Focus on digital health ｜ The wave of medical AI is reunited: How to start across the gap again?

The 21st Century Business Herald reporter Tang Wei Ke's intern Berlin reported that under the promotion of the country, the domestic medical AI development wave was resurrected.

Recently, the Department of Science and Technology of the Ministry of Industry and Information Technology and the Medical Devices Registration Department of the Ministry of Industry and Information Supervision and Administration released the "List of the Innovation of Artificial Intelligence Medical Devices Innovation Mission Unveiling", and a total of 221 projects were shortlisted. The unveiling task mainly faces the two directions of intelligent products and supporting the environment. It focuses on 8 types of tasks such as intelligent auxiliary diagnostic diagnostic products, intelligent auxiliary therapy products, and medical artificial intelligence databases. Many well -known enterprises in the industry are listed, such as Baidu, Mai Rui, HKUST Xunfei, Lianying Medical, Eagle Tong Technology, etc.

Although AI technology has continued to heat up in recent years, the story of medical AI has also been widely sought after -matching high -quality medical resources, improving hospital efficiency, strengthening quality control, and reducing misdiagnosis rates, but from scientific research to application to commercialization, medical AI still needs to cross the cross Several gaps.

Professor Xie Hongning, Director of the Obstetrics and Gynecology Ultrasound Department of the First Affiliated Hospital of Sun Yat -sen University, said to the 21st Century Business Herald reporter: "Doctors work in the first -tier clinical clinical, but also take into account students, medical education and scientific research. Do you have to do a lot of patients every day. From the doctor's side, we hope that medical artificial intelligence can really help us, improve our work efficiency, share work pressure, instead of only staying in the only technology and technology Study, you can really take out the product to help us. "

However, the strong supervision and complexity of the medical industry has determined the bumps of medical AI from concept to landing applications. At present, ultrasound is the most widely used specialty in medical imaging. AI has gradually provided auxiliary functions to doctors in front -line medical care.

Positioning switch after frustration

As early as the 1970s, the exploration between computer and disease began abroad, and it was mainly used for matching between diseases and diseases. In 1974, the University of Pittsburgh developed an expert system for Internist-I internal medicine. The knowledge base included 572 diseases, about 4,500 symptoms, and the connection between 100,000 diseases and diseases. The system entered a commercial stage in the 1980s, but was subject to the computing power and algorithms of the computer at that time, and entered the bottleneck period in the 1990s.

Then entered the AI ​​image era. With CT's clinical practical in the 1970s, the birth of digital equipment for medical images, the further development of image storage and transmission standards, AI images have a lot of attempts. However, it was restricted by the image resolution and algorithm at the time -a breakthrough in deep learning in 2006 -IBM Watson was a typical explorer in the field.

Theoretically, Watson can greatly improve medical efficiency. At present, medical documents will be twice as many as possible every 73 days. Watson can learn 2.67 million pages per second. In contrast, doctors need to spend 160 hours a week.

However, in fact, Watson's commercialization is not smooth, and the gap between revenue and investment is huge. In recent years, Watson has spent billions of mergers and acquisitions of machine learning, medical clinical data, population health data, medical image algorithm and other backgrounds. It spent nearly $ 4 billion in 2016. However, its financial report in 2020 shows that revenue is only 1.5 billion US dollars. In December 2020, it was reported that Watson would be sold by IBM.

As the medical AI stepped into the deep water area, the development trend of alternative doctors in the past has also become a further empowerment to assist doctors, helping doctors improve their work efficiency, and accelerate the development of medical treatment.

Professor Xie Hongning said: "In terms of ultrasound application, the automatic recognition function of artificial intelligence can help us quickly improve our work efficiency. Normally checking a pregnant woman needs to scan multiple parts of the fetus, including the fetal brain, heart, limbs of the fetus , The chest and abdomen organs, etc., but the long -term work of doctors is prone to tiredness, and it is easy to leak when we are busy. At this time, what we need is the medical AI that can effectively improve our own work efficiency to remind doctors every detection site. "

In addition, medical AI also has a larger application scenario in the sinking market. Professor Li Anhua Li Anhua, vice president of the Chinese Society of Ultrasonal Medical Engineering, pointed out to the 21st Century Business Herald that the current rainbow phenomenon of higher -level hospitals is serious. Talent. Taking breast cancer screening as an example, there is still a large screening gap. On the other hand, even if it is screened, how many women are still unknown to see a doctor. Many township hospitals have their own profit and loss, and even if there are no doctors in the machine, even a new B -ultrasound machine is placed in the corner. It is crucial to solve the gap in regional personnel through medical AI. In addition, how to use medical AI to introduce a standardized quality control system will also give new significance to the next step of breast cancer screening.

Medical AI uses two core technologies of machine learning and data mining to explore simulation, extension and expansion of human intelligence in medical scenarios. In clinical medical treatment, the medical AI can be shared by the repeated labor based on a large amount of standardized data. For example, diagnosis of diseases supported by objective data such as medical images and pathological interpretations.

In the hospital, doctors often get older and more popular, and medical AI deep learning database contains global clinical data and research documents. The volume of these data may be far more than the amount of medical data mastered by ordinary doctors in their lives. In other words, let computer learning medical big data after judgment, its efficiency and accuracy may surpass doctors. Of course, this is something.

With the collision of the wall again and again, we have to admit that the current medical AI role reality is the decision support that embeds the entire medical treatment and outputs the level of doctors, not replacing doctors. Where is the direction of the difficult breakthrough?

In the future, many difficulties faced by medical AI will still focus on data algorithms, landing and payment, and the combination of technology and clinical needs.

In addition, domestic medical AI companies generally lack self -venture algorithms, and computer computing power is also very constrained. "The current bottleneck is an algorithm. Chinese companies do not have their own algorithms, or they are taken from foreign algorithms. If you do not make a lot of re -compilation work, then the matching of the algorithm is not high." Li Anhua said, "Another another It is also important to combine the combination of technology and people who know clinical needs. "

Dr. Wang Nan, the founder of the CEO of Guangzhou Aidi Pregnancy, also pointed out: "People who can invent the algorithm are still mainly concentrated in Google, Microsoft, etc., but the original algorithm is basically identifying natural images. For enterprises, if you want to use it in the future In imaging or ultrasound, we must do a lot of compilation and optimization work on the bottom algorithm to allow it to adapt to the attributes of the application scenario. After the compilation is completed, a large number of tests are required. At present, Chinese companies are rarely done. "

Zhang Xi, Executive Vice President and Secretary -General of the Guangdong Artificial Intelligence Industry Association, told the 21st Century Business Herald reporter: "In the future, the medical AI industry needs to solve the problem of difficulty combining technical and medical talents. Intelligent in various segments of the medical industry also gradually integrated deeply, and intelligent medical care has gradually subdivided the auxiliary diagnostic system based on image recognition, digital hospitals based on big data processing and NRP technology, and drug research and development based on deep learning. Multiple subdivided tracks. "

The user group of medical AI can be divided into TO-C and To-B.

To-C mainly faces individual users, mainly in two types: intelligent consultation and health management. TO-B includes pharmaceutical companies and hospitals. The former is mainly used in drug research and development; the latter includes medical images, virtual assistants, medical research, hospital management, and gene sequencing.

Among them, AI medical imaging is currently a popular field of medical AI. AI medical imaging can help doctors' lesions screening, target area drawing, three -dimensional imaging, image analysis, quantitative analysis, etc. Compared to the artificial film that is repeatedly repeated according to experience, AI can be screened in batches according to the standards, with a short time reading time and a stable accuracy rate.

In addition, the image recognition algorithm is relatively mature, and the market demand is large (the images of the image of the image of the image -the average annual growth of the image data, and the average annual increase of radiologists by 4.1%). At present, medical imaging data accounts for more than 80%of all clinical data, and it is the cornerstone of clinical diagnosis, disease treatment and health management. Due to the uneven distribution of medical resources in my country, many doctors in remote areas have insufficient experience and equipment, and the analysis of medical images is not accurate and inefficient.

In front -line medical scenarios, the analysis of medical images is complicated and time -consuming. Doctors must combine multi -layer image data with fine decision -making processes and interpret evaluation. Wang Nan said: "Medical AI can help doctors solve some mechanical repeated labor in terms of medical imaging, release the doctor's time, and do some creative and decision -making work. This is the positioning of medical AI."

In the early industry, medical AI companies mainly wanted to solve these practical pain points in the clinical scene of the hospital. For example, in pushing medical care, they all adhered to the research and development ideas of "from clinical, to clinical", and the hospital with the hospital with scientific research cooperation. Perform algorithms and products. The development of artificial intelligence technology is expected to bring solutions. With the maturity of the product, the medical AI gradually enters the actual clinical scene of the hospital for the use of the department and other departments. At present, hospital scenes are the most mature scenes of medical AI, and they are also the most intense scenarios at present.

In terms of the cultivation of medical talents, medical AI also has corresponding empowerment scenes.

Xie Hongning pointed out that when training young doctors, high -age doctors often cannot take out time because of too much diagnosis patients. If you join the auxiliary of medical AI and the efficiency of work is improved, the mentor can talk to students one more lecture or talk with students in this case. The addition of medical AI in teaching assessment can also help standardize teaching and improve the quality of teaching. "Students will feedback artificial intelligence in practice. This structure is not right and can only score five points. For the assessment, it was easy to do the topic before. After all, the score is objective, but it is difficult to implement the operation test. Intelligent joining, students can score in a few seconds after practicing operations by themselves, which will greatly help teaching. "

- END -