Recently, Guangzhou Women's and Children's Medical Center developed an artificial intelligence system based on deep learning that can diagnose two major types of diseases: eye diseases and pneumonia, and the research was featured as a cover article in the Feb. 23 issue of the world's top journal Cell.

The AI achievement is able to give doctors diagnostic recommendations based on imaging information and explain the basis for the judgment. Comparison experiments found that the system was 96.6 percent accurate in diagnosing eye diseases; and 92.8 percent accurate in distinguishing between pneumonia and a healthy state, a level that rivals that of expert doctors with more than a decade of experience.

What's in the bag

Precise medication, second-by-second determination

Pneumonia is the leading cause of death from infections in children worldwide. Finding a lung nodule from a chest CT takes an average of 3-5 minutes for a trained doctor, while relying on AI takes just 3-5 seconds.

This is the AI platform developed by a group led by Prof. Zhang Kang at the Guangzhou Women's and Children's Medical Center and the University of California, San Diego.

It's not just about being fast, it's also about being accurate. A key factor in determining the prognosis of pneumonia is the ability to accurately administer medication based on the pathologic type of pneumonia. Traditional methods based on blood culture, sputum culture and biochemical tests are difficult to determine quickly and accurately. In contrast, the AI platform can realize the accurate determination of the pathogenic type of pneumonia in children in seconds based on children's chest X-rays.

This realizes the use of artificial intelligence to accurately guide the rational use of antibiotics, and the platform can be independent of the level of hospitals and regional restrictions, to achieve a wide coverage of community health care, family doctors, and specialty hospitals, to provide accurate medication solutions for pneumonia, a disaster area of antimicrobial abuse, to avoid the abuse of antimicrobials, and to promote the recovery of children with severe pneumonia.

The AI platform is of great clinical significance, and people expect AI with higher efficiency and better accuracy to become a good helper for doctors. In the pre-diagnosis of disease screening, prevention, medical image-assisted diagnosis at the time of consultation, test results analysis, surgical assistance, as well as post-diagnosis medical follow-up, chronic disease monitoring, rehabilitation assistance, health management and other aspects of AI will make a difference. It will even bring changes to basic research assistance, drug development, genetic screening and analysis, and medical training.

"Now our AI platform can allow more patients to be detected, diagnosed, and treated early, anywhere in the world, regardless of personnel or region." said Kang Zhang, a professor at the Shiley Eye Institute at the University of California, San Diego, who joined the Genetic Testing Center at the Guangzhou Women's and Children's Medical Center in 2016.

Is it worth trusting

High accuracy, visible process

Some people say, is it reliable for AI to see a doctor? Is it safe to trust your family's life to a robot?

The research team started with two diseases, macular degeneration and diabetic retinal macular edema, and had the AI system constantly learn from optical coherence tomography images of the eye. After learning image data from more than 200,000 cases, the platform diagnosed macular degeneration and macular edema with 96.6 percent accuracy and 97.8 percent sensitivity. Compared with the diagnostic results of five ophthalmologists, the confirmation platform can reach the level of trained ophthalmologists and decide whether a patient should receive treatment within 30 seconds.

The reporter learned that this AI system has deep learning capability. People are familiar with applications such as AlphaGo and autopilot, which are developed based on deep learning technology.

In this research and development process, the group applied a new algorithm based on the migration learning model, which greatly improves the learning efficiency of artificial intelligence, but also helps to achieve the goal of "a system to solve a variety of diseases".

"Traditional deep learning models generally require millions of high-quality labeled data of the same type in order to obtain a more stable and accurate output, but in reality, it is almost impossible to collect millions of high-quality labeled images for each disease, which makes it difficult to realize the wide coverage of diseases in the field of medical imaging by AI. " Zhang Kang introduction. Therefore, currently available medical AI generally a system can only target one disease.

Relatively speaking, this AI platform based on the migration learning model requires a very small amount of data, and the researchers only need a few thousand sheets to complete a cross-disease migration well.

For example, in this study, the group used only 5,000 chest X-ray images on top of the AI system trained on 200,000 eye image data to build an AI image diagnosis system for pneumonia through migration learning, which realized differential analysis and second-level determination of the pathogenic type of pneumonia in children. It was tested to be 92.8% accurate and 93.2% sensitive in distinguishing pneumonia from healthy states, and 90.7% accurate and 88.6% sensitive in distinguishing bacterial pneumonia from viral pneumonia.

In addition, in the past, research and products relying solely on deep learning technology gave only results in the report, but did not list the reasons for the judgment and the process, this "black box" type of diagnosis, even if the accuracy is very high, doctors do not dare to use. This AI platform overcomes this limitation to a certain extent, allowing people to "know what they know, but also why they know".

The group used the masking test thinking, through repeated learning, practice and improvement, the platform can show which area of the image from which it draws diagnostic results, to a certain extent, gives the reason for the judgment, thus making itself more credible.

Prospects have geometry

Systematic assessment, assisted decision-making

Artificial intelligence diagnosis of disease to be so efficient, robot doctors from our lives how far?

Zhang Kang said their AI system is currently in the United States and Latin American ophthalmology clinics for small-scale clinical trials. In addition, in subsequent studies, they will further increase the number of data-learning templates, as well as increase the types of diseases that can be diagnosed, and further optimize the system, among other things.

Back in 2015, Guangzhou Women's and Children's Medical Center launched the "Mimu Xiong" intelligent family research and development project based on medical big data and the integration of cutting-edge artificial intelligence technology.

"There are four bears in this family member, fever bear, imaging bear, diagnostic guide bear and nutrition bear." Liang Huiying, director of the hospital's clinical data center, introduced the "fever bear" to children's common fever-related diseases as the content of the study, based on authoritative guidelines, expert **** knowledge, more than 2 million copies of the massive medical records and other knowledge-based text, the integration of multi-source heterogeneous data integration technology, natural language processing technology and machine learning algorithms, after a year of training, has been able to successfully target 24 different types of children's diseases. After a year of training, it has been able to successfully carry out accurate auxiliary diagnosis for 24 common fever-related diseases in children, and become a caring assistant for outpatient doctors by seamlessly embedding into the electronic medical record system.

Based on the "chest X-ray film + microbial culture test big data" and deep learning algorithms, Imaging Bear can intelligently identify the microbial infection status of pneumonia (bacterial, viral, mixed infections) and provide decision support for the accurate application of antimicrobials, which has been practically applied to the auxiliary diagnosis of doctors. The data and technology developed in its practice have become an important foundation and component of the scientific research results of artificial intelligence systems.

The other two "bears" are also growing strong and are expected to meet the public in the near future.

The medical artificial intelligence research results published in the journal Cell are being used as a new starting point for the Guangzhou Women's and Children's Medical Center. The ultimate goal of the AI platform is to integrate multi-media data such as text-based medical record data, fully structured laboratory examination data, image data, and optoelectronic signals, to simulate a clinician's systematic assessment of a patient's condition, and to provide comprehensive assisted decision-making for medical staff," said Xia Huimin, director and president of the center. Instead of just providing a single aspect of assisted decision making for the imaging physician or a particular medical technologist."

"Therefore, the platform is still in the process of being enhanced." As an example, Huimin Xia said that, for example, in the field of intelligent discernment of the pathogenic type of pneumonia in children, the team is adding laboratory tests and learning of clinical symptoms to the systematic reading of X-rays, so as to more accurately determine the type of pathogenic bacteria in children's pneumonia.

"We hope that in the near future, this technology can be applied to primary care, community medicine, family doctors, specialty hospitals, and so on, forming a large-scale automated triage system." Xia Huimin said.

Links

This set of artificial intelligence is so "smart"

This set of artificial intelligence uses a transfer learning algorithm, which is to migrate the parameters of the trained model to the new model to help train the new model, that is, to use existing knowledge to learn the new knowledge to find the similarity between the existing knowledge and the new knowledge, and to use the existing knowledge and the new knowledge to learn the new knowledge, and to use the existing knowledge and the new knowledge to learn the new knowledge. In other words, the existing knowledge is used to learn the new knowledge, to find the similarity between the existing knowledge and the new knowledge, which is called "learning from one, learning from three" in idioms.

For example, if you have learned to play Go, you can learn chess by analogy; if you know how to play basketball, you can learn volleyball by analogy; if you know Chinese, you can learn English and Japanese by analogy. How to reasonably find the **** between different models, and then use this bridge to help learn new knowledge, is the core of "transfer learning". Transfer learning is considered an efficient technique, especially when faced with relatively limited training data.

In the case of medical image learning, for example, the system recognizes the characteristics of the images in the pre-system, and the researchers then go on to import a network system containing similar parameters and structures from the first layer of images to build the final layer.