Taking and analysing a brain scan is no simple feat. A typical magnetic resonance imaging (MRI), electroencephalography (EEG) or magnetoencephalography (MEG) brain scan require hundreds of images on average – and by themselves, these images are not of sufficient quality or resolution. Technicians need to manually regularize them in a process that can take up to 48 hours.
Enter artificial intelligence: Researchers are exploring ways of using machine learning to shorten and improve this process. Researchers from Carnegie Mellon University, led by Bin He (Professor of Biomedical Engineering), have developed an AI-based technology to improve the quality of brain scans with minimal human intervention. These trained deep neural networks promise to both capture images accurately and also reliably reconstruct the location, extent, and temporal dynamics of the brain.
“We first train deep neural networks with a large number of training sets, and once the training is completed dynamic brain imaging can be obtained by feeding collected EEG data to the network,” said He, who has been working on non-invasive brain imaging technologies for over a decade.
Technologies like functional MRI, EEG and MEG work by converting electromagnetic waves into images, allowing us to peer into the brain or other organs. These are widely used to study the function and development of brain cells, as well as to spot abnormalities or signs and symptoms of conditions like epilepsy, Parkinson’s Disease or Alzheimer’s.
Professor He added that this AI-based system removes the need for manual corrections to the EEG scans, and even outperforms current techniques to do this, improving efficiency and the speed of diagnosis, the team behind the project claims.
The tech has particular application for EEG imaging. The pursuit of EEG imaging involves translating scalp EEG data through signal processing and machine learning to provide images of brain activity over time. Although EEG imaging is typically faster and less expensive, choosing and fine-tuning recording-specific images requires specialised knowledge and skills from the user.
A pilot study
The technology that He’s group has developed was tested on 20 drug-resistant epilepsy patients. Using the AI-based system helped avoid using an invasive intracranial EEG, while also providing scans with better resolution, image quality and brain coverage.
The research team further emphasizes that the technology may be used to aid planning of surgical treatment or neuromodulation treatment of drug-resistant epilepsy patients. It may also be applied for precision diagnosis and treatment guidance of various neurological disorders such as pain, depression and stroke.
Dynamic brain imaging by means of deep neural networks and brain models can be extended for other diagnostics procedures as well.
These deep neural networks programme aims to make spatiotemporal dynamic human brain imaging a widely used technique, assisting in the clinical diagnosis and management of a range of neurological and mental illnesses.
The potential for AI in medical imaging is huge: A recent report by Frost & Sullivan estimated that integrating AI in medical imaging diagnostics as well as in hospital workflows could cut treatment costs by as much as 50% while also improving the treatment outcome by 30-40%.