|
Title
|
|
|
|
Exploring machine/deep learning applications for automation and optimization of patient individualized carepaths in radiotherapy
| |
|
Author
|
|
|
|
| |
|
Abstract
|
|
|
|
| The first part of this thesis focused on how Artificial Intelligence (AI) models can be correctly deployed in clinical environments with proper guidance on quality assurance (QA) for implementation and continuous monitoring. Given their unique role as a bridge between the clinical environment and new technologies, medical physicist experts (MPEs) are most likely at the frontier to implement these automated systems improving efficiency, quality, standardization, and acceleration of the workflow. However, as AI is a new field in radiotherapy (RT) and, no proper guidance was offered to assist MPEs in this complicated task, this work tried to fill in some of the current gaps by contributing to consensus guidelines in clinical implementation of AI in RT. The second part of this thesis illustrates challenges related to QA for AI models, more specific for auto-segmentations. The output of the contouring in RT can affect the subsequent steps in the workflow with an impact on the overall treatment. With the current implementation of adaptive radiation therapy, where treatments of patients are adapted to daily variations in anatomy, manual verifications of contours are no longer feasible to sustain an optimized workflow. The proposed method, which performs an automated secondary, independent verification of the auto-delineation, attempts to streamline the QA process. Other tools have been suggested already for QA of AI models, such as uncertainty quantification (quantifying the risk of failure), but in this thesis, a more pragmatic approach is proposed based on the existing methodology of independent verification. In the last part, AI solutions for QA procedures in RT were highlighted. Patient-specific QA needs to be performed to check if the treatment plan is suitable for irradiation, which is considered as a laborious and time-consuming task. To automate this procedure, an AI method was developed to reduce the burden of measurements and to optimise the analysis of the QA results. In the final chapter, we explored the potential of AI to assist in fundamental dosimetry. For Ultra-High-Dose Rate (UHDR) electron beams as a case of point, AI was used to predict the correction factor for recombination in ionization chambers for varying dose rates. The latter can not only be used to optimize the experimental results required for accurate dose assessments in UHDR FLASH RT, but can also help in optimizing the theoretical models behind recombination effects in different dosimeters. |
| |
|
Language
|
|
|
|
English
| |
|
Publication
|
|
|
|
Antwerp
:
University of Antwerp, Faculty of Medicine and Health Sciences, Centre for Oncological Research
,
2023
| |
|
Volume/pages
|
|
|
|
189 p.
| |
|
Note
|
|
|
|
:
Verellen, Dirk [Supervisor]
| |
|
Full text (open access)
|
|
|
|
| |
|