In-depth | AI medical care has been bumpy to implement, and 15 years later they are starting again

Leifeng.com: From gene detection to medical big data mining, from intelligent diagnosis to precision medicine, the realization of smart medicine is not something that can be accomplished unilaterally by the medical community or the scientific and technological community, genomics or imaging. Genetic data, imaging information to standardized big data is an unbreakable chain. As we all know, the value of artificial intelligence was discovered decades ago, but due to practical conditions, medical innovation pioneers have experienced breakthroughs and failures again and again. To date, AI medicine is still in its infancy, but the progress of the times has given us a new foundation. Today, this article will tell how innovators set sail again and revitalize artificial intelligence medicine from a global perspective through interviews with people from the medical, scientific and industrial communities.

In June 2016, Tian Jie's team from the Key Laboratory of Molecular Imaging, Institute of Automation, Chinese Academy of Sciences, and the Department of Radiology, Guangdong Provincial People's Hospital, used the emerging radiomics method to make important progress in the prediction of colorectal cancer lymph node metastasis. Compared with traditional CT imaging assessment, the radiomics prediction model increased the accuracy of preoperative lymph node prediction by 14.8%.

As early as 2003, China launched its first medical imaging big data research project. However, because the data was difficult to normalize at the time, the project failed. Nearly 15 years later, we have just officially had the opportunity to start doing this.

——Dr. Zhenyu Zhou, Senior Director of Clinical Science, Philips Greater China

Artificial intelligence is very popular now, and industries with huge stock data such as finance and medical care have become the first choice for the application of this technology. In addition to finance, which has taken the lead, BAT, start-ups, etc. have also begun to deploy artificial intelligence in medical care in the past one or two years.

It is worth mentioning that at the "Medical Artificial Intelligence Frontier Summit" hosted by Huiyi Huiying, a smart medical imaging platform startup, in Beijing a few days ago, Leifeng.com found through interviews with medical industry professionals from Philips, Siemens, front-line medical workers, academia, medical technology innovation, Intel, etc. that in terms of revitalizing artificial intelligence medical care, all walks of life have explored smart medical care and universal medical care much earlier than we thought; the joint collaboration within the industry is also beyond the reach of other industries that are also seeking AI innovation.

Uneven distribution of medical resources and policy obstruction: AI brings hope to China's medical predicament

In the current Chinese medical market, the state has invested a lot, but the contradiction in the allocation of medical resources is still prominent. According to public data, in the huge medical institution system, hospitals with high-quality medical resources only account for 0.1% of the total number of medical institutions under the National Health and Family Planning Commission.

A large number of patients come from the grassroots, but the few tertiary hospitals are crowded with people. In fact, patients' distrust of the grassroots is not about the facilities, but the people. According to statistics, the misdiagnosis rate of grassroots hospitals in my country is relatively high, and the number of misdiagnoses in the field of medical imaging reaches 57 million times per year; in general, the misdiagnosis rate of tumors in hospitals is as high as 60%. Not to mention the early years.

Therefore, it is difficult to implement the hierarchical diagnosis and treatment plan issued by the state. Whether it is doctors from tertiary hospitals going to the grassroots or telemedicine, it still requires excellent doctors to contribute their time, so it cannot solve the fundamental problem - the lack of excellent doctors and low service efficiency.

At the same time, Li Yadong, from Intel's Medical and Life Sciences Group, said that in China, the trend of an aging population and the growth of chronic diseases has driven the market's greater demand for AI. Therefore, AI has become the only way for medical innovation - empowering with productized artificial intelligence and making the capabilities of excellent doctors replicable.

This is a blue ocean market and also a career full of warmth.

The biggest resistance of smart medical care over the years

In fact, as early as October 1984, Academician Wei Yu, former Vice Minister of the Ministry of Education and President of Southeast University, founded the Department of Biological Sciences and Medical Engineering of Southeast University, which was the predecessor of the School of Biological Sciences and Medical Engineering established in August 2006. The establishment of the School of Biological Sciences and Medical Engineering aims at the leading disciplines of the 21st century - life sciences and electronic information sciences, and emphasizes the intersection and penetration of these two disciplines.

Dr. Zhou Zhenyu, senior director of clinical science at Philips Greater China and one of the earliest students to receive dual-degree education in medicine and engineering, told Leifeng.com (official account: Leifeng.com) about his generation's exploration of smart healthcare. He said, "The Department of Life Sciences Engineering was originally engaged in dual-degree education in medicine and engineering. Students were required to study engineering first and then medicine. At that time, we wanted to create such a big data platform for imaging."

It is understood that in 2003, China launched the first medical imaging big data research project, and Dr. Zhou Zhenyu also participated in it. However, at that time, although the project received funding support of up to 5 million RMB, it ultimately failed. This was because the algorithm problem was mature, but there were still many insurmountable challenges: the quality of equipment imaging, data, and the lag in computer capabilities.

With the participation of semiconductor manufacturers such as NVIDIA and Intel, the addition of CPU\GPU\FPGA has made great progress in computing processing capabilities. However, as Xing Lei, director of the Department of Medical Physics at Stanford University and a tenured professor, pointed out, the lack of data concentration and standardization is one of the biggest obstacles to the development of intelligent medicine. The problem of data normalization caused by standards, system compatibility and interoperability still exists. This is why Huiyi Huiying hopes to build a cross-device interconnected medical imaging cloud platform.

Professor Xing Lei said that the integration of medical images, case histories and other data to make comprehensive intelligent analysis and decisions is still in a very primitive stage.

"At present, hospitals are still far from doing enough to conduct systematic, comprehensive and intelligent analysis and decision-making on patients. For example, if a patient's MRI results come in today, they will be analyzed. But in fact, this patient may have left relevant MRI, CT and medical history results ten years ago. Can these historical data be integrated? With comprehensive intelligent analysis and decision-making, the effect will be much better."

Genetic data, image information to standardized big data is an unbreakable chain

So, when the perfection of the data end becomes a consensus, how to obtain data? What kind of data is the raw material for the realization of smart medical care?

On March 28, IBM signed a strategic cooperation agreement with Baiyang Pharmaceutical Group, a domestic Internet medical platform. Its subsidiary Baiyang Intelligent Technology will become a strategic partner of Watson Health in China.

Earlier media analysis said that the amount of data from Baiyang is what IBM values, because Watson's ability is based on Western patient databases and medical theories to provide evidence-based medical diagnosis and treatment. However, due to differences in racial genes and living environments, there are certain differences in the incidence and treatment of cancer diseases between Chinese and Western patients. Therefore, for Watson, it is necessary to successfully obtain sufficient case history data from Chinese medical institutions and conduct cyclical cognitive computing training.

In fact, industry insiders pointed out that IBM's data bottleneck is not that genetic differences affect diagnosis, but that data training is needed for genetic testing to promote precision medicine. As we all know, Watson also cooperates with Illumina, the world's gene therapy giant, to promote automatic interpretation of sequencing results and screening of targeted drugs based on the latter's latest sequencer.

Similarly, Dr. Zhou Zhenyu told Leifeng.com that human health issues start with genes, from genetic information to imaging information to standardized big data analysis - it is a straight line. "Without the previous information, it is incomplete to talk about smart medical care. From the most basic gene sequencing to the management of the entire human health system, such as chronic disease management, the whole concept is strung together to achieve big data intelligent diagnosis and research."

Therefore, in addition to the hospital, according to Huiyi Huiying CEO Chai Xiangfei, in order to unify the scattered data, Huiyi Huiying's cloud platform also needs to cooperate with Philips, Siemens and other equipment to open up. Outside the hospital, take Philips as an example. As one of the three major medical equipment giants that took the lead in transformation and landing in 2017, the company unified the process of linking genes, imaging diagnosis, treatment plans, etc., and integrated information. In the transformation of the company's medical equipment, Illumina and BGI are indispensable partners.

As for BGI, the institution operates the national gene bank with a wide coverage. Dr. Zhou Zhenyu revealed that in the future, BGI will have 100 million samples for basic (theoretical) research and 500 million samples for clinical research; and Philips is planning to deploy two large-scale imaging devices in the national gene bank, which will enable the collection of human genetic information and imaging information at the same location.

What's more exciting is that this work involving reading, storing and writing the entire data stream will produce what kind of effect? ​​Dr. Wang Xinjun, professor and party secretary of the Fifth Affiliated Hospital of Zhengzhou University, said that the perfect combination of neuroimaging and genomics has formed imaging genomics, which can more accurately diagnose diseases or predict the development of some diseases.

For example, glioma is the most stubborn disease in the nervous system and requires surgery, radiotherapy and chemotherapy. Radiotherapy is very effective, but it can cause the side effect of radionecrosis. During radionecrosis, it is still a difficult problem to judge whether the cells are recurrent or necrotic based on images alone. Therefore, imaging genomics based on big data and deep learning can combine the advantages of genomics and imaging for predictive diagnosis.

If we only rely on computers for auxiliary diagnosis, the world has had related technology as early as 1959, called "computer-aided diagnosis (CAD)"; however, although diseases have commonalities, the most important thing in diagnosis is that the onset of each individual is different. Zhou Zhenyu said that with the further development of reinforcement learning and deep learning today, in order to realize the process of computer self-training, more information needs to be continuously taken in.

Future Challenges: Feature Finding

In June 2016, the Radiology Department of Guangdong Provincial People's Hospital and the team of Researcher Tian Jie from the Key Laboratory of Molecular Imaging, Institute of Automation, Chinese Academy of Sciences, cooperated to make important progress in the prediction of colorectal cancer lymph node metastasis using the emerging radiomics method. The relevant research results have been published in the top journal of clinical oncology, JCO.

According to reports, radiomics is a new method that uses information technology such as data mining to extract and quantify massive tumor features from massive data such as imaging, pathology, and genes, and analyze the relationship between imaging and genetic and clinical information (typing, efficacy, and prognosis, etc.). In recent years, radiomics has become one of the most popular research hotspots and cutting-edge directions in the field of imaging.

The data showed that compared with traditional CT imaging evaluation, the imaging genomics prediction model increased the accuracy of preoperative lymph node prediction by 14.8%.

However, this type of AI medical results are currently being produced in affiliated hospitals of Zhejiang University, Jilin University, and other institutions such as the Chinese Academy of Sciences, but most of them are still in the scientific research stage.

As we all know, the immaturity of gene therapy is due to the fact that human doctors have not yet been able to find too many connections between genetic information and clinical manifestations. In the field of general image recognition in the industry, there are databases that teach machines to classify objects: Caltech 101 and ImageNet. The principle of the two databases is to let machines understand images through feature labels. Similarly, the field of AI medical imaging will also face such problems.

Zhang Huimao, director of the Radiology Department of the First Hospital of Jilin University, said that there are many imaging representations. The processing of 2D images of skin diseases is relatively smooth, but the diagnosis of other diseases becomes very complicated. "If the imaging and image processing cannot be standardized, then the data is useless, and there is no big data smart medical care today."

She said that how to extract effective data and what kind of data features to extract after standardized management require the collaboration of clinicians, genomics, imaging genomics, and equipment manufacturers to develop effective features in order to make artificial intelligence medical care more effective.

Of course, AI medical innovation is still facing the problem of lagging policy supervision, and it has also caused some people to worry about the "AI replacement theory". But as Huiyi Huiying co-founder Guo Na said,

What is more important for artificial intelligence is to respect and care for life.

When faced with the limits of life, everyone will succumb. When breath becomes air, all that is left for us is to move forward.