To develop an automated deep learning system for detecting the presence and location of disc hemorrhages in optic disc photographs.

Development and testing of a deep learning algorithm.

Optic disc photos (597 images with at least 1 disc hemorrhage and 1075 images without any disc hemorrhage from 1562 eyes) from 5 institutions were classified by expert graders based on the presence or absence of disc hemorrhage. The images were split into training (n = 1340), validation (n = 167), and test (n = 165) datasets. Two state-of-the-art deep learning algorithms based on either object-level detection or image-level classification were trained on the dataset. These models were compared to one another and against 2 independent glaucoma specialists. We evaluated model performance by the area under the receiver operating characteristic curve (AUC). AUCs were compared with the Hanley–McNeil method.

The object detection model achieved an AUC of 0.936 (95% CI = 0.857-0.964) across all held-out images (n = 165 photographs), which was significantly superior to the image classification model (AUC = 0.845, 95% CI = 0.740-0.912; P = .006). At an operating point selected for high specificity, the model achieved a specificity of 94.3% and a sensitivity of 70.0%, which was statistically indistinguishable from an expert clinician (P = .7). At an operating point selected for high sensitivity, the model achieves a sensitivity of 96.7% and a specificity of 73.3%.

An autonomous object detection model is superior to an image classification model for detecting disc hemorrhages, and performed comparably to 2 clinicians.