MSc Thesis: Evaluating LLM-based Agentic Systems for RT Plan Optimization

RaySearch develops innovative software solutions to improve cancer care. Over 1200 clinics in more than 54 countries use RaySearch software to improve treatments and quality of life for patients. RaySearch was founded in 2000 and is listed on Nasdaq Stockholm. The headquarters is located in Stockholm, with subsidiaries in the US, Europe, Asia and Australia & New Zealand. Today we are more than 485 employees with a common vision of improving cancer care with innovative software. Our great staff is crucial for our success and we offer a fantastic working environment in modern offices, flexibility and good opportunities for development. We believe in equal opportunities, value diversity and work actively to prevent discrimination. Can an AI agent create radiation therapy plans like an experienced dosimetrist? In this master thesis you will work with large language model (LLM) agents driving treatment plan optimization and measure how well they actually perform. Project description In this project, you will explore both quantitatively and qualitatively how well an agentic AI system can run and automate the optimization process: the agent reviews the results, updates the optimization problem or the automatic optimization workflows and repeats to improve the treatment plan. You will build an agentic harness that connects an LLM agent to RayStation. You will create its tools to let the agent read plans, adjust objectives and run optimizations. Your main tasks - Review research on LLM agents and automated treatment planning, and define the research questions and evaluation criteria. - Work on an agent prototype that reads plans, adjusts optimization objectives and deep learning planning model settings, and runs optimizations in RayStation. - Create a benchmark of treatment planning tasks to evaluate the agent. It includes patient cases, reference plans and scoring based on clinical goals, dose-volume metrics, number of iterations and time. - Design and compare agent strategies, such as step-by-step parameter sweeps versus free reasoning, across different LLMs. - Run the agent on a set of patient cases and compare its plans with reference plans. Use clinical goals, dose-volume metrics, number of iterations and time. - Review the agent's decisions with medical physicists, and present the results in a thesis report. Your Profile We are looking for a curious master's student who wants to work where AI meets cancer care. You enjoy combining programming with careful experiments, and you look critically at results instead of taking them at face value. You are independent and well organized, and you can explain your findings clearly to both software engineers and medical physicists. - Ongoing master's studies in engineering physics, medical physics, computer science, applied mathematics or a related field - Coursework or project experience in optimization and/or machine learning - Good programming skills in Python - Hands-on experience with LLMs, for example prompting, tool calling or building agents through APIs - Fluent written and spoken English Application Please apply to the position through the link below. Selection and interviews will be ongoing. We do not accept applications by e-mail. Application Link.