If signs of lung cancer can be found at a very early stage, then the patient has a survival rate of more than 90%. But for tumors that are close to advanced stages, the five-year survival rate may be around 12%. How can I help radiologists troubleshoot missed diagnoses? With the power of science and technology, we can change the status quo of imaging medical treatment in China!
I believe that you will have such a feeling after going to the domestic hospital, you may queue up for several hours, and go to the doctor for a few minutes to see it. Patients think they don't get the services they deserve, but for doctors, they get up at 5:30 in the morning, go to the hospital for a morning meeting at 7:30, and work until 10 p.m. may not be able to complete the day's diagnosis, and when they return home, they may have to carry out scientific research.
A ct scan of the lungs is a three-dimensional image, and from top to bottom, it takes two or three hundred images to see it. The radiologist's daily job is to read a lot of CT, X-ray, and mi radiographic images to look for early lesions. A case has to be looked back and forth three to four times, more than a thousand images, to confirm whether the patient has any problems. On this basis, a radiologist often faces more than a hundred patients a day.
It is precisely these early lesions that have a great impact on the survival rate of patients. Taking lung cancer as an example, a very early sign of lung cancer has a survival rate of more than 90%, and for a particularly early stage, there is even a survival rate of close to 100%. But for a nearly advanced tumor, the five-year survival rate may be around 12%. Therefore, radiologists need to work very hard to find these early lesions. If it cannot be found, it will greatly affect the efficiency of later treatment and diagnosis.
How can we better solve this problem? Technology with artificial intelligence is the most suitable. We apply the most advanced deep learning technology to the field of medical imaging diagnosis to help doctors' diagnosis become more effective and accurate.
From 2012 to 2014, I studied for my Ph.D. in Economics and Finance at the University of Chicago. In fact, the financial field is also one of the earliest areas where artificial intelligence technology was applied. In 2013, more than 70% of the U.S. trading market was bought and sold by machines. I was very interested in this at the time, so I went to see where the most advanced AI deep learning technology can play a role.
Until one day, I met a radiologist who was very interesting about the artificial intelligence and deep learning techniques he was showing, and asked me if I could use the same technology to help radiologists solve problems in the industry. We began to understand the industry, communicate with different people, and finally found that this is indeed a very big pain point. So I thought it was a career worth fighting for my life, so at the end of 2014, I gave up my phD in the United States and returned to China to establish Speculative Technology.
In early 2015, very few people in the healthcare industry had heard of AI and deep learning. We have communicated with many hospitals, most of whom are interested in this, but there are very few who are really willing to cooperate in depth. Among them, there are also people who have participated in artificial intelligence in the medical industry more than a decade ago, but at that time, the artificial intelligence methods based on traditional technology were not enough to meet clinical needs and could not make a suitable product, so they were disappointed in artificial intelligence. When I went looking for them again in 2015, they would say to me, "We've been doing this technology for twenty years, it's impossible to implement, it's just a fool's thing." ”
After visiting dozens of hospitals, we were fortunate enough to have an early partner. In fact, there are many challenges to the artificial intelligence of medical imaging, and for people who do not know much about medical imaging, every X-ray and ct is a very high-resolution image.
For example, X-ray, the resolution of the smallest image in the medical image is 3000×3000, while the resolution of ct is 512×512×250, which is the high resolution, high information and high dimension of medical images. We look for early lesions in this image, and the difficulty can be compared to that of asking you to find all the typos in hundreds of thousands of words in a minute.
A big feature of deep learning and artificial intelligence technology is that people may not be able to understand the results it makes. Just like when we see the alpha dog play chess, we will find that there is no way to understand why it is playing in that place, but in the end it can win this game, so it is unexplainable. But in the healthcare industry, interpretability is a very important part. For any diagnosis, you need to tell the patient rationally why I am making such a conclusion, and what you should do next.
There is also a very important link in doing deep learning and artificial intelligence, that is, high-quality training data. But in the medical industry, high-quality training data is difficult to obtain because not any one person can do a tool for medical imaging diagnosis and screening.
These are the difficulties and problems we face when applying deep learning artificial intelligence to medical imaging and radiology. So how do we solve it? This requires everyone to go to the hospital and go to work, work, work, and live with the doctor. Only in this way can we make a product and technology that truly meets and meets clinical needs.
This is a case we encountered when a hospital in Shanghai was just launched. At that time, our system was still in the testing stage, and it circled a point and thought there was a problem on it, but the doctor looked at it for half a day and thought it didn't look like a problem place. He thinks it's a false positive for our model, that our product gave a wrong result. But because it was during the testing phase, we recommended that this patient go for a more in-depth examination. After the CT can be seen very clearly, this is a sign of early lung cancer, and this in-depth examination, really let us see that the results given by ai are not wrong, it helps the doctor find the place where the missed diagnosis is.
When ai learns cases beyond any radiologist, it is possible to see lesions that are not easily seen by some people, helping doctors find missed cases. We want to help doctors and patients find out some of the lesions that may be missed due to fatigue and lack of experience, and save their health.
If we can put the efficiency of diagnosis to the premise, and the period of each lesion found to the premise, we can greatly improve the possibility of treatment of any disease, which is the value of early diagnosis and early treatment. We hope to use artificial intelligence to help complete early diagnosis and treatment, and ultimately help the entire industry realize value.
Q: How much impact will artificial intelligence have on the disruption of healthcare?
A: The impact of artificial intelligence in the medical industry must be subtle, and when everyone may not notice, it will be found that doctors are already using a variety of AI means to help them make diagnosis more efficient and accurate.
q: What do you think is the fastest application of artificial intelligence to break through the bottleneck of medicine?
a: First of all, it must be the medical imaging piece, I believe that in the future, there is no need for people to complete the repetitive and particularly tedious work, it may not be the huge breakthroughs imagined, but it can bring great value.
a. Artificial intelligence is not afraid of fatigue and is careful and fast, of course, it can replace radiologists!
b. Repetitive and tedious work can be handed over to ai, and doctors can better complete creative and associative work.
c. Artificial intelligence is invented by humans, which is inaccurate and cannot be replaced!
d. Who is a radiologist? I haven't even seen it before I know...
Special thanks to Chuangyebang, Lieyun Network, Geek Park, Pin Play, Silicon Valley Power, Fengsong, Zhonggu Cross-border Service Center, Jiangmen, Shanghai Shuangchuang Investment Center, and Event Line for their strong support for "Future Jane"!