Repository logo
 

Modelling brain tumours with network neurosciences and deep learning


Type

Thesis

Change log

Abstract

Glioma is one of the most common and lethal brain tumours. The clinical diagnosis and prognosis of glioma are challenging due to its heterogeneous nature. In the past, glioma was considered a focal disease with local disruption to the brain. However, recent studies suggested that tumour cells would migrate into normal-appearing brain regions through white matter tracts, implying the widespread impact of gliomas.

Meanwhile, quantitative MRI technology, including functional MRI and diffusion MRI, assisted in developing an emerging field called network neurosciences. The network neurosciences model the brain as a network consisting of macro-scale regional connectivity of the brain instead of the micro-scale networks of neurons. The novel brain network approach has the potential to characterise the global impact of brain tumours. This thesis aims to invoke specialised network neuroscience approaches for characterising brain tumours and systematically analyse the global tumour disruption across the brain.

First, I provided extensive reviews of brain tumours and network neurosciences. Then I discussed the limitations and current research gaps in the cross-disciplinary field of brain networks of brain tumours. The main finding implied that although there was potential for studying brain networks in brain tumours, the tumour disruption could hinder the robust application of the traditional methodology on brain tumour patients.

In Chapter 2, I developed a novel approach for constructing structural brain networks with diffusion MRI. This approach tackled the challenge of constructing brain networks using traditional tractography and brain atlas. For evaluating the proposed method, functional networks were generated as the evaluation baseline of the method. Results showed that brain networks generated from the proposed method had a stronger correlation with the functional networks compared to traditional methods, suggesting the robustness of the new approach.

In Chapter 3, I extensively analysed the disruption of brain networks in glioblastoma patients based on the methodology of Chapter 2. To address the heterogeneous disruption of glioblastomas, I classified the regions of brain networks into different categories according to their locations. In particular, the disruption in distantly connected regions was significantly associated with patient performance, cognition, and prognosis, which validated the clinical value of brain networks for glioblastomas and implied the early signs of tumour invasion into the normal-appearing brain.

In Chapter 4, to investigate the potential of early tumour invasion outside the tumour regions and tackle the unavailability of diffusion MRI, I developed a graph-based deep-learning model for estimating brain networks from anatomical MRI (e.g. T1, T2). Notably, the model was informed by the prior anatomical knowledge of white matter connectivity. The model successfully predicted IDH mutation status with higher performance than convolutional neural networks, which suggested the feasibility of estimating brain networks in retrospective datasets with only anatomical MRI.

In Chapter 5, I combined the prior knowledge of brain network disruption and the clinical significance of tumour geometrics to develop a multi-modal deep learning framework that learnt from brain networks, tumour image, and tumour geometrics simultaneously. The model was built on the hypothesis that multi-modal data respectively represented different views of the aggressiveness of the same tumour. Results showed that this model outperformed other state-of-the-art models in predicting the IDH genotype of glioma. Additionally, visual interpretation suggested the interrelation among multi-modal data.

In conclusion, this thesis presented feasible tools and experiments for applying network neurosciences for brain tumour patients, which revealed the potential of the quantified characterisation of tumour impact across the whole brain.

Description

Date

2023-01-04

Advisors

Price, Stephen

Keywords

Brain networks, Glioblastoma

Qualification

Doctor of Philosophy (PhD)

Awarding Institution

University of Cambridge