Neural networks give a new picture of brain disorders

The explosion of medical data in recent years has increased demand for faster and more accurate analysis methods in healthcare. At the Department of Information Technology, researchers are developing deep learning methods to improve the treatment of brain disorders.

From imagery of clinical cases, readings are being
taken to be used to build an artificial neural

The research project is led by Robin Strand from the Division of Visual Information and Interaction at Uppsala University. His role in the cooperation with neuroradiologists at Uppsala University Hospital is to develop interactive digital tools for support in image analysis of brain disorders. The focus is on glioblastoma, brain tumours and intracranial aneurysms, bulges in cerebral arterial branches.

“To keep aneurysms from rupturing and leading to bleeding and strokes, patients are treated by filling out the aneurysm with platinum wires. After that one wants to monitor and check if blood is still flowing in the aneurysm, and we are developing methods for this,” says Strand.

The methods are based on a variant of machine learning called deep learning. Readings taken from image data from clinical cases are used to build an artificial neural network. Several different layers are added to the network’s structure, with each layer corresponding to a certain resolution of the images. With the human brain as a template, small calculation units or nodes are then linked up, which vaguely resembles the structure of neurons in the brain.

“What we want is for the network to distinguish between remnants of an aneurysm or tumour and what is in the background. Given enough image data to practise on, we have a system that can make good guesses.”

However, the computer’s ability to make these distinctions is inadequate; the shapes of remnants of aneurysms and tumours may be irregular and difficult to distinguish. Other problems may be noise and other distortions of image data, explains Strand. This is why the crucial success factor is that the user is involved and controls the process.

“A neuroradiologist has to go in and draw a line – ‘this was right and this was wrong’ – and do a refined analysis. It is the interaction in the image processing that allows the treatment answers to be as accurate and certain as possible.”

According to Strand, the challenge is producing a system that does the calculations as accurately and fast as possible; the user’s time is valuable. The results from the research project are positive to date.

“We have seen that when we look at tumour remnants with the deep learning tools, we can do the quantification significantly faster and with less image data than with conventional analysis.”

The methods are best evaluated on different kinds of image data. But there is still a long way to go before the method becomes an ethically approved tool, according to Strand. However, the hope is to be able to broaden the applications for artificial neural networks.


1 maj 2019