Monte Carlo and machine learning approach to in-vivo transmission dosimetry for dynamic radiation treatments

Abstract

In this thesis the main aim was to improve patient-specific quality assurance (QA) in external beam radiotherapy for dynamic techniques such as intensity-modulated radiotherapy (IMRT) and volumetric-modulated arc therapy (VMAT) through the development of in vivo transmission dosimetry methods using the electronic portal imaging device (EPID). Five studies addressed interrelated challenges, ranging from the modelling of the radiation beam and dose calculations in the patient to the determination of post-treatment volume energy fluence and the prediction of transmission EPID images. In Study I an automated Monte Carlo (MC) system for pretreatment QA of VMAT, incorporating patient geometry and DICOM-compliant workflows was developed. A 3D dose comparison analysis software was developed and implemented together with diagnose-specific tolerance criteria for DVH parameters and 3D dose differences. Study II analysed MC-calculated lung dose sensitivity to tissue characterization during deep inspiration breath-hold (DIBH) for breast cancer treatments, identifying that very low-density lung regions require precise CT number to density conversion and tissue segmentation to avoid dose inaccuracies. Study III investigated the response of the CC13 ionization chamber under non-reference conditions, demonstrating its limited utility for narrow static fields but feasibility for dynamic, small-field geometries, provided measured and calculated doses agreed within 3% tolerance. Studies IV and V focused on EPID-based in vivo dosimetry. Study IV developed a hybrid MC and deep learning (DL) framework using U-Net architectures to predict EPID images from MC-generated exit phase space energy fluence data, demonstrating preliminary feasibility. Study V quantified EPID sensitivity to positional and material variations in a thorax phantom using gamma analysis and pixel difference histograms. While VMAT plans showed higher gamma passing rates than IMRT under positional errors, detection efficacy varied depending on error type and plan complexity. Collectively, this work proposes a strategy for in vivo EPID dosimetry combining MC-derived energy fluence predictions with DL-based image generation. The results highlight limitations in conventional QA methods, particularly for errors undetected by integrated dosimetry, and empha-size the need for robust error detection through post-patient energy fluence analysis. The devel-oped tools and methodologies propose a base for future work and clinical implementation.