On June 2nd, RSNA (North American Radiological Society) held the first issue of "Focus on AI Course: Radiology in AI Era" for radiologists from May 3rd to June 3rd. Through the two-day course, this paper attempts to introduce the technical origin, existing applications and how to understand the academic progress of AI medical imaging, hoping to help doctors adapt to the new era of close cooperation with emerging technologies.
After all, medical care has a large amount of data and technical requirements, and it is the first field to accept the impact of large-scale AI technology, and it is also one of the industries with the fastest application of many technologies.
This "AI Lecture Hall" includes a brief introduction of AI technology in the field of medical imaging, a discussion on its impact on better safeguarding human health, and how to access the AI system in their own medical practice. Each part invites outstanding people in the AI industry to discuss or give speeches. We extracted some key points:
In this course, the most clear point is that AI is already the most important technology in the field of radiation medicine. Medical imaging means such as CT, MRI and PET are important data for doctors to make diagnosis. AI's powerful data processing ability can help doctors on many levels.
Andrew Ng, a world-renowned expert in artificial intelligence and a professor at Stanford University, introduced the development of artificial intelligence and deep learning algorithms and the new progress of artificial intelligence imaging technology. In cooperation with Stanford Hospital, his laboratory has completed ChestXnet, Xray4all and other work. Use deep learning to understand images. These deep learning techniques can understand eleven different pathological manifestations in chest X-ray, detect abnormalities in knee MRI, and detect pathological manifestations of pointing aneurysms in head ct films.
"Deep learning has been able to complete all the basic tasks that human beings need to complete in one second. Of course, AI has a long way to go to completely replace doctors in diagnosis and judgment, and there are many breakthroughs to be made. " Andrew Ng said.
One of the organizers of this course, Professor Curtis Langlotz, deputy director of the Radiology Department of Stanford University School of Medicine, mentioned that he was not so pessimistic about the crisis that AI completely replaced the work of clinical radiologists. He emphasized that imaging doctors need to constantly change and learn more cutting-edge AI knowledge and skills, but AI is just another valuable new technology and development after new technologies such as CT, magnetic resonance and ultrasound encountered in clinical medicine. Clinicians need to apply new artificial intelligence technology to clinical work. "Some doctors feel that tasks are trivial, such as measuring the size of lesions and tracking the changes in the size of lesions in different disease cycles. These tasks are tasks that AI is good at but people don't like. So from one perspective, AI can make clinicians work better. " He said, "With the assistance of AI, clinicians can do some more interesting and cognitively challenging tasks."
Undeniably, doctors still face some new challenges. In the face of AI changing the status quo of the medical field, how can a doctor who is in close contact with patients and provides daily medical services adapt to this era?
First of all, doctors need to know more about the new technology and how it can be used in clinical diagnosis, surgical prognosis, early screening and other fields. During the course, many researchers of medical imaging artificial intelligence shared their new research in these fields.
"AI will not replace doctors, but doctors who use AI will replace doctors who don't use AI." Professor Curtis Langlotz frequently used golden sentences when discussing the application of AI in medical clinic.
Andrew Ng also said: "In the world of science and technology, great changes will take place in our work every five years. Nowadays, technology also makes all other industries change faster than before. In the past, many things that radiologists did were automated, but if these doctors are willing to think about what is really important, broaden their horizons and focus on jobs that are different from those that can be automated, they don't have to worry about anything. "
Secondly, the new technology itself can further improve the professional level of doctors.
Dr Hugh Harvey, a radiologist at Kheiron Medical in the UK, pointed out that radiologists need to know more about data science and technology. Radiologists need to know the basic knowledge such as data science and machine learning, especially the arrangement of data. He mentioned that AI technologies such as deep learning need a lot of data, but people often only pay attention to quantity and ignore quality when discussing. The data obtained directly from the clinical system is far from being really used in clinical AI research and application.
General data collation requires at least four levels of operation.
The first layer is the data obtained directly from the clinical system (PACS, Electronic Medical Record System). These data often contain sensitive information, which is large but complex and cannot be really used for research.
The second layer is the data reviewed by the ethics Committee and deleted from the patient's sensitive information. Doctors and researchers have limited access to it, but this kind of data is generally unstructured and directly used for research.
The third layer is to carry out further structural cleaning and visual inspection on these data, so as to ensure the quality of image data.
The fourth layer is to finally match these data with the corresponding clinical information, and manually or automatically label the data for AI research and analysis. But at the end of this layer, it is necessary to confirm whether the statistical value of the data is enough and whether there is a real standard to mark it. For example, the judgment of a patient's disease needs to be compared with the results of many doctors' reading pictures, and the disease can be diagnosed through the results of subsequent onset or follow-up.
For doctors, being open to technology, contacting and mastering emerging technologies through courses, activities and projects are likely to make future medical services "get twice the result with half the effort".
Professor Greg Zaharchuk, Ph.D. in Neuroimaging at Stanford University and director of the Frontier Neurofunctional Imaging Laboratory, who attended the meeting, said that such a course can explain the theory, application, development and limitations of AI to clinicians. He is glad to see that more and more imaging doctors are enthusiastic about artificial intelligence and hope to gain more knowledge in this field.
On the other hand, he also stressed that there is still a big gap between the research of clinical AI and the deployment of real clinical AI products. How to ensure the algorithm in different situations, equipment, scanning parameters, etc. These are all problems now and need to be solved step by step in the future.
"I am very efficient when I see so many imaging doctors and practitioners participating in this activity. This is the first artificial intelligence focus course organized by RSNA, and we hope to keep the communication between scientific research, clinical practice and industries. In addition, AI imaging companies, such as Shentou Medical, maintain academic reports and papers while commercializing. It is a good thing to rigorously analyze product performance and clinical value. " Professor Matthew Lungren, one of the organizers of this activity, said.
Radiologists are facing more opportunities and challenges in the AI era. For the wider public, technology can bring more protection and higher medical level.
In this activity, Pranav Rajpurkar, a doctoral student from Wu Enda Laboratory, showed the Xray4All platform on the spot: upload X-ray image photos intercepted by users, and after one or two seconds of transmission, you can get the results online, detect anomalies, and highlight the abnormal parts.
"The application scenario of this technology is particularly suitable for solving the shortage of clinician resources in developing countries and global health scenarios." Plana's introduction.
Arterys, another American AI imaging company that raised more than $45 million, also hosted a luncheon and introduced their future vision: further promoting their image analysis and AI products, and gradually expanding the platform. Through real-world data, we can provide medical decision-making for human beings around the world, automate daily medical tasks, further promote medical equality and democratization, and provide preventive analysis. Arterys emphasizes that its image analysis and AI products are all based on cloud computing, and emphasizes that cloud computing is actually faster, safer and more reliable than the calculation in the hospital internal computing system.
As one of the countries with the highest annual medical investment in the total government expenditure, the United States is at the forefront of the world in promoting AI medical technology. As a populous country, China is short of average medical resources and has a great demand for AI medical care.
In this course, China Imagination Technology, American Nuance, and deep penetration medical care, which rapidly popularized AI image processing in China and the United States, were invited to give lectures. With the theme of "Realizing AI: the Last Mile", the last key steps of clinical deployment of AI system industrialization were discussed.
Imagine Medical introduced that many of its products were exposed to millions of medical records in China and tested in four hospitals/imaging centers in the United States. Nuance has a large market share in speech recognition tools and picture reading and marking tools for clinical images in the United States, and it is also promoting its "Nuance AI market" medical image AI application store.
Shentou Medical is the only AI product among the three with FDA approval for commercialization. Dr. Gong Enhao, CEO of Shentou Medical Co., Ltd., introduced how to deploy its FDA-approved SubtlePET products in clinic, and conducted clinical trials on products such as SubtleMR in application.
Shentou Medical SubtlePET product is the first approved medical image enhancement application and the first AI application in the field of nuclear medicine. Its product value focuses on accelerating image acquisition by about 4 times with AI, and also provides a solution to reduce radiation and contrast agent consumption. This software scheme can make patients get more convenient, higher quality, safer and smarter clinical imaging examination. After the approval of FDA, commercial deployment and clinical cooperation have been carried out in 20 top hospitals and imaging centers in the United States and around the world.
In the United States, the threshold for hospitals to really apply AI and be willing to pay for it is very high. It is necessary to deeply integrate the hospital information system, confirm the system effect with clinicians, and demonstrate the return that the purchase of AI system can bring to the hospital.
"In the United States, the real deployment of hospitals requires communication with clinicians, heads of information systems, and hospital management and operation personnel. Taking Shentou Medical as an example, the company's clinical and sales leaders need to test the real data with the hospital quickly and effectively, and let the hospital conduct clinical tests with its own data in real time without affecting the existing operation of the hospital as much as possible. Through the actual test and the acceleration of real and considerable image inspection, the hospital can objectively see that AI brings new clinical value and economic value to the hospital, thus progressing to procurement and deployment. " Gong Enhao, CEO of Shentou Medical, said.
The CEO of TeraRecon, a medical image post-processing company, is also the CEO of Envoy, a medical image AI platform. Jeff Soreson and Professor Eliot Siegal, a famous imaging doctor and image AI promoter, also discussed how to optimize the workflow, deployment process and continuous verification of image AI in the form of mutual interviews.
"The deep clinical verification of AI algorithm is a key step to promote medical AI, and we are also developing towards this goal." Professor Eliot Seager stressed.
Although medical imaging is already one of the most suitable and fastest deployed fields in the AI field, we still face various challenges.
First of all, AI technology represented by deep learning is still a "black box". This means that technology can make medical image detection achieve high accuracy, but it is still difficult for AI to understand the real relationship between data and how to classify data.
"At Stanford, we hope to create a better attention map for medical image perception to avoid the black box effect." Safwan Halabi, a professor at Stanford Medical School, said. Recently, many studies and reports have discussed that the advanced serial attack algorithm based on data can make the AI that recognizes road signs fail to work properly. In medical AI, it is very important to ensure that AI is not misled, but the research in this field is not enough at present.
Dr Matthew Lungren, head of AIMI artificial intelligence medical imaging research project in Stanford University and one of the heads of undergraduate courses, also discussed the bias of clinical AI and its enlightenment to medical imaging AI. AI is likely to introduce data bias in actual clinical use. For example, a classifier used for medical image recognition is likely to recognize other markers in the image, rather than the lesion itself in the image. However, the current tools can not well understand the deviation between data and algorithm. The actual clinical application of AI must make people understand the reliability of the results in use. Considering human-computer interaction and confidence analysis given by AI algorithm in system design can greatly help people reduce possible deviation problems.
Professor Jayashree kalpathy, one of the heads of the Machine Learning Laboratory of Massachusetts General Hospital, discussed how to build a more robust model and how to share the trained deep learning AI model in multi-hospital cooperation projects through algorithms such as migration learning and federated learning, without sharing sensitive data for deep cooperation.
In the era of artificial intelligence, technology is constantly infiltrating and transforming all walks of life. Medicine is a field closely related to human life. In such a huge and important field, at the forefront of artificial intelligence application, we see more and more efforts to help technology and medical practice better combine.
For example, the first AI course in the field of medical imaging provided by RSNA attracted more than 200 doctors from top hospitals in the United States, and technicians in the industry were willing to provide more information to help doctors better understand AI. In addition, startups like Shentou try to make doctors "seamlessly" connect technology with past workflows through product design, without extra energy to adapt to products. Doctors know more about technology and entrepreneurs develop better products for doctors and patients.
In the future, there will be more technical support for human health, but the most important thing is that people in the industry will work together to bring a more efficient and effective medical care system.