Malignant pleural mesothelioma (MPM) is a rare cancer of the mesothelial cells of the visceral and parietal pleurae that is heterogeneous in terms of biology, prognosis and response to systemic anti-cancer therapy (SACT). The primary tumour forms an unusual, complex shape which makes survival prediction and response measurement uniquely challenging. Computed tomography (CT) imaging is the bedrock of radiological quantification and response assessment, but it has major limitations that translate into low sensitivity and high inter-observer variation when classifying response using Response Evaluation Classification In Solid Tumours (mRECIST) criteria. Magnetic resonance imaging (MRI) tools have been developed that overcome some of these problems but cost and availability of MRI mean that optimisation of CT and better use for data acquired by this method are important priorities in the short term. In this thesis, I conducted 3 studies focused on, 1) development of a semi-automated volumetric segmentation method for CT based on recently positive studies in MRI, 2) training and external validation of a deep learning artificial intelligence (AI) tool for fully automated volumetric segmentation based on CT data, and, 3) use of non-tumour imaging features available from CT related to altered body composition for development of new prognostic models, which could assist in selection of patients for treatment and improving tolerance to treatment by targeting the systemic consequences of MPM.

The aim of Chapter 3 is to develop a semi-automated MPM tumour volume segmentation method that would serve as the ground truth for the training of a fully automated AI algorithm. A semi-automated approach to pleural tumour segmentation has been developed using MRI scans which calculated volumetric measurements from seed points – defined by differential tumour enhancement – placed within a pre-defined volume of pleural tumour. I extrapolated this MRI method using contrast-enhanced CT scans in 23 patients with MPM. Radiodensity values – defined by Hounsfield units (HU) – were calculated for the different thoracic tissues by placing regions of interest (ROI) on visible areas of pleural tumour with similar ROIs placed on other thoracic tissues. Pleural volume contours were drawn on axial CT slices and propagated throughout the volume by linear interpolation using volumetric software (Myrian Intrasense® software v2.4.3 (Paris, France)). Seed points based on the radiodensity range of pleural tumour were placed on representative areas of tumour with regions grown. There were similarities in median thoracic tissue HU values: pleural tumour, 52 [IQR 46 to 60] HU; intercostal muscle, 20.4 [IQR 11.9 to 32.3] HU; diaphragm, 40.4 [IQR 26.4 to 56.4] HU and pleural fluid, 11.8 [IQR 8.3 to 17.8] HU. There was also reduced definition between MPM tumour and neighbouring structures. The mean time taken to complete semi-automated volumetric segmentations for the 8 CT scans examined was 25 (SD 7) minutes. The semi-automated CT volumes were larger than the MRI volumes with a mean difference between MRI and CT volumes of -457.6 cm3 (95% limits of agreement -2741 to +1826 cm3). The complex shape of MPM tumour and overlapping thoracic tissue HU values precluded HU threshold-based region growing and meant that semi-automated volumetry using CT was not possible in this thesis.

Chapter 4 describes a multicentre retrospective cohort study that developed and validated an automated AI algorithm – termed a deep learning Convolutional Neural Network (CNN) – for volumetric MPM tumour segmentation. Due to the limitations of the semi-automated approach described in Chapter 3, manually annotated tumour volumes were used to train the CNN. The manual segmentation method ensured that all the parietal pleural tumour was included in the respective volumes. Although the manual CT volumes were consistently smaller than semi-automated MRI volumes (average difference between AI and human volumes 74.8 cm3), they were moderately correlated (Pearson’s r=0.524, p=0.0103). There was strong correlation (external validation set r=0.851, p