ROLE OF IMAGING HETEROGENEITY IN ASSESSMENT OF CANCER BIOLOGY

Abstract

Purpose: To determine the most suitable tumour segmentation algorithm from 18-fluoro-2-deoxyglucose (18F-FDG) PET/CT scans and to evaluate the role of 18F-FDG PET-derived texture parameters in determining tumour micro-environment, tumour subtype, and patient prognosis. Methods: A) Three tumour segmentation algorithms were compared for inter-observer reproducibility of derived texture parameters using intraclass correlation coefficient (ICC), and prognostic capability using cox proportional hazards models. B) 18F-FDG PET-derived texture parameters were compared with changes in tumour volume and tracer uptake as biomarkers of early response to bevacizumab treatment in a case-control mouse model. C) Correlation analysis was performed between 18F-FDG PET-derived texture parameters and histopathological metrics of tumour micro-environment in human lung cancer specimens. D) Using semantic subjective parameters and CT-derived texture parameters, logistic regression models were developed to differentiate lung adenocarcinoma from squamous cell carcinoma. Results: A) Forty-percent of maximum intensity threshold was the most reproducible segmentation algorithm (median ICC: 0.9). Survival models were of equivalent quality regardless of segmentation algorithm used. B) There were no differences in tumour volumes between treated and control mice at 3 weeks. However, 1 texture parameter - grey-level size zone matrix-derived size-zone variability, was significantly different. C) Positive correlations were found between several histopathologic and 18F-FDG PET-derived metrics: between mean cell density (MCD) and mean standardised uptake value (SUVmean (rs: 0.55, p=0.007), and Pathology-lacunarity and SUV-lacunarity (rs: 0.5, p=0.018). Negative correlations existed between cell-poor proportion and metabolically active tumour volume (rs: -0.48, p=0.02), and MCD and both SUV-skewness and SUV-kurtosis (rs: -0.47, p=0.02). D) Best tumour classification performance was obtained with the combined texture and semantic-features model (area under receiver operating characteristics curve: 0.93) compared to semantic features-only (0.84) and texture features-only (0.825) models (p-value <0.05). Conclusions: 40P is a reproducible segmentation algorithm with acceptable survival predictive performance. Texture parameters may outperform size measurements in early response assessment. 18F-FDG PET-derived texture features are correlated with several histopathological metrics of tumour cellularity and heterogeneity. CT-derived texture parameters combined with human interpretation are accurate in classifying lung cancer.

Date of Award	30 Jun 2019
Original language	English
Awarding Institution	King's College London
Supervisor	Gary Cook (Supervisor) & Vicky Goh (Supervisor)