Issuu

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 12 Issue: 05 | May 2025

p-ISSN: 2395-0072

www.irjet.net

AI-Powered Detection of Diabetic Retinopathy Alla Krishna sai reddy1, Achanta Satya Sai Kishore2, Mudadla Koushik Manikanta Naidu3, Kharan Jagabathula4, Yogesh Raj5, Dalalappa gari Naveen6, Dr. Brajesh Kumar7 1,2,3,4,5,6UG Student, Dept. of School of Computer Science & Engineering, Lovely Professional University, Punjab, India 7Associate Professor, Dept of School of Computer Science & Engineering, Lovely Professional University, Punjab,

India ---------------------------------------------------------------------***--------------------------------------------------------------------

Abstract- Diabetic retinopathy (DR) is a serious

not only time-consuming and subjective but also often inaccessible in many regions [9].

complication of diabetes that can gradually lead to vision loss if not detected early. This document presents an automated system designed to identify the severity of DR using retinal fundus images. The system leverages a convolutional neural network based on a modified ResNet50 architecture to classify the disease into five stages—ranging from no signs of DR to the most severe, proliferative DR. To improve the model’s performance and address the challenge of class imbalance in the dataset, Mix-up data augmentation was employed. Our model was trained and evaluated using the APTOS 2019 dataset, and the results showed strong classification accuracy. By utilizing pretrained ImageNet weights and incorporating custom dense feature extraction layers, the system was able to effectively analyse retinal images. This approach offers a promising solution for automated DR screening, helping healthcare professionals detect the condition at an earlier stage and ultimately contributing to the prevention of vision loss on a global scale.

To bring consistency to diagnosis, the International Clinical Diabetic Retinopathy (ICDR) severity scale categorizes DR into five stages: No DR, mild nonproliferative DR (NPDR), moderate NPDR, severe NPDR, and proliferative DR (PDR) [4]. Each stage is marked by specific retinal abnormalities, such as microaneurysms, hemorrhages, hard exudates, cotton wool spots, and neovascularization [3]. Accurate detection and classification of these lesions demand both advanced expertise and substantial clinical experience. In this paper, we present a deep learning approach for automated detection and classification of DR using retinal fundus images. Our method employs a ResNet50based CNN architecture with transfer learning and data augmentation techniques to address the aforementioned challenges [12]. We utilize the APTOS 2019 Blindness Detection dataset, which contains a diverse collection of retinal images annotated by expert ophthalmologists according to the ICDR severity scale.

Key Words: Diabetic retinopathy, deep learning, convolutional neural networks, ResNet50, transfer learning, medical image classification, computer-aided diagnosis.

The primary contributions of this work include:

1.INTRODUCTION Diabetic retinopathy (DR) is a severe microvascular complication arising from diabetes mellitus that affects the retina and stands as one of the leading causes of preventable vision loss globally [1]. With the global prevalence of diabetes projected to reach 642 million by 2040 [2], the impact of vision impairment caused by diabetic retinopathy is projected to rise sharply, particularly in developing nations where routine screening remains largely inaccessible. DR progresses through various stages of severity, beginning with mild non-proliferative changes and potentially advancing to proliferative DR, which can lead to permanent vision loss if not treated in time. Early detection and timely intervention are crucial, as treating DR in its early stages can reduce the risk of severe vision loss by over 90% [6]. However, the current standard of diagnosis relies on manual inspection of retinal fundus images by trained ophthalmologists—a process that is

Impact Factor value: 8.315

Implementation of a deep learning framework based on ResNet50 architecture with customized dense layers for DR severity classification.

Application of the Mixup data augmentation technique to improve model generalization and address class imbalance.

Evaluation of the model's performance across different DR severity levels, providing insights into its clinical applicability.

Development of a practical prediction pipeline for single-image classification with confidence metrics.

The remainder of this paper is organized as follows: Section II discusses related work in the field of automated DR detection [5]. Section III details our methodology, including dataset preprocessing, model architecture, and training strategy. Section IV shows experimental results and analysis comparing every aspect [15]. Finally, Section V concludes the paper with a discussion of limitations and future directions [14].

ISO 9001:2008 Certified Journal

Page 111