Deep Learning-based Automated System for Identification and Gradation of Severity Levels in Diabetic Retinopathy

Diabetic retinopathy (DR) is a major cause of blindness in adults, but early detection can help to manage the condition effectively. Current methods for automated DR screening mostly focus on finding specific eye lesion. However, a new study takes a different approach in which instead of detecting individual lesions, multiple lesion are considered for DR classification. In the present work an innovative technique is used which extracts statistical and cnn image features from IDRID database to analyze retinal images for Diabetic retinopathy classification.. In this the input retinal fundus image is preprocessed using Gaussian filter and various lesions like microaneurysm, hemorrhages and exudates are segmented from which statistical and cnn based features are extracted for DR detection and classification. These features are further applied to convolution neural network for classification of Diabetic retinopathy which classify it into four classes' i.e. normal, mild, moderate and severe. The results are evaluated using performance parameters like accuracy (Acc), sensitivity (SN), and specificity (SP). The methodology attains 91.42% accuracy, reducing errors by almost 40% compared to conventional lesion-based methods.