Google’s AI diagnosis has “rolled over” in practice

According to statistics from the National Cancer Center of the Chinese Academy of Medical Sciences, there were 466,000 new cases of liver cancer in China in 2015. According to relevant data in 2017, the missed diagnosis rate of early liver cancer was as high as 40%. As of the data of 2019, the annual growth rate of my country’s medical imaging data is about 30%, while the annual growth rate of radiologists is only 4.2%, which means that in the future, the pressure on imaging doctors to process data will increase. Coincidentally, data from the Centers for Disease Control and Prevention (CDC) in 2018 showed that 14% of American adults suffer from diabetes, a chronic health problem that plagues about 30 million people. Among them, more than 24,000 people are blinded each year due to sugar net (sugar net is a serious complication of diabetes and one of the major eye diseases that cause blindness). The voice of AI diagnosis is getting higher and higher.

However, while more and more institutions have announced their continuous innovation in AI diagnosis accuracy, Google, the company with AlphaGo that defeats humans, has “turned over” in the practice of AI diagnosis.

The US FDA does not approve, the project is implemented in Thai hospitals
Google Health is a health department established by Google in 2018. At the beginning of 2020, they shared an AI model test for diagnosing breast cancer. After training more than 90,000 mammograms, the model test achieved better results than human radiologists. The project that made them hit the wall also came from Google Health, which is another core research of theirs-the AI ​​sugar web screening system.

In the context of huge demand, Google Health has established a data set consisting of 128,000 pictures, each of which records the diagnosis results of 3-7 ophthalmologists. These data are calculated by algorithms and then clinically verified. The test results are assessed to have reached the level of human experts, with an accuracy rate of over 90%. However, the US Food and Drug Administration (FDA) has not passed this project for the time being, and the project cannot be put into practical use in the United States.

In Thailand, on the other side of the globe, there is a huge gap in the screening of diabetes patients. In Thailand’s screening system, nurses first take fundus photos of patients and then send the photos to experts for diagnosis. This process often takes 5 to 6 weeks, or even 10 weeks. In Thailand, there are 4.5 million diabetic people and only 200 retinal experts. At this rate, it will take more than two hundred years for them to screen everyone, and there has been no increase during this period. In such an urgent situation, Thailand has chosen to cooperate with Google Health, which holds CE certification. Google has installed a screening system for 11 domestic households in Thailand. According to the data released by Google, the screening time can be shortened from a few weeks to a few. Minutes, but this is not the case. Google’s system adds a lot of trouble to the screening process.

First, the photos “recognized” by this AI system need to be taken with a non-mydriatic fundus camera, and it needs to be taken in a special dark room. Of the 11 clinics, only 2 have the conditions for such shooting. As a result, one-fifth of the fundus photos taken for the patients were rejected by the AI ​​system, and some of these rejected photos have obvious signs of sugar network. Second, the speed limit in Thailand makes the upload of photos very slow. Third, even if the AI ​​system check is smooth, Google will provide follow-up treatment plans. However, what is very unfriendly to patients is that if a patient agrees to AI screening, but his image is rejected by the system, these patients will not be allowed Do not seek another time for treatment. “Patients do not care about accuracy, but more about the experience,” the nurse told Google researchers, “40% to 50% of people will not join the study because they think they must go to the hospital.” For some reason, these patients do not think this AI system is useful.

Huge demand, AI screening appears in multiple types of testing
In life, we can see that the demand for AI screening is very large. In addition to Google Health’s screening research on Sugar.com, they have embarked on other projects, including screening for female breast cancer. Statistics show that for every 100,000 women, more than 42 have breast cancer. Like many other cancers, a major difficulty in the treatment of breast cancer is the issue of early screening. If the disease can be detected at an early stage, treatment can greatly reduce the risk. The screening system shared by Google has trained more than 90,000 mammograms and achieved better results than human radiologists.

On the other hand, the domestic Lenovo Research Institute is also trying to use artificial intelligence to help diagnose liver cancer and improve the diagnosis and treatment of liver cancer. At present, the accuracy of liver segmentation of this system has reached more than 90%, but it has not yet entered the clinical stage. In addition, under the new crown epidemic, Alibaba Dharma Institute and Alibaba Cloud jointly developed a set of AI diagnostic technology for new crown pneumonia. This set of technology can make a judgment on the CT images of suspected cases of new coronary pneumonia within 20 seconds, thereby providing clinicians with auxiliary diagnosis basis and greatly improving the diagnosis efficiency. At present, the accuracy of this AI diagnosis technology can reach more than 96%.

AI screening still leaves manual diagnosis
As far as the current situation is concerned, no matter how high the diagnostic accuracy of AI screening is, the judgment of the diagnosis result cannot leave manual verification. In other words, AI screening can only be an auxiliary tool. Insiders said that Google’s AI screening system has problems with many other AI research projects—the integration of scientific research and clinical practice is not close. From the perspective of the process, the AI ​​screening system will eventually reach the clinical stage, but the clinic has its particularities. The screening system requires that the relevant developers not only understand the technology, but also understand the testing process, the needs of the doctor, the acceptance of the patients and the things they care about. Only in this way can we build a system that really solves the problem.

In addition, some related people said that when screening for diseases, doctors and patients are not only concerned about “whether or not they are sick”, but also attach great importance to the stage of the disease, whether it is conscience or malignant, and what treatment options are needed. The problem. AI screening has a long way to go.