In a blog post published by Lily Peng and Varun Gulshan of Google, the Mountain View company says it has successfully used artificial intelligence to detect diabetic retinopathy (DR), a complication of diabetes mellitus.
Google, in the post, stated that research began a few years ago with a group of experts. It was to study the possibility of using deep learning to screen for diabetic retinopathy. This disease is considered the leading cause of blindness in people with diabetes with about 415 million patients exposed worldwide. However, treatment is possible if the diagnosis is made early enough; otherwise, blindness is irreversible.
To detect eye disease related to diabetes, the most used technique by the experts is the pictures of the back review of the eye. This helps to identify possible signs of the disease and to measure its severity. As shown in Figure below, the image on the left (A) shows a healthy retina, while the right one (B) has a number of red spots or bleeding; therefore, the retina at right display symptoms of diabetic retinopathy.
Google has published in the Journal of American Medical Association (JAMA) the conclusions of studies by researchers. In the report, researchers report that the deep learning algorithm they developed can diagnose diabetic retinopathy from retinal images. This feat, according to them, could be very beneficial for physicians as it will allow them to discuss the many patients as possible, especially those who live in areas that do not have enough specialists.
The machine learning
The Mountain View company says its researchers have worked closely with doctors from India and the United States. Together, they created neural network consists of 128,000 images and over which they conducted their tests. The test results via the deep learning algorithm were compared with those of another series of images reviewed by a panel of certified ophthalmologists.
It is evident from this comparison that the diagnostic algorithm is around those specialists. The Figure below is derived from a sample of 9963 images shows the test results obtained with the deep learning algorithm of Google and those made by specialized ophthalmologists. The results are represented by black dots for the Google algorithm and by colored dots for medical officers.
Google welcomed the quality of the results but said there is still much work to do. Indeed, in collaboration with specialists of the retina, it seeks to define more robust benchmarks that can be used to assess the performance of their algorithm better.
It states that the interpretation of a 2D retinal photography has been described in their report is only part of a process consisting of several stages. In some situations, the specialists use a 3D imaging technology called Optical Coherence Tomography (OCT) to examine in depth the different layers of the retina. Google says that the application of deep learning in the latter method is underway and their Deepmind colleagues manage the project.
Recall that the deep learning is a set of automatic learning methods to model with a high level of data abstraction through articulated architectures of different nonlinear transformations.