The once popular AI medical industry has now faded its mysterious color. The relatively mature fundus AI technology has highlighted the encirclement and has been active in people's sight. As we all know, the retina is the only tissue in the human body that can directly observe the changes of blood vessels and nerve cells in a noninvasive way, which can be used as an indicator of a variety of chronic diseases. Therefore, as a means of auxiliary examination, retinal imaging has always been listed as the research focus of medical imaging, and plays an important role in AI medical industry. According to the prospectus, the growth rate of artificial intelligence imaging technology in the composite market of Eagle medicine was 171.1% in 2019. AI medical has repeatedly encountered obstacles in the commercial battlefield and has been unable to touch the core of diagnosis and treatment. In addition to the development of technology that can not keep up with the rhythm of the times, in essence, it can not completely replace the dominant position of doctors. Even if the development prospect of fundus AI is broad, it has not yet walked out of the strange circle of loss year by year. "Burning money" has always been an indelible label of AI medical treatment. So, the medical AI industry is in this awkward position. Can it become a guiding light for emerging industries? How does fundus AI break the game? There is a consensus in the circle that the most difficult part to develop AI medical treatment is the establishment of database. After all, the data quality in each hospital's information system is uneven, and enterprises in the early stage of entrepreneurship can only reach cooperation through paid purchase or scientific research cooperation from one or several hospitals, but it is very likely that they can only obtain some incomplete data. There is no such worry in the field of fundus AI. According to the data, in March 2018, the fundus image database of the Chinese people's Procuratorate was completed. Based on this basic knowledge, in August 2020, Yingtong technology and silicon-based intelligent "sugar net auxiliary diagnosis software" passed the approval of the food and Drug Administration and obtained the certificate of class III medical devices. Then, the products of Zhiyuan huitu and Weiyi Medical Devices Co., Ltd. have also obtained certification, and several of them have obtained EU CE certification. It can be said that the development of fundus AI is steady, and a diagnosis system based on retinal images has been gradually established. According to our understanding, by capturing and analyzing retinal images, we can not only identify eye diseases such as retinal macular degeneration, pathological myopia, retinal detachment, but also diagnose systemic diseases such as diabetes, hypertension, cardiovascular and cerebrovascular diseases. Fundamentally speaking, although fundus AI can not completely replace ophthalmologists in diagnosis, it plays an obvious preventive role. According to the 2017 statistical yearbook of China's health and family planning, there are about 400 million patients with chronic eye diseases in China, while there are only 36000 ophthalmologists, with more than 110 million outpatients and more than 4.5 million inpatients. Combined with the whole medical environment, the emergence of fundus AI has enabled many chronic diseases to be "screened and treated early", and further combined with the diagnostic results of doctors, many patients have been treated early. In addition, the industry has its own specialties. Ophthalmologists' diagnosis and treatment of diabetes, endocrinology and other specialized diseases are not necessarily accurate and timely. Non ophthalmologists' ability to read films of fundus images is limited after all. In particular, the focus of fundus diseases is small and the differentiation between different focuses is low. Relatively few doctors nationwide have the ability to read films. The auxiliary diagnosis and treatment of fundus AI can become a "helper" of doctors to a certain extent. The reason why fundus AI can stand out in the artificial intelligence image market is also closely related to the diversity of enterprise landing scenes. Whether it is lung image, cardiovascular image or chest image, it will generally choose to land in large hospitals and bind hardware equipment for sales. The selection range of fundus AI is wider, which can either choose the third class hospital or make a small-scale breakthrough from the visual center. Just like the fundus screening software of Yingtong technology, we choose to use a low-cost fundus camera to achieve the penetration of the grass-roots medical system. However, the trend of fundus AI is good, and it is obvious that it does not make money. For example, according to the prospectus of Yingtong technology, in 2019, 2020 and the first half of 2021, its realized income was 30.415 million yuan, 47.672 million yuan and 49.477 million yuan. In November last year, Eagle pupil technology, the "first share of medical AI", successfully landed, but broke immediately after listing. In addition, if the emerging industry is in a downturn, it must be in a new state of development. If it wants to make a breakthrough in the new business model, it must be in a positive state first. Artificial intelligence + medical care, go hand in hand? Perhaps due to the influence of western movies, artificial intelligence has always been the representative of the ultimate form of human beings. In recent years, the development of artificial intelligence has also advanced by leaps and bounds. Its research shows that the amount of calculation used in the largest artificial intelligence model training increased exponentially from 2012 to 2018, of which the amount of calculation doubled in 3.5 months. Secondly, capital also continues to be optimistic about the development of China's artificial intelligence industry. By 2020, the financing scale of China's artificial intelligence industry will reach 140.2 billion yuan. In particular, traffic scenes and medical scenes have attracted extensive attention. In particular, enterprises have said that they will focus on artificial intelligence in the future. According to iimedia research, more than 80% of Chinese Internet users are optimistic about the future development prospects of artificial intelligence, of which traffic scenes and medical scenes account for 45.2% and 40.5% respectively. Eighty percent of the enterprises surveyed said they attached great importance to AI, and nearly 60 percent said they would focus on AI in the future. The arrival of the epidemic makes the scale of China's medical device market continue to expand, especially artificial intelligence plays a vital role in the medical industry. According to iimedia research, China's medical device market reached 734.1 billion yuan in 2020. Among them, intelligent medical tools such as surgical robots and AI clinical auxiliary systems play an important role in fighting the epidemic. At the same time, the number of patents related to artificial intelligence medical devices in China has further increased. It cannot be denied that the development of AI medical industry combines many forces such as medicine, science and industry. In addition to the boost of capital, it is also inseparable from the scientific research strength of relevant colleges and departments. On March 25, 2017, Zhejiang University Ruiyi artificial intelligence research center was established. Professor Wu Zhaohui, the leader, once said, "we will make use of the basic advantages of Zhejiang University, cooperate with major medical institutions and discipline leaders with an open attitude, open up data sources, establish a multi-disciplinary and multi agency coordination mechanism, and strive to make a breakthrough in the key technologies of medical artificial intelligence as soon as possible." However, the results did not develop as expected. The growth rate of the domestic medical AI market was extremely slow, even declining. According to statistics, the overall scale of China's medical AI market in 2020 was about 26.5 billion, and from 2015 to the first half of 2020, the total financing scale of medical AI exceeded 35 billion. In 2021, although medical AI enterprises ushered in the great mark of listing, the report card of two years of full commercialization is still sad for the eyes of the secondary market and lost one after another. In March, Keya medical submitted an IPO application to the Hong Kong stock exchange. By September, the listing status had become "invalid"; In June, Yingtong technology also submitted an IPO application. In November, the IPO broke; In July, the IPO status of Yitu technology became "terminated". In August, Yitu medical team was acquired by Shenrui. Not only did the domestic AI medical industry fail, but Google Health, which is "fully open" in the medical field, also fell into a lot of crises and had to lay off and restructure on a large scale. Google is already a giant in the AI field. After the establishment of Google Health, it also hired David Feinberg, a senior person in the medical field, as the director, and has not yet done Google's innovative business related to medical AI. According to the quarterly financial report of 2021 Q1, Google's innovative businesses, including artificial intelligence deepmind and intelligent medical verily, are still at a loss. The combination of artificial intelligence and medical scenes is not only a test of technology, but also a multi-faceted test of real situations. Unfortunately, such a large medical field can not save the application depth of artificial intelligence, not to mention that medical AI itself is struggling. AI medical, trapped in algorithms The characteristics of artificial intelligence itself determine that it cannot face complex medical scenes. It is difficult to achieve independent diagnosis and treatment without artificial intelligence. Turing Award winner yoshua bengio also said that in medical text documents, artificial intelligence systems cannot understand its fuzziness or the subtle clues noticed by human doctors. It can be said that the algorithm needs to be optimized. Only by strengthening the existing model with more comprehensive data can AI medical be gradually applied to more scenarios. Dating back to 2020, Wu Enda analyzed in his speech at the Stanford Hai seminar that the algorithm of AI research in the medical field is difficult to be put into production, because the model trained by some data is difficult to be generalized to other situations. This statement has been confirmed in the "diabetes retinopathy screening" business of Google health. Previously, Google mentioned in an article published in the Journal of the American Medical Association that AI algorithm makes the accuracy of "diabetes retinopathy screening" reach 90%, and theoretically the results can be obtained in a few seconds. However, when theory is combined with reality, there is a great contrast. In 2020, Google cooperated with the Thai public health department on this business. Finally, due to the high requirements of the algorithm for checking photos, the accuracy was far lower than expected. In addition, the time from uploading photos to the results also depends on whether the network signal of the local hospital is good. Patients usually take a long time to get the results. Obviously, AI healthcare relies heavily on external factors such as the environment and photo pixels. Once unexpected events occur, it will not be able to play its own role. Even the relatively mature fundus AI algorithm can not adapt to various situations, not to mention other fields with limited development. There are many limiting factors in the application of AI in medical treatment. First of all, due to the basic geographical restrictions, AI is not universal in the actual scene. In essence, AI with high applicability is basically customized research and development. For example, there are some problems in the diagnosis of children's bone age intelligent auxiliary diagnosis software. This software, which refers to the height in southern China as the standard, is not suitable for local areas with prominent average height. Secondly, based on the technical limitations, the diagnosis and treatment tools of AI in cardiovascular on the market focus on the image function. The technical difficulties of high threshold have deterred many enterprises. Compared with the scanning images of retina and lung, the images of extremely complex reticular structures of heart and coronary artery are more difficult to capture, and the image synthesis and three-dimensional reconstruction are very difficult. AI needs more sophisticated and huge algorithms in the cardiovascular field to diagnose and predict a very rich variety of heart diseases. Finally, AI medicine is applied to real medical scenes, and a large number of training data is the premise, but the accuracy and comprehensiveness of the data are not guaranteed. For example, according to the internal documents of IBM published by the American Medical Media stat, the training of Watson system uses the hypothetical data of virtual patients, and the recommended treatment scheme is based on the scheme commemorating the experts of Sloan Caitlin cancer center; And the training data is insufficient. Among the eight cancers, the highest amount of training data is only 635 cases of lung cancer and the lowest is only 106 cases of ovarian cancer. It is not difficult to see that the optimization of the algorithm can not keep up with the changes of the medical situation at all. Many medical AI are not used in the real operation