Identification of imaging biomarkers for personalisation of radiotherapy in lung cancer treatment

Student thesis: Phd


Stereotactic body radiotherapy (SABR) is standard of care for patients with inoperable early-stage non-small cell lung cancer. SABR is well-tolerated and successful, but treatment failures occur at approximate rates of 10% for local relapse (LR) and regional failure, and 20% for distant metastasis (DM). There is currently no cure for metastatic disease, and we hypothesise that prevention with personalised SABR delivery may improve patient outcomes. So far, predictors of treatment failure are not well developed, and little is known on the influence of incidental dose outside the gross tumour volume (GTV). Imaging biomarkers can quantify tumour and peritumour heterogeneity to describe risk of disease spread and possibly microscopic disease (MDE) and could be used to stratify patients for personalised SABR. This thesis aims to identify such biomarkers to predict LR, RF and DM using routine four-dimensional computed tomography (4D-CT) planning data. First, we studied tumour sphericity as a predictor of overall survival in lung cancer. It was determined that sphericity `weakly' correlates with: GTV volume, mean lung dose, and N-stage. Implementing common statistical pitfalls in radiomics (i.e., univariable analysis, dichotomisation, and omitting log-adjustment for GTV volume) gave a false-positive association between sphericity and overall survival, which was removed with multivariable analysis. From this, we recommend: to investigate imaging biomarkers individually, to test model assumptions, to not ignore weak correlations, and to not dichotomise continuous variables. Next, we developed an automated tool to generate a GTV from the iGTV. As the iGTV represents the GTV motion over the breathing cycle, the GTV was derived from tumour motion estimated with 4D phase registration successfully for 94% of patients. GTVs were within observer variation compared to an expert contour. A method to select the optimal 4D phase was then implemented, by selecting for each patient the phase with image features closest in value to those of neighbour phases. A radiomics model to predict DM using this approach out-performed models from the 50% phase or averaging of features. Feature selection on stability across all phases reduced model performance, and is not recommended for 4D-CT radiomics. The image features identified suggest tumour variability and higher peritumour density could promote DM. Finally, we developed a data mining methodology (Cox-per-radius) to investigate the spatial interaction between imaging biomarkers and dose in independent annuli around the GTV to predict treatment failure. Results demonstrate that high tumour variability and peritumour density are predictors of DM for patients with low dose ~3cm from the GTV. High peritumour density also interacts with dose outside the GTV to predict LR. This supports the hypothesis that imaging biomarkers describe risk of MDE, and for high-risk patients, incidental dose is beneficial. Importantly, the results of this thesis show that CT biomarkers interact with incidental dose to predict DM and LR following lung SABR. This discovery demonstrates that patients can be stratified based on pre-treatment 4D-CT for changes in margins or increased dose outside the GTV. For this analysis to be possible, an automated pipeline for imaging biomarker research on 4D-CT data was developed. This thesis ends with recommendations on how to make progress towards the goal of using routine imaging data to inform clinical practice.
Date of Award1 Aug 2022
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorAlan Mcwilliam (Supervisor), Corinne Faivre-Finn (Supervisor) & Marcel Van Herk (Supervisor)


  • statistical modelling
  • personalisation
  • imaging biomarkers
  • lung cancer
  • radiotherapy

Cite this