Development of revised ResNet-50 for diabetic retinopathy detection.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss due to retinal blood vessel damage.
• Early detection and intervention are crucial to prevent severe vision impairment and blindness.
• Deep learning and image recognition offer potential for accurate and rapid DR prediction.

Purpose of the Study:

• To enhance diabetic retinopathy (DR) prediction accuracy using the ResNet-50 model.
• To apply visualization and preprocessing techniques to improve ResNet-50's module calibration.
• To demonstrate the effectiveness of a standard operating procedure (SOP) for DR fundus image preprocessing.

Main Methods:

• A revised ResNet-50 architecture was developed, incorporating adaptive learning rates and regularization.
• Fundus images were preprocessed using a defined standard operating procedure (SOP).
• The performance of the revised ResNet-50 was compared against other Convolutional Neural Network (CNN) models.

Main Results:

• The revised ResNet-50 achieved a training accuracy of 0.8395 and a test accuracy of 0.7432.
• The proposed method outperformed other CNN models, effectively mitigating the overfitting phenomenon.
• The revised ResNet-50 demonstrated reduced loss values and minimized fluctuation compared to standard models.

Conclusions:

• A novel DR grading system was designed using an SOP for image preprocessing and a modified ResNet-50.
• The study highlighted the impact of SOP and visualization on the revised ResNet-50 model's performance.
• The findings offer insights into revising CNN structures for improved medical image analysis, particularly for DR detection.