ESTRO 2026 Congress Report I Physics Track 

At the ESTRO 2026 congress, a prominent highlight was the "Best Physics Paper" awarded to the multi-institutional team behind the SynthRAD2025 Grand Challenge. Their presentation, titled Benchmarking synthetic computed tomography for radiotherapy from head to abdomen: SynthRAD2025 Grand Challenge, outlined critical advancements in artificial intelligence (AI) and its direct, transformative application to synthetic CT (sCT) generation. 

Historically, acquiring conventional CTs for dose calculation exposes patients to ionising radiation, offers limited soft-tissue contrast, and introduces registration uncertainties when images are aligned with MRI or CBCT. Generating sCT directly from MRI or CBCT addresses these limitations head-on, enabling seamless MRI-only and CBCT-based adaptive workflows. Building upon its 2023 predecessor, SynthRAD2025 pushed the field's boundaries by tackling significantly more complex regions: the head and neck, thorax, and abdomen. These sites are notoriously challenging to model due to respiratory motion, intricate geometries, and high inter-patient variability. 

Presented onsite by Matteo Maspero (UMC Utrecht) on behalf of the ten organisers, the challenge utilised a massive dataset of 2,362 patients curated from five European medical centres. The benchmark was split into two tasks. Task 1 focused on MRI-to-CT conversion (890 cases) to facilitate MRI-guided treatments. Meanwhile, Task 2 targeted CBCT-to-CT conversion (1,472 cases) to support the growing need for rapid, daily adaptive radiotherapy workflows. 

The challenge attracted strong international participation, with 803 users testing various deep learning models. While the overall performance spread among the participating teams was large, the top-ranking submissions were remarkably close to one another. Notably, Task 2 (CBCT-to-CT) achieved better overall performance than Task 1 (MRI-to-CT), especially in the head and neck anatomy. Despite these task variations, the sCTs generated by the top teams across all three anatomies proved clinically acceptable for photon treatment planning. 

The presenters also demonstrated that while state-of-the-art models produced visually excellent sCTs—scoring highly on traditional image similarity metrics like SSIM—these metrics showed only a moderate correlation with actual dosimetric accuracy. This insight underscores a paradigm shift: image quality alone is an insufficient surrogate for treatment safety. A comprehensive dosimetric evaluation is strictly necessary before clinical translation. Finally, architectural trends for these deep learning models were highlighted. 

The SynthRAD2025 dataset and challenge framework will remain open, ensuring that future algorithms can be benchmarked consistently up to 2030. The platform and comprehensive results can be accessed online (https://synthrad2025.grand-challenge.org/), and the full report is available as a preprint (Viktor Rogowski et al. 2026, arXiv:2605.13555). 

 

Caption 

SynthRAD2025 + DLinRT.eu + COBRA2026 Grand Challenge merged to become the clan Overfitting Fjords for one night. Organisers of different initiatives channelling their inner Vikings in Stockholm to celebrate the success of SynthRAD2025. Because nothing fosters academic collaboration quite like tasting mead over dinner. 

 

 

Matteo Maspero 
UMC Utrecht, The Netherlands 

m.maspero@umcutrecht.nl 

https://www.linkedin.com/in/matteo-maspero/ 

 
Caption 

The SynthRAD2025 Grand Challenge, where participants were requested to develop and submit for a thorough evaluation a synthetic CT either from MRI or CBCT from head and neck, thorax and abdomen patients undergoing radiotherapy.