Neural search guided CNN–transformer for diabetic retinopathy prediction from fundus images

Diabetic retinopathy (DR), a leading cause of vision loss among diabetic patients, is a critical and progressive retinal disorder. The early identification of various abnormalities associated with DR is significant for the early intervention of blindness. This paper presents a novel enhancement-based ensemble framework for automatic gradation of DR. The proposed work highlights the significance of Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) in enhancing the image clarity and structural details, followed by contrast optimization using Ben Graham’s technique combined with Contrast-Limited Adaptive Histogram Equalization (CLAHE), highlighting superior local contrast enhancement compared to traditional CLAHE. Feature extraction of the enhanced dataset is achieved by the Neural Architecture Search Network (NASNet) in capturing minute complex retinal patterns and micro-lesions. The extracted minute features are fed to an ensemble of ConvNeXt and Swin Transformer architectures classifier. The ConvNeXt harnesses the hierarchical convolutional features, integrating the captures of long-range contextual dependencies by Swin Transformer. The ensemble model ensures the enhancement of both local and global structure analysis. The proposed model was validated on three benchmark retinal fundus datasets: APTOS 2019, DDR, and DIARETDB1. The proposed model achieved classification accuracies of 97.52%, 98.02%, and 98.55%, with corresponding precisions of 92.64%, 97.85%, and 95.56%, recalls of 93.89%, 97.60%, and 96.14%, and F1-scores of 93.23%, 97.70%, and 95.82%. The performance metrics highlight the model s robustness, generalization ability, and clinical reliability for automated DR detection and grading.