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

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucoma diagnosis models struggle with imbalanced datasets, where non-glaucoma cases dominate and rare conditions are underrepresented.
• This dual imbalance leads to poor classification of minority and rare glaucoma subtypes.

Purpose of the Study:

• To propose an adaptive re-balancing strategy, Self-Ensemble Dual-Curriculum learning (SEDC), to enhance glaucoma diagnosis accuracy on imbalanced data.
• To augment feature distribution for improved representation of both majority and rare cases.

Main Methods:

• SEDC employs self-ensembling (SEL) to distill features from majority cases, reinforcing minority class discriminability.
• Dual-curriculum learning (DCL) adaptively re-weights data in the feature space for a balanced decision function.
• The strategy augments feature distribution via feature distilling and re-weighting.

Main Results:

• SEDC achieved high performance on three challenging glaucoma datasets, with average Accuracy 0.9712, Sensitivity 0.9520, Specificity 0.9816, AUC 0.9928, and F2-score 0.9547.
• The method outperformed state-of-the-art and traditional re-balancing strategies in ablation and comparison studies.
• SEDC demonstrated a more effective training approach with optimal convergence for imbalanced data.

Conclusions:

• SEDC effectively handles dual data imbalances in glaucoma diagnosis by adaptively re-balancing feature distributions.
• The proposed method ensures fair representation of all cases, leading to accurate glaucoma diagnosis.
• SEDC offers a significant advantage for disease diagnosis tasks with imbalanced data.