Update on Google’s diabetic retinopathy project below! Take a look at the original post here.

Photo from Google AI Blog Post “Improving the Effectiveness of Diabetic Retinopathy Models”. Caption: “On the left is a fundus image graded as having proliferative (vision-threatening) DR by an adjudication panel of ophthalmologists (ground truth). On the top right is an illustration of our deep learning model’s predicted scores (“P” = proliferative, the most severe form of DR). On the bottom right is the set of grades given by physicians without assistance (“Unassisted”) and those who saw the model’s predictions (“Grades Only”).”


For the last several years Google has been working with clinics in India to develop a deep learning model to predict the severity of diabetic retinopathy (DR). In more advanced stages, DR can lead to vision loss and requires clinical intervention. Best practice suggests regular screening before the patient’s vision is impaired. In India, there are many patients that need this type of screening and a shortage of eye care specialists. That’s where deep learning comes in!

Back in 2016, Google trained a deep neural network classifier trained on anonymized retinal images to identify ‘Referrable DR’. The results were published in JAMA with the intent to improve the model and performance measurements. The last several years have shown improvements in model evaluation and proving out use cases. Several months ago the first prospective study for the project was published titled “Performance of a Deep-Learning Algorithm vs Manual Grading for Detecting Diabetic Retinopathy in India”.

The study utilized several different neural networks aimed at predicting the severity of DR from fundus images. During the study, additional improvements were made to the model with the addition of higher quality images, hyperparameter tuning and the use of the Inception-v4 neural network architecture.

During the course of the prospective data collection period, we made additional improvements to the model, including tuning the models with adjudicated data as reported by Krause et al. The improvements can be summarized as (1) larger training sets, (2) better hyperparameter exploration (tuning), (3) larger input image resolution, and (4) using the improved Inception-v4 neural network architecture. We graded the images using the model from Krause et al retrospectively at the conclusion of the study.
The study results indicate that the deep learning algorithm is able to automate DR grading to expand screening programs.
This study shows that the automated DR system generalizes to this population of Indian patients in a prospective setting and demonstrates the feasibility of using an automated DR grading system to expand screening programs.