Google DeepMind’s Revolution in AMDs
Age-related macular degeneration, or AMD, is the leading cause of blindness for those who are 60 years or older. It affects the macula, the part of the retina in our eyes that allows us to see fine detail. Most of the time, AMDs progress without pain and little vision loss, making it difficult for patients to notice its progression. There are two types of AMDs:
- Wet AMDs
- Dry AMDs
What is the difference between a wet and dry AMD?
Wet AMDs occur when abnormal blood vessels start to grow behind the macula. The fluids leaked from these vessels move the macula, causing blindness of the eye. A dry AMD is caused due to the thinning of the macula. The scariest thing about dry AMDs are that they progress from one eye to the other; you wouldn’t even know when it affects the other eye.
How is DeepMind solving this problem?
DeepMind uses an anonymous AMD dataset from The Moorfields Eye Hospital of patients with AMDs in one eye, and a high risk of developing it in the other eye. But AI can already make predictions, what is different about predicting a retinal disease progression?
They use a two-layer deep convolutional neural network to solve this problem.
Predictions from the raw eye-scans (using a digital OCT scan)
An Optical Coherence Tomography (OCT) is a non-invasive imaging test, which uses light waves to take cross-section pictures of the retina. While collaborating with Google Health, DeepMind developed a model that segments these eye scans into 13 anatomical categories. This basically means they are classifying the data to train ML models to predict a patients’ risk of developing exAMD in their other eye within the next six months.
Using both the raw data and the segmented data as inputs to train the AI model
What is interesting about this approach is that it gives the AI different views of the eye scans. The anatomical segmentation helps the system to identify known anatomical indicators — for example, it can detect the loss of retinal pigment epithelium(which helps to protect and feed the layers of the retina) or detect the growth of drusen (small fatty deposits). The raw scan, on the other hand, can allow the model to spot subtle changes that could become risks in the future. When combined, the system is able to predict if the eye will proceed into the exAMD stage, and if so, when and how fast the progression is. This way, patients can be informed of the disease and find better ways to prevent it.
Conclusion
In conclusion, AI is making its way to solve the most difficult problems in the medical world. With the data that is collected from hospitals, we can find ways to improve out healthcare system and find cures to even the most dangerous diseases.
