Back to search

University of Leeds

Investigating Treatment Response in Glioblastoma Using radiomic Evaluation (INTRIGUE)

Abstract

dc:description.abstract

Glioblastoma is the most aggressive primary brain tumour; despite maximal oncological management, median survival is at most 17 months, rising to 19 months if there is methylation of 6-O-Methylguanine-DNA Methyltransferase (MGMT) promoter (15 months otherwise). Presently, imaging biomarkers (IBs) used for prognostic stratification at the time of pre-operative, and suspected, diagnosis of glioblastoma are limited, principally, to the characteristics of the tumour on T1-weighted post-contrast (T1CE) magnetic resonance imaging (MRI) and the presence or absence of multifocal lesions. There is also a paucity of clinically-applicable prognostic imaging biomarkers (IBs) that characterise the peritumoural tumour habitat of glioblastoma, which is not typically the target of surgical debulking. Radiomics, a quantitative high throughput approach to image analysis, combined with machine learning (ML) analysis, has shown great promise in non-invasively characterising the whole (enhancing and non-enhancing) tumour volume in glioblastoma. However, the translation of radiomics-based models has been hampered by several factors including data heterogeneity due to multi-centre MRI acquisition. This PhD thesis outlines three studies undertaken to address some of the concerns regarding translation barriers of multi-centre radiomics models in Glioblastoma. (1) A systematic review of the evidence surrounding intensity standardisation techniques (ISTs) of MRI prior to the extraction of radiomic features was conducted. (2) A resampling study was conducted to investigate tumour volume as a prognostic radiomic IB in Glioblastoma, and examine whether non-linear transformation or sample size might contribute to heterogeneous results in other studies. (3) A modelling study was conducted to investigate the impact of ISTs and also of ComBat, a statistical model for realigning multi-centre radiomic features, on the performance of prognostic models. This included assessment of model stability and calibration, which are typically not assessed in proposed Glioblastoma survival models. The main findings from the studies included: (1) Three techniques, WhiteStripe (WS), Nyul's histogram matching (HM) and Z-Score (ZS) were the most commonly applied ISTs in the studies ($n=12$) included in the systematic review. There was no consensus on the optimal IST. (2) In a multi-centre cohort of patients with glioblastoma (n=259), whole tumour volume (WTV) and tumour diameter were found to be prognostic of overall survival (OS) in multivariable Cox proportional hazards models. Log-transformation of WTV and increasing sample size increased the chances of detecting a prognostic relationship during the bootstrap resampling experiment. (3) Increased batch size for ComBat realignment improved discrimination, relative model fit and explained variation of clinical and radiomic prognostic models. However, the calibration accuracy and model stability deteriorated. HM and WS tended to improve discrimination, fit and explained variation. There was limited evidence from the published literature for an optimal IST. In our multi-centre dataset HM and WS tended to improve some model performance metrics but this was inconsistent and model stability and calibration were not improved. ComBat also improved prognostic model performance but required larger batch sizes, which discarded a large proportion of data in this heterogeneous real-world dataset, and degraded model calibration and stability. Resampling experiments also suggest that variation in sample size and ignoring the possibility of non-linear relationships could be two reasons that prognostic studies show inconsistent prognostic relationship for tumour size and this could also be the case for other radiomic IB discovery studies. Future work will focus on exploring prognostic radiomic IBs in large, multi-centre and heterogeneous imaging data and evaluate any potential IBs across multiple performance metric including stability and calibration.

Degree

thesis:*
Name dc:type.qualificationname
Ph.D
Level dc:type.qualificationlevel
doctoral
Grantor dc:publisher.institution
University of Leeds
Year dc:date.issued
2024

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Fatania, Kavi Sudhansu
Advisors dc:contributor.advisor
  • Short, Susan C.
  • Currie, Stuart
  • Mistry, Hitesh
  • O'Connor, James
ORCID person:identifier.orcid

Rights

Language dc:language.dcterms_iso639_2
eng

Identifiers

dc:identifier.*
Dc Identifier Grantnumber
203914/Z/16/Z

Chain of custody

source
Harvested from
White Rose University Consortium
Base URL
etheses.whiterose.ac.uk/cgi/oai2
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
related terms
citation

Fatania, Kavi Sudhansu. Investigating Treatment Response in Glioblastoma Using radiomic Evaluation (INTRIGUE). doctoral thesis, University of Leeds, 2024.