Automatic Generation of Patient-specific Gamma Knife Treatment Plans for Vestibular Schwannoma Patients

Abstract: In this thesis a new fully automatic process for radiotherapy treatment planning with the Leksell Gamma Knife is implemented and evaluated: First, a machine learning algorithm is trained to predict the desired dose distribution, then a convex optimization problem is solved to find the optimal Gamma Knife configuration using the prediction as the optimization objective. The method is evaluated using Bayesian linear regression, Gaussian processes and convolutional neural networks for the prediction. Therefore, the quality of the generated treatment plans is compared to the clinical treatment plans and then the relationship between the prediction and optimization result is analyzed. The convolutional neural network model shows the best performance and predicts realistic treatment plans, which only change minimally under the optimization and are on the same quality level as the clinical plans. The Bayesian linear regression model generates plans on the same quality level, but is not able to predict realistic treatment plans, which leads to substantial changes to the plan under the optimization. The Gaussian process shows the worst performance and is not able to predict plans of the same quality as the clinical plans.