Detection of Diabetic Retinopathy from Ultra-Wide Field Scanning Laser Ophthalmoscope Images: A Multi-Center Deep-Learning Analysis.

ABSTRACT Purpose To develop a deep-learning (DL) system that can detect referable and vision-threatening diabetic retinopathy (RDR and VTDR) from images obtained on ultra-wide field scanning laser ophthalmoscope (UWF-SLO). Design Observational, cross-sectional study. Subjects A total of 9392 UWF-SLO images of 1903 eyes from 1022 subjects with diabetes from Hong Kong, the UK, India, and Argentina. Methods All images were labeled according to the presence or absence of RDR and the presence or absence of VTDR. Labeling was performed by retina specialists from fundus examination, according to the International Clinical Diabetic Retinopathy Disease Severity Scale. Three convolutional neural networks (ResNet50) were trained with transfer-learning procedure for assessing gradability and identifying VTDR and RDR. External validation was performed on four datasets spanning different geographical regions. Main outcome measures Area under the receiver operating characteristic curve (AUROC), area under the precision-recall curves (AUPRC), sensitivity, specificity, and accuracy of the DL system in gradability assessment and detection of RDR and VTDR. Results For gradability assessment, the system achieved an AUROC of 0.923 (95% CI, 0.892-0.947), sensitivity of 86.5% (77.6-92.8) and specificity of 82.1% (77.3-86.2) for the primary validation dataset, and >0.82 AUROCs, >79.6% sensitivity and >70.4% specificity for the geographical external validation datasets. For detecting RDR and VTDR, the AUROCs were 0.981 (0.977-0.984) and 0.966 (0.961-0.971), with sensitivities of 94.9% (92.3-97.9) and 87.2% (81.5-91.6), specificities of 95.1% (90.6-97.9) and 95.8% (93.3-97.6), and positive predictive values (PPVs) of 98.0% (96.1-99.0) and 91.1% (86.3-94.3) for the primary validation dataset. The AUROCs and accuracies for detecting both RDR and VTDR were >0.9 and >80% for the geographical external validation datasets. In addition, the AUPRCs were >0.9, and sensitivities, specificities, and PPVs were >80% for the geographical external validation datasets for RDR and VTDR detection. Conclusion The excellent performance achieved with this DL system for automated image quality assessment and detection of RDR and VTDR in UWF-SLO images highlights its potential as an efficient and effective diabetic retinopathy screening tool.