For more information on the Future of Healthcare Grand Rounds - Medical Imaging Topics series, please click on the hyperlink to read.
"Those who do AI don't understand medicine and those who study medicine don't understand AI ? The top "AI+Medical Imaging" course series is here
The following is an excerpt from a course taught by Professor L. W. Wang
Why did you choose the path of medical imaging?
Prof. Liwei Wang talked about three areas that will be closely integrated with deep learning, and hopes to choose one of them for in-depth research: one is driverless; one is in finance; and the other is in medicine.
He believes that driverless in computer vision recognition technology has matured and has the conditions that can be combined with machine learning in depth, but as an application direction closely related to automotive entities, the space for doing research within universities is relatively limited; in addition, driving as a dynamic activity is much more difficult and complex compared to static image recognition.
And although the financial sector has accumulated a very large amount of high-quality closed data, it is a very noisy signal, subject to sudden policies, human manipulation and not always strictly following objective laws.
AI technology is not the core technology in financial transactions; the security of transactions is the top priority. If you consider only the frequency of transactions, transactions are divided into low-frequency and high-frequency, "If it is low-frequency transactions, I do not think AI has too much use, because AI needs big data, low-frequency transaction data volume is too small, if you want AI technology to play technology, must be in the high-frequency field. But a problem at this stage is that there is a lot of analytical work in the high-frequency field that still needs to be done by people, such as refining factors or strategies that have an impact on trading. So combining several of the above factors, I ended up choosing to start with medical imaging. "
At this triple crossroads of applied research, Prof. Liwei Wang finally chose the field of medical imaging.
"Medical imaging is the area where artificial intelligence will have the most profound impact within the next few years."
On the one hand, medical images are static image recognition, which is more mature than video processing technology; on the other hand, relying on a series of affiliated hospitals of Peking University, Prof. Wang Liwei's research team of Peking University has a unique advantage in terms of data acquisition and system testing.
Professor Wang said that if the task the developer wants to perform is in a very closed environment and has little to do with human common sense, such a task is ideal for a machine to do, but if the task is tied to common sense, such as the understanding of natural language, it is very difficult for a machine. "Medical imaging is relatively a somewhat closed issue."
The real state of current AI medical imaging technology
Of course, it's not that medical imaging is somewhat closed and AI will be able to travel unhindered in the medical field. Professor Liwei Wang cited two cases where AI has produced significant images for medical imaging - the Stanford team in skin cancer detection and Google's DeepMind for glucose screenings.
Professor Wang's view is that. 1. The number of medical imaging treatments is very large, and the two team examples above address single-patient problems. In total, medical imaging can be divided into more than two thousand types of diseases. "Solving a single disease is no longer simple; the Stanford team has been working with top medical experts for several years to get the current results, and it's even harder to encompass more than 2,000 diseases. In addition, the degree of variation from disease to disease is so great that AI medical imaging requires the full collaboration of top medical experts and top machine learning academics to be successful. " 2. Differences in imaging equipment can produce multiple types of images. CT, MRI, X-ray, ultrasound, endoscopy, and pathology slides lack certain standards for these images. "Right now we're only making single point breakthroughs in some diseases, and medical imaging as a whole still has a long way to go."
From a technical point of view, today's machine learning and artificial intelligence techniques may do a better job for problems like detection, but global lesions and structural lesions are still difficult for machine learning.
The next few years will depend on who can win by choosing the right disease, and choosing the wrong one is a potential waste of time and energy.
Early screening of lung nodules is a popular area of AI+medical imaging, but lung nodules are diverse, morphologically diverse, and easily confused with other tissues in the lung (e.g., blood vessels, etc.).
These issues will also test entrepreneurs who cut their teeth in this space, such as the detection of nodules and ground glass nodules in the hilar region of the lung. Therefore, Prof. Wang's team has adopted a three-stage model framework to address such a problem.
The hilar region nodes are very prone to errors and missed examinations, and the nature of the hilar region nodes being completely connected to other structures makes the information difficult to distinguish. The core of Prof. Wang's team's technique is to combine multi-scale information and apply a kind of Feature Pyramid Network (FPN). By integrating multi-scale information, it effectively distinguishes nodules from normal tissue and other lesions and accurately detects nodules in the hilar region of the lung.
The ground glass nodules are also an easy part to miss, compared to other areas where the difference in brightness is smaller. By introducing a difficult case mining mechanism, the model is made to focus more on difficult samples and improve the accuracy of the detection of ground glass nodules.
In this problem, Prof. Liwei Wang's team uses the focal loss function (Focal Loss) to optimize the grinding glass hard case mining.
In terms of analysis of the clinical significance of nodules, Prof. L.W. Wang pointed out that one patient has nodules in only one place in the whole lung, and another patient has many nodules, and although there are the same nodules, the significance is different for the two individuals. We help physicians focus on the more clinically significant nodes by combining global information on all nodes within a case for nodal clinical significance analysis and determination.
"We want the final diagnosis to be helpful to physicians, and we must also analyze the clinical significance and ultimately submit the results." Context Net was used here to avoid looking at detected nodules in isolation, but rather to determine globally which ones are clinically significant, including using the system to determine probability values for nodule benignity and malignancy.
In Prof. Wang's opinion, the analysis is not a simple target test, but has to take into account the characteristics of medicine. Beginners will use general-purpose algorithms up front, but there are many details of machine learning that are not written in code and can only be appreciated more if you go many years deep and practice them yourself.
In addition, Prof. Liwei Wang's team has also made some breakthroughs in algorithm extensions.
"We also have some results on nodal property issues, model interpretability, etc. The nodal properties are helpful to the doctor in writing the report, and we can measure the nodal diameter and volume to give the doctor a more visual sense. In terms of the interpretability of the model, we use the model obtained by the pre-existing algorithm to detect lesions and determine benign and malignant properties. But doctors will ask what is the basis for the model's judgement, what is the logic, and we don't want the model to be a black box, thus affecting our confidence in using it. "
After the online live broadcast of Lei Feng AI Nuggets, Professor Wang Liwei answered some of the questions raised by the participants, the following are selected Q&A transcripts, Lei Feng AI Nuggets has been edited without changing the original meaning. Q: What are the most important issues to keep in mind during the process of working with physician hospitals?
A: First, be sure to get to know each other. It's pretty difficult for techs doing AI and doctors to communicate without knowing anything about each other's domain knowledge. These two areas are so different that you have to gradually adapt to each other's mindset. So if you can understand the other person's way of thinking, I think it will be easier and easier to communicate and to work together.
Second, as technical workers in artificial intelligence and machine learning, it is important not to think about how the technology works from your own perspective. The core and most fundamental thing about AI medical imaging is to consider what its real needs are from the physician's perspective. One thing I've learned the hard way is that in a sense, the demand should be physician-driven. So there is an in-depth communication between the two parties about which issues the doctor really needs assistance with and at the same time are likely to be solved by technology.
Q: If AI models are less reproducible, how can the credibility of the models be improved?
A: There is a lot of open source code available online, I mean are you able to reach the level of the algorithm designer after using this code. Code and models are two concepts; models are actually trained on data using code, but the training itself is a lot more skillful. If you don't understand the tricks of the trade and can't achieve the level of code developer, it just means that you get the code and can you train the same model with your own experience. The gap between a novice in machine learning, especially deep learning, and an experienced professional expert is still huge, and this takes many years of skill building.
Q: As a radiologist, what is the biggest concern about the automatic lung nodule detection product is false negatives, how can this problem be solved?
A: From a medical point of view, the performance indicators that people are concerned about should be two: specificity and sensitivity, and sensitivity actually refers to a kind of check-all rate, whether all the nodes are found. Specificity is the proportion of false positives. In a sense, these are two contradictory indicators, and if one is cranked up to the maximum, then surely the other will perform poorly, and everyone wants the two indicators to strike the best balance.
1、Create an intelligent digital 3D model library 3D models can be edited at will as if programmed which may solve the biggest pain point of 3D printing 2、How to manage NodeJS packages using npm 3、Awesome The next generation of artificial intelligence development plan these brains tell you 4、Using the command line with Python using YARA rules
5、The vivoX21 continues to sell like hotcakes With these points it is widely loved by consumers