To look into the future is to catch only a glimpse inside Simon Warfield’s radiology research lab at Boston Children’s Hospital. His team is pairing hyperfast imaging and deep learning to push the limits of medical imaging and artificial intelligence (AI) to identify, prevent and treat disease. He’s also eyeing ways AI will help as data sharing expands among research sites. “The research world needs to look forward to manage forward,” he says.
Simon Warfield, PhD, is focused on imaging pediatric disease that starts with fetuses in utero and stretches to kids of all ages, and even early adulthood. Over the last decade, the 30-person research team he directs at the Computational Radiology Laboratory (CRL) and Department of Radiology, has fine-tuned medical imaging powered by deep learning algorithms to identify key structures that enable physicians to see, treat and even rid patients of disease. They are making the previously impossible possible at this 404-bed comprehensive center for pediatric healthcare, one of the nation’s largest.
The solutions fall into two categories: faster, highly tuned imaging techniques and deep-learning algorithms that enable more comprehensive image interpretation. After 20 years in radiology research, Warfield sees the future shaping up like this: “The evolution of radiology research is shifting from a domain in which an expert looks at images and tells you what he or she sees to one in which machines capture data and imaging biomarkers of disease progression, disease activity, function and microstructure. These biomarkers are telling the story.”
Little patients make big data. The radiology department now sees 10 times more patients than they did a decade ago and every patient has 10 times more images, Warfield says. “But the number of radiologists looking at images is basically the same. So, every radiologist has 100 times less time to interpret every case. That’s where deep learning is helping to balance the load.”
Every year, Boston Children’s completes more than 5,000 brain MRI exams. They help the research team to understand normal and atypical variability. They also help to train deep learning algorithms and build the labeled data critical for deep learning interpretation of medical imaging data. While a number of the projects focus on the brain, others probe Crohn’s disease and congenital abnormalities of the kidney.
One example is autism spectrum disorder. Machine learning is helping to discover imaging biomarkers that characterize changes in the brains of children to understand prognosis and how to guide intervention. The team is working with a group of children in Boston and across a national network with rare genetic disorders that predispose them to a diagnosis of autism spectrum disorder. While all these children have a genetic change that expresses itself in the brain and can lead to autism spectrum disorder, only half of the children receive the diagnosis. Why only half? “We think that early brain maturation is associated with alterations in myelination and axonal guidance,” Warfield says. “When those changes go wrong early on, that leads to aberrations or mismaturation of the social communication system at a very early stage in life. If nothing is done about that, that leads to a series of behavioral deficits associated with not having the social communication system functioning well.”
By looking at the microstructure of the white matter of the brain with the help of machine learning, the team can tell who has disrupted axonal guidance or disrupted myelination and is thus predisposed to a diagnosis of autism spectrum disorder.
First-in-human data suggest that children who respond to drug therapy have cognitive improvements. Warfield’s research team is launching a trial that combines the drug with very early behavioral intervention. “The idea is the drug will reverse some of the consequences of the genetic alteration and in the early behavioral intervention,” he says, “we’ll stimulate the pathways in the brain to grow correctly.”
Deep learning also helps the research team find minute indicators of pediatric epilepsy. One of the most common causes of seizures is a malformation of the cortex called focal cortical dysplasia. If detected with MRI, early surgery to remove the focal cortical dysplasia can cure the epilepsy. “It is the most promising treatment in these patients,” Warfield notes.
Yet, it is far from easy to detect. Greater spatial resolution enhances subtle focal cortical dysplasia that looks like a blurring in the boundary between gray matter and white matter, with the problem area appearing thicker. Researchers have developed image acquisition strategies that allow them to zoom in on suspicious regions in the brain and generate very high-resolution images to identify focal cortical dysplasia or rule it out.
For radiologists, the case review is challenging due to the hundreds of images to review and the tiny, delicate abnormalities. The research team is developing deep learning and machine learning strategies to prereview the images and highlight potentially abnormal areas. The data are then compared to a normal control group and AI highlights atypical regions. “For the patient, this is a life-changer,” Warfield says. “When we can detect the problem, a physician can fix it. So, AI here has a significant impact on care.”
Driving data sharing
Warfield envisions that impact expanding even further as data-sharing opportunities emerge across the country and even around the globe. The National Institutes of Health are investing in large, publicly available imaging databases so labs can share and replicate results. Private sources are making data available as well, often aggregating multiple small studies of patients who are challenging to find. “I can import their experimental structure, apply our unique algorithms and run an analysis of my data or vice versa,” Warfield explains. “That is truly powerful to our research.”
Transformative is how Warfield describes the research evolution underway. “Radiology’s the domain where we can very rapidly transition from research idea to impact on patients,” he says.
Think of it as the next step in a second opinion. In the past, a radiologist with an interesting or challenging case would ask a colleague to take a look. Going forward, they might say: “I’ve got this interesting case and I’d like to look at other cases like it across the country,” Warfield says. “Give it five years, maybe less, and what we think is amazing in research, imaging and deep learning will be in the clinical domain.”