Release date: 2017-03-30
Recently, the “Medical Artificial Intelligence Frontier Summit†organized by Huizhi Huiying, a startup of the intelligent medical imaging platform, was held in Beijing. The conference gathered Philips and Siemens, as well as smart medical academics and industry experts from Stanford University and Tsinghua University to discuss the development of artificial intelligence imaging technology and the future medical scene imagination of artificial intelligence technology.
At this meeting, Lei Feng.com and Xing Lei, director of the Department of Medical Physics at Stanford University, talked about the status quo and future of AI medical research.
Xing Lei is a tenured professor at Stanford University and director of the Department of Medical Physics at Stanford University. He has worked in medical imaging, oncology, and bioinformatics for more than 20 years. At the same time, Professor Xing Lei is also an adjunct professor of electronic engineering, molecular imaging, bioinformatics and Bio-X at Stanford University. He has published nearly 300 papers in medical imaging and medical physics, and has received several NIH, Ministry of Defense, NSF, ACS, and major scientific research projects. He is a recipient of the Google Scholar Award, AAPM and AMBIE.
|| Deep learning algorithms are mostly general-purpose, but need to develop application-specific algorithms to improve efficiency and accuracy.
Is deep learning technology suitable for all types of medical imaging, is it more appropriate, and some are more suitable for other technologies?
Xing Lei: Deep learning is just a method of machine learning, but it has been very popular recently because it is widely used and very effective. Deep learning does have many applications in many medical problems, but some of them are relatively mature (such as deep learning for 2D image recognition, processing and segmentation), and some are in the stage of exploration (such as using deep learning to help doctors read). Explain 3D and 4D images). The work of Professor Sebastian Thrun from the Stanford Department of Computer Science on the use of deep learning for the diagnosis of skin cancer may be familiar to everyone.
At the same time, the algorithm is also improving. As time drifts, you will see various innovations to simplify the algorithm to form a more general algorithm. Of course, the general-purpose algorithm is effective in solving some specific applications. It may not be that high.
What difficulties does deep learning currently encounter in medical image processing? What are the academic problems that need to be overcome?
Xing Lei: Deep learning applications are very extensive. There are many people who do research and development in big data. The open source algorithm platform provided by Google has greatly reduced the threshold for entering artificial intelligence. Many people are rushing toward this research direction. Maybe in the next few years, people will see various applications like avalanches.
At the same time, there are of course many difficulties. First of all, the amount of calculation is very large. If the current computing power is not a big problem when dealing with two-dimensional images, but for three-dimensional or even four-dimensional image processing (time plus space), the computing and storage capabilities of modern computers become a Obstruction.
Therefore, in the future, some application-specific algorithm innovations will be needed to shorten the calculation time and improve the calculation efficiency.
What kind of breakthrough is the need for these difficulties?
Xing Lei: You can't just focus on developing general-purpose algorithms. Research on details and specific applications will help artificial intelligence to land and benefit humanity.
|| Intelligent medical decision-making is still in its infancy
The integration of medical imaging and case history data to make a comprehensive intelligent analysis decision, what level is it now?
Xing Lei: Very primitive stage. Now the hospital's systematic and comprehensive intelligent analysis and decision-making of patients is still not enough. For example, the nuclear magnetic result of a patient today is coming, and it is analyzed, but in fact, this patient may have been related 10 years ago. The results of nuclear magnetic, CT and case history, can these historical data be integrated? If you have a comprehensive intelligent analysis decision, the effect will be much better.
Of course, there are a lot of people who start to "think" in this area, but there are very few people who start to do it. Because it is very difficult to implement, first of all, there must be technology, doctors can not write their own programs, and there is a large amount of clinical data to prove the effectiveness of such practices, everyone will accept.
For artificial intelligence in the medical field, in the field of medical imaging, what is your vision and future vision?
Xing Lei: I personally think that in the future, every radiologist, mobile phone or computer terminal will have an APP for intelligent analysis and decision-making. That is to say, basically all of them have to go through the radiology department (including other departments). The patient will also pass this APP, especially the difficult condition, and artificial intelligence will assist in the analysis and decision-making. As we have seen at the meeting today, Huiyi Huiying has made amazing progress in building such a smart medical imaging platform.
So, will the final computer replace the radiologist?
In the near term, it is unlikely because people will need quality control and final decision making. However, technology does help a lot. In addition to improving efficiency, it can raise the quality and accuracy, prompting many doctors to think about many problems that were previously unseen. As the saying goes, the three stinkers are the top ones. Artificial intelligence is made up of three Zhuge Liang, and should be a super doctor. )
|| China's AI medical development is not inferior to foreign countries
How do the conditions for AI medical research in China compare with the international level?
Xing Lei: Overall, the difference is not big. In terms of hardware, domestic super-computing is second to none, but when it comes to high-performance computers, it may not be as popular in foreign countries. From a research perspective, it can be clearly seen that the popularity of this smart field in China is rising.
As I mentioned, the lack of data and non-standardization is one of the biggest obstacles to the development of smart medicine. There are still many advantages in this aspect in China. After all, the government can coordinate, negotiate and encourage this problem very effectively. It may be more efficient than abroad.
Â
|| The current standard of human learning as a machine learning is still reasonable
In general, measuring the level of AI medical images will compare it with human recognition ability as a standard. Do you think this is reasonable?
Xing Lei: Reasonable and unreasonable. The reasonable aspect is that in the early stage of R&D, there is no better standard in the industry. It is not just artificial intelligence technology. In the past, the development of imaging science often used the level of experts as the criterion for judgment, such as the segmentation of medical images, computers. Auxiliary breast cancer diagnosis, and more.
So, is it possible to use machine-learned and validated (super doctor's) results as a standard in the future?
Xing Lei: This is a question worth exploring, but in the end I think it is ultimately up to the clinician to participate.
In general, a new method must have a benchmark when it comes out of the application, and then the innovations that have been made must also look for new benchmarks.
|| Focus on automated medical decision making research
What is the main research you are doing now? What is the topic that you are most interested in?
Xing Lei: My laboratory has a wide range of research, from basic imaging equipment, molecular imaging, image reconstruction and processing, imaging and genomics, treatment planning, to clinical data collection and analysis. Many of these projects involve the application of machine learning and artificial intelligence. It can be said that artificial intelligence in the future will be an essential component in medical research and clinical applications. Our recent research in machine learning and artificial intelligence applications includes deep learning for image segmentation and processing, machine learning to automate the work that clinicians and doctors do, and artificial intelligence. Help doctors make clinical decisions, and use artificial intelligence to process and reconstruct images collected in some limited situations.
To give a simple example, in the process of implementing radiotherapy, doctors need to develop a treatment plan. The treatment plan is a technical machine optimization process. This process involves many decisions. It is very effective to use artificial intelligence to do this work. The work that can be done in hours or even days, the computer can be completed in ten minutes, and the quality can be guaranteed. Therefore, artificial intelligence can be of great help in improving efficiency and quality.
Can you tell Professor Xing to talk about what academic genres, including algorithms, are in the big data artificial intelligence, which we don't know, and what you see, or are studying?
Xing Lei: Actually, big data and artificial intelligence are not a new concept. Artificial intelligence was introduced in the 1950s, when scientists represented by John McCarthy of the Stanford University Computer Science Department clearly recognized the importance of "artificial intelligence." Perhaps because of many historical reasons, artificial intelligence has three relatively large so-called schools:
• The first is symbolism. Early AI operations used a symbolic operating system. Symbolism mainly focuses on artificial intelligence from the perspective of psychology. It believes that human cognition can be expressed through symbols to perform AI calculations. This method is still very active now, and the recent success of deep learning has injected new vitality into symbolism.
• The second is evolutionism, which is based primarily on cybernetics and perceptual-action systems. Evolutionism has also played an indelible role in computer simulation of human cognition.
• The third is the bionics school. The nerve-network that has recently been the hottest in the media is this school. Based on the brain is a basic model of a network of neurons and connected neurons, learning to train neuro-network seems to be a very intuitive approach. However, from the 1980s to the 1990s, the research progress of neural networks was not very significant, mainly because the computing speed and storage of computers could not keep up. Because the amount of calculation is too large, it goes without saying that deep learning is a difficult single-layer network.
There are three main parts to learning artificial intelligence: data, models, and algorithms.
• data, in addition to high-quality data requires a lot of outside, standardized data is also crucial. At Stanford University's Biomedical Information Program, one of the big labs is dedicated to standardizing terms. Expressing all medical, industrial and commercial terms in standard language is actually a huge project.
• models, the most worthy of the day I am afraid that if the depth of reinforcement learning (Deep Reinforcement Learning), reinforcement learning has existed for a long time, from some of the operations research (operation research) methods derived from development. Deep-enhanced learning is particularly well-suited to solve some of the problems associated with machine and environment interactions in artificial intelligence.
Source: Lei Feng Net
Automatic Air Hose Reel,Auto Rewind Air Hose Reel,Self Retracting Air Hose Reel,Air Compressor Retractable Hose Reel
NINGBO QIKAI ENVIRONMENTAL TECHNOLOGY CO.,LTD , https://www.water-hose-reel.com