By Mark Watts
The perceived attributes of innovation in medical imaging that AI can assist with are taking a long time to arrive. It has been labeled the trough of disillusionment on the maturity scale.
Consumer innovations such as the cellular phone and the VCR required only a few years to reach widespread adoption in the United States. Other new ideas – such as using the metric system or seat belts in cars – required decades to reach complete use. The characteristics of innovations, as perceived by individuals, help to explain the different rates of adoption.
Radiologist and radiology administrators have yet to appreciate the relative advantages that imaging AI offers. It has not proven itself better than the current workflow that it is trying to supersede.
The degree of relative advantage may be measured in economic terms, but social prestige factors, convenience and satisfaction are also important. It does not matter, so much, whether an innovation has a great ideal of “objective” advantage. What does matter is whether an individual perceives the innovation as advantageous. The greater the perceived relative advantage of an innovation the more rapidly its rate of adoption will be.
Imaging AI has a compatibility issue. It is an innovation that was framed as a replacement solution for radiologist and that conflicts with existing values, past experiences and the needs of the potential adopters. An idea that is incompatible with the values and norms of a social system will not be adopted as rapidly as innovations that are compatible.
The adoption of an incompatible innovation often requires the prior adoption of a new value system, which is a relatively slow process.
The complexity of imaging AI is perceived as difficult to understand and use. Some innovations are readily comprehended by most members of a society while others are more complicated and thus are adopted more slowly.
The trialability of imaging AI and the degree to which an innovation may be experimented with on a limited basis are factors. There is not an easy way to try an imaging AI solution. Some models are glass-like in deployment and fail unless supported by perfect non-real-world conditions.
The last issue is observability. Is it measurable? What is the degree of confidence that we have in the outcome it is providing?
When do we know that drift in results has occurred? What tools do we have to fix the AI when it has drifted? What does good look like? What does better look like?
Trust is the issue. To build trust more transparency and education need to occur. The solutions need to be wider in application and scope.
I spoke with renowned radiologist Dr. Curtis Langlotz, professor of radiology and biomedical informatics at Stanford University School of Medicine. We were discussing the killer application for medical imaging AI. The thing that will be the critical mass that launches it forward. He said, “I want a resident in a box.” Someone to pre read my exams. My solution was the power of the negative exam. A pre read by AI that lowers the priority of a normal exam and places an exam with a finding at the top of the worklist. The normal can be batch read later. The one in a thousand spontaneous pneumothorax is placed at the top of the worklist. The needle in the haystack is found by a super powerful magnate.
The future of AI in imaging is being built today, but it will look much different in 5 years.
Mark Watts is an experienced imaging professional who founded an AI company called Zenlike.ai.
