Deep learning emulation of physical processes in radiation therapy

Emerging and innovative approaches in Radiation Therapy (RT)—including microbeams, carbon ions, and FLASH-RT—have the potential to significantly improve cancer treatment, leading to substantial clinical and social impact. However, current methodologies for dose calculation and treatment optimization may slow down or limit the adoption of these new techniques. Deep Learning (DL), which has emerged over the past decade as a powerful tool capable of revolutionizing entire fields of study, holds the promise to both improve existing methodologies and facilitate the development of novel approaches. This thesis explores the potential of DL algorithms, specifically Graph Neural Network (GNN) architectures, to emulate physical processes across various use cases in novel RT modalities. These use cases include emulating dose or energy deposition in patients undergoing external beam RT as a function of beam parameters, as well as the development of hybrid low-energy nuclear interaction models. Dose emulation was tested for two radiotherapy modalities currently at the research stage: Very High Energy Electron RT and Microbeam Radiation Therapy. In both of these modalities, a fast and reliable dose computation and treatment planning systems remain unmet needs. GNNs were trained to emulate Monte Carlo (MC) simulated dose distributions and demonstrated the ability to quickly and accurately compute dose distributions both in simple materials and patients’ computed tomographies. The potential of DL-based dose engines as a basis for treatment planning optimization was also explored. Leveraging the intrinsic differentiability of DL models, the feasibility of a differentiable, gradient-based plan optimization approach was investigated in the context of Very High Energy Electron RT. Although limited by computational constraints, the developed demonstrator has the potential to optimize all beam parameters simultaneously as continuous degrees of freedom. Finally, the construction of hybrid—classical and DL—nuclear reaction models was explored in the context of carbon ion therapy. Extensive literature has highlighted the limitations of nuclear interaction models in commonly used MC codes for simulating beam-patient interactions, while the most reliable models are too slow for practical application. By using physics-informed neural networks to emulate the slowest sections of the classical models, two hybrid models were developed, interfaced with classical simulations, and tested. These models demonstrated their ability to reconstruct major physical observables and have the potential to accelerate or improve classical models. Overall, this work demonstrates the potential of using deep learning models to emulate various physical processes in radiation therapy, achieving the necessary accuracy, enabling substantial speed-ups, and paving the way for new opportunities in treatment and model optimization.