Effective automatic classification methods via deep learning for myopic maculopathy.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Pathologic myopia (PM) and myopic maculopathy (MM) cause significant visual impairment, particularly in East Asia.
• Early detection and classification of fundus lesions are crucial for managing PM.
• Automated diagnostic tools are needed due to the limitations of manual analysis.

Purpose of the Study:

• To develop and evaluate a deep learning system for classifying five types of MM from color fundus photographs.
• To assess the performance of various deep learning architectures and an ensemble approach.

Main Methods:

• Utilized ResNet50, EfficientNet-B0, Vision Transformer (ViT), CLIP, and RETFound architectures.
• Employed an ensemble learning approach with weighted voting for enhanced performance.
• Trained and evaluated models on 2,159 annotated fundus images.

Main Results:

• The ensemble model achieved high accuracy (95.4%), sensitivity (95.4%), specificity (98.9%), F1-Score (95.3%), Kappa (0.976), and AUC (0.995).
• The ensemble method demonstrated robustness and superior generalization in classifying complex MM lesions.
• Performance metrics significantly outperformed individual models.

Conclusions:

• The deep learning ensemble system enhances accuracy and reliability in MM classification.
• This system can assist ophthalmologists in early detection and precise diagnosis of MM.
• Future work includes dataset expansion and algorithm optimization for broader applicability.