Screening of Common Retinal Diseases Using Six-Category Models Based on EfficientNet.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Early detection of common retinal diseases is crucial for preventing vision loss.
• Primary care physicians require efficient tools for preliminary screening of retinal conditions.
• Deep learning models show promise in analyzing medical images for disease diagnosis.

Purpose of the Study:

• To propose a six-category deep learning model for screening common retinal diseases.
• To evaluate the performance of the EfficientNet-B4 model in classifying normal and diseased fundus images.
• To provide a tool for primary medical institutions to aid in the preliminary screening of five common retinal diseases.

Main Methods:

• Trained two six-category deep learning models (EfficientNet-B4 and ResNet50) on 2,400 fundus images.
• Compared the performance of the six-category models with a previous five-category model.
• Evaluated models using sensitivity, specificity, F1-score, AUC, kappa, and accuracy on 1,315 test images.

Main Results:

• The EfficientNet-B4 model achieved a diagnostic accuracy of 95.59% and a kappa value of 94.61%.
• Area under the curve (AUC) for all diagnoses exceeded 0.95.
• High sensitivity, specificity, and F1-scores were reported for normal fundus, RVO, high myopia, glaucoma, DR, and MD diagnoses.

Conclusions:

• The EfficientNet-B4 based six-category model effectively diagnoses normal fundus and five common retinal diseases.
• This AI tool can assist primary care physicians in screening retinal diseases and recommending timely referrals.
• Implementing this model can enhance diagnostic efficiency, particularly in rural areas, and prevent treatment delays.