Among the many developments in imaging informatics this past year, several stand out more than the rest:
- Continual improvement in supervised learning, which relies on human annotation
- Growth in sophisticated, unsupervised machine-learning algorithms that do not require human annotation, allowing for the use of considerably larger datasets
- More use-case reinforcement learning, in which algorithms begin autocorrecting their own errors as they learn
- Natural language processing of various types of nonimaging data, including using an artificial intelligence algorithm to examine surveillance CT reports and then predict whether a patient will require further imaging down the road
- The U.S. Food and Drug Administration (FDA) regulatory proposal to shift its focus on vetting the ability of institutions and vendors to produce high-quality algorithms, rather than evaluating every minor adjustment to algorithms
Every year, developments that were initially thought to be reserved for research purposes only move on to become cleared and approved by the FDA and find their way into clinical practice, Chen noted.
"What I expect to see coming in to 2020 is some of the so-called research -- unsupervised algorithms, reinforcement algorithms, or generative adversarial networks -- actually making their way into the FDA-clearance, FDA-approval process, and start getting attention in the clinical world," he said.
Copyright © 2019 AuntMinnie.com