dc.contributor.advisor | Koh D-M | |
dc.contributor.author | Kalantar, R | |
dc.contributor.editor | Koh, D-M | |
dc.date.accessioned | 2024-07-25T13:36:07Z | |
dc.date.available | 2024-07-25T13:36:07Z | |
dc.date.issued | 2024-07-25 | |
dc.identifier.citation | 2024 | en_US |
dc.identifier.uri | https://repository.icr.ac.uk/handle/internal/6316 | |
dc.description.abstract | Cervical cancer is a threatening global health concern for women, typically treated with radiation therapy (RT) in advanced stages, traditionally planned using computed tomography (CT) scans. However, advancements in medical imaging, particularly the incorporation of magnetic resonance imaging (MRI), have offered enhanced clarity in RT treatments due to superior soft-tissue contrast, exemplified by MRI-based linear accelerators (MR-Linac or MRL). Despite such advancements, manual delineation of regions of interest (ROIs) in treatment plans remains a limitation due to its time-consuming nature and susceptibility to variabilities. Addressing this, this thesis explores the synthesis of MRL images from CT scans through a patch-based cycle-consistent generative adversarial network (Cycle-GAN), aimed at enhancing the segmentation of organs-at-risk (OARs) for MRgRT. By leveraging an existing CT-based repository (n=212) at the Royal Marsden Hospital (RMH), this methodology broadens the training dataset significantly, achieving improved segmentation outcomes with acceptance rates of up to 100% and 98%, as score by two radiation oncologists, for treatment planning on MRL data. Furthermore, segmentation challenges due to the intricacies in the tumour's appearance and its unclear boundaries necessitate robust models and extensive datasets for efficient learning of the representation of disease patterns. This research, therefore, employs a transfer learning approach to enhance the segmentation of cervical cancer by applying semantic knowledge from a substantial external dataset (n=206) to a smaller, in-house dataset (n=21), demonstrating the viability and efficacy of this technique in boosting segmentation accuracy. A novel multi-head architecture, utilising b1000 diffusion-weighted images (DWI) and apparent diffusion coefficient (ADC) maps, was introduced, indicating improved performance in segmentation tasks and yielding a Dice similarity coefficient (DSC) of 0.82, compared to 0.79 from conventional multi-channel training. Moreover, a service evaluation employing magnetic resonance fingerprinting (MRF) was conducted to discern the changes in T1 and T2 relaxation times before and after treatment in 7 patients, revealing no significant shifts in values across the general population but significant changes in 4 individual cases. The contours generated by the segmentation algorithm were in strong agreement with the expert-defined contours. Although these findings are preliminary, they lay the groundwork for integrating quantitative MRI with AI for enhanced characterisation of pelvic malignancies. In conclusion, this thesis makes pivotal contributions towards refining medical image segmentation technologies and methodologies, offering advanced representation learning to bridge semantic differences between varying diseases and acquisition protocols. These developments pave the way for future research, expanding the prospects of integrating more refined and accurate AI-driven technologies in clinical decision-making processes and applications beyond pelvic malignancies. | |
dc.language.iso | eng | en_US |
dc.publisher | Institute of Cancer Research (University Of London) | en_US |
dc.rights.uri | https://www.rioxx.net/licenses/all-rights-reserved | en_US |
dc.title | Deep Learning for Automatic Segmentation, Treatment Planning and MRI Quantitative Analysis of Cervical Cancer | en_US |
dc.type | Thesis or Dissertation | |
dcterms.accessRights | Public | |
dc.date.updated | 2024-07-25T13:34:56Z | |
rioxxterms.version | AO | en_US |
rioxxterms.licenseref.uri | https://www.rioxx.net/licenses/all-rights-reserved | en_US |
rioxxterms.licenseref.startdate | 2024-07-25 | |
rioxxterms.type | Thesis | en_US |
pubs.organisational-group | ICR | |
pubs.organisational-group | ICR/Primary Group | |
pubs.organisational-group | ICR/Primary Group/ICR Divisions | |
pubs.organisational-group | ICR/Primary Group/ICR Divisions/Radiotherapy and Imaging | |
pubs.organisational-group | ICR/Primary Group/ICR Divisions/Radiotherapy and Imaging/Computational Imaging | |
pubs.organisational-group | ICR/Students | |
pubs.organisational-group | ICR/Students/PhD and MPhil | |
pubs.organisational-group | ICR/Students/PhD and MPhil/19/20 Starting Cohort | |
icr.researchteam | Computational Imaging | en_US |
dc.contributor.icrauthor | Kalantar, Reza | |
uketdterms.institution | Institute of Cancer Research | |
uketdterms.qualificationlevel | Doctoral | |
uketdterms.qualificationname | Ph.D | |
icr.provenance | Deposited by Mr Barry Jenkins (impersonating Mr Reza Kalantar) on 2024-07-25. Deposit type is initial. No. of files: 1. Files: R Kalantar PhD_Thesis.pdf | |
dc.type.qualificationlevel | Doctoral | |
dc.type.qualificationname | Ph.D | |