The ESTRO 2025 congress in Vienna provided a comprehensive perspective on AI's expanding role in radiotherapy. Curated by the AI in Radiotherapy (AI in RT) Focus Group, the event featured diverse sessions, including lectures, symposia, mini-orals, debates, and a community meetup. Key developments and challenges were highlighted, from technical advances to clinical implementation and regulatory considerations.

Key Scientific Themes

1. Foundation Models in Radiotherapy

The congress began with a discussion on foundation models, such as the Segment Anything Model (SAM) and MedSAM, which are being explored for automated segmentation in radiotherapy. Presenters emphasized the importance of robust validation, transparency, and uncertainty quantification before clinical implementation.

2. Clinical Implementation: Autosegmentation and Autoplanning

Several sessions highlighted mature deep learning applications in autosegmentation and autoplanning. Demonstrations included integrating AI-generated contours into treatment planning systems, visualizing inter-observer variability, and developing quality assurance (QA) dashboards. The focus was on enhancing consistency, safety, and efficiency, not just automation..

3. AI Guidelines and Quality Assurance

A key milestone was the introduction of ESTRO–AAPM AI guidelines, providing structured recommendations for AI deployment in clinical settings. Topics covered included automation bias, detection of out-of-distribution cases, and domain-specific “model cards” to document performance, limitations, and training datasets.

4. Communicating Uncertainty in AI Predictions

A symposium addressed communicating and managing AI-related uncertainty, balancing technical methods like predictive confidence maps with real-world usability. The importance of clinicians interpreting and acting on uncertainty data was emphasized.

5. Workflow Optimization

Real-world studies indicated that while AI reduces the time for tasks like segmentation, it doesn't automatically lead to more efficient workflows. The need for holistic integration, where AI contributes across the entire radiotherapy chain, was a recurring theme.

6. AI in MRI-Guided Radiotherapy (MRIgRT)

Multiple sessions covered AI's role in MRI-guided workflows, including auto-contouring, synthetic CT generation, online adaptation, and real-time tracking. The MESCAL project aimed to define clinical benchmarks for sCT generation in MRI-only workflows. The TrackRAD2025 challenge introduced a multicenter dataset and benchmarking platform for developing real-time tumor tracking algorithms using cine-MRI, paving the way for continuous tracking-based delivery.

Community Reflections and Takeaways

A well-attended Meet & Greet session revealed the community's optimism and enthusiasm for AI in radiotherapy, tempered by calls for better training, clearer guidelines, and stronger data sharing practices. The common message was that AI should enable more intelligent, adaptive, and patient-centered radiotherapy, not just speed.

Text inspired by the focus’group notes as well as the posts of our social media ambassadors Ana Maria Barragan Montero, Peter van Ooijen