Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Automatic Planning of Spine Surgery using Reinforcement Learning Algorithms
The goal of this work is to implement a pipeline that enables fully automated surgical planning using reinforcement learning algorithms.
Keywords: Machine Learning, ML, AI, Artificial Intelligence, Surgical Planning, Python, Reinforcement Learning, RL, Genetic Algorithms, Medical Image Processing, Spine Surgery
Due to their complexity, orthopedic spine surgeries often require a complex preoperative planning to ensure optimal patient outcome. This planning procedure is often performed manually based on pre-acquired patient medical images (e.g, CT scans) through which the surgeons precisely define the desired surgical plans (e.g., where inside the anatomy should the surgical hardware be placed). Once the preoperative plan is available, it is visualized to the surgeon during the operation on a screen or, thanks to recent efforts, with Augmented Reality (AR) glasses (HoloLens). Apart from the time and manual efforts required for creating the said preoperative plans, such plans are often static, meaning that they cannot adapt to the unforeseen changes during the surgery, which are frequent.
The Research in Orthopedic Computer Science group at the Balgrist University Hospital is working on an augmented reality solution that aims to provide automatic, live, patient-specific and adaptive planning during surgery based on data collected during the procedure. These models are generated autonomously based on the images collected from the patients during the surgery and can adapt to the changes during the surgery. As an example use case, we are developing surgical planning algorithms for spinal fusion procedures in which metal implants are placed inside the patients’ spine with the goal of connecting the adjacent vertebrae. Due to the vicinity to the vital organs such as the spinal cord, the planning for this intervention is of utmost importance to define the exact location and trajectory of the implants.
Due to their complexity, orthopedic spine surgeries often require a complex preoperative planning to ensure optimal patient outcome. This planning procedure is often performed manually based on pre-acquired patient medical images (e.g, CT scans) through which the surgeons precisely define the desired surgical plans (e.g., where inside the anatomy should the surgical hardware be placed). Once the preoperative plan is available, it is visualized to the surgeon during the operation on a screen or, thanks to recent efforts, with Augmented Reality (AR) glasses (HoloLens). Apart from the time and manual efforts required for creating the said preoperative plans, such plans are often static, meaning that they cannot adapt to the unforeseen changes during the surgery, which are frequent. The Research in Orthopedic Computer Science group at the Balgrist University Hospital is working on an augmented reality solution that aims to provide automatic, live, patient-specific and adaptive planning during surgery based on data collected during the procedure. These models are generated autonomously based on the images collected from the patients during the surgery and can adapt to the changes during the surgery. As an example use case, we are developing surgical planning algorithms for spinal fusion procedures in which metal implants are placed inside the patients’ spine with the goal of connecting the adjacent vertebrae. Due to the vicinity to the vital organs such as the spinal cord, the planning for this intervention is of utmost importance to define the exact location and trajectory of the implants.
The goal of this masters’ thesis is to automate the planning of ideal drilling trajectories using machine learning algorithms. Based on a 3D model of the patient's spine generated during the surgery, the ideal drilling trajectory shall be calculated. You will investigate the feasibility of the emerging reinforcement methods for this purpose. As common when designing a reinforcement learning algorithm, your project will start by creating a virtual planning environment (i.e., game environment) in which an agent (planner) will be trained to place pedicle screws. Later, taking into account the clinical criteria, you will design a reward policy so that the agent can learn from its actions in the environment. As a final step, a comparison between automatically planned trajectories and human-generated plans shall be performed in order to evaluate the performance of the developed method.
Throughout the development of these algorithms, you will have direct communication access to your technical supervisors and peers with background in artificial intelligence and computer-aided surgery as well as medical experts in the field of orthopedic surgery.
The goal of this masters’ thesis is to automate the planning of ideal drilling trajectories using machine learning algorithms. Based on a 3D model of the patient's spine generated during the surgery, the ideal drilling trajectory shall be calculated. You will investigate the feasibility of the emerging reinforcement methods for this purpose. As common when designing a reinforcement learning algorithm, your project will start by creating a virtual planning environment (i.e., game environment) in which an agent (planner) will be trained to place pedicle screws. Later, taking into account the clinical criteria, you will design a reward policy so that the agent can learn from its actions in the environment. As a final step, a comparison between automatically planned trajectories and human-generated plans shall be performed in order to evaluate the performance of the developed method. Throughout the development of these algorithms, you will have direct communication access to your technical supervisors and peers with background in artificial intelligence and computer-aided surgery as well as medical experts in the field of orthopedic surgery.
Prof. Philipp Fürnstahl, Head of ROCS Group,
University Hospital Balgrist, Zurich, Switzerland;
philipp.fuernstahl@balgrist.ch
Dr. Hooman Esfandiari, PostDoctoral Fellow,
ROCS Group;
hooman.esfandiari@balgrist.ch
Sascha Jecklin, PhD Student
ROCS Group, University Hospital Balgrist, University of Zurich, Balgrist CAMPUS;
sascha.jecklin@balgrist.ch
Prof. Philipp Fürnstahl, Head of ROCS Group, University Hospital Balgrist, Zurich, Switzerland; philipp.fuernstahl@balgrist.ch
Dr. Hooman Esfandiari, PostDoctoral Fellow, ROCS Group; hooman.esfandiari@balgrist.ch
Sascha Jecklin, PhD Student ROCS Group, University Hospital Balgrist, University of Zurich, Balgrist CAMPUS; sascha.jecklin@balgrist.ch