Generative Artificial Intelligence for Retinal Image Translation to Improve Glaucoma Screening With Deep Learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Automated glaucoma detection relies on high-quality fundus images.
• Scanning laser ophthalmoscopy (SLO) images are underutilized for glaucoma classification.
• Domain shift between SLO and color fundus (CF) images hinders model performance.

Purpose of the Study:

• To enhance automated glaucoma detection by using generative adversarial networks (GANs).
• To translate underutilized SLO fundus images into synthetic CF photographs.
• To improve the performance of deep learning models for glaucoma classification.

Main Methods:

• A Cycle-Consistent GAN (CycleGAN) framework was employed to generate synthetic CF images from 16,936 SLO images.
• Five deep learning models were trained using real CF, synthetic CF, SLO, and combined datasets.
• Model performance was assessed using area under the operating characteristic curve (AUC) and sensitivity at specificities of 90% and 95%.

Main Results:

• The "GAN+CFP" model, trained on real and synthetic CF images, achieved the highest AUC (0.94) and superior sensitivity (0.83 at 90% specificity, 0.77 at 0.77).
• This model outperformed models trained solely on real CF (AUC=0.89) or SLO images (AUC=0.82).
• The "GAN+CFP" model demonstrated consistent performance across diverse racial and ethnic groups.

Conclusions:

• GANs effectively translate SLO images into synthetic CF photographs, addressing domain shifts.
• This translation increases dataset size and improves the consistency of training data.
• GANs offer a viable method to enhance automated glaucoma detection models.