Computing on the brain: where MRI meets epilepsy treatment

Building an accurate brain model is computationally demanding. JamesJam
Building an accurate brain model is computationally demanding. JamesJam

It’s been referred to, somewhat disparagingly, as blobology, but MRI technology has the potential to improve treatment for epilepsy – in part thanks to developments in computing.

Identifying where seizures occur in the brain is a fundamental problem in epilepsy research. But by applying image processing methods to high-resolution MRI scans – an expanding field of research, in which I work – we can localise the source of seizure activity and plan the best treatment strategy for patients.

In short, we can use computers to objectively identify brain regions that are associated with epilepsy.

Because of this, subtle neuroanatomical features observed by a radiologist can be quantified and objectively investigated using statistics; and quantitative methods (measuring properties such as volume or thickness) can be used to assist the radiologist in identifying brain regions related to a patient’s epilepsy.

Turn up the volume

One of the earliest applications of the quantitative approach involved identifying hippocampal sclerosis in temporal lobe epilepsy.

Individuals with hippocampal sclerosis often have epilepsy that is resistant to treatment with medication, and a very effective strategy for treating those patients is to surgically remove the affected hippocampus.

Quantitative analysis of hippocampal volume was used to prove that MRI could detect hippocampal sclerosis with high sensitivity and specificity.

A number of Australian researchers were instrumental in pioneering this discovery, including Professors Graeme Jackson and Alan Connelly from the Florey Insitute of Neuroscience and Mental Health, and Professor Sam Berkovic from the University of Melbourne.

Although measuring the volume of the hippocampus is useful for surgical planning and multiple other applications, the old-school way of measuring hippocampal volumes involves manually outlining the hippocampus on MRI scans.

This is highly labour-intensive and requires a great time commitment, in particular for studies involving large numbers of subjects. More recent computational methods have been developed that can obtain these measurements automatically.

Modelling the brain

The computational methods work by using high-resolution MRI scans to build up a detailed model of the structure of the brain, in much the same way that children (and some adults) build models of classic cars or dinosaurs.

Once a model is built, features of the brain can be quantified. In addition to hippocampal volume, examples of important features include the thickness, curvature and structural integrity of the cortex, the outer layer of the brain.

These features are then statistically compared between the epilepsy patients and healthy controls (people who don’t have epilepsy) and epilepsy-related brain differences are identified objectively.

The process of building an accurate brain model is computationally demanding, for the simple reason that an MRI scan contains a vast amount of information.

A typical scan might consist of approximately 10 million elements. The computational problem essentially reduces to figuring out which brain region each of these elements belongs to.

When I first began experimenting with these techniques in 2008, I processed MRI scans using a desktop computer that sat in the corner of my apartment. It took four months to process ten epilepsy cases and ten healthy controls.

Two years ago, my colleagues and I were granted time on a cluster computer, “Bruce”, provided by the Victorian Life Science Computation Initiative based at the University of Melbourne.

Access to Bruce greatly expanded the scope of my research: now I could process hundreds of MRI scans over a weekend.

More recently I have been carrying out these computational analyses using the Amazon Elastic Compute Cloud (Amazon EC2) – a web-based service that allows users to use as many computing instances as they require, for as long as they require them.

Services such as these democratise high-powered computing – the vast computational power is available to any member of the public.

Computing the odds

There are a number of biomedical problems that benefit from the computational approach to analysis of MRI scans.

Research studies that I work on include:

  • the Australian CANVAS study, led by Dr Amy Brodtmann, which aims to predict which people who have had a stroke will end up with dementia
  • work on the human epilepsy project, trying to predict outcomes and plan treatment for new onset epilepsy
  • evaluating the safety and efficacy of anticonvulsant medication in individuals with epilepsy and related disorders

Although great advances have been made in the treatment of neurological disorders, there are still many fundamental questions to be answered.

The computational approach to analysis of neuroimaging data is likely to be a key step for answering these questions.(Heath Pardoe/The Conversation)



Comments are closed.