Feature fusion and optimization integrated refined deep residual network for diabetic retinopathy severity classification using fundus image

Diabetic retinopathy (DR) is a lasting diabetic condition that causes vision damage, and if left untreated, the progression of DR may result in permanent blindness. Therefore, it is obligatory to identify the pathological changes in the retina to detect the severity level of DR. Manual observation of retinal disorders makes the detection process more intricate, tedious, and error prone due to the prevalence of subtle pathologies in the retina. Deep learning-enabled optimization techniques have recently made promising progress in DR detection and classification tasks. Therefore, the current work suggests an enhanced classification approach based on metaheuristic optimization for severity grading in fundus images. Initially, a pre-processing step eradicates noise and enhances the image contrast. Following the optic disc segmentation using dilated convolution-based spatial attention U-Net, blood vessels are segmented using an entropy-based hybrid technique. Subsequently, the lesion area is segmented using the attention-based fusion U-Net model with a weighted focal loss. Then, features are extracted and fused using a two-layer fusion network. Based on the fused features, the DR stages are classified with refined deep residual network (RDRN) integrating a squeeze-excitation module and tunicate swarm spider monkey optimization (TSSMO) algorithm. The fundus images from the DIARETDB1 and DIARETDB0 datasets were used to evaluate the performance of the suggested method using different measures. The comparative performance reveals the efficacy of the developed method in classifying the DR severity stages.