Mixing Synthetic and Real Images Improves Artificial Intelligence-Based Detection of the Pupil, Iris, and Sclera: A Cross-Domain Validation Study.
Authors
Affiliations (4)
Affiliations (4)
- Ophthalmology, All India Institute of Medical Sciences, New Delhi, New Delhi, IND.
- Ophthalmology, National Cancer Institute, All India Institute of Medical Sciences, New Delhi, Jhajjar Campus, New Delhi, IND.
- Volunteer, Manthan Eye Healthcare Foundation, Gurgaon, IND.
- Medical Physics, All India Institute of Medical Sciences, New Delhi, New Delhi, IND.
Abstract
Background This study aimed to evaluate whether mixing synthetic and real eye images for artificial intelligence (AI) training improves cross-domain segmentation of the sclera, iris, and pupil compared to single-domain datasets. Methodology Four Roboflow 3.0 instance segmentation models were trained: (1) 100 AI-generated images, (2) 100 real images, (3) 50:50 mixed, n = 100, and (4) 50:50 mixed, n = 200. All were tested on five AI-generated and five real eye images. Detection accuracy was calculated per structure. Performance was compared using paired t-tests and two-way analysis of variance. Results Mixed models eliminated domain-specific failures. Pupil accuracy showed a significant training × test domain interaction (p = 0.003), with single-domain models failing on opposite domains: AI-trained 0% on one real image; real-trained 50% on one AI image. The 100-Mix model achieved 88.5% ± 5.9% pupil accuracy with no failures, with a standard deviation of <6% versus 29.8% for AI-only. Doubling mixed data to 200 images gave no added benefit (p = 0.95). Conclusions Hybrid training with 50:50 synthetic-real images achieves robust, domain-stable detection of the pupil, iris, and sclera. Mixed datasets, not larger datasets, are key for clinically deployable ocular AI.