General deep learning model for detecting diabetic retinopathy.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) detection relies on retinal ophthalmoscopy, with deep learning (DL) aiding diagnosis and workflow.
• Traditional DL methods for DR use 80/20 data splits and Synthetic Minority Oversampling Technique (SMOTE), risking model overfitting and inaccurate predictions.
• Overfitting in DR models can distort variables, leading to erroneous predictions despite high reported accuracy (90%-99%).

Purpose of the Study:

• To address the overfitting problem in deep learning models for diabetic retinopathy detection.
• To develop a robust deep learning model capable of accurate DR classification.
• To propose a generalizable method for handling imbalanced medical image datasets.

Main Methods:

• A two-stage training approach was implemented, involving two learning modules.
• Learning module 1 identified DR vs. no-DR, while module 2 classified DR severity (mild, moderate, severe, proliferative) using SMOTE synthetic data.
• Early stopping and data division techniques were employed to minimize overfitting during training.

Main Results:

• The model was evaluated on multiple datasets (DIARETDB0, DIARETDB1, eOphtha, MESSIDOR, DRIVE).
• Prediction accuracies ranged from 84.27% to 92.5% across the evaluated datasets.
• The developed model demonstrated effective performance in detecting and classifying diabetic retinopathy.

Conclusions:

• A general deep learning model for DR detection was successfully developed and validated across diverse DR databases.
• A straightforward method for managing imbalanced DR datasets was presented.
• The proposed method shows potential for application with other medical imaging datasets.