Self-adaptive stacking ensemble approach with attention based deep neural network models for diabetic retinopathy severity prediction

Diabetic Retinopathy (DR) is a chronic eye disease that is common in people who have had diabetes for a long time. If the disease is not treated during the early stages, it leads to complete vision loss, which can be avoided by treating in the early stages. The development of effective tools is critical for carrying out large-scale diagnostics at low cost while avoiding human bias. In this work, we propose a self-adaptive ensemble approach for retinopathy severity grading by stacking multiple dual attention based approaches. The proposed dual attention model leverages two distinct attention mechanisms. The model can focus on lesion-specific regions with the first level of attention, while the second level of attention allows it to learn correlations between spatial descriptors. The proposed model effectively predicts the severity level of retinopathy with dual levels of attention. We also present a self-adaptive meta learner for effectively stacking multiple dual attention models. Experimental studies on the benchmark APTOS 2019 dataset reveals that the proposed approach outperforms several existing models by achieving an accuracy of 86.22%. The proposed model exhibits generalization not only in terms of accuracy but also in terms of other evaluation measures, and achieves a quadratic kappa score of 89.65% and an AUC score of 96.47%.