Heart Smart: How HeartFlow Uses AI to Detect Heart DiseaseJuly 6, 2017
Coronary heart disease is the world’s biggest killer, responsible for nearly 9 million deaths worldwide and diagnosed in 12 million to 13 million Americans each year. Personalized medical technology company HeartFlow uses GPU-accelerated deep learning to find a better solution.
Heart disease is so devastating because it often goes undetected. Tragically, the disease is often misdiagnosed, especially in women. Heart disease is difficult for doctors to diagnose because, until recently, the best test for it was an angiogram, which is an invasive and costly procedure.
HeartFlow provides a non-invasive alternative. It combines standard CT scans — available at tens of thousands of healthcare facilities worldwide — with complex fluid dynamics and deep learning algorithms. The result is a 3D map of the patient’s heart that gives doctors a detailed view of blockages and blood flow on which to base a diagnosis.
This approach enables clinicians to provide the appropriate treatment for each patient, with the potential to greatly improve quality of life. It also means 60 percent of patients can avoid an angiogram, reducing healthcare system costs by 25 percent.
Seeing Clearly with AI
Building personalized models of the heart is a complex challenge. Besides creating a subvoxel-accurate model for each patient, HeartFlow’s system must simulate the flow of blood, vessel by vessel. Diagnosis is also time-sensitive in fast-paced emergency care departments.
“When a patient presents to the emergency room with suspected coronary heart disease, they need a diagnosis fast,” says Leo Grady, senior vice president of engineering at HeartFlow.
To solve this problem, HeartFlow turned to deep learning, accelerated by NVIDIA GPUs. The company adapted conventional deep learning to the analysis of blood vessels using a novel vessel-specific architecture. Its computer vision algorithms analyze medical imaging data to create a personalized 3D model of the patient’s heart and coronary arteries by understanding the information contained in a CT scan.
This model is then meticulously assessed by a team of trained professionals who make any changes necessary to ensure the precision and accuracy of the models based on the imaging data. These changes enable the algorithms to learn and improve, so that the more images the algorithms process, the more accurate they become.
Saving Lives Worldwide
Healthcare organizations around the world have embraced HeartFlow’s solution. The U.S. Food and Drug Administration has cleared it. It’s also received the backing of the U.K.’s National Institute for Health and Clinical Excellence.
“Using GPU-accelerated deep learning, we help physicians make definitive and personalized decisions more quickly,” says Grady. “That means better outcomes for patients — and saving the healthcare system money.”