ESTRO 2025 - Abstract Book

S101

Invited Speaker

ESTRO 2025

4816

Speaker Abstracts Foundation vision models for radiotherapy Ana Maria Barragan Montero MIRO, UCLouvain, Brussels, Belgium

Abstract:

The advent of foundation models—large-scale, pre-trained neural networks capable of adapting to diverse downstream tasks—has marked a paradigm shift in artificial intelligence. These models, initially designed for natural language processing, have demonstrated remarkable potential across domains, including computer vision and multimodal applications. In the context of radiation oncology, foundation vision models can further automate and enhance tasks where smaller AI models, such as UNet, are already used—e.g. segmentation, classification, and treatment planning—while also enabling AI to tackle more complex challenges. Compared to traditional AI models such as UNet or task-specific neural networks, foundation models excel in transfer learning, zero-shot learning, and adaptability. Their ability to generalize across tasks reduces the need for extensive task-specific training data, making them particularly valuable in medical imaging where labeled datasets are limited. Additionally, emerging multimodal models, which integrate information from multiple sources (e.g., text, images, audio), promise to enhance decision support in cancer care by combining imaging, clinical, and genomic data. However, deploying foundation models in clinical environments poses significant challenges. A critical question is whether such complex models are always necessary or if classical architectures can deliver sufficient performance with reduced computational and environmental costs. Moreover, ensuring model reliability, robustness, and uncertainty quantification remains essential in safety-critical applications. Trustworthiness, interpretability, and rigorous validation are vital to avoid clinical risks, which may become more challenging as model size increases. This talk will provide a comprehensive overview of foundation vision models and their transformative potential, with examples in cancer care and radiotherapy, alongside a balanced discussion of their limitations. We will explore future directions for integrating these models into clinical workflows while emphasizing the need for responsible and sustainable AI development.

4817

Speaker Abstracts Data science competitions as a pathway to innovative AI approaches in radiation oncology Kareem A. Wahid Imaging Physics, The University of Texas MD Anderson Cancer Center, Houston, USA

Abstract:

Data science competitions, i.e., structured online events where participants attempt to solve well-defined computational problems using publicly available datasets, have accelerated recent advancements in artificial intelligence (AI). By offering transparent evaluation frameworks and publicly accessible data, these competitions foster rigorous benchmarking, methodological innovation, and reproducibility, collectively driving rapid progress across diverse application domains, particularly within the field of computer vision. This impact is notably pronounced in healthcare settings, where annotated imaging datasets are scarce and costly to curate, making data science competitions vital platforms for collaborative advancement in medical AI. In radiation oncology, the