Current Events in AI: DeepMind Diagnosis Eye Diseases
DeepMind has announced a product it plans to release as Google parents company Alphabets first medical product. This product aims to diagnose diseases which occur in the eye. It is currently capable of detecting a range of diseases including glaucoma, macular degeneration, and diabetic retinopathy. This benefit of this product is in providing early detection to those who may not have the time, money, or resources to make routine visits to their eye doctor and free up doctors as well to focus on treatment more while handing off some diagnostic capabilities to this DeepMind product. Additionally, specialists could focus and prioritize those with the greatest need due to the severity of their disease.
With their partnership with eye hospitals, DeepMind was able to train its model to predict diseases more accurately than leading specialists. In order to make predictions, the product uses two different types of neural networks connected together. One network is a segmentation network. This type of network assigns each pixel in the image an object class in oder to understand an image at the pixel level. Unlike classification this allows for dense pixel based predictions. This neural net is used to analyze OCT scans (3D images of the back of the eye) in order to map parts of the eye and features of diseases in the images. The second network is a classification network which uses the map from the previous network to present diagnoses and recommendations with a percentage of confidence. With this method, clinicians can see the methodology behind a diagnosis using the first neural network.
This technology could mean drastic changes in early diagnosis and treatment of ocular diseases to limit their effects. However, it should be noted, this is not a replacement to the job of medical professionals but rather a tool to assist. Additionally, there are ethical implications of products of this nature including patient confidentiality. This is due to the method of retrieving training data for the model. DeepMind has, unfortunately in the past, attained training data for other products with methods that violated patient rights who were not informed that their data would be used for their select test purposes.
For more on this product, you can view DeepMinds blog: Predicting eye Disease with Moorfields Eye Hospital.
