervisor name: 
Mustafa Suphi Erden
Supervisor and Deputy email addresses: 
m.s.erden@hw.ac.uk