As AI in medical imaging continues to advance, opportunities for machine learning applications continue to expand across the spectrum of neurological disease imaging.
"Machine learning is a broad field that has many potentially powerful methods and techniques," said moderator Yvonne Lui, MD, professor and vice chair of research in the Department of Radiology at NYU Langone Health/Grossman School of Medicine in New York.
In a Monday education session, three presentations centered on current and future considerations for AI in clinical neuroimaging.
Making the Call for More, Better Automation
Opening the discussion, Christopher Filippi, MD, chair of the Department of Radiology at Tufts University School of Medicine in Medford, MA, shared several studies demonstrating applications of AI in glioma diagnosis, treatment and follow-up. He noted that deep learning use in the field of glioma is exploding, but more can be done to improve AI tools for radiologists.
Ho, Lui, Chow, Filippi
Making an appeal to vendors and AI developers, Dr. Filippi said segmentation tools that automatically extract imaging features and provide tumor classification and other useful information would be most helpful. "This could easily be incorporated onto a PACS environment and should be done in a way that's automated," he said.
Dr. Filippi called on radiologists to be better at data sharing and working with PACS providers to develop tools that can be made universally available. He emphasized the importance of focusing on clinical efficacy when using AI models. Looking ahead, he said work should be done to push automated PACS forward and increase collaboration between imaging groups to remove siloes. Metrics for efficacy must be determined and arranged. "It's all about process, and we still have a lot of work to do," Dr. Filippi said.
Defining Measures of Success in AI
How do we define the correct measures of success in using AI in clinical practice? In his session, Daniel Chow, MD, MBA vice chair of innovation and entrepreneurship at University of California, Irvine (UCI), and chief of neuroradiology and co-director at the UCI Center for AI in Diagnostic Imaging, questioned whether success should be measured on a given tool's sensitivity and specificity or on its clinical benefit.
"A lot of the work right now has really been centered on diagnostic and technical efficacy but there's been a relative lack of effort in trying to figure out if it's actually improving patient outcomes," Dr. Chow said.
Focusing on applications of AI in stroke and vascular imaging, Dr. Chow highlighted the accelerated increase in AI products on the market. "We've seen very rapid turnover," he said. "It's not about having enough solutions but about what is the solution that makes the most sense."
To ensure that solutions add real value, radiologists have the opportunity to position themselves as leaders in the space. "One of the things that we are starting to struggle with is that there may be a paradigm shift. How we practice radiology right now might not be the future of the way we practice," Dr. Chow said.
Shifting from Big Data to "Small and Wide"
"Children are not little adults," began Mai-Lan Ho, MD, director of radiology research and director of advanced neuroimaging core and lead for imaging genomics at Nationwide Children's Hospital in Columbus, OH.
In a talk that drew a parallel between limited data availability in pediatric medical innovation and limitations of small data sets in AI applications, Dr. Ho noted that compared to many other areas of industry and development, health care data sets are relatively small. "There's been a shift from this idea of big data to small and wide," Dr. Ho said.
To overcome common challenges presented by small data, like overfitting and scalability, Dr. Ho shared solutions including feature extraction, transfer learning, federated learning and meta learning.
She provided a literature review of the use of AI across several pediatric applications including in prenatal, tumor, white matter, epilepsy, trauma and stroke imaging. She emphasized the importance of bringing together genomics and imaging to deliver precision health care.
Looking ahead, Dr. Ho said the next stage of the information age is human/machine collaboration. "So the future radiologist will be a driver of tech innovation providing quality, safety and value to patients and will incite integration in terms of activities between specialties and between countries across the world," Dr. Ho said.
Access the presentation, "AI: Intelligent Clinical Neuroimaging On the Horizon," (M3-CNR10) on demand at Meeting.RSNA.org.
