The companies say the first target for their AI-enabled ultrasound will be finding veins for IV (intravenous) needle insertion. The technology would enable technicians to hold a simple ultrasound wand over the skin while software on a connected mobile device locates the vein for them.

 

For this project, Fujifilm SonoSite tapped the Allen Institute for Artificial Intelligence (AI2), which has an incubator for AI startup companies. “Not only do we have to come up with a very accurate model to analyze the ultrasound videos, but on top of that, we have to make sure the model is working effectively on the limited resources of an android tablet or phone,” says Vu Ha, technical director of the AI2 Incubator.

 

In an interview with IEEE Spectrum, Ha did not share the name of the startup that will be taking on the task, saying the fledgling company is still in “stealth mode.”

 

Ha tells the AI2 startup will take on the project in two stages: First, it will train a model on ultrasound images without any resource restraints, with the purpose of making it as reliable as possible. Then, the startup will go through a collection of experiments to simplify the model by controlling the number of hidden layers in the network, and by trimming and compressing the network until it is not difficult to operate on a mobile phone. The trick will be to shrink the model without sacrificing too much accuracy, Ha says.

 

If successful, the device could help clinicians diminish the number of unsuccessful attempts at finding a vein, and enable less trained technicians to start IVs as well. Hospitals that do a big volume of IVs often have highly trained staff capable of eyeballing ultrasound videos and using those images to help them to find small blood vessels. But the number of these highly trained clinicians is very small, says Ha.

 

“My hope is that with this technology, a less trained person will be able to find veins more reliably” using ultrasound, he says. That could broaden the availability of portable ultrasound to rural and resource-poor areas. 

 

SonoSite and AI2 are homes to two of the numerous groups of researchers putting AI to work on medical imaging and diagnostics. The U.S. Food and Drug Administration (FDA) has certified for commercial use a deep learning algorithm to analyze MRI images of the heart, an AI system that looks for signs of diabetic retinopathy in the images of the retina, an algorithm that analyzes X-ray images for signs of wrist fracture, and software that looks for indicators of stroke in CT images of the brain, to name a few.  

 

More importantly, the FDA in 2017 also approved for commercial use smartphone-based ultrasound technology distributed by Butterfly. The device, which costs less than $2000, can be used to take sonograms for 13 various clinical applications, including blood vessels. Butterfly has announced publicly that it is developing deep learning–based AI that will assist clinicians with image interpretation. But the company has not yet commercially launched the technology. 

 

Not less than four other portable or mobile device–based ultrasound technologies have been recognized by the FDA, including that of Fujifilm SonoSite, and the Lumify from Philips. But the adoption of these devices has been relatively slow. As Eric Topol, director of the Scripps Research Translational Institute, told Spectrum recently, the smartphone ultrasound is a “brilliant engineering advance” that’s “hardly used at all” in the health care system. Complex challenges such as reimbursement, training, and the old habits of clinicians often hinder the uptake of new gadgets, despite engineers’ best efforts.

 

TRONSERVE