Cancer imaging and image analysis methods in whole-body MRI and PET/MRI

Abstract: Diagnostic medical imaging techniques such as magnetic resonance imaging (MRI) and positron emission tomography (PET) can provide structural and functional assessments of the whole body. This has great value for potentially systemic diseases such as cancer. To take advantage of the enormous amount of data provided by current imaging systems, improvements in whole-body imaging protocols and advancements in image analysis methods are however needed. This thesis aims to develop advanced imaging and image analysis methods for the purpose of tumour characterisation in MRI and combined PET/MRI whole-body image datasets. Early prediction of progression free survival (PFS) and overall survival (OS) in patients with relapsed/refractory (r/r) large B-cell lymphoma (LBCL) undergoing chimeric antigen receptor (CAR) T-cell therapy was assessed using whole-body PET/MRI pre- and post-therapy. Reference standard manual segmentations of tumours and non-malignant lymphoid tissue were used, and an extended set of semi-quantitative and quantitative PET/MRI metrics was extracted. Predictive PET/MRI metrics included the metabolic tumour volume (MTV), tumour apparent diffusion coefficient (ADC) and 18F-fluorodeoxyglucose (FDG) uptake in non-malignant bone marrow. To enable automated image analysis, deformable image registration was used to create multiparametric normal atlases of healthy volunteers examined with whole-body FDG PET, diffusion weighted imaging (DWI) MRI and water-fat MRI. To improve the geometric accuracy of DWI in the normal atlas, the reverse polarity gradient (RPG) distortion correction method was evaluated. RPG increased the geometrical alignment between DWI and structural images acquired in the same scan session, with little effect on healthy tissue ADC. It was further shown that healthy tissue assessments in atlas space was possible, with the normal atlas employed to study voxel-wise correlations between ADC and age across the whole body, confirming results from a manual segmentation approach. As proof of concept, a probabilistic atlas based approach was successfully used for segmentation of suspected malignant disease in FDG PET data and detection of liver fat infiltration in fat fraction (FF) MRI data. Lastly, using a cohort of r/r LBCL patients, statistical deviations between patient and normal atlas DWI data included as input in a deep learning based model, improved its performance for automated tumour segmentation.