Automated Identification of Referable Retinal Pathology in Teleophthalmology Setting

AI Generated Summary*

Retrospective evaluation of a deep learning system for flagging referable retinal pathology from paired optical coherence tomography and color fundus photographs obtained in primary care, including ungradable images. The dataset comprised 1148 image pairs from 647 patients with diabetes; retina specialists labeled 924 eyes pathology negative and 224 positive. A dual-input convolutional neural network was trained with an alternate gradient descent scheme that limits the influence of uninterpretable inputs rather than discarding them. On 114 held-out eyes, accuracy reached 88.60%, sensitivity 87.72%, specificity 89.47%, and area under the ROC curve 92.74%, with accuracy and AUC exceeding three baseline approaches. Removing ungradable images from testing left performance unchanged, whereas two baselines improved markedly. The model does not distinguish specific diseases.

*This summary was generated by AI and is published unedited. Oku does not alter these summaries. It may contain errors or omissions and is provided for general informational purposes only. Oku does not guarantee its accuracy, completeness, or reliability. For authoritative information, please refer to the original, peer-reviewed article.