Automated Multi-Class Classification of Retinal Pathologies: A Deep Learning Approach to Unified Ophthalmic...

Area of Science:

• Ophthalmic artificial intelligence (AI)
• Deep learning in medical imaging
• Retinal pathology classification

Background:

• Current ophthalmic AI models are often disease-specific, limiting their clinical utility for differential diagnosis.
• A unified framework is needed to address the complexity of multiple retinal conditions.
• This study introduces a novel approach to multi-class retinal pathology classification.

Purpose of the Study:

• To develop and validate a unified deep learning framework for multi-class classification of retinal pathologies.
• To create a comprehensive AI screening tool for various eye conditions from fundus photographs.
• To move beyond single-disease AI models in ophthalmology.

Main Methods:

• A curated dataset of 1841 fundus images across nine classes was used.
• Data augmentation techniques were applied to address class imbalance.
• Three CNN architectures (ResNet-152, EfficientNetV2, YOLOv11-based) were evaluated using transfer learning.

Main Results:

• The YOLOv11-based classifier achieved superior performance.
• The model attained an accuracy of 0.861 and a macro-F1 score of 0.861 on the test set.
• A validation AUC of 0.961 was achieved, significantly outperforming ResNet-152 and EfficientNetV2.

Conclusions:

• A unified deep learning framework using a YOLOv11 backbone can accurately classify nine distinct retinal conditions.
• This holistic AI approach surpasses single-disease limitations.
• The framework shows promise as a comprehensive screening tool to aid clinical decision-making in ophthalmology.