RFormer

RFormer restores real clinical fundus images using a Transformer-based generative adversarial network to improve image clarity and support downstream ophthalmic analyses.

Key Features:

• Clinical Dataset (Real Fundus, RF): A benchmark comprising 120 pairs of low- and high-quality (LQ/HQ) fundus images for evaluating real fundus image restoration methods.
• Transformer-based Generative Adversarial Network: A GAN architecture built with Transformer components specifically to restore degradation in clinical fundus images.
• Window-based Self-Attention Block (WSAB): WSAB captures non-local self-similarity and long-range dependencies within fundus images to enhance restoration performance.
• Transformer-based Discriminator: A Transformer-based discriminator enforces visual fidelity and realistic appearance in the restored images.

Scientific Applications:

• Real fundus image restoration: Demonstrated improvements over state-of-the-art (SOTA) methods in restoring real fundus images.
• Benchmark evaluation: Validated and benchmarked on the Real Fundus (RF) dataset of 120 LQ–HQ image pairs.
• Vessel segmentation: Improves performance of vessel segmentation by providing higher-quality input images.
• Optic disc and cup detection: Enhances optic disc and cup detection accuracy through improved image clarity.

Methodology:

Employs a Transformer-based GAN incorporating a Window-based Self-Attention Block (WSAB) and a Transformer-based discriminator, evaluated on the Real Fundus (RF) dataset of 120 low- and high-quality image pairs.