A panel of three physician experts in developing AI imaging platforms and one manufacturing CEO shared their insight in a lively discussion at a Wednesday morning RSNA session. Titled "Medical Imaging Analytics & AI: Technologies and Solutions for Better Healthcare Today and in the Future," the event was presented by Intel.

A moderator posed a series of questions to three radiologists with ties to the AI sector. The panel included:

- Dr. Eliot Siegel, chief of Radiation Oncology for the VA Maryland Healthcare System
- Dr. Lei Xing, professor and director of Medical Physics at Stanford
- Dr. Greg Zaharchuk, professor of Radiology at Stanford University and founder of Subtle Medical
- Gene Saragnese, CEO MaxQ and past CEO of Philips Imaging

Here, slightly paraphrased, are a few of the perspectives and insights the panelists brought to the conversation.

Why does medical AI imaging matter?

Dr. Xing pointed out that radiology is an essential element of healthcare, with some 75 percent of clinical decisions based on imaging. He said that deep learning using biological data allows for much prior experience to augment the decisions made by radiologists, which could potentially add considerable value to care delivery.

Rapid image reconstruction will reduce radiation dose to patients, according to Dr. Siegel. That is a clear benefit and a reason why AI matters. He also drew a comparison between AI and imaging residents at his facility. "More than one can read faster than I can and perhaps one of them has recently reviewed a journal article about the medical condition we're diagnosing," he said. "Those residents are not as good as me, the attending, but they help me get through information much faster. AI offers the same kind of benefit.

Saragnese pointed out that one radiologist – even the most highly trained – is not as good as the collective experience of 10,000 radiologists.

What types of AI will be the first to penetrate the market?

According to Dr. Zaharchuk, it will be applications that are transformational, that get images faster and enter workflow to provide efficiency. Patients want shorter scans and radiologists are struggling with burnout, he pointed out. Patients and radiologists alike want to reduce dose and gadolinium exposure; AI can do that. Zaharchuk contended that those are the first places where successful AI solutions will emerge, adding that it will take a while for radiologists to trust these systems.

Saragnese said that apps that work on both low-end and high-end systems will be the most desirable and therefore the first to penetrate the market.

A killer AI app would be one that counts lung nodules under low dose CT, said Siegel, citing the process as "tedious" and requiring a lot of work to submit the report to a registry for reimbursement. Ushering AI into that element of workflow would reduce cost and make the procedure more widely available, he said.

Machines can see a tumor earlier, with more accuracy than humans – we know that, said Xing. However, he added, the best AI app will take that info and combine it with other patient data from other clinical sources to fundamentally change value. It might predict, for example, which patients will respond to certain treatments based in part on genomic information coupled with the image analysis.

The biggest black box is the human brain, said Zaharchuk, but AI can help. You hear many vendors talk about achieving 95 percent accuracy with their AI apps, but that's not good enough, he cautioned.

"If I missed five out of every 100 brain bleeds, I'd lose my job. The potential of AI might make us feel better, but it has to work well."

Artificial Intelligence Homepage