Deep Ensemble Learning for Retinal Image Classification

PURPOSE: Vision impairment affects 2.2 billion people worldwide, half of which is preventable with early detection and treatment. Currently, automatic screening of ocular pathologies using convolutional neural networks (CNNs) on retinal fundus photographs is limited to a few pathologies. Simultaneous detection of multiple ophthalmic pathologies would increase clinical usability and uptake. METHODS: Two thousand five hundred sixty images were used from the Retinal Fundus Multi-Disease Image Dataset (RFMiD). Models were trained (n = 1920) and validated (n = 640). Five selected CNN architectures were trained to predict the presence of any pathology and categorize the 28 pathologies. All models were trained to minimize asymmetric loss, a modified form of binary cross-entropy. Individual model predictions were averaged to obtain a final ensembled model and assessed for mean area under the receiver-operator characteristic curve (AUROC) for disease screening (healthy versus pathologic image) and classification (AUROC for each class). RESULTS: The ensemble network achieved a disease screening (healthy versus pathologic) AUROC score of 0.9613. The highest single network score was 0.9586 using the SE-ResNeXt architecture. For individual disease classification, the average AUROC score for each class was 0.9295. CONCLUSIONS: Retinal fundus images analyzed by an ensemble of CNNs trained to minimize asymmetric loss were effective in detection and classification of ocular pathologies than individual models. External validation is needed to translate machine learning models to diverse clinical contexts. TRANSLATIONAL RELEVANCE: This study demonstrates the potential benefit of ensemble-based deep learning methods on improving automatic screening and diagnosis of multiple ocular pathologies from fundoscopy imaging.