%0 Thesis
%9 Doctoral
%A Pinnock, Mark A
%B Medical Physics & Biomedical Engineering
%D 2024
%E Barratt, Dean C
%E Hu, Yipeng
%E Bandula, Steve
%F discovery:10197643
%I UCL (University College London)
%K Medical imaging, Computed tomography, Deep learning, Convolutional neural networks, Super-resolution, Contrast enhancement, Interventional radiology
%P 179
%T Applied Deep Learning in  Interventional Computed Tomography
%U https://discovery.ucl.ac.uk/id/eprint/10197643/
%X Computed tomography (CT)-guided treatments for renal carcinoma offer a favourable complication profile and rapid recovery but have the drawback of exposing the patient to ionising radiation. In addition, iodinated contrast agents are commonly used, which come with a risk of allergic and nephrotoxic complications. Reducing the X-ray dose results in a lower risk to the patient, but increases noise and decreases spatial resolution, while avoiding contrast agents decreases the anatomical contrast, making needle insertion more challenging.    The purpose of this thesis is to propose deep learning-based techniques to solve the twin problems of super-resolution (SR) and synthetic contrast enhancement (sCE). A U-Net-type network is evaluated on SR, de-noising and partial volume effect correction and is shown to perform well on real interventional CT images after training on simulated data. Moving to sCE, a framework for generating multiple contrast phases is shown to be competitive with that of separately trained networks, and also shown to generalise to a continuous range of contrast phases. Lastly, a multi-task learning formulation is created to jointly optimise SR and sCE networks using vector quantization (VQ). An extensive ablation study demonstrates the efficacy of VQ in reducing blurring seen in some cases, and suggests architectural considerations that can make deep learning models more robust when jointly training on related tasks.    These techniques are all shown to be robust when being evaluated on out-of-distribution data containing features not seen during training. As many of the challenges addressed are ubiquitous in medical imaging, the contributions in this work can likely be carried forward to other applications and image modalities.
%Z Copyright © The Author 2024. 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.