Chai Xiangfei, founder of Chinese artificial intelligence startup HuiyiHuiying, is modest and painfully honest, something that is rare among Chinese entrepreneurs.
When asked about the future of applying AI technology to the healthcare sector, Chai did not talk about billion-dollar market potential. "I always tell my team that…the most important thing is who can survive longer. If one can survive for five years, he/she will be the winner," Chai told China Money Network in an interview in HuiyiHuiying’s headquarters in Beijing last week.
The honesty may come from Chai’s experience as a postdoctoral researcher at the Department of Radiation Oncology at Stanford University, as well as working at the Holland Cancer Research Center and the radiology department in Leuven University. Before founding the Beijing-based medical imaging AI startup in 2015, Chai had worked in fields that rely on precision and accuracy.
But venturing into the tech startup world in China for almost three years has altered Chai’s mentality, for sure. "I changed from the mindset of an engineer and a scientist to a businesses person. Before, we only thought about the technology and products. But now the more important things are: How can you make money? Who are the real payers for your products?" Chai confessed.
This focus on the end user has led HuiyiHuiying to create products with maximum value potential. Since hospitals are the paying clients for HuiyiHuiying’s medical imaging AI products in China, the company has focused on hospitals’ needs on key diseases: major vascular diseases and cancers including breast cancer, lung cancer and liver cancer. These two types of diseases accounted for roughly 66% of the causes for death in China in 2016, the biggest killer diseases by a wide margin, according to the World Health Organization.
Because hospitals are organized around disease types, HuiyiHuiying has designed its products to assist single disease throughout the diagnosis and treatment process. In November, the company launched two full cycle health management platforms for breast cancer and cardiac diseases.
"Different hospitals always have different processes that need to be adjusted to. For some big customers, we always need to adjust, even the algorithms, to do the fine tuning," said Chai.
If the future is the survival of the fittest for healthcare AI startups, HuiyiHuiying is currently among the strongest in China. Last month, Intel Capital led a strategic investment round in the company, with participation from Beijing Singularity Power Investment Fund, a state-backed firm.
Even though the company did not disclose how much it raised, people with knowledge of the matter said it is around RMB200 million (US$29 million) to RMB300 million (US$43 million). In January this year, Chinese private equity firm CDH Investments invested an undisclosed amount in the company.
In Chai’s own words, Chinese venture capital markets has cooled during the second half of 2018, even though "AI was still very hot" during the first half. To be able to secure financing in today’s crucial environment is a testament of HuiyiHuiying’s team, but also its intense focus on execution and business fundamentals.
Read an interview Q&A below. Also subscribe to China Money Podcast for free in the iTunes store, or subscribe to our weekly newsletter.
Below is an edited version of the interview.
Q: HuiyiHuiying last month raised strategic investments from Intel Capital and Beijing Singularity Power Investment Fund. Was it easy to close this funding round in the so-called "VC winter"?
A: The Chinese VC market, especially during the second half of 2018, fundraising has become more difficult in general, even for the AI industry. During the first half, fundraising in other industries was difficult but AI was still very hot.
I guess investors have become more selective in companies, with a stronger focus on if you have a viable business model, a competitive product and team, a niche market focus and strong execution. That was different from 2017, when investors focused mainly on the team, on if you have great data scientists and AI scientist.
Q: How long did it take for this round to close?
A: It took around four to five months, from knowing each other, talking to each other, conducting due-diligence and finalizing the term sheet through negotiations.
Almost two years ago, we and Intel Capital started a joint lab together. In two years, we have become a major user of Intel’s CPU, and we were exploring some of their new devices like FPGA. So we had a long relationship already.
Q: Beijing Singularity Power Investment Fund was established by the National IC Industry Investment Fund and BOE Technology Group. Does it mean that HuiyiHuiying is going into the chip sector?
A: I don’t think we’re going into the chip sector, but Intel and BOE are both chip manufacturers. They are realizing that outside of building chips to support computational power, they need to go into more application-based segments.
Both Intel and BOE have done many investments in other companies with the idea of diving deeper into specific fields, especially the medical field. The medical field is definitely one of the fields they are very interested in but they want to enter the sector by investing, instead of doing it all by themselves.
Q: Does that mean you could partner with them to design some kind of AI chips specifically for the healthcare industry?
A: Like I mentioned, specifically-designed chips are probably the next thing, but right now we want to use more common chips like the CPU. Most of the medical devices only have CPUs. Usually, we use the GPU to do all the computations for AI algorithms. After Intel’s investment, they will provide supporting teams to help us migrate all the AI computation to the CPUs. That could be a big step forward as it means we can directly empower existing medical devices with our AI algorithms.
In addition, we are working on FPGA, which is more programmable and related to real-time computing. It’s great for ultra-sound and real-time movie-style medical images, which requires real-time computational capability. With the support from Intel, we will focus more on generic computational power-based applications and then aim to develop FPGA-based applications.
Q: How is it different dealing with a state-owned investment fund compared to a market-oriented VC fund?
A: I guess state-owned funds care more about long-term returns and how a company aligns with government policy. They have more patience compared to private funds that have to achieve returns in three to five years.
But I don’t think there’s a big difference. They care more about policy directions, but also about fundamentals like revenues and profits.
Q: What is HuiyiHuiying’s main objective?
A: We apply image recognition technology to the medical imaging field. What we are doing can assist radiologists to read medical images better and more efficiently, for example. Our technology can provide hints and detection alerts, reduce the number of misdiagnosis and improve efficiency.
Q: You studied at Stanford before coming back to China to set up this company. What changes did you make in the initial stages of the company?
A: The biggest change is that we become more business-focused. It’s not just the difference between the U.S. and China, but I changed from the mindset of an engineer and a scientist to a businesses person.
Before, we only thought about the technology and products. But now the more important things are: How can you make money? Who are the real payers for your products? Now I start to think about everything from the end point (clients), then back to the starting point of products.
Hospitals were the buyers in the very beginning, and it is still so at the current stage in China. In the U.S., the biggest buyers are insurance companies. But Chinese insurance companies are mostly state-owned companies, and it takes much longer to make decisions.
So I have to say that for some time, hospitals will probably still be the biggest buyer for medical AI products in China. There is a chance to see the government and third-parties like medical equipment companies and drug makers become big buyers in China too.
Q: The U.S. and China healthcare AI market, which one has bigger potential?
A: I just came back from the U.S., where I talked to a lot of U.S. companies doing similar businesses as we do. They all have this perception that China has more opportunities.
I do think so too, because at the very beginning of the industry, the actual buyers in China are top-tier hospitals. These hospitals with a strong focus on academics want to try new technologies for research. Another place where healthcare AI products can bring great value is hospitals in small cities and rural areas. These areas have a shortage of doctors, and healthcare AI products can help alleviate China’s unbalanced healthcare systems. On the contrary, U.S. major hospitals are all private hospitals and have less willingness to change.
Moreover, in terms of data, there is much more data one can acquire in China at a much lower cost than the U.S. So, I would say there are more opportunities in China and also other developing countries to accept healthcare AI products.
Q: HuiyiHuiying’s products have been used in over 600 hospitals in China. What are the most urgent needs for these hospitals?
A: The top-tier hospitals need the most pioneer devices and new algorithms to conduct research themselves and for academic publication purpose. In hospitals in smaller cities, they need something that can complete single tasks to help with their doctor shortage.
Q: How are you designing your products and solutions to meet these different needs?
A: Right now, what AI can do is still quite limited. They can only do some single test diagnoses like screening, TB testing or bone fracture diagnosis. Starting from collaborations with top hospitals for research and development, we will do testing and clinic trials. After that, we have the core algorithm that can be deployed to our IT system. Different hospitals always have different processes that need to be adjusted to. For some big customers, we always need to adjust, even the algorithms, to do the fine tuning.
Q: Take breast cancer scans as an example, how large the data-set should be to train the algorithm? Millions?
A: No, not that many. That’s something quite different for medical imaging. Actually we don’t have that large amount of data as in the facial recognition, in which one million is the starting point. But for us, we generally work with a few thousand to ten thousand images.
I have to say that (data) is a big challenge for us compared to other AI companies. We have to deal with small amount of data because we don’t have millions of images for diseases. For breast cancer, which is a big disease with around 200,000 patients every year, but that’s only 200,000 and we can’t get all of that data because data is distributed around different hospitals. For other diseases, maybe there are only 20,000 patients every year. So, we have to do the training based on a small amount of data.
Q: Right now, the accuracy rate of correctly identifying an anomaly is around 90% to 94%. Do you think the accuracy rate will get better?
A: The accuracy number is not that important for medical applications. It cannot be 100% because even doctors cannot be 100% sure what the disease is. So that’s something quite special and complicated. We usually launch something when we get a certain degree of accuracy, and then we continuously (improve the product) by getting more feedback.
We let users start using the software and give us feedback on whether the results are right or wrong. Sometimes, even doctors in the radiology department cannot decide whether it’s a benign or malignant tumor, or what stage the tumor is. So you have to wait to get the pathology data. That’s why we also developed research platforms and follow-up study platforms to collect feedback data.
Q: Outside of image recognition, what other AI technologies you feel should be applied in the healthcare sector? Such as voice recognition and knowledge map?
A: I do not feel strongly about voice recognition’s application in healthcare, but we do have knowledge map-based diagnoses. For us, we have about 20% of our products using knowledge map.
We are also working on specific diseases to create whole cycle solutions. For mammography, we have mammography management tools including mammography ultrasound, breast MRI and guidelines on making decisions for mammography treatment. Beside image information, we include lab information and clinical data to do the whole treatment.
Q: What are some future changes you see in your products?
A: In the future, we are more interested in not just focusing on medical images, but also working on specific diseases in which images are very useful for the diagnoses, such as artery disease and cancers, to do more full cycle management. That is our future direction.
Right now we already have four products for the radiology department, and two products for the single disease full cycle management. In the next two years, I think probably the number would be doubled because we already have a couple of prototypes for single disease management, and the main goal for the next year is commercialization. We will focus on two areas: major vascular diseases and cancers including breast cancer, lung cancer and liver cancer.
Q: What do you think will be the biggest challenge in this process?
A: The challenge is the more things we are working on, the more diversified the products are from different departments with different processes. The cycle of medical AI devices or medical AI software is much longer than other AI products.
Q: Looking out into the future, what major trends do you see in the healthcare AI sector in general?
A: I think the main trend is that AI technology will be everywhere, not just for decision-making, but also for creating data, reconstructing images and reconstructing signals. Also, everything becomes more digitized and quantitative. Before, things were experience-based. If you go to different doctors and different hospitals, it’s very common for them to give you different recommendations. The reason is that their decisions are mainly based on experience.
Q: Recently we heard from Waymo CEO saying that ubiquitous autonomous driving vehicles are still decades away. The difficulty of implementing and commercializing certain AI technologies have just been understood more fully. Do you feel the same for healthcare AI?
A: Maybe not decades, but at least three to five years, or even longer for sure. I always tell my team that we don’t need to look at what competitors are doing. The most important thing is who can survive longer. If one can survive for five years, he/she will be the winner.