Medical AI Startup Quickly Shifts Strategy to Fight COVID-19 Pandemic

The COVID-19 pandemic has disrupted the world like few events before it.
But for Shukun Technology, a response required “a minor change in our strategy,” according to its chief technology officer, Chao Zheng.
That’s because Shukun, a startup founded by some of China’s brightest AI and medical minds, was busy refining its AI-powered platform to diagnose heart disease and strokes when the global pandemic struck.
The company quickly shifted resources to develop a system that analyzes chest CT scans to help speed up diagnoses of COVID-19 patients. That system, called Lung Doc – pneumonia edition, has already been rolled out to 30 hospitals in China over the past few months, where it will grow more accurate as it learns from more data.
Powered by NVIDIA GPUs, the system is proving so effective, according to Zheng, that Shukun is working on beefing it up for use in the many other countries that have inquired about using it.
Super-Charging Radiologists
Prior to Lung Doc, the three-year-old company had introduced two suites of products focused on heart ailments and strokes. At its essence, Shukun’s technology serves a very specific and valuable function: Shortening the time radiologists spend reconstructing and analyzing 2D and 3D images, and making diagnoses based on their findings.
Typically, these tasks require about a half-hour of a skilled radiologist’s time. But radiologists use varying techniques and approaches, so some take longer. Shukun’s AI adds consistency and speed, requiring just a minute to perform both the reconstruction and diagnosis processes.
“It’s a very efficient tool that helps radiologists deliver quicker, more accurate results,” said Zheng.
Little public data is available for life-threatening diseases, so in building its AI models, Shukun has turned to a vast network of hospital, academic and research partners to obtain private datasets totaling more than 100,000 cases, each of which typically contains 200-300 images.
As the company has worked on its models, it’s relied on blending imaging data to support segmentation and classification work. Transfer learning has helped speed the development process by applying lessons from one disease across others.
GPUs Delivering Results
NVIDIA GPUs have played a critical role for Shukun, with a combination of NVIDIA V100 Tensor Core and P100 GPUs (more than 500 in all) used for training, while NVIDIA T4 data center GPUs handle inference locally. The company can set up hospitals to run the system completely on premises or in a private cloud-like environment.
Zheng said each of these GPUs is delivering 20 times the performance of previous generations of GPUs and CPUs, and heart disease and stroke products have achieved accuracy rates that rival human detection abilities.
While future GPUs are expected to continue speeding up the training process, and thus...