Djupinlärning för prediktion av maskinparametrar för strålbehandling

This thesis investigates the application of deep learning to automate and enhance Volumetric Modulated Arc Therapy (VMAT) treatment planning. Traditional VMAT planning relies on solving a dose-mimicking optimization problem to determine machine parameters, a process that is highly sensitive to initial conditions due to its non-convex nature. Commonly, initialization involves fluence map optimization (FMO) followed by leaf conversion, which often results in suboptimal machine parameters for VMAT. Recent studies in deep learning for VMAT machine parameter prediction have explored architectures like 3D U-Nets and encoder-decoder models to directly predict MLC leaf positions and monitor units (MUs). While promising, these approaches face challenges such as data scarcity and limited generalizability. The reliance on 3D inputs, which require large datasets of segmented anatomical structures or dose distributions, has restricted their applicability to single cancer types, often necessitating separate models for each site. The proposed method employs a 2D U-Net model to predict multi-leaf collimator (MLC) apertures directly from beam’s eye view (BEV) dose slices. The predicted apertures are then used to initialize MU-only optimization, producing machine parameters for dose computation or further optimization. This thesis addresses four key research questions related to mitigating data scarcity challenges, assessing model generalizability across cancer sites, and evaluating the resulting plans in terms of dose conformity and optimization efficiency. The methodology involved curating a dataset of breast and prostate cancer plans, implementing a tailored feature extraction process, and training the 2D U-Net model. Model performance was evaluated using metrics such as Dose Mean Absolute Error and Gamma Passing Rates (3%/3 mm). Additionally, the study analyzed optimization efficiency by monitoring objective value curves during dose-mimicking optimization. The results demonstrate that the proposed approach significantly improves dose conformity compared to FMO-based initialization. Plans initialized with predicted MLC apertures achieved higher gamma passing rates and lower dose MAE. The General model, trained on combined breast and prostate data, effectively predicted machine parameters for both cancer sites, demonstrating its capacity to generalize across diverse anatomies. This directly tackles the reliance on 3D model inputs in recent studies, which demand extensive datasets and have limited adaptability across multiple cancer sites. Furthermore, analysis of objective value curves suggests that the proposed initialization method could potentially reduce the number of optimization iterations needed to achieve clinically acceptable dose distributions, improving planning efficiency. While the study is limited to generated, non-clinically approved plans and focuses on a single treatment machine, the findings highlight the potential of this approach to set new benchmarks in VMAT planning. Future research should prioritize validation on clinical plans, extending the method to additional cancer sites and multi-arc treatments, and conducting comparative studies with recent 3D deep learning approaches. This thesis contributes to the growing body of research exploring AI- driven radiotherapy solutions, providing a foundation for more accurate, efficient, and adaptable VMAT treatment planning systems.