Germany
Deepeye empowers ophthalmologists by using an AI model trained on thousands of images. It helps doctors understand the optimal time and frequency to give injections that can help prevent blindness. This in-depth article discusses the process of how three-dimensional images from eye scans trained artificial intelligence systems, why the term “augmented intelligence” may be more appropriate to describe advancements in AI and where further applications of this tech could be in the future.
On the latest episode of 15 Minutes With The Doctor, Manuel Opitz a serial healthtech entrepreneur joins Vinay on the show again. After continued growth at Mecuris, a 3D prosthetics startup, he has now started work on a Deepeye. It’s an AI platform that supports ophthalmologists make important therapeutic decisions in blindness causing health conditions such as age-related macular degeneration and diabetic retinopathy. Learn how he got involved, the specifics of the platform, gaining investment and contracts and his personal story. In this special article, they both explore in depth how optical coherence tomography or pictures from the back of the eye were used to build neural networks from millions of images, they discuss how augmented intelligence may be a more appropriate term for artificial intelligence and how they can see their AI being adapted for use in Asia.
I wanted to touch briefly on the AI. I’ve had previous conversations with guests focused on diagnosis and screening and not therapeutics. It strikes me this specific area in ophthalmology treatment is excellent to train the AI because you’ve got images and you’ve got years and years of data, and you’ve got significant numbers of the population where the condition occurs. How was the AI developed? Could you share some information on that?
So we just actually last week, we published a scientific paper on this topic. It took a very long time on publishing, but finally, the results are out. What’s important to understand is, as you just said, I mean, AI is just the means to an end. It’s not about replacing anyone… I prefer this definition of the American Association of Physicians; they say, “AI means augmented intelligence and not artificial intelligence,” because in the end, it’s not really about having artificial intelligence, but more about learning from patterns and cases that other doctors have seen. So we provide superpowers to doctors, and those superpowers are basically the knowledge of all those other doctors out there and all those cases they have seen.
We don’t provide it just in one big AI model. We have trained several neural networks, and we have access to a proprietary database, stretching back almost ten years of treating age-related macular degeneration and diabetic retinopathy. There are several hundred thousand actually images. And of course, we don’t use all of them. We pick those who are most likely to provide some value for the training of the algorithm. And we have several thousand cases and patients involved. And actually, we take only patients who are in treatment for a long time, so they need at least 20 images taken. So usually, they’ll stretch this over four to five or even six, seven years.
And it’s imperative to have this longitudinal data as well. So it’s not very easy to train an algorithm, especially if you have only some clinical trial data, which stretches over 12 or 24 months. So this really provides our neural networks a possibility to learn how this disease will evolve and how it will react to an injection or not receiving an injection of anti-VEGF.
Additionally, we have monthly follow-up retinal imaging, which is unique and rich within this field. Most study centres don’t have this, and this is why we know how patients progress in between injections when they don’t get an injection. And again, this is a constructive input for our algorithms and those AI models we trained. And it’s not just one model, but several models. So, for instance, one model just for the decision, should I treat now? Yes or no. Then the second model can say, “Hey, this is a patient who is likely to need more or fewer injections.” And then we have a third algorithm that can decide, “should I ask the patient to come in the next time, just for follow up, or should I ask the patient to already come for an injection?” And then a fourth algorithm to say, “Okay, when should this next appointment be?” So we try to provide counselling for the doctor on every step around this therapy planning and decision making.
Vinay:
Fascinating. I often say to my GP colleagues that as a doctor in primary care, we are pattern recognising and risk stratifying in our heads all the time subconsciously, and through experience doing this faster and better hopefully each time. These neural networks are literally this on a mass scale. And I do like the definition and your description of augmented intelligence. I think it’s important to also bear in mind that some parts of the world and some institutions might not have the ophthalmologist’s significant, special knowledge to treat these conditions. This tool could also be used to support those places where that type of resource isn’t available!
Manuel:
Definitely, in countries beyond the USA, Europe, and the UK, the ration of retina specialists to patients is much lower. Most people there are focused on refractive surgery, i.e. cataract surgery. So, of course, having those possibilities and maybe providing the knowledge of thousands of doctors on millions of images and thousands of cases to some practitioners in other areas will be a huge thing.
Looking forward, maybe not till 2030, but by 2050, most older adults will not be living anymore in just the US and the EU. But by then, of course, the majority of elderly people will be living in Asia. And that’s why, we are very, very excited that now such solutions are possible. Of course, we would need to de-bias our algorithm, which now would not achieve the same accuracy as it can achieve on European population retinal scans on Asian people. But if you manage to do so over the next ten years, this will be a huge opportunity to provide to those doctors in China, India, and all those countries where a lot of people will pass the threshold of 55 or 60 years old and will be in imminent danger of having such a retinal disease themselves.
