Since Hinton and his students published a paper entitled "Reducing the Dimensionality of Data with Neural Networks" in "Science" in 2006, although the concept of neural networks has returned to the public's field of vision, this technology has not yet been large-scale. usage of. Because of the unsatisfactory results, many scholars, including academics, have doubts. Until the birth of the 2012 ImageNet contest champion AlexNet, the door to deep learning of this "alchemy" was really opened (to pay tribute to Ali Rahimi who won the "Time of Inspection" Test of Time Award at NIPS2017). This event has aroused great enthusiasm in the industry. Many start-up companies with deep learning technology have sprung up in Silicon Valley, Israel and China.
The development of deep learning technology has directly promoted the advancement of the technical direction of natural language processing (NLP) and computer vision (CV). There have been many successful and mature applications in speech recognition, machine translation, image processing and recognition. Medical image analysis, as a branch of computer vision technology in the field of image, has also become a research hotspot. A document published in Medical Image Analysis in mid-2017 published statistics on deep learning techniques in the field of medical image analysis. The results are shown below:
It can be seen that after deep breakthroughs in the field of natural images in 2012, deep learning technology began to enter the medical imaging field on a large scale. Several main techniques such as target detection, instance segmentation and image classification in computer vision are applied in medical image analysis, and cover different modal data such as MRI, CT, X-ray, Ultrasound, etc. Different parts. Not only that, the above is the distribution of the number of academic papers. In terms of the influence of the papers, the application of deep learning techniques in medical image analysis has also been greatly recognized, and some important researches in the past two years are simply listed. The result is evident.
For example, in January 2017, Stanford University's interdisciplinary research team published "Nature" in "Corrigendum: Dermatologist-level classification of skin cancer with deep neural networks". This article uses deep learning technology and uses nearly 130,000 clinical cases. The data was trained and tested for performance on biopsy-confirmed clinical images under the supervision of 21 certified dermatologists. In this experiment, the deep convolutional neural network has reached the level of all test experts in the most common cancer identification and the most deadly skin cancer identification. It proves that the artificial intelligence has reached the level of skin cancer identification. Comparable to the level of dermatologists.
Another example is the "Identifying Medical Diagnoses and Treatable Diseases by Image-Based Deep Learning" paper published by Professor Zhang Kang of the University of California, San Diego in February 2018. He uses about 100,000 accurately labeled retinal optical coherence faults. The imaging images were trained to achieve 96.6% accuracy in the diagnosis of eye diseases, with a sensitivity of 97.8% and a specificity of 97.4%. This work introduces migration learning techniques that can be used to diagnose diseases other than retinal diseases and further training with pre-trained ophthalmic AI diagnostic models using 5000 chest X-ray images to distinguish between pneumonia and health status At the time, the accuracy can reach 92.8%.
At the same time, a large number of artificial intelligence medical imaging startups emerged. According to public statistics, in 2017, the domestic medical imaging AI track has raised more than 2 billion yuan, and the number of single financing has exceeded 100 million. It is one of the most active directions of equity investment and financing in the field of medical artificial intelligence, so why is medicine? Image AI can be recognized by many entrepreneurs and investors, attracting so many capitals and talents to gather in this field?
Here are a few preconditions:
Technology upgrade
The breakthrough of deep learning technology in the field of images makes it possible to apply AI technology for medical diagnosis in medical imaging;
Supply and demand imbalance
With the aging of the population and the improvement of people's health awareness, the number of medical imaging examinations increases by more than 30% every year, while the growth rate of imaging doctors is less than 5% per year. There is a serious imbalance between supply and demand.
Market size
According to the "Market Map of Medical Imaging and Industry Development Analysis", the scale of China's medical imaging market will reach RMB 60-800 billion in 2020, of which the diagnostic link accounts for about 20%, but it is also a market of 100 billion. The US market for medical imaging diagnosis in 2018 is over 10 billion US dollars.
Then, the market and value are very clear. What problems can medical image AI solve at this stage?
At present, the main application mode of medical imaging AI products is to assist doctors in clinical diagnosis. It belongs to tool-type products from product classification, and the core of tool-based products is to efficiently solve user needs.
Then we simply analyze the pros and cons of the tool-based products. The advantage of the traditional tool-type products is that the user needs are clear, the use scene is pure, the landing is easy, and the experience is extremely easy. However, tool-based products also have great disadvantages, because the usage scenario is single, so users use less frequently; because the user needs are clear, once the user's needs are met, the use stops. Therefore, such products have short use time and poor user stickiness, which makes tool products often develop very fast in the early stage and can quickly occupy the market, but when it reaches a certain stage, it will be limited by large-scale commercial realization.
In the medical field, the modifiable model of tool-based products is very clear. At present, most of the products in the medical field belong to tool-type products, such as medical devices. As long as the product quality is excellent, it can solve the actual clinical needs, and can be cut into the clinical path, it can be commercialized, so the medical imaging AI products mainly need to cross. The threshold is to find clear user needs and usage scenarios, and to land.
Let us analyze the user needs and usage scenarios, first look at the statistics of malignant tumors in China.
There are more than 14 million new cancer cases every year in the world. The number of new cases in China is about 4.3 million per year, which is equivalent to more than 10,000 people diagnosed with cancer every day. Among them, lung cancer and breast cancer are the largest cancers in men and women, respectively, and lung cancer is the cancer with the highest incidence and the highest mortality rate in China and even in the world. Because of the significant difference in the five-year survival rates between lung cancer and breast cancer in early and late stages, early diagnosis and early treatment can significantly reduce mortality, which is difficult to measure for patients themselves, their families and society.
But unfortunately, the incidence of lung cancer is concealed, and it is late in the patient's clinical symptoms (such as cough, blood in the sputum, chest pain, fever, shortness of breath, hoarseness, etc.). The data show that about 75% of lung cancer patients in China are in advanced stage at the time of diagnosis, the five-year survival rate is only 15%, and the five-year survival rate of lung cancer patients with distant metastasis is less than 5%, more than half of lung cancer patients Death within one year after diagnosis. Therefore, early diagnosis of lung cancer is extremely important for treatment. NLST (National Lung Screening Trial) results show that lung cancer screening for high-risk populations can effectively reduce mortality (Reduced lung-cancer mortality with low dose computed tomographic screening, New Engl J Med 2013).
Compared with lung cancer, breast cancer has a better therapeutic effect and is less prone to recurrence. According to the US Cancer Center, the five-year survival rate of breast cancer is 89%, and the 5-year survival rate of 0-I is close to 100%. However, in China, due to the lack of a nationwide breast cancer screening program, compared with most patients in the United States, the diagnosis is 0 phase, phase I, the majority of breast cancer patients in China are phase II, and the proportion of stage III and IV is also higher. United States. According to statistics, even in Beijing, 82.1% of women found obvious symptoms when they developed breast cancer. The proportion of patients with stage 0 and stage I was only 32%, and the age of onset of breast cancer in China was significantly lower than that of foreign countries. Two-thirds of the patients are young and middle-aged women under the age of 45, and breast cancer screening for large-scale population is imminent.
Sugar nets (diabetic retinopathy) also have an urgent need for screening. As of 2015, the number of diabetic patients in China is as high as 110 million, ranking first in the world. Sugar net is one of the common chronic complications of diabetes, and it is also the most important eye disease caused by diabetes. The incidence rate is about 31.7%. At the same time, there are often no clinical symptoms in the early stage of the disease. Once the symptoms are present, the condition is more serious and it is easy to miss the best treatment opportunity. Studies have shown that diabetic patients undergo a fundus examination once a year, which can reduce the incidence of blindness by 94.4%.
It can be seen that there are large-scale population screening needs in the above three major diseases, which requires a large amount of human and financial investment, and the application of AI technology to early screening of major diseases can solve human resources well. And the problem of funds, while improving the overall screening efficiency, from the national level can also help the quality of medical services to sink, to achieve early diagnosis and treatment of major diseases, reducing medical and social costs. This is the main landing scene of the current image AI application, and it is also a problem that the current medical image AI can solve.
With the groping and understanding of medical AI in the past two years, the products and models of various companies have been converging. According to the public information, most of the products released by various companies are focused on the screening and early diagnosis of major diseases such as lung cancer, sugar nets and breast cancer. Although medical imaging AI has high technical and resource thresholds, the competition is the same. Very intense, but also attracted many mature companies to join, such as Alibaba, Tencent, Baidu, Ping An Technology, Keda Xunfei and so on.
Companies in this field have their own characteristics, whether they are from the non-medical field cross-border giants, or from the traditional medical field, or from 0 to 1 startups, in the field of medical imaging AI Continuously test the company's technical capabilities, product capabilities and business capabilities, any shortcomings will seriously restrict the company's development. Medical imaging AI is a very promising and valuable field. Focusing on technological innovation, starting from product experience, pragmatically solving the efficiency problems faced by doctors in clinical work, and getting practical applications in clinical work, and then forming Dependence is the most practical path at the moment.
Dried Wakame Stalks,Dried Wakame,Wakame Stem,Alga Wakame
Shandong Haizhibao Ocean Science and Technology Co.,Ltd. , https://www.haizhibaoseafood.com