Analysis of medical images is a time sink for radiologists, but crucial for diagnosing all sorts of ailments.
When screening for lung cancer, radiologists spend up to three minutes analyzing a normal image and as long as 10 minutes analyzing one with an abnormality.
Moreover, the varied levels of skill and expertise of radiologists can result in differing conclusions and occasional misreads, leading to unnecessary procedures.
In developing nations like Mozambique, where there are only a few radiologists serving a population approaching 30 million, the challenges are even greater.
Into this fray steps Arterys, which in 2017 became the first healthcare company to receive clearance from the U.S. Food and Drug Administration for cloud-based clinical solutions that use deep learning.
The startup’s first product, CardioAI, enables clinicians to use a web browser to automatically measure cardiac parameters using AI. One such parameter is the so-called ejection fraction, a proxy for the pumping ability of the heart.
Previously, measuring the ejection fraction required manual contouring of images, which could take 30 or more minutes. With CardioAI, this is done from magnetic resonance images in a matter of seconds.
Launched last year, CardioAI already boasts a network of 170 physicians and 100 hospitals that have used it to diagnose over 30,000 cases.
“Clinicians are becoming more convinced that this is really a viable technology that can help them with their workflow,” said Dan Golden, director of machine learning at Arterys.
Creating a Healthcare AI Suite
This acceptance may give Arterys a crucial head start in a market that’s expected to get crowded. Hospitals are projected to spend some $2 billion a year on AI for medical imaging by 2023, according to a recent report.
Arterys, a member of NVIDIA’s Inception virtual accelerator program, figures to benefit from this trend. In addition to its hallmark FDA-cleared product CardioAI, the company is introducing a cloud-based oncology suite featuring LiverAI and LungAI, for cancer diagnostics.
Arterys trained its CardioAI network, which can process CT and MRI scans, on NVIDIA TITAN X GPUs running locally and on Tesla GPU accelerators running in Google Cloud Platform. Both were supported by the Keras and TensorFlow deep learning libraries. Inference occurs on Tesla GPUs running in the Amazon cloud.
While the company hasn’t done a performance analysis, Golden said GPUs were providing about 10x the performance of CPUs a few years ago, and today “that gap has increased.”
For its oncology suite, Arterys has trained its models using a public dataset of more than 1,000 CT scans containing 1,300 nodules. The most recent inference tests had the LungAI system returning results in 90 seconds, all performed on study upload to ensure that inference is complete before the users begin reading the study.
Seeking Even More
Golden looks forward to using the newest GPUs to bring that latency down, and he’s excited about the added memory each generation of GPUs brings.
“Additional memory has given us more ability to work with images that are higher resolution, which means more robust models,” he said.
In the meantime, Arterys is also looking to expand its training efforts by collaborating with pharmaceutical companies to help them unlock the vast reservoirs of protected data collected during clinical studies.
That might be the key to open the healthcare AI floodgates.