Diagnosing glaucoma on imbalanced data with self-ensemble dual-curriculum learning

Glaucoma diagnosis often suffers from two types of data imbalances: 1) class imbalance, i.e., the non-glaucoma majority cases occupy most of the data; 2) rare cases, i.e., few cases present the uncommon retinopathy e.g., bayoneting or physiologic cupping. This dual-imbalances make glaucoma diagnosis model easy to be dominated by the majority cases but cannot correctly classify the minority and/or rare ones. In this paper, we propose an adaptive re-balancing strategy in the feature space, Self-Ensemble Dual-Curriculum learning (SEDC), to improve the glaucoma diagnosis on imbalanced data by augmenting feature distribution with feature distilling and feature re-weighting. Firstly, the self-ensembling (SEL) is developed to reinforce the discriminative ability of feature representations for the minority class and rare cases by distilling the features learned from the abundant majority cases. Secondly, the dual-curriculum (DCL) is designed to adaptively re-weight the imbalanced data in the feature space to learn a balanced decision function for accurate glaucoma diagnosis. Benefiting from feature distilling and re-weighting, the proposed SEDC fairly represents fundus images, regardless of the majority or rare cases, by augmenting the feature distribution to obtains the optimal decision boundary for accurate glaucoma diagnosis on the imbalanced dataset. Experimental results on three challenging glaucoma datasets show that our SEDC successfully delivers accurate glaucoma diagnosis by the adaptive re-balancing strategy, with the average mean value of Accuracy 0.9712, Sensitivity 0.9520, Specificity 0.9816, AUC 0.9928, F2-score 0.9547. Ablation and comparison studies demonstrate that our method outperforms state-of-the-art methods and traditional re-balancing strategies. The experiment also shows that the adaptive re-balancing strategy proposed in our method provides a more effective training approach with optimal convergence performance. It endows our SEDC a great advantage to handle the disease diagnosis on imbalanced data distribution.