Why does AI imaging medical care get together for lung nodule screening？

　　"88% of departments use AI products for lung nodule screening, 6% for coronary artery analysis, and the rest are bone age, breast and prostate intelligent diagnosis".

　　A few days ago, China Medical Imaging AI Industry-University-Research released the "China Medical Imaging AI White Paper", which made the legendary phenomenon of AI medical clustering lung nodule screening more intuitive.

　　AI medical treatment is the most widely used and mature field of AI technology, but with the growing controversy, optimists believe that AI can help solve the problem of lack of medical resources and improve the work efficiency of doctors, thus creating value; The singer questioned that its homogenization was serious and the landing process was extremely slow. Is AI medical care going into a dead end?

　　status

　　Lung nodule screening becomes a necessary topic for AI medical enterprises.

　　"In my contacts and peer exchanges, I found that companies that usually do AI medical care have lung nodule screening products." Professor Zhang Shuixing, director of the Radiology Department of the First Affiliated Hospital of Jinan University, also had similar feelings. "Like our department, in addition to AI-assisted lung nodule screening, there are also breast cancer molybdenum target screening and bone age screening. At the same time, thyroid nodule screening research is being carried out with AI companies. " In the past two years, many AI medical companies have come to promote their own image products, and another one has knocked at the door during the interview.

　　Liu Shiyuan, chairman of China Medical Imaging AI Industry-University-Research Innovation Alliance, once said: Comparing the business layout of AI medical companies in China and the United States, it can be seen that the medical AI enterprises in the United States are widely distributed, while those in China are mostly focused on imaging.

　　According to estimates, there are currently more than 130 companies engaged in medical imaging among more than 1,500 AI companies in China. There are many players on the track of lung nodule screening and diagnosis, including industry giants such as Ali and Tencent, as well as head startups such as Zero Krypton Technology, Shenrui Medical, Tuma Shenwei, Huiyi Huiying, etc.

　　cause

　　For science and technology enterprises, pulmonary nodule screening is easy to get started.

　　Why is everyone getting together for pulmonary nodule screening? In Zhang Shuixing’s view, for technology companies, lung nodule screening can be described as an entry level.

　　Pulmonary nodules mainly refer to single or multiple circular or quasi-circular nodules with a diameter of less than 3cm in the lung parenchyma, and the possibility of early lung cancer cannot be ruled out. "China Expert Consensus on Diagnosis and Treatment of Pulmonary Nodules (2018 Edition)", the incidence of pulmonary nodules was 35.5%, of which 0.54% was diagnosed as lung cancer.

　　The malignant ratio of pulmonary nodules is not high, but it is closely related to lung cancer, which also makes it easy for the public to "talk about the discoloration of nodules."

　　This scene must have been familiar to many people: Aunt Zhang’s physical examination showed that there was a small nodule of 8mm on her lung. The doctor told her that there were benign and malignant pulmonary nodules. If it was benign, it could be controlled by taking medicine or even without treatment. However, if it was malignant, it needed to be removed immediately, which might lead to life-threatening lung cancer. What is puzzling is that I am afraid of mistaking lung cancer as a benign disease and delaying the best treatment opportunity; I am also afraid that I will get a knife for nothing, which will affect my lung function and reduce my quality of life. A nodule made Aunt Zhang worried all day and couldn’t sleep well.

　　A practitioner in the field of image AI revealed to reporters that with the widespread use of high-resolution CT and the public’s attention to physical examination, pulmonary nodules have become more and more common clinical problems, and there are more public data. If it is a malignant lung nodule, it indicates the risk of cancer. Therefore, it is easy for AI enterprises to produce products in the field of pulmonary nodules, and the short period of effect is more promising to convert them into economic benefits. Like liver cancer, few people touch it. First, the image manifestations of liver cancer are diverse, such as massive type, nodular type, diffuse type and small cancer type; Second, liver cancer is accompanied by complex liver diseases, such as cirrhosis, fatty liver, alcoholic hepatitis, etc. The data of all kinds of liver cancer in the competition training set are unbalanced, which increases the difficulty of judgment.

　　AI medical enterprises are burning money at this stage. There are various algorithms, and the accuracy rate needs to be improved. The payment scenario has not yet been realized. Zhang Shuixing bluntly said that in this case, it is rare to sink into research and development. Many enterprises are just anxious to sell products to doctors.

　　question

　　AI medical products have high sensitivity and low specificity.

　　In a research article entitled "Deep learning as a tool to improve the accuracy and efficiency of histopathological diagnosis" published by the top international journal Nature in recent years, researchers indicated that the sensitivity of deep learning algorithm reached 100%, but the false positive rate was as high as 40%.

　　"Indeed, AI can find smaller and smaller lung nodules. When doctors have a heavy workload and work fatigue, AI can easily help find lung nodules below 3mm, but there are also many false positives, which requires doctors to check again." Zhang Shuixing said: "Many of our AI medical products are highly sensitive and not specific."

　　The so-called sensitivity refers to the chance of not missing diagnosis when diagnosing a disease, and the so-called specificity refers to the chance of not misdiagnosis when diagnosing a disease.

　　Liang Changhong, director of the Department of Imaging Medicine and Radiology of Guangdong Provincial People’s Hospital, recently shared a case at the 4th Global Summit on Artificial Intelligence and Robotics in 2019: Last year, the US Food and Drug Administration approved a software based on artificial intelligence. The application of this software Viz.AI Contact will analyze the results of computed tomography (CT) and notify neurovascular experts by SMS to judge the cerebral infarction.

　　However, the editor-in-chief of the authoritative magazine Radiology strongly criticized this software system. Because this software system has not published much data so far, the only available result comes from the report of the American Heart Association’s 2018 International Stroke Conference. The training and verification data of the system come from ALADIN non-public data set established by doctors at Emory University. In addition, the names of two neuroradiologists are provided, one of whom works in this software system company, and the other can’t be identified through formal channels. The most important thing is that its specificity is only 52%, which is similar to coin toss. This reminds our AI company to pay attention to the integrity and traceability of the data.

　　prospect

　　There is still a distance from beautiful laboratory data to clinical application.

　　Luo Ligang, co-founder and CTO of Zero Krypton Technology, pointed out that some medical AI products now pay too much attention to algorithms and ignore doctors’ work habits and system stability. "They perform well in the established training data scenarios, but the clinical environment is more complicated, which may not meet the clinical needs and can only be shelved."

　　Luo Ligang admits that if an enterprise wants to develop an AI auxiliary diagnosis system, it is necessary to have the ability to process multimodal data such as images and genes. In addition, as a system science, the logic, modeling and layering of medical data are very complex and rich. Only when algorithm experts and medical experts are deeply integrated and understand each other’s "languages" can we cross the threshold of medical AI.

　　"No matter how beautiful the algorithm is, it doesn’t mean that it will work in the actual clinic," Zhang Shuixing said with emotion. "I have been in contact with AI medical images since 2014 and participated in research and development. It is far more difficult for AI enterprises to do medical treatment than everyone thinks. AI medical treatment is not only a scientific and technological field, but also has high requirements for medical majors." In the application of AI, doctors need to constantly communicate with AI enterprise technicians to improve products, so as to better serve the people.

　　Nanfang Daily reporter Roger Lee correspondent Zhang Cancheng