ESTRO 2025 - Abstract Book

S93

Invited Speaker

ESTRO 2025

4802

Speaker Abstracts Current status of AI autosegmentation in clinical RT workflows Rajesh Jena Department of Oncology, University of Cambridge, Cambridge, United Kingdom

Abstract:

Autosegmentation of organs at risk for radiotherapy planning has been a leading example of the use of deep learning in cancer care, with multiple research and commercial solutions for AI based autosegmentation (AIBAS) in existence around the world. In this session, we will briefly review the motivation for introducing AIBAS to the clinic as an example of a non-diagnostic workflow acceleration tool. We review the strong alignment between AIBAS and computer vision research that led to the current state-of-the art tools in the clinic, and consider the nature of clincial evaluation that is needed to demonstrate utility and understand it's role in the overall quality management of the radiotherapy treatment workflow.

We review the issue of post-marketing suveillance of AIBAS, and some of the important issues arising from implementation such as automation and anchoring bias and direct attenion fatigue.

Finally we consider the curent state of the art for tumour autosegmentation, which remains at present in the research setting.

4803

Speaker Abstracts Automation in treatment planning: approaching reality? Stine Korreman Danish Center for Particle Therapy, Aarhus University Hospital, Aarhus, Denmark

Abstract:

With developments in automation and Artificial Intelligence over the recent years, solutions are becoming available for all steps of the treatment planning process. While autosegmentation is most mature, development is also promising for image reconstruction, dose prediction, plan generation and dose calculation. Bringing together all these steps, in principle we could enable automation of the entire chain, getting directly from scan to plan in one automated process. In connection with the ESTRO physics workshop 2023 “AI for the fully automated radiotherapy treatment chain”, a challenge was conducted to test a fully automated scan-to-plan approach for prostate cancer. Several of the entries in the challenge – including the winner - fulfilled the criterion of being a fully automated procedure with no human intervention and no human oversight required. This suggests that we technically are at a stage where full automation of the treatment planning process is indeed possible. In this presentation, we will discuss the potential of this future perspective, as well as limitations, obstacles, and risks.

We will also touch on further perspectives and opportunities, including integration of generative AI, multi-scenario exploration, and interpretability of AI predictions.