Robust and interpretable graph neural networks for the analysis of MRI and EEG to classify epilepsy subtypes and predict patient outcomes

   School of Electrical Engineering, Electronics and Computer Science

This project is no longer listed on and may not be available.

Click here to search for PhD studentship opportunities
  Prof Xiaowei Huang, Dr Simon Keller, Prof T Marson  No more applications being accepted  Funded PhD Project (UK Students Only)

About the Project

Epilepsy is one of the most common serious neurological disorders. Magnetic resonance imaging (MRI) and electroencephalography (EEG) are routinely used for the clinical assessment of patients with epilepsy to identify abnormalities in brain structure and physiology that may be causing epileptic seizures. MRI and EEG are digital technologies that may provide crucial information for patient diagnosis, prognosis prediction and therapy choice in selected cases. However, for the vast majority of patients, treatment outcomes are unpredictable on the basis of qualitative assessment of scans; there is also diagnostic uncertainty in many cases. We propose to apply Artificial Intelligence (AI) methods to MRI and EEG data that has been reconstructed into brain network matrices in patients with epilepsy to improve diagnostic classification of epilepsy subtypes and predict treatment outcomes. We will apply brain network approaches as epilepsy is a network disorder [1]. If successful, these approaches have the real potential to be incorporated into routine clinical care as supplementary neuroradiological and neurophysiological investigation methods. This work will offer a new and unique collaboration between the departments of computer science and pharmacology and therapeutics at the University of Liverpool and showcase how AI and digital imaging research have the potential to translate to clinical practice.

Outline work plan. Brain network graphs will be generated from all imaging modalities. From MRI, network nodes will be parcellated from structural MRI scans and network edges will be considered as measures of structural or functional connectivity between nodes using the structural, diffusion and functional MRI data. Using EEG, nodes will be electrodes and edges being correlations between functional activity at each node. The Liverpool BRAIN lab has considerable expertise in the reconstruction and analysis of brain networks from MRI/EEG data. The CS department at Liverpool has a long history, and significant strength, on AI. Related to this project, Dr Huang, who directs the autonomous cyber physical systems lab [6], is an expert on safety and trustworthiness of deep learning. There will be the following two major techniques to be developed in this project.

Handling structured data. Graph neural networks (GNNs), which are deep learning models that capture the dependence of graphs via message passing between the nodes of graphs, will be considered as the primary technique. For diagnostic purposes, a key technical question needs to be addressed -- interpretability. We will develop interpretability techniques to support each diagnosis result with an explanation (or evidence), by e.g., highlighting the nodes of the graphs which play important role in leading to the diagnosis results. Such level of explanation may reassure the doctor that the decision making by the deep learning is trustable.

Prognostics. We will construct a deep learning architecture to take as input these data and make the GNN one of its components. For prognostics in healthcare context, a key technical question needs to be addressed -- uncertainty estimation. It has been known that a deep learning model may predict a result firmly, but the prediction cannot be trusted due to high uncertainty (such as epistemic uncertainty and aleatoric uncertainty) from either the data or the model itself. A quantification or estimation of how significant these uncertainties are will provide essential information to assist the doctor in making the final prediction.

For any enquiries please contact:

Dr Xiaowei Huang [Email Address Removed]

Dr Simon Keller [Email Address Removed]

To apply for this opportunity please visit:

Funding Notes

This studentship is funded by the EPSRC DTP scheme and is offered for 3.5years in total. It provides full tuition fees and a stipend of approx. £15,609 tax free per year for living costs. The stipend costs quoted are for students starting from 1st October 2021 and will rise slightly each year with inflation.
The funding for this studentship also comes with a budget for research and training expenses of £1000 per year, and for those that are eligible, a disabled students allowance to cover the costs of any additional support that is required.


[1] Berhardt et al. Epilepsy Behav, 2015;50:162-70;
[7] Meng et. al. “CNN-GCN aggregation enabled boundary regression for biomedical image segmentation”, MICCAI2020; [8] Sun et. al. “Explaining Image Classifiers using Statistical Fault Localization”, ECCV2020.
PhD saved successfully
View saved PhDs