Enhancing Proton Therapy Planning Through Uncertainty Quantification Maps

Abstract

This project discusses the evaluation of a state-of-the-art deep learning-based model for improving the efficacy of the preparation of a proton therapy treatment. This technique is known due to its precision in targeting tumoral areas, which requires a thorough planning to minimize damage of healthy surrounding tissues. The importance of Artificial Intelligence in medical imaging led to the development of the nnUNet model, considered the gold standard for image segmentation. Thus, this work focuses on the study of this model to better understand the process of automatic segmentation of critical organs in CT examinations compared to a ground truth of segmentation by radiation oncologists, by performing a full study on the results. Initial evaluations were conducted, ending up with a Dice Coefficient of 0.810 and Hausdorff Distance of 3.381mm, which led to an uncertainty quantification study to assess the robustness of the model with the Test Time Augmentation method. This study concluded that adaptive architectures are clinically more effective than complex deep learning models. Due to the robustness nature of the nnUNet model, its usability for proton therapy planning is proven.