Efficient Ensemble Learning with Curriculum-Based Masked Autoencoders for Retinal OCT Classification.

Area of Science:

• Ophthalmic Imaging Analysis
• Medical Image Classification
• Deep Learning in Ophthalmology

Background:

• Retinal optical coherence tomography (OCT) is crucial for diagnosing eye diseases.
• Developing accurate multiclass classifiers for OCT is hindered by limited labeled data and high computational costs of self-supervised pretraining.
• Existing methods struggle with efficiency and performance in complex OCT classification tasks.

Purpose of the Study:

• To introduce a curriculum-based self-supervised framework (CurriMAE) for improved representation learning in OCT classification.
• To reduce the computational burden associated with self-supervised pretraining for OCT analysis.
• To enhance the performance of multiclass classifiers for retinal OCT images.

Main Methods:

• Developed two ensemble strategies, CurriMAE-Soup and CurriMAE-Greedy, utilizing progressive masked autoencoder (MAE) pretraining.
• Employed a curriculum-guided MAE pretraining run with progressive masking ratios, avoiding repeated training.
• Evaluated methods on Kermany and OCTDL datasets, comparing against standard MAE and supervised baselines (ResNet-34, ViT-S).

Main Results:

• Both CurriMAE methods significantly outperformed standard MAE and supervised baselines on the OCTDL dataset (seven retinal classes).
• CurriMAE-Greedy achieved the highest performance: 0.995 AUC and 93.32% accuracy.
• CurriMAE-Soup offered competitive accuracy with substantially lower inference complexity and reduced model storage.

Conclusions:

• The proposed curriculum-based self-supervised ensemble framework (CurriMAE) is an effective and resource-efficient solution for multiclass retinal OCT classification.
• CurriMAE methods demonstrate high performance with reduced computational costs through progressive masking and model fusion.
• This framework shows significant potential for real-world ophthalmic imaging applications with limited data and computational resources.