Ocampo, Jeremy;
(2023)
Designing Convolutional Neural Networks for Scintillation Photography and General Applications.
Doctoral thesis (Ph.D), UCL (University College London).
Preview |
Text
thesis_jeremyocampo_corrected.pdf - Other Download (29MB) | Preview |
Abstract
Modern radiotherapy treatments provide complex dose distributions which are difficult to measure and verify. This is due to beams having high-dose gradients, timevarying intensity and sizes reducing to millimetre scales. A dosimeter was designed to provide improvements over standard detectors that are unsuitable for measuring complex small fields. The proposed detector system consists of an irradiated scintillator sheet that is photographed, from which the dose is reconstructed. This provides a cheap, fast, and high-resolution solution. But scintillation images come with a variety of visual artefacts that need correcting. Convolutional neural networks (CNN) have been shown to have excellent accuracy in extracting useful information from noisy images. This requires thousands to millions of training images. In scintillation photography, there is not enough data to achieve an acceptable performance for CNNs. A novel method using domain randomisation was used to solve this issue, where thousands of images were simulated with varying physical parameters. This data can be used to train the CNN to be robust to visual artefacts. These CNNs are designed to assist in our image processing by predicting relative dose distributions and (un)known physical parameters, which gives confidence that the measured images are correct. Results showed that CNNs performed better than classical methods and could provide dose distributions that are suitable for routine QA. This work was extended by designing a novel CNN layer which can be generalised to non-Euclidean domains while maintaining scalability, e.g. the sphere which has many applications. This is done by developing modern methods in group convolutions and helping them scale to high resolution. This method leverages symmetries in the data, which improves the CNNs ability to generalize to “unseen” data with only a few thousand training examples. The models were tested on spherical data benchmarks for which state-of-the-art performance was achieved.
| Type: | Thesis (Doctoral) |
|---|---|
| Qualification: | Ph.D |
| Title: | Designing Convolutional Neural Networks for Scintillation Photography and General Applications |
| Open access status: | An open access version is available from UCL Discovery |
| Language: | English |
| Additional information: | Copyright © The Author 2023. Original content in this thesis is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) Licence (https://creativecommons.org/licenses/by-nc/4.0/). Any third-party copyright material present remains the property of its respective owner(s) and is licensed under its existing terms. Access may initially be restricted at the author’s request. |
| UCL classification: | UCL UCL > Provost and Vice Provost Offices > UCL BEAMS UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Maths and Physical Sciences UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Maths and Physical Sciences > Dept of Physics and Astronomy |
| URI: | https://discovery.ucl.ac.uk/id/eprint/10173798 |
Archive Staff Only
 |
View Item |
