Identifying Retinal Features Using a Self-Configuring CNN for Clinical Intervention.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal diseases are a major cause of blindness globally.
• Optical coherence tomography (OCT) is crucial for diagnosing retinal conditions.
• Limited annotated OCT datasets hinder AI development for retinal disease diagnosis.

Purpose of the Study:

• To address the scarcity of annotated 3D OCT datasets.
• To introduce OCTAVE, a comprehensive 3D OCT dataset with pixel-level annotations.
• To provide annotated public OCT datasets for external validation.

Main Methods:

• Developed the OCTAVE dataset with 198 training and 221 validation volumes.
• Annotated anatomic and pathological structures at the pixel level.
• Trained a deep learning segmentation model (nnU-Net) and evaluated on external datasets.

Main Results:

• The OCTAVE dataset contains 198 OCT volumes (3762 B-scans) for training.
• 221 OCT volumes (4109 B-scans) were used for external validation.
• The deep learning model achieved clinically significant performance across all retinal structures.

Conclusions:

• Demonstrated robust segmentation performance and generalizability of AI models.
• OCTAVE dataset supports AI tool development for precise retinal disease detection and monitoring.
• This resource can improve clinical outcomes and advance AI-driven retinal disease management.