Deep learning approach for diabetic retinopathy screening

Purpose Diabetic retinopathy (DR) is the major cause of blindness in the working‐age population. With an increasing number of diabetic patients worldwide, automated screening tools become indispensable. Recent progress in machine learning and image analysis enables efficient automated screening. Methods DreamUp Vision uses state‐of the art technology based on deep‐learning. Our algorithm was trained on over 70,000 labeled retinal images. Images were graded by ophthalmologists as follows: 0 (no retinopathy), 1 (mild non proliferative DR), 2 (moderate non proliferative DR), 3 (severe non proliferative DR) and 4 (proliferative retinopathy). Each patient in the dataset is represented by two images of left and right eyes. Grading is done for each eye image separately. Our algorithm performs quick and reliable detection of anomalies in retinal images, diagnoses their stage of diabetic retinopathy and provides the location of the anomalies detected in the pictures. We consider a patient as referable if the DR stage is between 2 and 4, otherwise we consider the patient as non‐referable. We evaluate our model on over 10,000 fundus images from 5,000 patients taken from the Kaggle DR Detection Challenge dataset, provided by California Healthcare Foundation. Results Our algorithm achieves an area under the receiver operating characteristic curve AUROC of 0.946 with 96.2% sensitivity (95% CI: 95.8–96.5) and 66.6% specificity (95% CI: 65.7–67.5) for identifying referable DR on the Kaggle dataset. Conclusions The performances we have obtained enable a reliable automated DR screening. As the amount of available labeled data grows and given our technology's ability to learn from labeled images, we believe that significant performance improvement can be achieved. The same process can be applied to the detection of other eye diseases as well.