Longitudinal Forecasting of Retinal Structure and Function Using a Multimodal StyleGAN-Based Architecture.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate forecasting of Optical Coherence Tomography (OCT) images and best-corrected visual acuity (BCVA) is crucial for patient monitoring and personalized treatment in ophthalmology.
• Generative Adversarial Networks (GANs) show promise for medical image synthesis and clinical outcome prediction.

Purpose of the Study:

• To develop a multimodal GAN for synthesizing OCT images and predicting visual acuity and retinal biomarkers.
• To enhance patient monitoring and personalize treatment planning through accurate forecasting of retinal morphology and functional outcomes.

Main Methods:

• A multimodal GAN, inspired by StyleGAN, incorporated super-resolution, a multi-scale patch discriminator, and temporal attention.
• A hybrid deep-shallow LSTM model predicted logMAR values, and an EfficientNet classifier predicted 16 retinal biomarkers.
• 3-fold patient-level cross-validation ensured subject independence.

Main Results:

• The GAN achieved high performance in OCT forecasting (SSIM: 0.9264, FID: 11.9, PSNR: 38.1 dB).
• The logMAR prediction module yielded an MAE of 0.052, and the biomarker classifier achieved a macro-F1 score of 0.81.
• Patient outcome categorization (Winner, Stabilizer, Loser) based on logMAR change forecasting reached an F1 score of 0.84.

Conclusions:

• The proposed multimodal GAN effectively forecasts retinal morphology and functional outcomes in ophthalmology.
• This approach offers valuable predictive insights for proactive clinical decision-making in retinal health management.
• The study demonstrates the potential of advanced AI models for improving patient care in ophthalmology.