Investigation of Quantitative MRI Radiomics Techniques for Glioblastoma

Glioblastoma is a primary brain cancer characterised by a high degree of both intra- and inter-tumoural heterogeneity. Despite significant effort, oncological therapies have not led to improvements in survival outcomes. Current treatment regimens do not address physiological heterogeneity either within or between different tumours. Treatment adaptations for Glioblastoma should consider physiological heterogeneity both within and between tumours.Magnetic Resonance Imaging (MRI) is used routinely for diagnosis, treatment planning and response assessment of Glioblastoma. MRI is valued for its high contrast soft tissue images and physiological imaging capabilities. Quantitative MRI (qMRI) is a group of MRI techniques that measure physiological quantities in tissue. qMRI is potentially valuable as a prognostic tool for Glioblastoma as it may allow measurements of physiological tumour properties. Perfusion weighted imaging (PWI) and Diffusion Weighted Imaging (DWI) are two forms of qMRI that have shown utility in prognostication of Glioblastoma patients. Radiomics is a rapidly emerging field of research that utilizes many quantitative imaging features related to the shape, intensity, and texture of radiological images. The underlying hypothesis of radiomics is that these radiomic features are indicative of physiological properties of tissue and can be used as imaging biomarkers. Radiomics has a natural synergy with qMRI as both techniques investigate physiological tissue properties using imaging.The work presented in this thesis investigated the development of a quantitative MRI radiomics model for outcome prediction in Glioblastoma, focussing on radiomic features derived from Dynamic Contrast Enhanced (DCE) and Apparent Diffusion Coefficient (ADC) MR images. The process of developing a prognostic radiomics model for clinical use was investigated through the repeatability of qMRI radiomics features, the robustness of machine learning techniques for multicentre radiomics survival classification models and the utility of implementing serial qMRI into radiomics survival classification models.Chapter 3 covers an investigation of the repeatability of radiomic features derived from qMRI of Glioblastoma. The impact of feature class and image standardisation on radiomic feature repeatability for features derived from multiple qMRI techniques was assessed. In this study it was found that among features derived from non standardised ADC images, the features in each feature class had a high mean ICC (>0.85). No standardisation method significantly improved the repeatability of either first or second order radiomic features derived from ADC images. In non standardised DCE images (CBV, Ktrans, Ve and Kep) shape features were the most repeatable, followed by second order features then by first order features. Intensity normalisation helped to improve the repeatability of first order features derived from DCE parametric maps, however, the method that was most effective at improving their repeatability varied by image type and patient cohort.In chapter 4 the relative efficacy of feature selection and machine learning methods for developing a Glioblastoma survival classification model were assessed. Radiomics survival classification models trained using publicly available data were developed using varying machine learning techniques and then tested on similar data acquired from other public datasets. The impact of feature selection techniques, machine learning model and number of features on model transferability was investigated. Feature repeatability filtering followed by recursive feature elimination was found to be a robust method of feature selection. Increasing the number of features beyond 20 features did not produce a consistent improvement in the performance of models within the training set and reduced their performance within the test datasets. Random Forest had the best performance of the tested machine learning classification algorithms in both the training and test datasets.In chapter 5 the relative utility of longitudinal serial MRI radiomics as markers of treatment response was assessed. qMRI of 12 patients with newly diagnosed Glioblastoma were acquired throughout the six week course of chemo-radiation therapy at radiotherapy simulation, mid radiotherapy and end of radiotherapy. Radiomic and delta-radiomic features were computed. For a given machine learning classification algorithm, the accuracy of follow up 2 models was the highest, followed by follow up 1 models and then the baseline models. The follow up 2 models showed near perfect classification accuracy and selected only delta radiomic features. The small size of the patient cohort limits the statistical confidence of these findings. The high performance of models developed using delta-radiomic features suggests that these features contain valuable information, investigation on an expanded dataset is needed to validate the results of this study.