Multi-class classification of central and non-central geographic atrophy using Optical Coherence Tomography

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• Geographic atrophy (GA) is an advanced form of age-related macular degeneration (AMD) that leads to irreversible vision loss.
• Accurate classification of GA subtypes is crucial for understanding disease progression and developing targeted therapies.
• Optical Coherence Tomography (OCT) is a key imaging modality for visualizing retinal structures and diagnosing macular diseases.

Purpose of the Study:

• To develop and validate deep learning (DL) models for classifying geographic atrophy (GA) subtypes using Optical Coherence Tomography (OCT) scans.
• To evaluate the performance of different DL architectures (ResNet50, MobileNetV2, Vision Transformer) across four clinical classification tasks.
• To assess the impact of two experimental approaches (using all B-scans vs. selective foveal B-scans) on model accuracy.

Main Methods:

• A retrospective study utilizing 455 OCT volumes from 104 patients, including Central GA (CGA), Non-Central GA (NCGA), and no GA (NGA) cases.
• Implementation of ResNet50, MobileNetV2, and Vision Transformer (ViT-B/16) architectures with transfer learning and data augmentation.
• Models were trained and tested using patient-level data splitting (70:15:15), evaluating performance with AUC-ROC and accuracy metrics.

Main Results:

• The Vision Transformer (ViT-B/16) consistently outperformed other architectures across all classification tasks.
• ViT-B/16 achieved high accuracy in distinguishing GA subtypes (e.g., CGA vs. NCGA: AUC-ROC 0.728, accuracy 0.831) and GA from NGA (AUC-ROC 0.950, accuracy 0.873) using selective B-scans.
• All models demonstrated excellent performance in differentiating GA from other AMD forms (AUC-ROC > 0.998).

Conclusions:

• Deep learning models, particularly ViT-B/16, can successfully classify GA subtypes from OCT scans with clinically relevant accuracy.
• The ViT-B/16 architecture excels due to its ability to capture spatial relationships, improving diagnostic accuracy.
• Utilizing selective B-scans containing foveal regions enhances diagnostic accuracy while reducing computational load, aligning with clinical workflows.