OPTNet: Optimized Pixel-Transformer Model for Adaptive Retinal Fundus Image Enhancement

Retinal low-quality images present significant challenges for accurate diagnosis and monitoring of eye diseases by obscuring critical anatomical features and reducing analytical precision. This study introduces OPTNet, an optimized pixel-wise transformer model designed to efficiently enhance degraded or low-quality retinal images. The proposed approach consists of three main stages: 1) pre-processing to standardize image dimensions and balance color channels, 2) model development, in which a lightweight ANN-based feature extractor learns retinal structures and generates self-measured quality labels, and 3) pixel-level transformation guided by these predicted labels to perform localized enhancement. The performance of OPTNet was evaluated using statistical metrics across various architectures during training and testing, and benchmarked on six public retinal datasets: DRIVE, CHASE-DB1, HRF, DRHAGIS, FIRE, and FIVES. A comprehensive evaluation was conducted using both full-reference and no-reference quality assessment (QA) metrics, supported by qualitative analysis. OPTNet achieved competitive results including a 21.3% improvement in NIQE and 17.8% reduction in BRISQUE compared with existing methods. The final scores included SSIM (0.1925), VIF (0.1911), BIF (1.3018), EME (11.1704), NIQE (4.0730), and BRISQUE (30.3003), indicating perceptual and structural enhancement. Additionally, it effectively preserved brightness and anatomical fidelity while minimizing distortion (CD = 0.4214), blur (0.0889), and artifacts (0.2903). In conclusion, OPTNet outperforms state-of-the-art enhancement techniques by striking a robust balance between quality improvement and artifact suppression, demonstrating its strong potential for integration into clinical ophthalmic diagnostic pipelines.