OCTA-based AMD Stage Grading Enhancement via Class-Conditioned Style Transfer.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Optical Coherence Tomography Angiography (OCTA) is crucial for visualizing retinal vasculature in age-related macular degeneration (AMD).
• Deep neural networks (DNNs) show potential for AMD diagnosis using OCTA, but data scarcity is a significant limitation.

Purpose of the Study:

• To investigate the efficacy of DNN classifiers trained on OCTA images for AMD diagnosis.
• To address limited OCTA data by employing style transfer techniques.
• To enhance OCTA image datasets for improved classification accuracy.

Main Methods:

• Combined OCTA data from Heidelberg and Optovue instruments.
• Utilized CycleGAN for domain adaptation between OCTA instruments.
• Proposed class-conditioned CycleGAN with supervised/unsupervised constraints for enhanced feature learning.
• Trained DNN classifiers on augmented OCTA datasets.

Main Results:

• Style transfer using CycleGAN enriched OCTA datasets.
• Class-conditioned CycleGAN effectively improved data diversity.
• DNN classification accuracy was significantly elevated in both OCTA domains.
• The proposed method demonstrated effectiveness in enhancing AMD diagnosis.

Conclusions:

• Class-conditioned CycleGAN is a viable strategy for augmenting OCTA data.
• This approach effectively improves DNN-based AMD diagnostic accuracy.
• The study highlights the potential of AI and OCTA for non-invasive AMD detection.