A Patch-based Transformer Framework for Robust Cross-Dataset Retinal Vessel Segmentation

Retinal vessel segmentation plays a crucial role in automated diagnosis systems for ophthalmological diseases. However, achieving high accuracy across different datasets remains challenging due to variations in imaging conditions, noise, and vessel complexity. This work proposes a transformer-enhanced deep learning framework that combines patchbased learning, Gaussian-weighted reconstruction, and test-time augmentation (TTA) to improve both accuracy and robustness. The model is trained on the DRIVE dataset and evaluated on unseen datasets such as CHASE_DB1, STARE, and HRF without retraining. The results demonstrate that the proposed approach not only achieves competitive segmentation accuracy but also maintains consistent performance across datasets, highlighting its ability to generalize well in real-world clinical scenarios.