Automated Identification of Diabetic Retinopathy Using Deep Learning

• Published in: Ophthalmology (Rochester, Minn.) Vol. 124; no. 7; pp. 962 - 969
• Main Authors: Gargeya, Rishab, Leng, Theodore, MD, MS
• Format: Journal Article
• Language: English
• Published: United States Elsevier Inc 01.07.2017
• Subjects: Algorithms; Diabetic Retinopathy - diagnosis; Diagnostic Techniques, Ophthalmological; Humans; Image Interpretation, Computer-Assisted - methods; Machine Learning; Neural Networks (Computer); Ophthalmologists; Ophthalmology; Reproducibility of Results; ROC Curve; area under the receiver operating characteristic curve; diabetic retinopathy; DR; AUC
• ISSN: 0161-6420; 1549-4713
• DOI: 10.1016/j.ophtha.2017.02.008

Purpose Diabetic retinopathy (DR) is one of the leading causes of preventable blindness globally. Performing retinal screening examinations on all diabetic patients is an unmet need, and there are many undiagnosed and untreated cases of DR. The objective of this study was to develop robust diagnostic technology to automate DR screening. Referral of eyes with DR to an ophthalmologist for further evaluation and treatment would aid in reducing the rate of vision loss, enabling timely and accurate diagnoses. Design We developed and evaluated a data-driven deep learning algorithm as a novel diagnostic tool for automated DR detection. The algorithm processed color fundus images and classified them as healthy (no retinopathy) or having DR, identifying relevant cases for medical referral. Methods A total of 75 137 publicly available fundus images from diabetic patients were used to train and test an artificial intelligence model to differentiate healthy fundi from those with DR. A panel of retinal specialists determined the ground truth for our data set before experimentation. We also tested our model using the public MESSIDOR 2 and E-Ophtha databases for external validation. Information learned in our automated method was visualized readily through an automatically generated abnormality heatmap, highlighting subregions within each input fundus image for further clinical review. Main Outcome Measures We used area under the receiver operating characteristic curve (AUC) as a metric to measure the precision–recall trade-off of our algorithm, reporting associated sensitivity and specificity metrics on the receiver operating characteristic curve. Results Our model achieved a 0.97 AUC with a 94% and 98% sensitivity and specificity, respectively, on 5-fold cross-validation using our local data set. Testing against the independent MESSIDOR 2 and E-Ophtha databases achieved a 0.94 and 0.95 AUC score, respectively. Conclusions A fully data-driven artificial intelligence–based grading algorithm can be used to screen fundus photographs obtained from diabetic patients and to identify, with high reliability, which cases should be referred to an ophthalmologist for further evaluation and treatment. The implementation of such an algorithm on a global basis could reduce drastically the rate of vision loss attributed to DR.