Computational science-enabled radiological pathology for the non-invasive mapping of tumour heterogeneity in childhood neuroblastoma

Loading...
Thumbnail Image

Embargo End Date

Authors

Zormpas Petridis, K

Document Type

Thesis or Dissertation

Date

2021-01-31

Date Accepted

Abstract

Neuroblastoma is a common childhood solid tumour that accounts for 15% of all cancer paediatric deaths. This thesis addresses key deficiencies in our ability to define, monitor and predict neuroblastoma heterogeneity for precision medicine. I used computational science to integrate the spatially-encoded phenotypic information provided by multi-parametric magnetic resonance imaging (MRI) with digital histopathology, demonstrating that MRI can provide non-invasive pathology to characterise neuroblastoma heterogeneity and provide biomarkers of response in clinically-relevant transgenic mouse models of high-risk disease. I first developed and demonstrated the application of novel computational pathology methodologies to enhance the quantitative assessment of tumour components from H&E-stained whole-slide images (WSI). These include two frameworks: SuperCRF, which fuses traditional machine learning with deep learning to model the way pathologists incorporate large-scale tissue architecture and context across spatial spaces to significantly improve single-cell classification and, SuperHistopath, which combines the application of the SLIC superpixels algorithm on low-magnification WSIs (5x) with a convolutional neural network (CNN) for superpixels classification to accurately map tumour heterogeneity from low-resolution histology. I then developed an MRI-histopathology cross-validation pipeline which provides the rigorous validation needed to support the deployment of novel MRI scans in the neuroblastoma clinic. Using this platform, I demonstrated the sensitivity of susceptibility-, T1-Mapping- and diffusion-weighted-MRI to the cellular and microenvironmental hallmarks of high-risk neuroblastoma and their modulation by either vascular- or MYCN-targeted therapies. Finally, I used supervised machine learning classification- and regression-based approaches to show proof-of-concept that habitat imaging derived from these three scans can non-invasively provide quantitative data typically acquired from histological analysis, such as densities of specific cell populations. This thesis demonstrates the potential of multi-parametric MRI to deliver non-invasive "virtual" biopsies to enhance diagnostic and treatment monitoring for children with neuroblastoma and pave new ways in studying tumour as an evolving ecosystem.

Citation

2021

DOI

Source Title

Publisher

Institute of Cancer Research (University Of London)

ISSN

eISSN

Research Team

Computational Imaging

Notes