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

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• Diabetic retinopathy (DR) is a primary cause of vision loss, necessitating early detection for effective management.
• Automated image analysis using deep learning offers a scalable approach for timely DR diagnosis.
• Convolutional Neural Networks (CNNs) show promise for DR detection in retinal fundus images.

Purpose of the Study:

• To develop an ensemble deep learning framework (EfficientNetB0 and DenseNet121) for five-stage DR classification.
• To evaluate the impact of image preprocessing on diagnostic performance.
• To integrate Grad-CAM for lesion localization and ensure computational efficiency for screening.

Main Methods:

• A dataset of 53,412 fundus images was curated from multiple sources.
• Image preprocessing included CLAHE, artifact removal, and normalization.
• Transfer learning with EfficientNetB0 and DenseNet121 backbones was employed, followed by hybrid ensemble and Grad-CAM visualization.

Main Results:

• The hybrid ensemble model achieved 91.2% accuracy, 0.961 macro-AUC, 92.1% sensitivity, and 0.913 F1-score.
• Image preprocessing enhanced performance by 3-4%.
• The ensemble model outperformed standalone CNNs, with Grad-CAM confirming accurate lesion localization.

Conclusions:

• A clinically viable and explainable deep learning model for DR detection was developed.
• The ensemble approach demonstrates superior performance and efficiency for DR screening.
• Future work includes external validation and optimization for point-of-care applications.