Weakly-Supervised Detection and Quantification of Vitreous Opacities in OCT Images Using Synthetic Data Augmentation

Vitreous opacities colloquially known as eye floaters represent a clinically underserved visual condition affecting an estimated 25% of the global population. Despite causing significant impairment in quality of life, no standardized objective measurement tool exists in clinical practice; assessment relies entirely on subjective patient reporting, limiting diagnostic precision and treatment evaluation. This paper presents a proof-of-concept computational framework for the weakly-supervised detection and quantification of vitreous opacities in optical coherence tomography (OCT) images. We introduce a physics-informed synthetic floater generation method that simulates realistic opacity patterns including hypo-reflective collagen aggregates and posterior acoustic shadowing directly within OCT B-scans, circumventing the fundamental obstacle of absent annotated clinical datasets. A U-Net architecture with a ResNet34 encoder pretrained on ImageNet is trained on this augmented synthetic dataset using a combined Dice-BCE loss. On the synthetic validation set the model achieves a best Dice coefficient of 0.9038, with a mean Dice of 0.5936 ± 0.3974 and mean IoU of 0.5269 ± 0.3738, demonstrating feasibility on simulated data. We further introduce the Floater Severity Index (FSI), a novel four-component composite metric combining opacity area coverage, intensity-based density, foveal proximity weighting, and fragment count into a clinically interpretable 0-100 score. To our knowledge, this is the first computational framework specifically designed for objective vitreous opacity detection and quantification from OCT images.