DIABETES RETINOPATHY SEVERITY GRADING USING MULTI- SCALE IMAGE PYRAMID TECHNIQUES

Abstract

Diabetic Retinopathy (DR) is a frequent diabetic complication that affects blood vessels in the retina, which is a light-sensitive tissue. It is one of the most prevalent causes of vision loss in diabetic individuals, and in older adults. DR severity level diagnosis is important since early therapy can significantly reduce or perhaps prevent vision loss. The majority of today's DR detection and classification algorithms focus on single fixed scale picture properties for prediction, ignoring the multi-scale nature of images. Although items in medical pictures typically exist in diverse forms and sizes, the constant input scale will limit the efficacy of spatial feature integration and will fail to collect scale-dependent details. These problems can be overcome by merging scale- dependent information from the DR image with features from multiple scales. As a result, this study presents a multi-scale feature descriptor technique for DR classification that addresses the disadvantages of a single-scale approach while also improving classification performance. The proposed multi-scale features were created utilising a gaussian pyramid to create a mixture of three different scales. The Local Binary Pattern (LBP) extractor was used to extract the features of the three created scales. The collected features were fed into Decision Tree (DT) and Error-Correcting Output Codes (ECOC) classifiers for training and prediction. The IDRiD dataset was utilized to test the proposed Multi-Scale Feature Descriptor (MSFD) approach. In comparison to the performance of single-scale features, the suggested MSFD technique produced an accuracy of 66.35%, which was higher than the accuracy of 54.80% achieved by the single-scale feature for DT prediction. The new technique also outperformed earlier studies using the IDRiD dataset. The values of the accuracy, precision, recall and f-score obtained imply that the suggested approach can detect and assess DR severity levels.