Radiomics for survival prediction in non-small cell lung cancer patients undergoing radical-intent radiotherapy

Al Battat, Khamael 2026. Radiomics for survival prediction in non-small cell lung cancer patients undergoing radical-intent radiotherapy. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF (Thesis) - Accepted Post-Print Version Download (8MB) | Preview
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (168kB)

Abstract

Lung cancer remains the leading cause of cancer-related mortality worldwide, with non-small cell lung cancer (NSCLC) accounting for approximately 85% of cases. This thesis focused exclusively on NSCLC because it represents the predominant lung cancer subtype in clinical practice and because the available retrospective cohort predominantly comprised patients with histologically confirmed NSCLC treated with radical-intent radiotherapy. Restricting the analysis to NSCLC also avoided the additional biological and radiomic heterogeneity that would arise from combining fundamentally different lung cancer subtypes, such as small-cell lung cancer, which differ substantially in tumour biology, metabolic behaviour, treatment strategies, and prognosis. Consequently, a homogeneous NSCLC cohort was used to improve the reliability and interpretability of the radiomic analyses. Despite advances in imaging technology and therapeutic approaches, current staging systems inadequately capture the biological heterogeneity that drives clinical outcomes, resulting in substantial survival variability among patients with identical stage classifications. This thesis investigated whether radiomic features extracted from routine ¹⁸F-FDG PET imaging could enhance prognostic stratification beyond conventional staging in NSCLC patients undergoing radical-intent radiotherapy. The research examined 383 patients with stage I–IIIB NSCLC undergoing radical-intent radiotherapy, employing standardised feature extraction following international guidelines. The central hypothesis, that radiomic analysis could provide prognostic information beyond conventional staging, received partial support, though not always in the manner initially anticipated. The findings revealed that the value of radiomics depends critically on thoughtful methodological choices regarding feature selection, lesion aggregation, and anatomical site prioritisation. Rather than supporting indiscriminate extraction of features from all available imaging data, the results advocate for biologically-informed, streamlined approaches that are more robust and clinically implementable. For predicting lymph node metastasis from primary tumour characteristics, integrating RFs with established clinical and PET parameters achieved superior discrimination (AUC: 0.816 training, 0.771 test) compared to either approach alone. Metabolic tumour volume emerged as the strongest predictor, while texture features capturing spatial heterogeneity demonstrated consistent prognostic value across training and validation cohorts, confirming that radiomics provides complementary information to conventional imaging parameters. A critical finding emerged from the systematic comparison of lesion aggregation strategies for patients with multiple metastatic lymph nodes. Selecting RFs from the single lymph node with the largest metabolic tumour volume substantially outperformed both multi-lesions averaging and the highest standardised uptake value selection approaches (AUC: 0.912 versus 0.856 and 0.818 for overall survival). This finding challenges the conventional assumption that comprehensive characterisation of all disease sites maximises prognostic information, instead reflecting the biological reality that the volumetrically dominant metastatic deposit likely harbours the most aggressive tumour clone, while aggregating features across multiple nodes dilutes this prognostically dominant signal. The investigation of combined primary tumour and lymph node radiomics yielded another counterintuitive result: integrating features from additional anatomical sites, other than the primary tumour, failed to improve survival prediction beyond primary tumour features alone. Analysis revealed substantial discordance between tumour based and node-based predictions in a significant proportion of patients, suggesting these sites encode fundamentally different—and potentially conflicting—biological information that cannot be resolved through simple feature concatenation. However, lymph node radiomics demonstrated utility for predicting short-term disease progression, suggesting site-specific applications where nodal features complement rather than replace primary tumour characterisation. The thesis makes several original contributions: the first systematic within-cohort comparison of lesion aggregation strategies establishing metabolic tumour volumebased selection as optimal for multi-lesion radiomic analysis; demonstration that primary tumour radiomics provide superior prognostic information compared to combined tumour-nodal approaches; identification of differential prognostic utility of nodal radiomics for progression-free versus overall survival; and establishment of an evidence-based hybrid segmentation protocol balancing reproducibility with clinical accuracy. Together, these findings underscore the value of biologically informed radiomic approaches in NSCLC prognostication, where primary tumour characterisation forms the foundation and selective nodal feature incorporation addresses specific clinical questions. This integrated framework of optimal segmentation strategies, lesion aggregation methods, and site-specific prognostic applications advances the field toward clinically implementable decision support tools and informs future multicentre validation efforts aimed at personalised treatment stratification in lung cancer management.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Medicine
Date of First Compliant Deposit:	31 July 2026
Last Modified:	31 Jul 2026 13:12
URI:	https://orca.cardiff.ac.uk/id/eprint/188450

Actions (repository staff only)

Edit Item