Robust Deep Learning Framework for Early Diabetic Retinopathy Detection Using Preprocessed Fundus Images and Optimized CNNs

Summary

This study presents a describes a deep learning-based workflow for early detection of diabetic retinopathy using preprocessed fundus images and optimized convolutional neural network architectures, enabling accurate disease classification and explainable lesion localization for scalable clinical application.

Abstract

Diabetic retinopathy (DR) is a leading cause of vision loss globally, particularly among patients with poorly controlled diabetes. Early detection through automated image analysis has emerged as a scalable solution to support timely intervention. Deep learning-based models, particularly convolutional neural networks (CNNs), have shown significant promise in DR detection from retinal fundus images. This study aimed to (i) develop an ensemble deep learning framework using EfficientNetB0 and DenseNet121 for five-stage DR classification, (ii) systematically evaluate the impact of fundus image preprocessing on diagnostic performance, (iii) incorporate Grad-CAM-based visual explanations for lesion localization, and (iv) achieve high diagnostic accuracy with computational efficiency suitable for scalable screening. A dataset comprising 53,412 fundus images from EyePACS, APTOS 2019, and a tertiary-care centre in China was curated. Preprocessing steps included contrast-limited adaptive histogram equalization (CLAHE), artifact removal, and normalization. Transfer learning was applied using EfficientNetB0 and DenseNet121 backbones, followed by hybrid assembling. The models were evaluated using accuracy, macro-AUC, sensitivity, specificity, and F1-score. Grad-CAM was used to visualize lesion localization. The hybrid ensemble model achieved 91.2% accuracy, 0.961 macro-AUC, 92.1% sensitivity, and an F1-score of 0.913. Preprocessing improved performance by 3-4%, and the ensemble approach outperformed standalone CNNs. Grad-CAM overlays confirmed accurate lesion localization. Model performance was evaluated for both five-stage DR grading and binary referable DR detection to reflect clinical screening requirements. This study presents a clinically viable, explainable deep learning model for DR detection. Future work will include external validation on independent datasets, prospective real-world evaluation, and model optimization (e.g., pruning and quantization) for mobile and point-of-care screening applications.

Introduction

DR is a microvascular complication of diabetes mellitus and remains one of the leading causes of avoidable visual impairment worldwide¹. Chronic hyperglycemia induces damage to retinal capillaries, leading to increased vascular permeability, microaneurysm formation, hemorrhages, exudates, and in advanced stages, neovascular proliferation. Without timely intervention, these changes may progress to vitreous haemorrhage, tractional retinal detachment, and irreversible vision loss^1,2. Epidemiological studies estimate that approximately 35 percent of individuals with diabetes exhibit some de....

Protocol

This study involved a retrospective analysis. In accordance with institutional and national regulations, formal ethics committee approval and informed consent were waived, as no identifiable patient information was accessed or used.

1. Data sources and acquisition

1. Re-label all images according to the International Clinical Diabetic Retinopathy (ICDR) Severity Scale (Grades 0-4)¹⁰, to ensure consistency across datasets obtained from EyePACS, APTOS 2019, and the tertiary-care cohort.
2. Map the original labels for datasets with existing annotations to the ICDR scale to maintain unifo....

Results

Dataset summary and preprocessing effects

The curated dataset consisted of 53,412 color fundus images distributed across the five DR severity classes: No DR (0) - 39.2%, Mild (1) - 18.4%, Moderate (2) - 22.5%, Severe (3) - 12.1%, and Proliferative DR (4) - 7.8%. Data originated from EyePACS, APTOS 2019, and a tertiary Chinese ophthalmology center, ensuring diversity in ethnicity, imaging conditions, and camera specifications.

Preprocessing significa.......

Discussion

In this study, we developed and rigorously evaluated a hybrid ensemble model combining EfficientNetB0 and DenseNet121, fine-tuned using transfer learning and enhanced with classical preprocessing steps, CLAHE, artifact removal, and intensity normalization, for five-class DR grading. The ensemble achieved outstanding performance: accuracy 91.2%, macroAUC 0.961, and F₁score 0.913 for referable DR detection, and sensitivity 92.1% as reported in Section 6. These metrics substantially outperformed both a baseline CNN (a.......

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Fundus Camera (Topcon, Canon, Zeiss)	Topcon / Canon / Zeiss		Used to capture high-resolution color fundus photographs under varying illumination conditions
NVIDIA Tesla V100 GPU	NVIDIA Corporation	Tesla V100	32 GB VRAM; used for training deep learning models
Python (OpenCV, imgaug libraries)	Open-source community		Image preprocessing including CLAHE, artifact removal, normalization, augmentation
TensorFlow 2.10 + Keras	Google Brain		Deep learning framework used for CNN model building and training
EfficientNetB0 & DenseNet121 pretrained backbones	Open-source (TensorFlow/Keras Applications)		Transfer learning backbones pretrained on ImageNet used for DR classification