Glaucoma is a leading cause of irreversible blindness, and early diagnosis is essential to prevent disease progression. Deep learning models using color fundus images have achieved performance comparable to ophthalmologists, yet cross-institutional deployment remains challenging due to variations in imaging devices, acquisition protocols, and patient populations. In federated learning settings, such heterogeneity degrades performance and introduces disparities in sensitivity across institutions, raising fairness concerns. We propose a preprocessing strategy that suppresses non-diagnostic variation while emphasizing clinically relevant regions, thereby reducing cross-institutional variability. We evaluate the method on six multi-institutional fundus datasets from Taiwan, India, China, and Bangladesh using FedAvg with EfficientNetV2-B0. Results show improved fairness, with Fairness Difference increasing (0.7825 → 0.9247) and AAOD reduced by 53.7%. The method is model-agnostic and can be integrated into existing federated learning pipelines as a lightweight module.