Deep Learning-Based Detection of Reticular Pseudodrusen in Age-Related Macular Degeneration

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Reticular pseudodrusen (RPD) are a significant phenotype associated with vision loss in age-related macular degeneration (AMD).
• Accurate detection of RPD is crucial for understanding and managing AMD progression.
• Current diagnostic methods may benefit from advanced computational approaches.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for the automated detection of RPD.
• To assess the performance of the DL model against expert human graders using optical coherence tomography (OCT) scans.
• To provide a tool that could support clinical decision-making in AMD.

Main Methods:

• Manual segmentation of RPD in 9800 OCT B-scans from a multicentre trial.
• Development of a DL model for RPD instance segmentation.
• External validation of the DL model's performance on OCT volumes from multiple datasets, comparing it to retinal specialists.

Main Results:

• The DL model demonstrated superior agreement (Dice Similarity Coefficient [DSC] = 0.76) compared to inter-specialist agreement (DSC = 0.68) in an internal dataset.
• In external datasets, the DL model achieved performance comparable to retinal specialists in detecting RPD (Area Under the ROC Curve [AUC] = 0.94-0.96).
• The model's performance was robust across diverse external test datasets.

Conclusions:

• A DL model for automatic RPD detection has been developed with expert-level performance.
• This AI tool shows potential for clinical application in supporting the management of AMD.
• The publicly available model encourages further research into the RPD phenotype in AMD.