这篇文章来自 nvidia.com。原始 url 是: https://blogs.nvidia.com/blog/2018/07/23/ai-lowers-costs-risks-medical-scans/
以下内容由机器翻译生成。如果您觉得可读性不好, 请阅读原文或 点击这里.
With $2,000, you could take a round trip to Argentina, Hong Kong or your local hospital’s MRI machine.
In dramatic ways, Subtle Medical, a Silicon Valley startup, is using AI to reduce the costs associated with medical imaging scans — in terms of money, time and radiation exposure.
PET scans and MRIs are used to produce detailed images of a patient’s organs and internal structures, which help doctors diagnose medical conditions. PET scans typically take 45 minutes to capture the full human body, while MRIs can take up to three hours. This means that over 35 million patients each year, in the U.S. alone, collectively spend countless hours and billions of dollars on medical scans.
Subtle Medical, an NVIDIA Inception program award winner, aims to improve the inefficiencies of the process. And by changing the process, they hope to drastically improve the productivity of the hospitals and the experiences of the patients.
Greg Zaharchuk and Enhao Gong, Subtle Medical’s founders, met at Stanford University, where Zaharchuk advises Gong in his doctoral research. Their company uses deep learning to improve image quality, which can reduce the duration of MRI and PET scans by three-quarters.
In addition to reducing the length of time people spend inside MRI scanner, their technology can improve medical scans’ safety. Gadolinium, a potentially harmful metal, is deposited in the body during MRI contrast scans, and its alleged side effects have led to lawsuits. Subtle Medical cuts the amount of contrast dose needed by up to tenfold.
The company uses NVIDIA GPUs and CUDA to train its deep learning model. GPU computing allowed its research team to accelerate their deep learning process from 1-10 minutes per image to one 第二 per image.
While Subtle Medical could drastically reduce the length of an MRI scan, studies found that accuracy also improved — by up to three times.
The company trained its deep learning model using thousands of images from patients through Stanford. Leveraging this work, Zaharchuck and Gong were able to enhance scans from patients who had received only 10 percent of the typical radiation dosage, and generated results that matched the quality of scans of those who had received full-contrast radiation.
These same findings can also be applied to PET scans, which are commonly used for Alzheimer’s disease and cancer diagnosis. The Subtle Medical team managed to produce the same quality of imaging with an AI-enhanced five-minute exam while also reducing the required radiation dosage up to 200-fold.
There are many practical applications for Subtle Medical’s research: faster scans mean improved productivity and faster diagnosis. This can lower needed radiation doses, meaning safer conditions for patients. And an improved workflow means better management of patient needs, along with money and time savings and, most importantly, more diagnosed and treated patients.
Going forward, the company hopes to expand to clinical trials, where their findings can directly help patients and hospitals. Despite having been founded just a year ago, Subtle Medical expects to apply for FDA clearance this year.
“We want to bring this technology into clinics so that all the patients and hospitals can benefit from this technology,” Gong said. “We want to empower medical imaging with AI infrastructure to make it more accessible.”
Editor’s note: Post has been updated to clarify sources of radiation, per comments below.