ESTRO 2025 - Abstract Book

S100

Invited Speaker

ESTRO 2025

Abstract:

Large Language Models (LLMs) have rapidly transformed multiple areas of healthcare, presenting particular promise within radiation oncology. This talk reviews the state-of-the-art advancements and emerging applications of LLMs within radiation oncology. First, we will explore LLMs as decision-support tools, focusing on their ability to encode radiation oncology knowledge and improve access to scientific evidence. Recent developments in reasoning models and LLM agents, such as those employed in deep research tasks, enable fast synthesis of available evidence, offering clinicians actionable insights grounded in the latest literature. Next, we explore the role of locally fine-tuned LLMs in automating documentation tasks, such as physician letter writing, ensuring both efficiency and compliance with data privacy standards. The integration of LLMs within existing Oncology Information Systems could provide streamlined clinical workflows, reduced administrative burden, and improved communication quality. Lastly, we address emerging LLM applications including standardizing RT structure names and automating parameter extraction from unstructured text, which could significantly accelerate radiation oncology research. Multimodal LLMs, interactive chatbots, and LLM-based avatars could enable transformative applications for patient interaction.

In summary, this review aims to provide a comprehensive overview of current advancements and future directions for leveraging LLM technology in radiation oncology.

4815

Speaker Abstracts Physics-informed neural networks for sCT generation and beyond Matteo Maspero Radiotherapy, UMC Utrecht, Utrecht, Netherlands

Abstract: Physics-informed neural networks (PINNs) offer a promising approach for improving the accuracy and applicability of synthetic CT (sCT) generation from MRI, overcoming the challenge of the lack of direct physical correlation between MRI and CT. This presentation investigates the use of PINNs to integrate physical priors, such as tissue density and X-ray attenuation models, to enhance the fidelity of sCT generation from MRI data. While MRI and CT rely on different physics principles, PINNs can incorporate known relationships from modalities like CBCT, where X ray attenuation is directly related to tissue composition, to guide the model in generating more accurate sCT images. Beyond sCT generation, we explore the potential of PINNs in other areas of medical imaging, including quantitative MRI, image registration, and radiotherapy treatment planning, where physical constraints can improve accuracy in dose calculation and treatment verification. The results highlight the potential of PINNs to integrate domain-specific knowledge, opening new avenues for multimodal imaging and personalized treatment in radiotherapy.