Airdoc CEO Ray Zhang on the Challenges Facing AI in Medicine
Airdoc, a leader in the field of artificial intelligence in the medical field, has been focusing on the application of artificial intelligence in medical image recognition. They have been working to solve the problem of uneven distribution of medical imaging resources. At the Medical Data and Medical Artificial Intelligence Summit Forum in Wuhan, China, Airdoc’s founder and CEO, Ray Zhang, shared some of their experience working with intelligent imaging, as well as some of the challenges facing artificial intelligence in medical image recognition.
More Intelligent, Better Care
“The distribution of medical resources is uneven, many people have no access to proper medical care. Artificial intelligence can learn from the experience of medical experts, and when applied at the community level, can help enhance and scale doctors’ work with disease identification.” Zhang described this as the original intent of Airdoc.
In recent years, artificial intelligence image recognition technology has developed rapidly and its abilities have surpassed humans in certain areas. Medical imaging as a tool for disease diagnosis has become quite prevalent. Of the various roles AI can play in healthcare, such as biotechnology, auxiliary diagnosis, pharmaceutical drug development and others, medical image recognition is the most widely used application.
“The AI industry is thriving, but bringing AI to practical applications still has many challenges.” Zhang mentioned a few common practices and challenges artificial intelligence in medical image recognition currently face.
According to Zhang, current AI applications of medical image recognition fall into three main types: classification, detection and segmentation.
Classification is the simplest, where a huge training sample size is used to differentiate weather an image is normal or not.
Detection or identification, of what is contained in the image. The sample size required is generally less than that for classification, but requires additional labeling effort. A doctor needs to look at and mark the overall problem areas in each image. Anything missed in this marking process will affect the overall detection accuracy.
Segmentation requires the greatest amount of manual effort, where doctors need to outline all lesion areas in each image. Certain types of lesions are relatively easy to label, but some are more difficult. It is possible that a single image may contain hundreds of lesions, marking them all manually would be too unrealistically big a workload, when in fact a doctor would usually simply look at the image to determine its severity.
Therefore, detection and segmentation methods are generally used when the training sample size is relatively small. At present, there are some ways to create data for rare or small sample data, and the data that looks out is no different, but the data is cleaned.
Each of the above methods can solve different problems though common tasks are a combination of these 3 methods.
1. Labeling is often the bottleneck
“Labeling is very tedious. A lot of people would think that artificial intelligence practitioners just write algorithms. Actually 80% of the time is spent on data pre-processing. Every bit of effort spent on data pre-processing will see its corresponding return show up in the end result. It is a critical part of the process, but the labeling of medical data usually requires medical experts’ time and effort, creating a bottleneck.
Zhang believes that in the next 2 to 5 years, small-sample learning will make a breakthrough, and Airdoc has been exploring this in hopes of making certain headway. However, there have been no recent advancements in this area, leaving the task of labeling to doctors.
2. Data quality issues
Whether in the United States or China, data from most hospitals are not standardized. There has never been a way to effectively and comprehensively transfer all the data that a doctor has ever seen to another doctor; a doctor’s data is highly personalized. Therein lies the problem: for any one disease, ten doctors will give several different diagnostic advice. It is often hard to say who is right and who is wrong. So the common practice is to go with the consensus, which is still far from ideal.
While examining an image, a doctor could designate a severity of 4a. However it is very common for the same doctor to diagnose that same patient 3 months later as 4b. These discrepancies and other factors, if not handled well, can lead to data corruption.
3. Interactive issues
In general, the more the patient and the doctor interact, the worse the quality of the algorithm.
Chinese medicine tends towards holistic treatment. Conversely an algorithm strives to isolate and pinpoint a single area in order to increase its ability to discern that area. While creating an AI scenario, the less the interaction between the doctor and the patient the better. Best is if the doctor does not need to see the patient to make a diagnosis, to minimize irrelevant bias, “because the algorithm is controllable, the interaction is not.”
4. Weak generalized artificial intelligence
Generalized Artificial Intelligence is very weak. For example, if someone asks “how are you?”, a human would very easily and naturally determine this as just a greeting.
In 2015, Airdoc collected all extensive literature and knowledge material to build a cardiovascular-related common sense library, but eventually found that it is still less effective than a doctor’s knowledge base.
5. There is a huge gap between regulatory development and technological development
Regulations change very slowly, while technology changes very fast. The iterative speed of technology and that of the medical system are obviously complete opposites.
Zhang said that it’s fundamentally impossible to obtain FDA approval by following their existing rules. The FDA stipulates that the final product must be exactly the same as what was specified in its application. However AI algorithms are self-learning and become more and more intelligent than its original design.
6. The market is still in development
Zhang asserts that the current market is still slowly developing.
“It is a very long process from the time something new becomes accepted, then adopted, and then inducted into an entire medical system. This is why those in this field must work together tirelessly to built up the industry.”