Deep learning based binary classification of diabetic retinopathy images using transfer learning approach.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of blindness globally, necessitating early detection for vision preservation.
• Timely diagnosis of DR is critical to prevent irreversible vision loss in diabetes patients.

Purpose of the Study:

• To develop and evaluate a deep learning model for accurate binary classification of diabetic retinopathy (DR) images.
• To fine-tune pre-trained neural networks for identifying healthy versus unhealthy retina images.

Main Methods:

• Utilized a combined dataset from DRD-EyePACS, IDRiD, and APTOS-2019, comprising annotated DR images.
• Applied pre-processing techniques including denoising, normalization, and data augmentation to enhance model robustness.
• Fine-tuned 20 pre-trained deep learning networks (Series, DAG, lightweight) using transfer learning.

Main Results:

• The ResNet101 model, within the DAG category, demonstrated superior performance in classifying DR images.
• Achieved a classification accuracy of 97.33% using the ResNet101 model.
• Evaluated model efficiency using metrics such as accuracy, sensitivity, specificity, precision, F1-score, and ROC-AUC.

Conclusions:

• The proposed ResNet101 model effectively detects diabetic retinopathy from radiological images.
• This deep learning approach provides a valuable tool for healthcare professionals in early DR detection.
• Offers a supplementary diagnostic opinion for patients and experts, facilitating timely intervention.