Artificial Intelligence: Moving With the Speed of Light Into Medical Research and Practice

Michael Dwyer, PhD

What is the role of AI in multiple sclerosis (MS) management? Is this technology poised to replace the jobs of neurologists and neuroradiologists? Or, is it inherently flawed, like the now-infamous tale of an AI platform suggesting the use of glue to stick the toppings on a pizza? At the 2024 CMSC annual meeting, experts in the medical applications of AI discuss how this technology might be applied in MS.

“Unless you’ve been hiding in a shelter for the past few years, you have probably interacted with AI,” said Michael Dwyer, PhD, associate professor of neurology and biomedical informatics at Jacobs School of Medicine, University at Buffalo.

AI is not ready to take over the world yet, Dwyer noted. Humans are actually still much smarter. An AI platform such as ChatGPT memorizes a lot of different problems and then tries to solve them by pulling the most relevant “recipes” and averaging or interpolating between them, Dwyer explained. AI deep learning tools have about 80 million “neurons” (an AI neuron represents a connection point in an artificial network), while a house mouse has a comparable 70 million neurons.

“So, think of it as like a mouse with the factual knowledge of the entire internet,” Dwyer suggested. Humans, in contrast, have over 86 trillion neurons, along with about 150 trillion synapses compared with AI’s 1.7 trillion. These synapses are important in terms of the complexity a system can handle. “Even a year or two ago, AI had just a billion synapses and now it's a trillion,” he said. “So we're moving a lot faster than evolution is capable of doing. This is Star Trek sort of stuff that was once the realm of science fiction.”

AI can enhance MS management without replacing the neurologist—yet

What does this mean for medicine, and more specifically multiple sclerosis (MS)? Currently, many of the advantages lie in the ability of AI to enhance neuroimaging capabilities.

Francesca Bagnato, MD, PhD, associate professor of neurology and associate vice chair of research at Vanderbilt University Medical Center, presented examples of how AI is being applied to address unmet needs in MS based on the shortfalls of conventional MRI. For example, AI may enable MS radiologists to use lower-resolution MR equipment (1.5 or 3T) to achieve comparable results to a very high-resolution 7T device, Bagnato said. This has value in identifying cortical lesions, which are not visible on standard imaging. Double image recovery is another advanced MR approach that can identify cortical lesions. Investigators are training AI models to use this technology to extend the ability of standard MRI to identify cortical pathology, which is useful for MS diagnosis, identifying cognitive deficits, and other clinical applications.

Bagnato also presented data suggesting that AI may help to eliminate or reduce the need for gadolinium contrast agents. Gadolinium use has been waning in recent years due to safety concerns about buildup in the brain over time, but the contrast agent adds valuable information in identifying acutely active white matter lesions. A study by Narayana et al fed MR images acquired between 2005 and 2009 from 1008 patients into an AI program designed to compare brain pathology and found that the AI system had moderate to high accuracy in identifying enhancing MS lesions from the unenhanced multiparametric MRI.

Being able to identify whether a lesion is demyelinated or remyelinating would also be of great value in MS management. “T2 weighted lesions are quite nonspecific because they all look hyperintense, so we cannot say from these images how much myelin is preserved,” Bagnato explained. Positron emission tomography (PET) technology with an amyloid beta tracer can identify myelin concentration, but these tests are expensive and invasive due to the injection of a radioactive tracer. AI technology is being used to correlate PET myelination data with multisequence MRI data to predict myelination using standard imaging.

The speakers discussed some of the pitfalls that still exist in AI applications and emphasized that when it comes to assisting with the management of MS, AI still has a lot to learn.