Implementation of probabilistic robust optimization with contour propagation uncertainty in adaptive radiotherapy

Details

Supervisors

Sterpin, Edmond

Degree label

Master [120] en sciences physiques, à finalité spécialisée: physique médicale

Abstract

Background : Radiotherapy aims to effectively target tumor volumes with adequate doses while preserving nearby healthy cells. It is challenged by the dynamic nature of patient anatomy, leading to significant uncertainties in the position of structures, which impact treatment quality. Traditional approaches, such as using conservative margins to account for positional errors, tend to be overly conservative. Adaptive radiotherapy provides a dynamic approach, adjusting to anatomical changes. However, it requires significant time and resources for frequent re-contouring structures for each optimization phase. Methods : The use of Deformable Image Registration (DIR) through a neural network significantly speeds up the contouring process. Uncertainties related to these DIRs can be taken into account to form probability maps (PM) of the positions of the structures. We have implemented a fully probabilistic optimization method which weights the optimization cost function according to the probability of a setup error (the only error considered in this work) and considers probabilistic masks of structures. We compared the fully probabilistic workflow with different optimization methods: classic and worst case and with different structures definition. We also simulate classic non-adaptive workflow. Results : The classic workflow without optimization shows its limitations by not fulfilling 5 of the 14 constraints/objectives. The structure propagation appears to be validated as it resembles the structures segmented by physicians. The fully probabilistic method appears promising, delivering almost always the lowest doses to organs at risk. However, the tumor itself is not irradiated enough, indicating a need to modify weights, monitor units and indicating the difficulty of weighting constraints for PM targets. Conclusion : The outcomes confirm the applicability of a neural network in facilitating unsupervised structure propagation. Probabilistic optimization seems to work correctly but should be used with different constraint/objective weights compared to other methods. This is because PM also generate a weighting of dose constraints and objectives.