Reference Source: http://www.sohu.com/a/209212040_114877
From the AI excavation Jinzhi and mu School Joint Video Course, Course address: http://www.mooc.ai/course/339/thread?page=2
I did not watch the video, or reluctant to spend the money ah. Record some of the highlights from the Internet.
Expert Introduction: Professor Wang Liwei of Peking University, the first Asian scholar to receive AI ' s to watch award, and also the Luna and Tianchi AI Medical competition champion team Mentor
A case study by Professor Wang Artificial intelligence pulmonary nodule detection and benign and malignant diagnosis: a three-stage model framework is adopted to solve the problem of multiple pulmonary nodules, which are different in shape and easily confused with blood vessels and other tissues.
1. Pulmonary Portal Area nodules
Pulmonary Hilar regional nodules are very prone to errors and missing spots, and the characteristics of the Pulmonary portal area nodules and other structures are completely interconnected to make the information difficult to distinguish. The technical core of Prof. Wang's team is to combine multi-scale information and apply a feature pyramid network (FPN).
2. Frosted glass nodules
The grinding glass nodule is also an easy to be missed part, compared to the grinding glass nodules and other areas of the brightness difference is small. By introducing the difficult mining mechanism, the model pays more attention to the difficult samples and improves the detection accuracy of the grinding glass nodules.
3. Clinical significance of nodules
In the clinical analysis of nodules, Professor Wang Liwei points out that only one place in a patient's entire lungs has nodules, and the other has many nodules, although the same nodules have different meanings for two of people. We analyzed and judged the clinical significance of nodules by combining the global information of all the nodules in the case to help doctors focus on more clinical nodules. The whole process is:
Note: The context Net is used here. Instead of looking at the detected nodules in isolation, it is a global judgment of what is clinically significant, including the probability of using the system to determine the benign and malignant nodules.
In his view, medical imaging problems can be divided into two categories:
1. Improve efficiency. For doctors, it's easy to detect a pulmonary nodule, and the artificial intelligence blood drug does it to improve efficiency.
2. Overcoming Medical Problems
For the future in the AI Medical imaging field of workers, the most important thing is to have both artificial intelligence technology and medical knowledge, both to become machine learning experts, but also to continue to learn medical knowledge. AI workers need to spend a lot of time developing their own reading skills, like a junior physician to continue to learn the image of knowledge. It's better to be a two-way person who understands medical knowledge and machine learning.
Some questions and answers to students:
Q: As a radiologist, for the lung nodules automatic detection products, the most concern is false negative, how to solve this problem.
A: From the medical point of view, we are concerned about the performance indicators should be two: specificity and sensitivity, sensitivity actually refers to a recall, is not all the nodules have been found. Specificity refers to the percentage of false positives. In a sense, this is two contradictory indicators, if one indicator to the maximum, then the performance of another indicator will become very poor, we all hope that these two indicators to achieve a best balance.
Q: The current depth learning technology, can compare high-precision detection of how small nodules, about five mm.
A: Our current system detects about five millimeters of nodules without any problems. Of course, from the medical point of view, the state has also issued a guide to the manual on such screening. What we really need to care about is three millimeters or more knots, three millimeters. The following can now not be reflected in the report, our system in more than three mm detection is not a problem.