Medical imaging AI is hot, but there are a series of problems.

At the end of 2017, Zheng Zhongxi attended an artificial intelligence seminar at Huaxi Hospital. Many departments talked about medical imaging AI. Doctors believe that AI needs to be intelligent based on a large number of accurate professional logos.

In July 2017, Huaxi Hospital announced the establishment of a medical artificial intelligence R&D center. In the digestive endoscopic artificial intelligence demonstration, 12 images were uploaded through the cloud, and polyps, new organisms (cancer) and veins were screened in less than 10 seconds. The common results of three digestive endoscopy examinations were 92.7%, 93.9% and 96.8%, respectively.

At that time, Ali Health and Wan Liyun jointly launched the medical AI product “Doctor You”, which announced that the lung nodules were correctly identified by more than 90%. One month later, Tencent launched the medical imaging AI product “觅影” to screen for early esophageal cancer. Accuracy is as high as 90%.

"Accuracy rate exceeds 90%", "speed beats doctors", one keyword seems to make AI become the key to medical mineral deposits, AI medical imaging, AI auxiliary diagnosis and treatment, AI drug development, AI health management, people waiting for people Swing "iron shovel".

"Now many companies are also artificial intelligence, and that is also artificial intelligence. In fact, there are two places that really need artificial intelligence. One is difficult to diagnose even people, such as pathology; the other is that people who have worked hard have already done it. Zheng Zhongxi, a professor at Huaxi Hospital, told Titan Media, "In the field of cancer diagnosis and pathology, we are very much looking forward to embracing artificial intelligence."

As the fastest application area of ​​AI in the medical field, medical imaging AI has poured more than 4 billion yuan in 2017. According to the potential investment statistics of Titanium Media (see the end of the table for details), the highest amount of financing is Lianjiao 3.333 billion yuan. Round of financing; Imagine Technology, Tuoma Shenwei, Shenrui Medical, Vision Medical, etc. all obtained two rounds of financing in 2017; in the first half of 2018, Shenrui Medical and Airdoc obtained B round financing, Huihui Huiying, Pushing Technology Enter the C round stage one after another.

Medical imaging AI presents a facet of China's Internet business environment: the influx of hot money, the surge of entrepreneurs, the serious homogenization of products, the emergence of bubbles, and the questioning of business models. Tencent and Ali entered the game, and AI medical images were completely “ignited”. ". However, the uniqueness of the industry lies in the fact that medical imaging AI also has Philips, GE, and Neusoft Medical, which are originally located in the medical imaging industry and are hidden under the Internet.

According to the calculation of Professor Liu Shiyuan of the Chinese Medical Association Radiology Branch, the medical imaging AI should be fired for about two years. "The heat is now very high and has entered a critical stage. The development of AI has also entered the deep water area. What kind of problem is solved? What areas does the product focus on? How do the upstream and downstream industries closely integrate? How do products solve clinical practical problems? There are actually a series of problems."

Sad data off

“2017 can be named as the lung nodule year.” Fan Wei, director of the Ali Health Artificial Intelligence Laboratory, told the Titanium media that “the lung nodule is a process of barbaric admission. Many people are doing lung nodules and can get a lot of Information, fast entry."

Compared with the startups entering the mid-term, Tencent and Ali are not too early. Chang Jia, the head of Tencent Internet + medical business, believes that not only BAT, but also the entire medical image is highly integrated in product and pathology. This is mainly the industry. The problem of getting started.

The open data set is the direct cause of the emergence of lung nodule products, and it also provides opportunities for latecomers to “over the corner”. But the lung nodules are the first step in AI's cutting into the medical “long march”.

For products that also claim 90% accuracy, Chang Jia mentioned, “Now many people claim that their accuracy, test and data levels are homologous, and the accuracy is high, but different source conditions are more critical.”

In more than 100 top three hospitals where the shadows landed, Chang Jia found that the anti-noise requirements were very high. Different equipments and different doctors had different operating habits. Some hospitals did very good results and did not apply to other hospitals. It is now being solved through extensive testing, but it is a big difficulty and requires data richness."

He Guowei, CEO of Philips Greater China, shares the same view. “Different devices use different data at different clinical nodes. Simplify human repetitive behavior, the problem is data flow, and there must be a complete data flow in conjunction with clinical. No clinical application. When you are structured, the data that goes into deep learning is garbage."

But the more critical issue is that many platforms are doing medical AI, but what is the fight? Qiao Youlin, a professor at the Cancer Hospital of Peking Union Medical College, told the Titanium media that “it is possible to get medical high-quality, gold-standard materials, even if you take thousands of films, you won’t get the correct answer.”

From public data sets, data richness to gold standard data, the difficulty is gradually increasing. Compared with the demand for medical imaging AI products in the top three hospitals, the demand for medical imaging AI products bundled hospitals is more intense. How to get hospital data to train AI is the key to achieving differentiated competition among various products. He Guowei said frankly, "There are many necessary conditions for AI to land in the medical system."

Grab the hospital and go to the grassroots

But for the current AI medical imaging products, the biggest problem with entering the hospital is "not easy to use."

Yan Weixin, a professor at Shanghai Jiaotong University, told the titanium media that “the Renji Hospital has about 180 people in the radiology department, and the number of outpatients in a day is nearly 3,000. I have calculated that artificial intelligence can help him save 30% of his manpower a year. It is tens of millions of grades, but why is it useless? It is not easy to use, so it is easy to use."

Yan Weixin believes that artificial intelligence is still only a "six-seven-year-old child", which can only help clinicians to outline simple and rude activities, and can't take up the work of clinical doctors.

"Medical, Internet + AI are two different worlds. Simply speaking, there are two different languages. Your language and my language don't have much intersection." Zheng Zhongxi told Titan Media, "Doctors are more 'Use', but from the perspective of artificial intelligence, there must be integration. You can understand what you say, and you can understand it when I say it."

Compared with the accuracy rate, doctors are more concerned about whether it is good or not.

What is the standard for doctors to judge? Wu Wenzhao, a medical practitioner, told Titan Media that from the perspective of doctors, whether AI products meet the needs of the medical industry depends mainly on solving clinical problems. If an AI product is finally designed according to the needs of doctors, it will achieve certain results according to doctors' plans. It will definitely be a product that doctors feel is very good, and the close integration with doctors' research and development is the ultimate development of AI in the future.

As a doctor in a cancer hospital, Qiao Youlin put forward a more specific demand. "When applying artificial intelligence, we must use gold standard materials and develop a set of products. It is like a very difficult 'grey zone', a plausible place. We divide cervical cancer into five levels, normal, cancer, and there are three levels in the middle. These three are the key. It is simple to identify whether cancer is, but which level is difficult."

When the products are different and there is no unified measurement indicator, Tsuma’s CEO, Zhong Wei, believes that “the measurement standard has two big blocks. One is that the connection is convenient and inconvenient, the doctor is comfortable to use, and the other is the function. Incomplete, performance is not the best."

Although entering the hospital has become the key to winning, but Qiao Youlin told the titanium media that the most urgent scene for medical imaging AI products is at the grassroots level. "I hope that patients with intractable diseases will come up, and other minor injuries will not come over. You also hurt people. The problem is that doctors at the grassroots level do not know which diseases should go up, and patients do not know."

Wang Wei, chief marketing officer of GE Healthcare Greater China, introduced to the titanium media that the traditional liver site of the tumor interventional embolization surgery, the tumor oxygen supply blood vessel embolization, tumor necrosis, but many times there are multiple oxygen vessels, experience Doctors are easier to judge, but for experienced doctors, using artificial intelligence can help them to display all the oxygenated blood vessels at once.

On the one hand, it is highly tied to the top three hospitals that own the data, and the products are polished; the other side needs to go deep into the grassroots that have urgent needs and get the first application. It seems that the two-way advantage is the device manufacturer that was originally in the upstream of medical imaging. In the market environment where the high-end equipment of the top three hospitals is saturated, Philips, GE, and Neusoft Medical have taken the grassroots strategy in recent years, and thus established an image cloud platform and updated equipment software to achieve remote diagnosis and treatment.

Giants make platforms, entrepreneurs are vertically segmented

According to the observation of the titanium media, although the two types of giants represented by Philips and Tencent have different focuses, they are all moving toward the platform, while the founding enterprises are deeply immersed in the subdivision. Faced with the huge market of AI medical imaging, the emergence of giants is not reflected in the direct crushing of the founding enterprises. For the giants, it may be more crucial to formulate the rules of the industry game, but in the platform ecology that the giant wants to create, The growth space of enterprises must face more challenges.

In November 2017, the Ministry of Science and Technology held a new generation of artificial intelligence development planning and major science and technology project kick-off meeting, and announced the first four national artificial intelligence open innovation platforms, one of which is to build a national artificial intelligence open innovation platform based on Tencent. .

“The scope of medical AI is very large, screening, diagnosis advice, and robot navigation are available. This is not something that can be done by a certain company. We hope that the whole industry will work together to develop Tencent’s resources in Shuangchuang.” Chang Jia Tell Titanium Media, the idea of ​​Yingying is to build a platform. On the basis of the platform, Tencent invested in AI drug research and development company Jingtai Technology, Atomwise, led the AI ​​medical company voxel technology, AI health management carbon cloud intelligence.

Compared with the “Medical AI Supermarket” created by 觅影, medical device manufacturers such as Philips are more concerned with the overall solution for specialist diseases. At this year's CMEF exhibition, Philips released the Shenfei cloud imaging platform, and AI appeared as one of the modules.

At present, Philips is working on AI medical care, and 60% of research and development costs are invested in software and artificial intelligence. He Guowei told Titan Media that Philips wants to build an ecosystem. "We have an open software platform. Whether it is a lung nodule or other diseases, startups can use our platform. Finally, we will explore each future based on each company. What is the difference in the direction of development?"

Why does the AI ​​medical imaging industry need a platform? In Zheng Zhongxi's view, "The significance of an open platform is to integrate information. Image diagnosis is not just a conclusion about a certain feature of this figure, especially after cytology and histology, there is a lot of information to be integrated."

On this point, Wu Wenzhao gave an example. "The vast majority of products in China are pulmonary nodules, but after the lung nodule screening, what can the system do to get a diagnosis to see if it is malignant early lung cancer? If it is early stage lung cancer, after the operation plan is treated, the follow-up needs to be done. All the following things are blank for all current AI companies.” Wu Wenzhao summed up the key points of the medical imaging AI product wins as two points, “ First, there is a large amount of data. Second, it is enough to mobilize enough people to standardize and structure these data, and to bundle them with clinicians and imaging doctors."

The data requires a lot of capital and resources, and the business model is still unclear. This road that does not see the end point will become a barrier for the giants to build up. "After 17 years, 18 years is definitely a process of differentiation, and the barbarians have begun to become regular troops. Companies will slowly move in different directions." Fan said.

“I think everyone is a starting line in a certain segment.” Zhong Wei believes that the advantage of big companies is that they can be opened up. Small companies will do more on a few things. “Last they are Face, we are sharp. Sometimes the face will not hide the awl, we are a point, then form the diffusion of the point, and then form the group of points, but some come up is a multi-point group. Finally Everyone will do multiple product lines, which is the goal of all companies."

Although most companies in the industry currently have lung nodule products, there are also startup companies that study areas such as retinopathy and cardiovascular disease. Zhong Wei told Titanium Media, "After a long time, you will find that everyone's products still have There are big differences, and some products are always stuck in testing and two-dimensional testing. Everyone has a variety of differences."

He Guowei revealed to Titan Media: “For startups, financing may be more important, but we are more important about how to provide more effective oncology solutions, patient experience, and clinical value; we have accumulated a lot in the past few decades. With the data and experience, we feel that our clinical experience is stronger than others."

However, in Chang Jia's view, the advantage of the device manufacturer is that the data has an advantage, but the disadvantage is that it is too limited. "When doctors operate, many hospital departments do not use only one department. So we think that the effect and quality There is a need for a long-term development. In fact, everyone can discuss how to improve quality, instead of discussing how to do the division of labor."

Embracing the group for commercialization

“In the next two years, medical AI will die. At least in my opinion, this will be a high probability event.” After watching more than 30 medical AI projects, Yuan Min’s capital Tian Min said in an interview with the media. . In March of this year, Caijing magazine also issued a document saying that China Medical AI Company encountered the "C-round death" curse, how to achieve commercialization, and became a question that medical imaging AI products need to answer in 2018.

These questions for startups revolve around the point that medical imaging AI technology can achieve business value?

Liu Jiren, chairman of Neusoft Group, believes that “when a company with a single AI technology is fully popularized and shared, and there are some free opportunities for business models, this is the risk of investing in AI technology.” Wan Liyun CEO Huang Jiaxiang was accepting The media interview said, "For AI companies, we have a view that the value of the application scenario is greater than the value of the data, and the value of the data is greater than the value of the algorithm itself. If it is the startup of AI, the only thing in the core is the algorithm, then this thing It will be very dangerous."

In April 2017, Tuma was awarded a multi-million dollar A round of financing. Seven months later, it received a 200 million B round of financing led by Softbank China. Zhong Rong told Titan Media that in the A and B rounds of financing, the focus of capital is very different. “It is no longer a product primary form or just a potential sales model. The B round financing should see if we have Some implementation, advance the speed of the hospital, some feedback from the hospital in the hospital, and whether the business development has been implemented, this may be a more important point in the B round."

On August 31, 2017, the State Food and Drug Administration (hereinafter referred to as CFDA) released a new edition of the “Medical Device Classification Catalogue”, adding a category corresponding to AI-assisted diagnosis, which is embodied in the catalogue for medical imaging and Analysis and processing of pathological images.

If the diagnostic software provides diagnostic advice through an algorithm, only the auxiliary diagnostic function does not directly give a diagnosis conclusion, then the second type of medical device is declared. If the lesion is automatically identified and a clear diagnosis prompt is provided, the third type of medical device is followed. management.

This means that if medical imaging AI products want to take the road of hospital procurement, they must pass the corresponding certification. However, at present, there is no medical AI product in China that has been certified. The arterial network summarizes this phenomenon into six core issues, including “the standard database for approval is gradually being established.”

In the Suzhou Artificial Intelligence Conference on May 11, Liu Shiyuan mentioned that there are currently nine companies applying for CFDA certification, but what standards and specifications are still under discussion. Zhong Rong told Titan Media, "Getting CFDA is a very important node for every medical AI."

In the process of interviewing with titanium media, different types of participants will mention a key question. Is the AI ​​medical imaging service paid for by the hospital or by the patient? Since the “Artificial Intelligence Reading Fee” is not included in the list of medical treatment charges, and considering the comprehensive factors such as safety, many hospitals still have no incentive to purchase artificial intelligence.

For medical device manufacturers, AI is an added value, regardless of its commercialization as a separate product. Liu Jiren told Titanium Media that “AI technology is not highly likely to obtain value when solving medical problems, but To get value through other carriers. Neusoft does AI technology, we don't need to make money through a single AI technology. For example, all our CT scans tell him what is wrong, we only need to sell equipment to make money."

In this experiment of exploring the commercialization of the group, Fan Wei believes that the government plays an important role. "The business model must be jointly promoted at the government level, the hospital level and the company level. The improvement of medical efficiency is a national proposition, but only in There is no way to go through the specific strategy, and it needs to be iterated."

After experiencing the “barbarians admission”, it proves that the existence of value is no longer the top priority of medical imaging AI. The more important question is, what kind of industry ecology will the startup companies, Internet giants and medical device manufacturers finally form? In what way is the search for commercialization? Which medical imaging AI ventures can run out? These questions remain to be verified by time.

"At present, in the Warring States era, the future may be unified." He Guowei said.