Hi Hiroki,
You are Artificial Intelligence (AI) scientist and founder of the Start-Up Xeno-Hoc. Please tell us a little bit about yourself and the mission of your company.
Our mission is to communicate engineering in a way that is easy for doctors to understand. Engineers have the scope of engineering, and doctors have the scope of doctors. This makes it impossible for doctors and engineers to communicate smoothly. We want to be a bridge between doctors and engineers, not just an AI company.
We also worked together on several projects regarding image recognition by artificial intelligence. Moreover, you already published many studies, where you developed algorithms for solving specific ophthalmological problems. To your knowledge, what are limitations of Artificial intelligence?
The limitation of artificial intelligence is its inability to deal with unknown data. Artificial intelligence can only act within the scope of the data it is given and cannot deal with unknown risks. Therefore, I believe that it can improve efficiency within a limited scope, but it cannot replace doctors who have a very broad vision.
When we worked together, I was always wondering, that your company was able to develop very accurate algorithms, even without really large amounts of imaging data in the context of big data. How much data do you need for the development of really accurate algorithms and what does this mean for the detection for example of rare diseases in ophthalmology, where we don’t have so much data?
How much data do we need to develop AI?" is a question we are often asked. In fact, we won't know how much data we need until after the development is complete. In general, we can analyze a few hundred pieces of data at the research paper level. However, there is a possibility of image bias (e.g., bias in the background of the images depending on the disease). Especially in retrospective studies. To develop without any bias, the quantity of data (tens of thousands of images or more) is naturally important, but the quality of the data (minimal bias) is also important.
Discovering rare diseases is one of the things that AI is not good at. The reason is that data cannot be collected, and the possibility of bias in the data is very high. To collect data, we need to collaborate with many institutions and collect as much as possible, but it may still be difficult. Another strategy is to collect normal images and detect abnormalities. However, anomaly detection also has its drawbacks. It is less accurate than image classification, and it is impossible to know how abnormal it is (no disease name can be given).
In Europe, but also in many other countries, ophthalmologists and also physicians of other disciplines are always frightened, that data scientist as you are one, have the aim to get rid of doctors and establish a completely algorithm based medicine. What is your opinion on this? Will we have doctors in 20 years? How does the perfect symbiosis between practitioners and algorithms looks like in your opinion?
I don't think algorithms can eliminate doctors. At least not during the 21st century. As mentioned above, data science, such as machine learning, cannot deal with unknown risks. The more I learn about data science, the more I realize this. Doctors can use their imagination to deal with diseases that they have only seen in textbooks but have never actually seen before. Artificial intelligence has tremendous processing power, but no imagination. Also, the knowledge of physicians is essential in setting up the topics to be covered by machine learning. AI developed with data scientists taking the lead in deciding on themes can be frustrated by social problems. We believe that AI will be assistant of doctors by improving the efficiency of tests and making test results easier to read in the future.
When we developed algorithms together, we always had two problems. 1. The problem of the ground truth (e.g. that OCT of graft detachments on DMEK have been choosen for rebubbling by one or two specific surgeons) and 2. that we cannot really predict what happens, if we transfer algorithms, that have been developed with a specific surgeon, on a specific group of patients or on a specific device, cannot be transferred to patients of other surgeons, in other countries or when using other imaging devices (e.g. Heidelberg Engineering vs. Tomey etc.). So, the transferability of these algorithms is very limited. How can we solve this problem?
As for the ground truth problem, the same problem occurs in Grading for diabetes, etc. If you have enough resources, you can use multiple doctors (e.g., 3 doctors) to verify the results and decide the correct answer by majority vote. However, while there are ways to increase the likelihood of Ground truth being correct, there is no way to make it 100%.
Low transferability is another problem that often arises in studies of small scale. Regarding facilities, the problem can be solved by collaborating with many facilities to create a database. Regarding race, no major problems are expected to arise for images with relatively small differences in appearance between races, such as OCT. On the other hand, it will be difficult to deal with images with large racial differences such as iris by the same AI model. Regarding devices (e.g., Heidelberg vs. Tomey), the differences of the images taken by the different devices are too large. The range of imaging is different, and the appearance is also very different. Therefore, the transfer would be very difficult. Even if a multi-center, multi-racial study can be done, I believe that AI models should be developed for each device.
Thank you very, very much Hiroki for answering all these questions!
Links
Hiroki Masumoto on Researchgate
Associated Research Group of Prof. Hitoshi Tabuchi
References
Masumoto, Hiroki, et al. "Deep-learning classifier with an ultrawide-field scanning laser ophthalmoscope detects glaucoma visual field severity." Journal of glaucoma 27.7 (2018): 647-652.
Sogawa, Takahiro, et al. "Accuracy of a deep convolutional neural network in the detection of myopic macular diseases using swept-source optical coherence tomography." Plos one15.4 (2020): e0227240.
Nagasato, Daisuke, et al. "Automated detection of a nonperfusion area caused by retinal vein occlusion in optical coherence tomography angiography images using deep learning." PloS one 14.11 (2019): e0223965.
Kato, Naoko, et al. "Predicting Keratoconus Progression and Need for Corneal Crosslinking Using Deep Learning." Journal of clinical medicine 10.4 (2021): 844.
Hayashi, Takahiko, et al. "A Deep Learning Approach for Successful Big-bubble Formation Prediction in Deep Anterior Lamellar Keratoplasty." (2021).
