Doctors will also be unemployed? Chinese companies develop artificial intelligence technology to diagnose cancer

NVIDIA actively cultivates deep learning development talents and cooperates with industry, government, and academia to provide practical training courses; So can't other chip vendors, such as Xilinx, which is also promoting the use of its FPGAs in deep learning, offer similar training programs?

Kevin Krewell, chief analyst at market research firm Tirias Research, believes that this is not entirely true: "FPGAs are still too complex for machine learning programming, and there are some advantages to using FPGAs (or designing ASICs like Google's TPUs), but GPUs are generally available, readily available, and have a variety of functions that can be used to perform displays or machine learning. ”

NVDIA promotes real-world success stories of deep learning

Nvidia specifically introduced companies that have developed deep learning programs/products on the company's platform, such as Infervision, a Chinese startup that aims to develop artificial intelligence tomography (CT) diagnostic solutions for lung cancer diagnosis.

CK Chen, founder and CEO of Inference Technology, is one of the leading figures in the AI wave, and his program will show how the new technology can help medical radiographers read CT scans and X-ray results to detect suspicious lesions and nodules in lung cancer patients earlier and more efficiently.

Chen Kuan did not participate in Nvidia's training, but in 2012, when he was majoring in economics and finance at the University of Chicago, he happened to see an introduction to Nvidia's deep learning platform: "A friend of mine showed it to me, and I was fascinated." ”

During the 2012 U.S. presidential election, he collaborated with other students from the University of Chicago and the Massachusetts Institute of Technology (MIT) to develop a program that can classify posts by bipartisan candidates Barack Obama and Mitt Romney on Twitter to detect public perception of candidates; This is Chen Kuan's first investment in the field of deep learning.

In 2014, Chen Kuan, who was still a doctoral student, returned to China to look for AI business opportunities in different industries, and after many interviews, a radiology technician working in China's top hospital gave him an inspiration about the possibility of developing deep learning cancer detection technology, which gave birth to inference technology. Chen Kuan can be said to have met a nobleman.

The adoption of professional physicians is a key factor in the continuous improvement of the process of inferring technology, and Chen Kuan said that more than 100 hospitals in China are now working with the company to import data captured by tomography and X-ray equipment and compare the results.

The watershed moment of Chen Kuanzhi's deep learning products occurred when AlphaGo, developed by Google's artificial intelligence company Deep Mind, defeated human Go masters in 2015; AlphaGo won again in 2016 in a duel against a human chess player. "Since then, those in the Chinese medical community who are still skeptical of AI have changed their attitudes. Otherwise, no one really trusts deep learning software. Physicians at Tongji Hospital in Wuhan, China, are using a program developed by Inference Technology (Source: Inference Technology)

Let the machine learn on its own

Chen Kuan said that physicians have been using traditional computer-aided machine vision software, such as R2, since the 1990s; However, R2 is different from the new generation of deep learning software programs of Speculation Technology, where doctors must first tell the machine what to look for and describe the characteristics of the object it is looking for.

Inference technology allows machines to learn what to look for: "The machine will learn the actual area to pay attention to and the characteristics of the object to look for; However, Chen Kuan emphasized that such learning relies on a large amount of data collected from various medical institutions over a long period of time.

Fortunately, since the SARS pandemic broke out in 2002, the Chinese government has been actively promoting the installation of new generation IT equipment in large hospitals. Chen Kuan said that many first-tier hospitals already have their own data centers to store all image data. Of course, the stored images are not always perfect: "If the resolution is too poor, it becomes a classic case of GIGO (garbage in, garbage out)." ”

It is currently preparing to complete test results from radiographers participating in early-stage adoption projects, and the company is also awaiting approval of its software from the China Food and Drug Administration (CFDA) to scale its business.

Chen Kuan said that so far, the results of comparative studies between human radiographers and computers are "quite promising", and the two can find cancerous nodules larger than 6mm at the same time; The computer performed better in the search for nodules of 3~6mm or smaller. However, he also admitted that scientists cannot explain how computers can draw specific conclusions, which is a disadvantage of deep learning.

He also emphasized that the purpose of deep learning software is not to replace radiologists, but to ensure that human experts work with computers to verify the correct results.