A phase-aware cross-modal framework is presented that synthesizes UWF_FA from non-invasive UWF_RI for diabetic retinopathy (DR) stratification. A curated cohort of 1198 patients (2915 UWF_RI and 17,854 UWF_FA images) with strict registration quality supports training across three angiographic phases (initial, mid, final). The generator is based on a modified pix2pixHD with an added Gradient Variance Loss to better preserve microvasculature, and is evaluated using MAE, PSNR, SSIM, and MS-SSIM on held-out pairs. Quantitatively, the mid phase achieves the lowest MAE (98.76 $\pm$ 42.67), while SSIM remains high across phases. Expert review shows substantial agreement (Cohen’s $\kappa$ = 0.78–0.82) and Turing-style misclassification of 50%–70% of synthetic images as real, indicating strong perceptual realism. For downstream DR stratification, fusing multi-phase synthetic UWF_FA with UWF_RI in a Swin Transformer classifier yields significant gains over a UWF_RI-only baseline, with the full-phase setting (Set D) reaching AUC = 0.910 and accuracy = 0.829. These results support synthetic UWF_FA as a scalable, non-invasive complement to dye-based angiography that enhances screening accuracy while avoiding injection-related risks.

Metaheuristic optimization methods are iterative search processes that aim to efficiently solve complex optimization problems. These basically find the solution space very efficiently, often without utilizing the gradient information, and are inspired by the bio-inspired and socially motivated heuristics. Metaheuristic optimization algorithms are increasingly applied to complex feature selection problems in high-dimensional medical datasets. Among these, Teaching-Learning-Based optimization (TLBO) has proven effective for continuous design tasks by balancing exploration and exploitation phases. However, its binary version (BTLBO) suffers from limited exploitation ability, often converging prematurely or getting trapped in local optima, particularly when applied to discrete feature selection tasks. Previous studies reported that BTLBO yields lower classification accuracy and higher feature subset variance compared to other hybrid methods in benchmark tests, motivating the development of hybrid approaches. This study proposes a novel hybrid algorithm, BTLBO-Cheetah Optimizer (BTLBO-CO), which integrates the global exploration strength of BTLBO with the local exploitation efficiency of the Cheetah Optimization (CO) algorithm. The objective is to enhance the feature selection process for cancer classification tasks involving high-dimensional data. The proposed BTLBO-CO algorithm was evaluated on six benchmark cancer datasets: 11 tumors (T), Lung Cancer (LUC), Leukemia (LEU), Small Round Blue Cell Tumor or SRBCT (SR), Diffuse Large B-cell Lymphoma or DLBCL (DL), and Prostate Tumor (PT). The results demonstrate superior classification accuracy across all six datasets, achieving 93.71%, 96.12%, 98.13%, 97.11%, 98.44%, and 98.84%, respectively. These results validate the effectiveness of the hybrid approach in addressing diverse feature selection challenges using a Support Vector Machine (SVM) classifier.