An Imbalance-Aware AI/ML Framework for Five-Stage Diabetic Retinopathy Classification Using Retinal Fundus Images
International Journal of Drug Delivery Technology · 04.04.2026 · 16 (87s)
DOI: 10.25258/ijddt.16.87s.183 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Drug Delivery Technology, online erschienen am 04.04.2026. Verfasst von M. Srinivas und 3 weiteren Autor:innen.
Abstract (Originalsprache)
Diabetic Retinopathy (DR) is a progressive eye disease related to diabetes that may lead to loss of vision if not detected early. Artificial intelligence has benefits as it can independently evaluate ocular images, which makes it easier to screen large populations and reduces the workload for physicians doing this manually. However, the indicators are so subtle that it is hard for computers to categorize diabetic retinopathy accurately at various stages. Furthermore, the available datasets often lack a balanced set of cases, which makes the problem even more complex. An accurate system for diagnosing diabetic retinopathy from retinal fundus pictures is important for diabetic retinopathy classification. To address this challenge, an artificial intelligence and machine learning–based framework was developed to automatically classify diabetic retinopathy into five clinically relevant stages. The data set contains 3662 images, each labelled with one of five diagnostic categories: no diabetic retinopathy, mild, moderate, severe or proliferative. A simple and efficient pipeline was developed to analyse the retinal fundus images for diabetic retinopathy classification. The pipeline resizes the images, enhances the contrast using a method called contrast-limited adaptive histogram equalization and describes the images using a combination of features such as Histogram of Oriented Gradients, RGB colour histograms and low-resolution pixel features. The extracted features were subsequently evaluated using four machine learning classifiers, namely Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Multilayer Perceptron (MLP), to assess their effectiveness in diabetic retinopathy classification. To mitigate the effects of class imbalance across the different diagnostic grades in the dataset, class-weighted and balanced training strategies were adopted during model training. Finally, the predictions generated by the individual classifiers were combined using a soft-voting ensemble approach, thereby leveraging the complementary strengths of the individual classifiers. The proposed approach aims to achieve more accurate and reliable classification of diabetic retinopathy from retinal fundus images. This approach has potential to improve diagnosis and management of diabetic retinopathy and further explore its potential. The goal is to build systems that assist physicians and patients in making informed decisions and enhancing healthcare outcomes, whether for diabetic retinopathy or other diseases. Experimental results show that the proposed ensemble achieved an accuracy of 74.18%, balanced accuracy of 46.54%, macro-F1 of 0.4748, weighted-F1 of 0.7080, Cohen’s kappa of 0.5892 and ROC-AUC of 0.8458. The group performed very well especially in the no-DR and moderate-DR categories with sensitivities of 95.94% and 80.00% respectively. However, the detection of severe diabetic retinopathy was challenging, with a sensitivity of 6.90%. Class-weighted SVM improved severe-DR sensitivity from 6.90% to 27.59%, suggesting that focusing on minority-class identification may reduce overall accuracy but improve clinically significant class-sensitive metrics. The results show the potential and limitations of the lightweight AI/ML approaches for automated diabetic retinopathy grading and provide a reproducible benchmark for future deep-learning studies. Keywords: diabetic retinopathy, retinal fundus images, artificial intelligence, machine learning, SVM, random forest, XGBoost, multilayer perceptron, ensemble learning, class imbalance, medical image classification.
Zitieren
APA
Srinivas, M., Kuchulakanti, H., Thirupathanna, K., & Tejaswi, K. L. (2026). An Imbalance-Aware AI/ML Framework for Five-Stage Diabetic Retinopathy Classification Using Retinal Fundus Images. International Journal of Drug Delivery Technology, 16(87s). https://doi.org/10.25258/ijddt.16.87s.183
Vancouver
Srinivas M, Kuchulakanti H, Thirupathanna K, Tejaswi KL. An Imbalance-Aware AI/ML Framework for Five-Stage Diabetic Retinopathy Classification Using Retinal Fundus Images. International Journal of Drug Delivery Technology. 2026;16(87s). doi:10.25258/ijddt.16.87s.183
BibTeX
@article{Srinivas2026,
author = {Srinivas, M. and Kuchulakanti, Harish and Thirupathanna, K. and Tejaswi, Koutharapu Lakshmi},
title = {An Imbalance-Aware AI/ML Framework for Five-Stage Diabetic Retinopathy Classification Using Retinal Fundus Images},
journal = {International Journal of Drug Delivery Technology},
year = {2026},
volume = {16},
number = {87s},
doi = {10.25258/ijddt.16.87s.183}
}
EUVIMED