25995 Automated assesment of surgical skill in robotic surgery using deep learning techniques
Begeleider(s): ir. Mathieu De Coster en Pieter De Backer

Richtingen: Master of Science in Biomedical Engineering, Master of Science in Computer Science Engineering , Master of Science in de informatica

Probleemstelling:

Prostate cancer is the second most common cancer in males. Treatment includes surgical removal of the prostate gland, which is often performed in a minimal-invasive way through robotic surgery.

Orsi academy is the world's largest training centre in robotic surgery and has a broad expertise in minimal invasive surgery. As such Orsi has developed and validated a state-of-the art training methodology called "Proficience Based Progression" in which surgeons are trained to proficiency before they can start real life surgery.

The specific metrics used in this methodology objectivate surgical performance through a standardised approach. However, this assesment is done purely manual by experienced surgeons. Current computer vision methods only allow for basic disctinctions in scoring allowing basic scales such as "novice - intermediate - expert" and are rarely trained on real life images due to a lack of qualitative data.

It is evident that this approach is inaccurate to provide certification.

Hence, Orsi academy is heavily involved in automating this type of assessment through computer vision techniques such as deep learning. As mentioned earlier, one major bottleneck in this rapidly developing field is the lack of qualitative data.

However, due to it's central teaching role in the robotic surgery landscape, Orsi has very large labeled and unlabeled datasets in house to apply new computer vision techniques, unparalleled in current literature.

Intuitive robotic system: Da Vinci Xi


Doelstelling:

This master topic will focus on applying computer vision techniques to robot-assisted radical prostatectomy training videos in order to assess surgical performance.

The student will focus on one or more subtasks of our automated assessment in which surgical action-recognition takes on a central role. These include grasping a needle, making a byte and knot-tying. Large labeled datasets are readily available as well as an extensive unlabeled video database.

We will look at using self-supervised learning to pretrain the network based on similar but unlabelled data. In this approach, unlabelled data can be used by extracting labels from the data itself and the feature extraction layers are pre-trained for that task.

The student is required to have a profound knowledge of Python and a background in machine learning.

Training model images