Deep-RPD-Net: A 3D Deep Network for Detection of Reticular Pseudodrusen on Optical Coherence Tomography Scans

Abstract

ObjectiveTo propose Deep-RPD-Net, a 3D deep learning network with semi-supervised learning (SSL) for the detection of reticular pseudodrusen (RPD) on spectral-domain optical coherence tomography (OCT) scans, explain its decision-making, and compare it with baseline methods. DesignDeep learning model development. Participants315 participants from the Age-Related Eye Disease Study 2 Ancillary OCT Study (AREDS2) and 161 participants from the Dark Adaptation in Age-related Macular Degeneration Study (DAAMD). MethodsTwo datasets comprising of 1,304 (826 labeled) and 1,479 (1366 labeled) OCT scans were used to develop and evaluate Deep-RPD-Net and baseline models. AREDS2 RPD labels were transferred from fundus autofluorescence images, which were captured at the same visit for OCT scans and DAAMD RPD labels were obtained from the Wisconsin reading center. The datasets were divided into 70%, 10%, and 20% at the participant level for training, validation, and test sets, respectively, for the baseline model. Then, SSL was used with the unlabeled OCT scans to improve the trained model. The performance of Deep-RPD-Net was compared to that of three retina specialists on a subset of 50 OCT scans for each dataset. En face and B-scan heatmaps of all networks were visualized and graded on 25 OCT scans with positive labels, using a scale of 1-4, to explore the models' decision-making. Main Outcome MeasuresAccuracy, area under ROC curve (AUROC). ResultsDeep-RPD-Net achieved the highest performance metrics, with accuracy and AUROC of 0.81 (95% Confidence Interval “CI”: 0.76-0.87) and 0.91 (95% CI: 0.86-0.95) on the AREDS2 dataset and 0.80 (95% CI: 0.75-0.84) and 0.86 (95% CI: 0.79-0.91) on the DAAMD dataset. On the subjective test, it achieved accuracy of 0.84 compared to 0.76 for the most accurate retina specialist on the AREDS2 dataset and 0.82 compared to 0.84 on the DAAMD dataset. It also achieved the highest visualization grades, of 3.26 and 3.32 for en face and B-scan heatmaps, respectively. ConclusionsDeep-RPD-Net was able to detect RPD accurately from OCT scans. The visualizations of Deep-RPD-Net were the most explainable to the retina specialist with the highest accuracy. The code and pretrained models are publicly available at https://github.com/ncbi-nlp/Deep-RPD-Net.