Detecting Retinopathies for Diabetic patients by algorithms of Artificial bee Colonies and Information Augmentation

Abstract

With the aid of retinal fundus image analysis (RFIA), valuable assessments concerning a diabetic patient's vulnerability to blindness are regularly extracted from DR scans. This technology functions as a discreet ally to eye care specialists, also known as ophthalmologists, by quietly supporting their efforts in identifying vision-threatening conditions. In doing so, it alleviates some of the weight associated with such crucial detection tasks. The timely identification and intervention for a sight-threatening complication of diabetes, often referred to as diabetic retinopathy, are crucial to prevent vision loss. However, the associated lesions, which are small and subtle, develop hidden behind the eye's structure, making their monitoring challenging. This study introduces a novel technique for automatically identifying stages of a sight-impairing eye disease (likely referring to diabetic retinopathy) using a sophisticated architecture driven by active deep learning principles. To pinpoint potentially harmful abnormalities within the eye imagery, the system first employs a segmentation process. This process leverages an algorithm inspired by the behaviour of bee colonies, along with a carefully chosen threshold based on the image's intensity distribution analysis. The researchers also created a unique image analysis method, specifically a type of neural network architecture, to automatically grab important characteristics from the segmented retinal areas. This method prioritizes using resources efficiently, making it a more suitable option for certain situations. Our system achieved impressive results in separating different regions within the pictures, with success rates of 93.19% and 91.50% as measured by two different metrics. Additionally, its ability to categorize different stages excelled at 99.20%, outperforming previous methods. This was further confirmed by the consistent improvement in performance on both practice and evaluation data, indicating reliable learning.