Radiomics enhanced machine learning-based classifier to improve survival estimation in glioblastoma multiforme

Duman, Abdulkerim 2025. Radiomics enhanced machine learning-based classifier to improve survival estimation in glioblastoma multiforme. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF (Thesis) - Accepted Post-Print Version Download (29MB) | Preview
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (197kB)

Abstract

In the field of neuro-oncology, precision oncology is advancing to improve patient survival outcomes. Radiomics, mainly through the use of standardised engineered (hand-crafted) features, has recently been utilised in neuro-oncology research and holds potential as biomarkers in the diagnosis and treatment planning of glioblastoma multiforme (GBM). However, the use of multiparametric Magnetic Resonance Imaging (mpMRI) data and the inclusion of datasets from multiple institutions pose substantial challenges for reproducibility. Therefore, establishing a standardised preprocessing pipeline and developing interpretable radiomic models could enhance the transition of radiomic studies in clinical settings. This thesis investigated the optimisation of preprocessing pipelines to enhance reproducibility. The research addressed artefacts in registration and resampling on a widely used preprocessing pipeline from the Brain Tumor Segmentation Challenge (BraTS) and proposed an optimised version of this pipeline. For our domain-specific dataset (STORM_GLIO) with clinically defined contours used in radiotherapy treatment planning, we designed a preprocessing pipeline that excludes registration to a comprehensive Magnetic Resonance Imaging (MRI)-based reference of normal adult human brain anatomy and integrates a state-of-the-art brain extraction tool to improve accuracy and consistency. The proposed pipeline was assessed by our clinicians. Results demonstrated that the widely adopted preprocessing pipeline can be reliably reproduced through these optimisations, thereby ensuring consistency with our domain-specific clinical requirements. In addition, a resource-efficient strategy, Region-Focused Selection Plus (RFS+), for enhanced automated tumour segmentation was implemented using state-of-the-art models to improve generalisability. By introducing weighted ensemble learning alongside different normalisation techniques (such as Z-score and Nyul), RFS+ enhanced segmentation performance and model generalisability when the model training process utilised three segmentation approaches (Multi-label, Binary class, and Multiclass), incorporating tumour-specific characteristics such as overlapping and non-overlapping regions. Also, the strategy demonstrated competitive results, resource-efficiency, flexibility, as it can be applied to different models, including U-Net and nnU-Net. The radiomic analysis studies focused on evaluating the effectiveness of using a limited number of radiomic features (RFs) from MRI sequences, in accordance with current radiomic study guidelines. For the radiomic analysis, RFs were combined with a single clinical variable due to incomplete clinical information across datasets. These studies were developed for overall survival (OS) prediction in GBM under two different settings while maintaining model interpretability by limiting the feature set to a maximum of 10 features. First, on BraTS 2020 and RHUH-GBM datasets, which used the same contouring format from the BraTS Challenge and were pre-processed through the widely used pipeline, we developed a novel hybrid feature selection method (LASSO-PSO). LASSO-PSO boosted radiomic model performance and achieved generalisable, state-of-the-art results, supported by external validation. Second, a radiomic model was developed using as few as two robust and reproducible RFs since many RFs extracted from the BraTS 2020 and STORM_GLIO datasets showed higher instability. This instability stemmed from differences in preprocessing pipelines and contouring formats across datasets. The radiomic model achieved moderate C-index performance when utilising a single contour and MRI sequence, suggesting potential applicability across different clinical challenges and limitations.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Engineering
Uncontrolled Keywords:	1. magnetic resonance imaging (MRI) 2. radiomics 3. machine learning 4. clinical applications 5. glioblastoma multiforme 6. brain tumour segmentation
Date of First Compliant Deposit:	29 September 2025
Last Modified:	29 Sep 2025 12:54
URI:	https://orca.cardiff.ac.uk/id/eprint/181369

Actions (repository staff only)

Edit Item