Inside Advanced Scale Challenges|Sunday, June 25, 2017
  • Subscribe to EnterpriseTech Weekly Updates: Subscribe by email

GPU-driven Mobile Device for Deep Learning Medical Diagnostics – at Scale, at the Edge, at ‘the Roof of the World’ 

By bringing to India, Bangladesh and Nepal a low-cost, GPU-driven mobile device trained to recognize chronic disease “signatures” in medical images, a healthcare start-up is building a distributed cognitive computing capability that provides advanced diagnostics to a resource-poor, population-dense region of the world.

If the eyes are windows into the soul, they also can help provide a view of a patient’s overall health. Breakdowns and leaks in the microscopic capillaries in the retina can be early indicators of diabetes, high blood pressure and vascular disease long before patients and doctors notice symptoms. Early detection enables patients to take preventive measures related to life-style issues, such as diet, smoking and exercise to head off the life-threatening toll that chronic diseases take.

MARVIN

MARVIN

Start-up SocialEyes has developed MARVIN (Mobile Autonomous Retinal Evaluation), an Android tablet based on the NVIDIA Shield mobile device, powered by the NVIDIA Tegra K1 processor with a 192-core NVIDIA Kepler GPU and 2.2 GHz quad-core CPU. Originally developed for gaming, Shield is also effective for medical image processing, according to SocialEyes CEO Nicholas Bedworth.

“They’re designed for the extraordinarily compute-intensive experience of gaming,” said Bedworth in a phone interview with EnterpriseTech from Jabalpur, India. “More or less by chance, we found that the kind of computation that the GPUs do extremely well for gaming software is also the kind of computation you need for scientific image processing and machine learning. It’s a terrific medical computation platform. And it’s cheap.”

Bedworth is working with an international team of scientists, physicians and engineers to develop GPU-enabled apps that deliver healthcare at scale in places where doctors are scarce. Using two NVIDIA machine learning support technologies – the CUDA parallel computing platform and programming model and the DIGITS DevBox for machine learning training – SocialEyes developers used convolutional neural network techniques and thousands of images of healthy and unhealthy retinas to program MARVIN to analyze images taken by a fundus camera (a camera combined with a microscope) and find retinal markers for disease that would otherwise require a medical specialist to identify.

Bedworth said MARVIN delivers a combination of qualities needed for SocialEyes’ distributed diagnostics delivery strategy: high performance throughput to provide real-time results in a small package along with good battery life; and they are ruggedized.

“The basic idea is that we want to embody the diagnostic expertise of an ophthalmologist into a mobile device,” Bedworth said. “The device costs $200, and an ophthalmologist costs six figures and quite a few years to train.”

Global improvements in life expectancy mean more people are living longer and coming down with chronic disease. Some 415 million adults have diabetes, a number forecast to rise by 50 percent in 25 years, according to the World Diabetes Foundation. Most new cases occur in low-income countries where medical resources are strained and many people don’t become aware of their condition until a crisis occurs. MARVIN is designed to help identify hidden cases at low cost and early, when intervention can be most effective.

“We want to go where people are in desperate need of better healthcare,” Bedworth said. “We’re going where there are extraordinary dense populations and scarce resources.”

Under SocialEye’s business model, medical facilities in India, Bangladesh and Nepal (the site of Mount Everest, called “the roof of the world”) will purchase tablets from SocialEyes along with a fundus camera on an amortized basis that spreads CAPEX costs over a number of years. The medical practitioners in the field charge between $1 and $3 per retinal exam (those unable to pay receive free exams while those with higher incomes can pay up to $20 for premium service, enabling cross-subsidization). This compares with $30-$50 per retinal exam using conventional methods, according to Bedworth, a cost beyond the means of many people in South Central Asia.

MARVIN's low cost and mobility means diagnostic services can be distributed to rural areas, where the majority of people in the developing world live.

SocialEyes founder Nicholas Bedworth demonstrates a tablet in Nepal

SocialEyes founder Nicholas Bedworth demonstrates a tablet in Nepal

“Most ophthalmologists are in big cities,” Bedworth said, “but maybe 60 percent of these populations live outside the cities, and travel in these countries is difficult. Our tablet lets you move ophthalmological expertise out to the edge, out to villages and small towns.”

MARVIN processes three to five cases per hour, each exam generates 10-15 Mb of data and the tablet has the storage capacity for several hundred cases. The SocialEyes app assesses the retina’s condition and treatment can begin immediately, rather than waiting days for a report to come back from a remote tele-health grading center.

Still in pilot phase, SocialEyes initially plans to have 250 MARVINs in operation at 150 sites. Within three to four years, Bedworth said, the company plans to have 25,000 tablets in the field conducting 20-25 million retinal exams per year in the three countries.

“Then we’ll be doing what needs to be done,” Bedworth said. “Everyone who needs care will be getting it.”

Longer term, SocialEyes plans to monetize the data it collects from millions of retinal images. The company has a patent pending on its retinography diagnostics methodology, and Bedworth said SocialEyes eventually intends to seek FDA approval based in part on the results of its work in Asia over the next several years.

“Using the pooled data will help us identify disease signatures that can be paired with millions of cases,” Bedworth said. “We believe the pooled data and the compute associated with it will allow us to come up with new diagnostic tools.”

Add a Comment

Share This