Deep-Reticular Pseudodrusen-Net: A 3-Dimensional Deep Network for Detection of Reticular Pseudodrusen on OCT Scans

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Reticular pseudodrusen (RPD) are associated with age-related macular degeneration.
• Accurate detection of RPD from optical coherence tomography (OCT) scans is crucial for diagnosis and monitoring.
• Current detection methods may lack precision and explainability.

Purpose of the Study:

• To introduce Deep-RPD-Net, a 3D deep learning network utilizing semisupervised learning (SSL).
• To evaluate Deep-RPD-Net's efficacy in detecting RPD on spectral-domain OCT (SD-OCT) scans.
• To compare Deep-RPD-Net's performance and decision-making explainability against baseline models and human experts.

Main Methods:

• Development of Deep-RPD-Net using two large OCT datasets (AREDS2 and DAAMD) with labeled and unlabeled scans.
• Application of SSL to enhance model performance by leveraging unlabeled data.
• Comparative analysis against baseline models and retina specialists, incorporating visualization techniques (heatmaps) for explainability assessment.

Main Results:

• Deep-RPD-Net demonstrated superior performance, achieving high accuracy and AUROC values on both datasets.
• The model's accuracy in subjective testing (0.84 on AREDS2, 0.82 on DAAMD) surpassed that of the most accurate retina specialist.
• Visualization heatmaps of Deep-RPD-Net received higher grading for explainability compared to other networks.

Conclusions:

• Deep-RPD-Net effectively detects RPD from OCT scans with high accuracy.
• The model's transparent decision-making process, visualized through heatmaps, enhances its clinical utility.
• The developed model and code are publicly accessible, promoting further research and application.