Generative adversarial networks (GAN) for pre-dilation retinal photograph quality enhancement

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Retinal image quality is crucial for diagnostic AI performance.
• Pupil dilation improves image quality but is not always feasible in real-world settings.
• Suboptimal image quality in undilated retinas challenges AI deployment.

Purpose of the Study:

• To fine-tune a CofeNet model for enhancing the quality of undilated retinal images.
• To evaluate the effectiveness of the model-enhanced images compared to ground truth dilated images.
• To assess the potential of AI-driven image enhancement as an alternative to pupil dilation.

Main Methods:

• A generative CofeNet model was fine-tuned using 313 paired undilated and dilated retinal images.
• Pixel-level alignment was performed for accurate model training.
• Model performance was evaluated using peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and image gradability assessments by multiple graders.

Main Results:

• PSNR and SSIM significantly improved in model-enhanced images compared to undilated images (p < 0.001).
• Image gradability increased from 41% to 76% after enhancement with the CofeNet model.
• While image quality improved, inter-rater agreement for gradability showed a decrease.

Conclusions:

• The CofeNet model successfully enhanced undilated retinal image quality and structural similarity to dilated images.
• AI-based image enhancement shows promise as an alternative to pupil dilation in clinical settings.
• Further validation is required to establish the clinical utility of this AI approach.