by Greg Freiherr

A.I. Inside: How machine intelligence might save lives

Feb 01, 2016
Emerging TechnologyHealth and Fitness SoftwareHealthcare Industry

Machine intelligence built into existing medical machines may be how A.I. can soon start impacting healthcare.

Much as Intel processors are the power inside PCs, A.I. algorithms may emerge as the power behind some advanced medical applications.

Virtual Radiologic, a U.S. company that provides diagnostic services to thousands of hospitals and medical facilities around the world, is developing a deep learning algorithm to review CT images taken in the emergency room. The company, also known as vRAD (an affiliate of MEDNAX), is grooming the algorithm to find signs of intracranial hemorrhage.  The algorithm is designed not to diagnose but to prioritize CT images for review by neuroradiologists in the vRAD network. Patients with intracranial hemorrhage, for example, due to stroke or trauma, such as a car accident, require immediate treatment to prevent brain damage or death.

vRAD’s algorithm could be helping to save lives in the near future.  The company plans to integrate the DL algorithm into its existing PACS, its picture archiving and communications system, which has a 510(k) clearance from the FDA.

“We will be running it through the FDA to accommodate the use of deep learning technology as part of the PACS,” said Shannon Werb, vRAD chief information officer, explaining that the vRAD’s patent filing will provide the basis for its submission to the FDA.

Werb says the company has begun putting together an FDA application for the algorithm, Werb told me. He did not indicate when it would be submitted.

If and when the algorithm is cleared for routine use, it could have a big effect on patient outcomes.  Werb explains that the algorithm, when integrated into vRAD’s PACS, would prioritize CT images for interpretation. If, for example, there were a dozen vRAD neuroradiologists, all licensed in the state where the CT images were taken, the algorithm would channel the suspected intracranial hemorrhage case to the top of each of these physicians’, he told me, “so that the next available neuroradiologist would immediately open the (suspected intracranial hemorrhage) study.”

Werb estimates that doing so would cut in half the average time of “getting eyeballs on images,” reducing the current 10 or 15 minutes that elapse between exam and interpretation to five or 10 minutes.

Development of deep learning is part of the company’s broader effort to move technology close to the patient” so as to “tap the doctor on the shoulder a lot more rapidly.”

vRAD has developed a global network of radiologists, many of them specialized in specific fields such as neuroradiology. The company typically partners with local radiology practices, Werb said, to provide preliminary interpretations during the day, as well as specialty and after-hours final interpretations.

The company acquires images from provider facilities and assigns them through a worklist that ushers cases to radiologists in this network on the basis of their availability or special training. The goal behind developing the intracranial hemorrhage algorithm is to reduce the time needed for the diagnosis and, thereby, accelerate the time to treatment.

When cleared for routine use, the DL algorithm will function as part of the PACS as an aide in distributing images to physicians for their review. As such, it would work as a cog in the company’s existing machinery for distributing images among the radiologists in its network.

Today vRAD works with 40% of the healthcare facilities in the United States. Cases suspected of intracranial hemorrhage, due to stroke or trauma, are prioritized by physicians.  The use of this DL algorithm now in development by vRAD would dramatically cut the time needed for this process. Future iterations might even show where in the images radiologists should look for the hemorrhage, Werb said.

In the future vRAD plans to develop DL algorithms to accelerate the interpretation of other high priority patients, such as those suspected of pulmonary embolism or aortic tears. Collaborations with the equipment makers, he told me, might even lead to algorithms built into the CT scanners themselves, flagging patients who must be evaluated immediately because they may have life threatening conditions.

“Imagine the scanner with an embedded deep learning algorithm, identifying the possibility of an intracranial hemorrhage while the patient is still on the CT table,” he said.   “That (would) move the care of the patient much closer to the point at which they are being seen.”