The use of machine learning and artificial intelligence to analyze medical imaging data has grown significantly in recent years – with 60 new products approved by the U.S. Food and Drug Administration in 2020 alone.
But AI scaling, particularly in the medical field, can face unique challenges, said Elad Benjamin, general manager of Radiology and AI Informatics at Philips, during an Amazon Web Services presentation this week.
“The AI field is experiencing an amazing resurgence of applications and tools and products and companies,” said Benjamin.
The question many companies are seeking to answer is: “How do we use machine learning and deep learning to analyze medical imaging data and identify relevant clinical findings that should be highlighted to radiologists or other imaging providers to provide them with the decision support tools they need?” Benjamin asked.
He outlined three main business models being pursued:
- Triage involves identifying time-critical findings and managing workflows to ensure prompt responses to urgent matters. “For example … a tear in the lung, or a bleed in your brain, or stroke are critical things where time matters, and if IDed early can really save a patient’s life,” explained Benjamin.
- Population health, meanwhile, involves risk-stratifying populations to uncover people in danger of high-impact, high-cost diseases. “Prevention is often easier and cheaper than treatment,” noted Benjamin.
- Decision support in this context means providing real-time analysis of imaging data to assist with reading and interpretation.
Benjamin described common challenges and bottlenecks in the process of developing and marketing AI tools, noting that some were specifically hard to tackle in healthcare.
Gathering data at scale is one hurdle, he noted, and diversity of information is critical and sometimes difficult to achieve.
And labeling data, for instance, is the most expensive and time-consuming process, and requires a professional’s perspective (as opposed to in other industries, when a layperson could label an image as a “car” or a “street” without too much trouble).
Receiving feedback and monitoring are critical too.
“You need to understand how your AI tools are behaving in the real world,” Benjamin said. “Are there certain subpopulations where they are less effective? Are they slowly reducing in their quality because of a new scanner or a different patient population that has suddenly come into the fold?”
“Without solving these challenges it is difficult to scale AI in the healthcare domain,” he said.