Autonomous Screening for Diabetic Macular Edema Using Deep Learning Processing of Retinal Images.

Objective: To develop and validate a deep learning model for diabetic macular edema (DME) detection using color fundus imaging, which is applicable in a diverse, multidevice clinical setting. Design: Evaluation of diagnostic test or technology. Subjects: A deep learning model was trained for DME detection using the EyePACS dataset, consisting of 32 049 images from 15 892 patients. The average age was 55.02%, and 51% of the patients were women. Methods: Data were randomly assigned, by participant, into development (n = 14 246) and validation (n = 1583) sets. Analysis was conducted on the single image, eye, and patient levels. Model performance was evaluated using sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC). Independent validation was further performed on the Indian Diabetic Retinopathy Image Dataset, as well as on new data. Main Outcome Measures: Sensitivity, specificity, and AUC. Results: At the image level, a sensitivity of 0.889 (95% confidence interval [CI]: 0.878, 0.900), a specificity of 0.889 (95% CI: 0.877, 0.900), and an AUC of 0.954 (95% CI: 0.949, 0.959) were achieved. At the eye level, a sensitivity of 0.905 (95% CI: 0.890, 0.920), a specificity of 0.902 (95% CI: 0.890, 0.913), and an AUC of 0.964 (95% CI: 0.958, 0.969) were achieved. At the patient level, a sensitivity of 0.900 (95% CI: 0.879, 0.917), a specificity of 0.900 (95% CI: 0.883, 0.911), and an AUC of 0.962 (95% CI: 0.955, 0.968) were achieved. Conclusions: Diabetic macular edema can be detected from color fundus imaging with high performance on all analysis metrics. Automatic DME detection may simplify screening, leading to more encompassing screening for diabetic patients. Further prospective studies are necessary. Financial Disclosures: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.