ESTRO 2025 - Abstract Book

S85

Invited Speaker

ESTRO 2025

Abstract: Achieving high quality radiotherapy plans requires precise dose calculation and dose planning. Dose calculation methods provide solutions of varying accuracy to the Boltzmann particle transport equation, ranging from computationally expensive gold standard Monte Carlo and deterministic solvers to less accurate, but faster methods like pencil beam algorithms. Once the dose distribution from the many possible individual small beamlets (also called the dose deposition matrix) is known, dose planning can be performed such that a clinically optimal treatment plan can be constructed, consisting of machine settings (gantry angles, multi-leaf collimator (MLC) shapes and intensities for photons and beamlet intensities for protons/ions) that would deliver this desired dose to the patient. Dose planning can be further divided into a dose prediction task aiming to guess a dose distribution that has a clinically optimal balance between competing objectives of high target dose and low organ doses and is (close-to) physically achievable; an intensity/fluence optimization task yielding the intensities of the small beamlets that would result in the chosen balance between objectives; and a machine setting optimization task (sometimes also called dose mimicking) yielding the actual settings (MLC leaf positions, aperture intensities, pencil beam intensities) that are fully deliverable by the treatment machine, respecting all constraints like maximum leaf motion speed or minimum monitor unit values. As all of these tasks are time-consuming – especially in adaptive and online/real-time adaptive workflows – significant efforts have been put into their improvement. Artificial Intelligence (AI) approaches are increasingly driving these efforts. For dose calculation , the primary motivation is that AI can provide better accuracy vs. computational time compromise than traditional approaches, and might be the only feasible option in real-time settings. For dose prediction , the aim is to exploit AI‘s power to quickly guide planners towards achievable high quality plans and minimize time-consuming (and potentially futile) trial-and-error attempts. For both intensity/fluence optimization and machine setting optimization AI could potentially find better (i.e., closer to optimal) solutions faster than traditional optimizers, although these applications have been less explored so far. While AI tools offered by commercial vendors and developed in-house are increasingly used for these and other tasks in real clinical settings, their limitations and the uncertainties associated with their predictions are less understood. Although methodologically challenging, special AI methods that can yield estimates of their own uncertainty do already exist, but their applications to RT tasks has been limited. At the same time, uncertainties and how to handle them in the treatment planning workflow are nothing new to radiotherapy. Uncertainties related to different dose calculation methods, target and organ contouring, and treatment delivery (patient setup, image guidance, immobilization, etc.) have all been extensively studied. The current practice is to enlarge the true target into a Planning Target Volume (PTV) in photon and to use robust optimization explicitly incorporating potential error scenarios in proton/heavy ion treatment planning to deal with these uncertainties. While well established, both practices have some shortcomings, motivating the increasing use of accurate uncertainty quantification (UQ) methods to estimate quantitative statistical metrics (like expectation values or percentiles) in order to provide clinicians a more realistic assessment of the impact uncertainties on patient dose. The challenge UQ methods face is that they are typically computationally expensive and yield extensive amounts of data, which provides lot of insight but is not so easy to interpret and visualize for clinicians. These findings show that there is a clear need for more systematic and quantitative evaluation of uncertainties in radiotherapy AI models. In this presentation I will first give an overview of the current state-of-the-art in radiotherapy AI solutions focusing on dose calculation and dose prediction . Then, I will introduce some efficient and accurate uncertainty quantification methods and discuss the challenges of presenting statistical information for clinical decision making. I will also review methods aiming to quantify uncertainties of AI model predictions, as well as their limitations, finishing with some outstanding challenges of AI in radiation oncology.