Deep Learning for Comprehensive Analysis of Retinal Fundus Images: Detection of Systemic and Ocular Conditions

Area of Science:

• Ophthalmology and Artificial Intelligence
• Medical Image Analysis
• Deep Learning in Healthcare

Background:

• The retina serves as a crucial indicator for both ocular and systemic health.
• There is a growing need for advanced, AI-driven tools for efficient disease screening and risk assessment.
• Evaluating the performance of deep neural networks on diverse ophthalmological datasets is essential for clinical translation.

Purpose of the Study:

• To comprehensively evaluate six state-of-the-art deep neural networks on the Brazilian Multilabel Ophthalmological Dataset (BRSET).
• To assess model performance across seven distinct classification tasks, including diabetes, diabetic retinopathy, hypertension, hypertensive retinopathy, drusen, and sex.
• To explore the explainability of deep learning models using gradient-based saliency maps.

Main Methods:

• Utilized the BRSET dataset, containing 16,266 fundus images with multiple clinical and demographic annotations.
• Evaluated convolutional neural networks and vision transformer architectures on seven classification tasks.
• Employed metrics such as precision, recall, F1-score, accuracy, and Area Under the Curve (AUC) for performance assessment.

Main Results:

• The Swin-L model demonstrated superior performance across most tasks, achieving high AUC and weighted F1-scores for diabetes, diabetic retinopathy (2- and 3-class), hypertension, hypertensive retinopathy, drusen detection, and sex classification.
• Diagnostic performance ranged from excellent to outstanding, highlighting the efficacy of the evaluated models.
• Gradient-based saliency maps provided insights into decision-making processes and identified relevant retinal features.

Conclusions:

• Deep learning models, particularly vision transformers, show significant potential for accurate and interpretable screening of retinal and systemic diseases.
• The findings support the development of AI-powered tools for clinical applications in ophthalmology and broader healthcare.
• The study underscores the clinical utility of AI in identifying key features for disease detection from fundus images.