Abstract / Summary
Abstract Diabetic retinopathy (DR) is a progressive retinal disease and is a major contributor to irreversible blindness in diabetic patients globally. Recognizing retinal lesions early and correctly will be vital to preventing vision impairment and for timely clinical intervention. This study introduces an intelligent optimization-based framework for automated detection of diabetic retinopathy, where Improved Grey Wolf Optimization (IGWO) is used to preprocess the retinal images, Improved Whale Optimization Algorithm (IWOA) to segment lesions in the retinal images, Improved Artificial Bee Colony (IABC) to select features from the retinal images, and an Improved Fuzzy Support Vector Machine (IFSVM) to detect the diabetic retina. First, the images of the retina undergo enhancement for image contrast, reduction of noise and illumination variations with IGWO. The improved images are then segmented applying IWOA for precise estimation of the lesion regions. The image segments of the lesions are then used to extract comprehensive colour, texture and shape features, which are then optimized using IABC to remove redundant information while retaining highly discriminative features. Lastly, the optimized feature subset is classified by using the proposed IFSVM. The preprocessing stage of IGWO had a PSNR of 38.64 dB, SSIM of 0.986, MSE of 0.0042, entropy of 7.82, contrast improvement index of 2.06 and processing time of 0.78 s. The IWOA segmentation algorithm proposed in this paper achieved a Dice coefficient of 98.42%, IoU of 96.89%, sensitivity of 98.51%, specificity of 98.83%, precision of 98.46%, and overall segmentation accuracy of 98.68%. The IABC algorithm has only 82 informative features (with a feature reduction of 65.6%), and only 6.41s to optimize. The proposed Improved Fuzzy Support Vector Machine demonstrated superior classification accuracy (98.55%), precision (98.48%), recall (98.51%), F1-score (98.49%) and AUC (99.02%) compared to the rest of the conventional machine learning and deep learning methods. The robust performance of the proposed framework was also statistically validated by 5-fold cross validation with a mean accuracy of 98.55% with a standard deviation of 0.13% which resulted in statistically significant p-values less than 0.005 and narrow confidence intervals.