Options
Automatic grading of non-proliferative diabetic retinopathy
ISSN
24464732
Date Issued
2023-09-01
Author(s)
Suman, Supriya
Tiwari, Anil Kumar
Singh, Kuldeep
DOI
10.1007/s42600-023-00298-4
Abstract
Purpose: Diabetic Retinopathy (DR) is a progressive retinal disease caused by long-term diabetes. Non-proliferative Diabetic Retinopathy (NPDR), an early stage of DR, damages retinal blood vessels, often leads to swelling and leakage of blood. This results in the formation of microaneurysms (MAs) and hemorrhages (HAMs). These changes in the retina can cause vision problems, and an early diagnosis and management are crucial to prevent the progression of the disease. Methods: This paper presents a reliable method for severity grading of NPDR into normal, mild, moderate, and severe classes using fundus images. The proposed method consists of image enhancement, masking out the optic disc region and blood vessel elimination for initial candidate extraction of MAs and HAMs, features extraction, and classification. Fundus image enhancement includes denoising, contrast enhancement, shade correction, and image normalization. In this study, we construct a hybrid feature set based on multiple descriptors, including shape, color, texture, and statistics, to enhance the classification process. For texture-based information, the performance of the gray-level co-occurrence matrix (GLCM) and Gabor-based feature descriptor is thoroughly analyzed. The proposed method is evaluated using two datasets APTOS and MESSIDOR, which are divided into four NPDR classes, each of which suffers from class imbalance, where the number of samples in one class significantly outweighs the other. Such an imbalance can adversely affect machine learning classification models, as they tend to over-predict the majority class and under-predict the minority class. To address this issue, the synthetic minority oversampling technique (SMOTE) is utilized. To grade the images into one of the severity classes (normal, mild, moderate, and severe) and to further improve the performance for class imbalance, we present an ensemble learning-based random forest (RF) classifier. Results: The proposed method achieved a weighted average accuracy of 98.6%, a sensitivity of 97.2%, a specificity of 98.3%, an F1-score of 97.2%, and a precision of 97.2% on the MESSIDOR dataset. For the APTOS dataset, the proposed method achieved an average accuracy of 98.9%, a sensitivity of 97.6%, a specificity of 98.8%, an F1-score of 97.6%, and a precision of 97.6%. Conclusion: The performance evaluation results demonstrate the effectiveness of the proposed method, which will aid in the early diagnosis, regular screening, and effective management of NPDR.