Virtual patient-specific treatment verification using machine learning methods to assist the dose deliverability evaluation of radiotherapy prostate plans

Quintero Mejía, Paulo Alejandro

Physics
October 2022

Thesis or dissertation


Rights
© 2022 Paulo Alejandro Quintero Mejía. All rights reserved. No part of this publication may be reproduced without the written permission of the copyright holder.
Abstract

Machine Learning (ML) methods represent a potential tool to support and optimize virtual patient-specific plan verifications within radiotherapy workflows. However, previously reported applications did not consider the actual physical implications in the predictor’s quality and model
performance and did not report the implementation pertinence nor their limitations. Therefore, the main goal of this thesis was to predict dose deliverability using different ML models and input predictor features, analysing the physical aspects involved in the predictions to propose a
reliable decision-support tool for virtual patient-specific plan verification protocols.
Among the principal predictors explored in this thesis, numerical and high-dimensional features based on modulation complexity, treatment-unit parameters, and dosimetric plan parameters were all implemented by designing random forest (RF), extreme gradient boosting (XG-Boost), neural networks (NN), and convolutional neural networks (CNN) models to predict gamma passing rates (GPR) for prostate treatments. Accordingly, this research highlights three principal findings. (1) The dataset composition's heterogeneity directly impacts the quality of the predictor features and, subsequently, the model performance. (2) The models based on automatic extracted features methods (CNN models) of multi-leaf-collimator modulation maps (MM) presented a more independent and transferable prediction performance. Furthermore, (3) ML algorithms incorporated in radiotherapy workflows for virtual plan verification are required to retrieve treatment plan parameters associated with the prediction to support the
model's reliability and stability. Finally, this thesis presents how the most relevant automatically extracted features from the activation maps were considered to suggest an alternative decision support tool to comprehensively evaluate the causes of the predicted dose deliverability.

Publisher
Department of Physics, The University of Hull
Qualification level
Doctoral
Qualification name
PhD
Language
English
Extent
12 MB
Identifier
hull:18747
QR Code