Two million cells a minute. That’s how quickly brain damage happen when the cells get no oxygen in a stroke or in some brain injuries.
Both can have tragic consequences — paralysis, memory loss, speech difficulties and even death. But doctors can’t start treatment without an initial diagnosis, and that requires reading a CT scan as soon as the test’s completed.
Unfortunately, that’s not what usually happens, said Prashant Warier, co-founder of Qure.ai, a member of our Inception startup accelerator program.
“Radiologists typically have a backlog of cases,” he said. “They may have 30 cases to read, and the injured patient with bleeding in the brain is the 31st. They have to go through all the others before they get to that one.”
Qure.ai is changing that with GPU-accelerated deep learning technology that detects critical problems in head CT scans in less than 10 seconds. It’s designed to help doctors prioritize cases quickly, Warier said.
Tech Reads the Bleed
Qure.ai recently launched its automated CT scan reader to identify defects like brain bleeding, head fractures and blood clots in patients with strokes or traumatic brain injuries. It also distinguishes among five types of hemorrhages, pinpoints their locations and determines the extent of the bleeding.
All of that is crucial to deciding how to treat a patient. For example, strokes are usually caused by blood clots, so the standard treatment is a blood thinner. But that could be deadly if there’s bleeding in the brain.
Knowing the location and amount of bleeding shapes other treatment options, said Qure.ai co-founder Dr. Pooja Rao. As blood accumulates in the skull, it puts extra pressure on the brain, which cuts off oxygen. To relieve that pressure or stop the bleeding, doctors may perform surgery to remove a part of the skull.
With Qure.ai’s automated scanner, doctors can quickly prioritize and treat the most serious cases to protect precious brain cells.
Rivaling Accuracy of Radiologists
Qure.ai trained its deep learning model on more than 300,000 head CT scans and reports, using NVIDIA TITAN X and GeForce GTX 1080 GPUs with the cuDNN-accelerated Pytorch deep learning framework. In tests, they reached an average accuracy rate of more than 95 percent, compared with 97 percent by a panel of three senior radiologists.
The two-year-old startup deployed its model — a process known as inference — using our Tesla GPUs in the Amazon Web Services cloud.
The company’s also working on chest X-ray analysis, aided by NVIDIA GPUs. Among other things, Warier wants to be able to diagnose tuberculosis more quickly to avoid its spread and treat patients promptly. Left untreated, TB can be fatal; it’s one of the top 10 causes of death worldwide, according to the World Health Organization.
The big goal for Qure.ai is to get radiological services to regions where radiologists are scarce, Warier said. In Kenya, for example, there’s just one radiologist for every 275,000 people, and India has one radiologist for every 65,000 people, he added.
Even some highly developed countries are struggling. The U.K. has just 7.5 radiologists per 100,000, according to the Royal College of Radiologists — a situation the country’s National Health Service describes as “desperate.”
“We want to use AI to help make radiology less expensive and more accessible,” Warier said
Qure.ai. has rolled out its CT scan technology in India, where some radiology centers are using it in daily practice. The company is testing it in three U.S. centers while seeking FDA approval, and is planning tests in the European Union as well.
For more information about how Qure.ai developed its scan technology, read the paper, Development and Validation of Deep Learning Algorithms for Detection of Critical Findings in Head CT Scans.
* Main image for this story shows automated CT scans of bleeding inside the skull. On the left, there’s blood in the functional tissue of the brain. On the right is bleeding outside the brain, which is usually caused by a head injury. Images courtesy of Qure.ai.