PRODIGIO · PROton therapy DIfferentiable Gradient-based Inverse Optimization

PRODIGIO (PROton therapy DIfferentiable Gradient-based Inverse Optimization) aims to develop the first uncertainty-aware auto-planning framework for Intensity-Modulated Proton Therapy (IMPT).This project addresses a critical gap in current treatment planning systems, which optimise only beam fluences while keeping orientations and energy layers fixed and manually chosen by the operator, and account for uncertainties through simplistic global margins. These limitations restrict plan quality, increase computation times, and lead to strong operator dependence.At its core, PRODIGIO will develop a deep-learning dose engine trained on Monte Carlo simulations, capable of producing millisecond-accurate dose maps. Differentiable estimates of motion- and density-related uncertainties will be incorporated using advanced methods such as polynomial chaos expansion. This engine will be embedded in a novel optimisation pipeline where fluences, energies, and orientations are simultaneously optimised with gradient-based algorithms that explicitly account for uncertainties.The project is highly interdisciplinary, positioned at the intersection of medical physics, numerical optimisation, and clinical practice. It involves three world-leading centres: the host institution (UniOvi), under the supervision of Pietro Vischia, an expert in differentiable optimisation and particle physics; a secondment at TU/e on differentiable uncertainty quantification; and a secondment at MAASTRO Clinic for clinical IMPT practice and validation.In summary, PRODIGIO will deliver a significant impact on both the researcher’s career and on cancer care. On the one hand, it will provide extensive multidisciplinary training, placing the researcher in a leading position in the field. On the other, it will deliver the first uncertainty-aware, operator-independent auto-planning system for proton therapy, with the potential to transform radiotherapy treatment planning in Europe and worldwide.