Shulou(Shulou.com)06/02 Report--

In the AI medical industry, medical imaging has always been a relatively hot field. The model is established through the sum analysis of medical image data by deep learning algorithm, and the intelligent analysis of medical image is realized. Now every year in the academic forum of computer vision, there are image data analysis competitions about various diseases, and in many fields, such as oncology, cardiovascular disease and so on, the accuracy and speed of AI has exceeded that of human beings.

However, in a series of AI medical imaging projects, AI fundus imaging has always been a very special existence. Compared with other projects, the landing speed of AI fundus screening software is faster than many people imagine. For example, the AI fundus screening technology jointly launched by DeepMind and Moorfields Eye Hospital in London can identify several diseases within 30 seconds, which is known as DeepMind's first commercial AI project. Baidu has also donated hundreds of AI fundus screening machines to grass-roots hospitals in a high profile, allowing the technology to fall directly into the application scene in the form of packaging. AI Medical start-up company voxel Technology's "VoxelCloud-Retina fundus screening solution" has been launched in more than 130 MMC (National standardized Metabolic Disease Management Center) across the country. It is not too much to be called the "king of landing" for this kind of popularization efficiency and speed.

The question we are going to discuss today is how AI fundus screening has come from the laboratory to reality, and what will be the trend in the future?

Why do we need more efficient fundus screening from fundus to body?

Through fundus imaging examination, a large number of fundus diseases can be judged. It can not only reflect a series of fundus diseases such as diabetic retinopathy, glaucoma and senile macula, but also reflect the pathological changes and accumulation of chronic diseases such as hypertension and diabetes.

However, the popularity of fundus imaging examination does not match the value of this technology.

The reason lies in two aspects. One is that the focus of fundus disease is small and the differentiation between different lesions is low. At present, there are relatively few doctors who have the ability to read films in the whole country. At the same time, in the architecture of most hospitals, fundus screening is classified as ophthalmology, and it is difficult for ophthalmologists to make follow-up diagnosis and treatment of diseases such as sugar net disease, hypertension and endocrinology, but after leaving ophthalmology, it is very difficult for doctors in other departments to have the ability to read fundus images.

In addition, according to the 2017 China Health and Family Planning Statistical Yearbook, China has about 400 million patients with chronic ophthalmopathy, while there are only 36000 ophthalmologists, with more than 110 million outpatients and more than 4.5 million hospitalized patients. Patients with diabetic retinopathy are often unable to know their condition in time, so early screening and early treatment is the key to the management of chronic diseases, but now medical resources can not meet the demand, resulting in many patients can not find treatment until the late stage, but it is often too late. It brings a huge unnecessary burden to the national health insurance, seriously affecting the working ability of the population.

In the case of increasing shortage of medical resources, if we can make efficient use of the ability of fundus screening, it can not only help patients discover their condition in time, but also assist chronic disease management more efficiently, so that more people can enjoy efficient medical experience.

Voxel technology finds opportunities in such contradictions.

The nurturance Plan of Doctor AI: how is fundus screening intelligent?

Voxel Technology, like many AI medical start-ups, has found the value of fundus screening technology, and there are some open fundus image data sets on the market for companies to analyze and mine, Voxel Technology's ophthalmic product director said in an interview.

Specifically, the realization process of AI fundus screening is to label the data through professional medical knowledge, mining and training the disease classification network to form a classification model and disease recognition pattern for different lesions. But the question is, how do you build technical differentiation when everyone is faced with the same public dataset?

When it comes to the reasons for this, the director of voxel technology ophthalmology products told us a number of details.

First of all, voxel technology independently expands the scope of obtaining fundus image data. In addition to the application of open data sets, voxel technology continues to expand the scope of data acquisition in the process of constantly trying to land. Especially as a multinational start-up enterprise across Shanghai and Los Angeles, voxel technology can access rich fundus data of different people, and enhance the accuracy and wide applicability of the model.

While other enterprises focus on the research and development of single disease (sugar net disease), voxel technology has begun to do the research and development of the whole disease. General ophthalmologists get the fundus photos of patients, after reading the film, they can basically judge which diseases the patient is suffering from, so if AI technology can really assist doctors in clinical practice, they must have the ability to read films of all diseases. Voxel technology has been continuously increasing the data labeling of fundus images, which are classified into hundreds of disease combinations according to the location and type of lesions. This realizes the training of the multi-task model from the balanced data of the whole disease distribution. At present, the AI fundus screening software of voxel technology can not only identify diabetic retinopathy, diabetic macular edema, glaucoma, cataract, age-related macular degeneration and other lesions, but also identify more than ten kinds of lesions, including microhemangioma, intraretinal hemorrhage, hard exudation, cotton spots, external retinal blood vessels, neovascularization, laser spots, vitreous membrane warts, vitreous opacity and so on. The whole disease system of voxel technology has greatly expanded the scope of application of AI fundus screening. After all, many patients may not know what disease they are suffering from when they go to the hospital. If AI fundus screening technology can cover as many diseases as possible, it will undoubtedly improve the efficiency of medical treatment.

At the same time, in the process of landing, voxel technology found that because the shooting of fundus photos is man-made, there will be a lot of uncertainties, such as the intensity and weakness of light, the image quality of different fundus photography equipment, and so on, which may lead to the misjudgment or even complete failure of AI. Generally speaking, the robustness of the algorithm model is poor, and the requirement for the application environment is high. After the discovery of this problem, voxel technology while training the quality control model to generalize the model, while strengthening the collection of shooting data from different equipment, and then constantly improve the coping ability of the model when landing.

Among them, the director of ophthalmology products of voxel technology also mentioned a very interesting point: when many start-ups do AI fundus screening projects, they often focus on improving the accuracy of algorithmic models in a single data set. Voxel technology is more concerned about the performance of algorithms in real situations, constantly finding and solving problems.

The result is that the AI algorithm that focuses on the performance of a single dataset will have some "high scores and low performance" and will naturally be eliminated by the market because of its poor performance when it hits the ground. Voxel technology has a sensitivity of 97% in third-party tests, and has received tens of millions of dollars of investment led by Sequoia Capital and 100 million yuan of investment from Tencent within three months, and another $50 million investment from Hongtai Capital in September 2018.

Breaking the Quadrant: the Health Fairness brought by AI fundus screening

From the intensive attention of capital, we can also see the landing value reflected by AI fundus screening.

From the story of voxel technology to create "Doctor AI", we can find that an important feature of AI fundus screening is that it opens the door to intelligent diagnosis for many diseases. For example, chronic diseases such as high blood pressure and diabetes, or eye diseases such as glaucoma, which are difficult to detect at the initial stage, have previously been very difficult to diagnose, treat and follow up through types of medical images. unable to manage through medical images, which means that efficiency cannot be improved through computer vision.

However, with the continuous mining and utilization of the data of the impact of AI fundus screening on fundus screening, associated with more diseases, it means breaking the original quadrant of fundus screening only used in ophthalmology, expanding the scope of the ability of this disease screening method, and enabling the diagnosis and treatment of chronic diseases with a large base, such as hypertension and diabetes.

In addition to its own powerful ability, AI fundus screening is also faster on landing. In the form of an all-in-one fundus screening machine, operators can get started quickly, not only by other departments other than ophthalmology, but also by AI fundus screening technology to assist diagnosis and treatment in some non-top-three grass-roots hospitals and community hospitals, even if they do not have highly professional medical personnel. Thus, it will greatly improve the therapeutic effect of patients with sugar net disease and reduce the damage of sugar net to vision.

In this way, AI fundus screening not only breaks the efficiency limit brought by the limited medical resources, but also breaks the difficulty of seeking medical treatment caused by the distribution of medical resources, and reduces the cost of patients. Many patients in third-and fourth-tier cities and towns do not need to come to the big hospitals in Beijing, Shanghai and Guangzhou, but can directly enjoy convenient fundus disease screening on an all-in-one machine. Because the whole process is fast and convenient, patients can also increase the frequency of examination and manage their own chronic diseases. The resulting health fairness is also achieved by many other AI medical imaging technologies. Ning Guang, academician of the Chinese Academy of Engineering and president of the Medical College of Shanghai Jiaotong University, is also very positive about the AI fundus screening technology of voxel technology. Voxel technology not only brings the expansion of primary medical diagnosis and analysis capacity, but also greatly improves people's quality of life and the ability to work for the elderly, in which the potential social value is also very valuable.

That is why voxel technology has been able to raise tens of millions of dollars again in 2018, known as the economic cold wave. The "healthy China" strategy proposes to build a 16 trillion-yuan health management industry by 2030. In the process of achieving this goal, AI fundus screening is a key technology to reduce the cost of chronic disease management.

The research on fundus screening technology will be projected into the construction of the whole health cause. Apart from the economic effect, the public welfare effect can not be ignored. The so-called ten thousand years at a glance, today's AI for fundus screening data accumulation mining and technology landing promotion, is changing the way people manage their own health in the future.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.