Yoshihiro Yonekawa, MD
Mass Eye & Ear / Boston Children’s
RETINA Roundup Co-Editor
One of the highlights from Friday morning was a special lecture by Dale Webster, Ph.D., a software engineer for Google. Rick Spaide introduced Dr. Webster, who spoke about machine interpretation of fundus images.
He started his fascinating talk by defining common terminologies that may be confusing:
- Artificial intelligence is the science of making things smart.
- Machine learning is machines that learn to be smarter – this represents one of the most promising techniques of artificial intelligence, where the machine is programed to learn from its self.
- Deep learning is a particular kind of machine learning.
Such artificial “neural networks” have been around since the 1980s, but it never caught on due to limited computing power. Now we can train machines to perform highly complex tasks, and this is will continue to grow.
The most mature application in ophthalmology is for screening patients for diabetic retinopathy. In India, there is an estimated shortage of 127,000 eye doctors, and 45% of patients suffer vision loss before diagnosis of diabetic retinopathy. This is completely preventable with improved access to care. Telescreening with automated diagnosis by machine learning can solve this issue.
There are three pillars to deep learning:
- Many images need to be collected
- Accurate diagnoses need to be assigned to each image
- The trained model needs to be generalizable to the test population
Dr. Webster then discussed how the machines are trained to diagnose diabetic retinopathy. More than 50,000 images are usually needed for good performance of the model. Then you need high quality data of the diagnoses themselves. Google worked with 130,000 images, which were graded by 54 ophthalmologists, to no retinopathy, mild, moderate, severe, and proliferative.
Multiple ophthalmologists graded each image, but the consensus was still inadequate. It was decided to have three ophthalmologists grade the images, and if there was any discordance, there was an active discussion between the three graders to come to a consensus.
The third pillar of generalization is very important. A common failure of implementing machine learning technologies is where the team spends a lot of time training the system, but it breaks down when you apply it to a specific population that you are interested in. One of the key challenges in diabetic retinopathy deep learning was the presence of various ethnicities and coloration of fundi, but recent papers have shown that it can be overcome to provide accurate grading.
Dr. Webster indicated that up until this point, his lecture had focused on trying to replicate diagnoses. But what about things that doctors cannot diagnose, or think about diagnosing?
For example, the deep learning algorithms that Google developed was able to determine the refractive error of an eye based on the fundus image. This was accurate to half a diopter. Just by looking at a fundus image.
Another exciting development is in the risk prediction of cardiovascular disease based on the fundus image. A deep learning model trained to look for various cardiovascular risk factors in fundus images. The results still need to be refined in its applicability, but it would be very exciting to be able to perform cardiovascular risk assessments, noninvasively.
Finally, Dr. Webster discussed the future of artificial intelligence in our field. So far, we have been using deep learning for image interpretation. Google actually has a “reverse image search” where you can upload an image and find images that look alike. What if we were able to uploaded fundus images and use artificial intelligence technologies to determine the diagnosis? It would be very helpful for rare and difficult diagnoses.
Dr. Webster concluded his talk by expressing his excitement about being able to improve care and access for patients in the future.