Hybrid Model Structure for Diabetic Retinopathy Classification.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of blindness and a common diabetes complication.
• Delayed diagnosis of DR often leads to irreversible vision impairment due to resource limitations.
• Automated DR classification using neural networks offers improved efficiency and cost-effectiveness.

Purpose of the Study:

• To enhance the performance of diabetic retinopathy classification models.
• To introduce an improved loss function and novel hybrid model architectures for DR detection.
• To leverage deep learning for more efficient and accurate early diagnosis of DR.

Main Methods:

• Utilized EfficientNetB4, EfficientNetB5, NASNetLarge, Xception, and InceptionResNetV2 as base Convolutional Neural Network (CNN) models.
• Trained base models using both standard cross-entropy loss and an enhanced cross-entropy loss function.
• Developed and evaluated three hybrid model structures (Hybrid-a, Hybrid-f, Hybrid-c) using outputs from the base models.

Main Results:

• The enhanced cross-entropy loss accelerated training and improved model performance across various metrics.
• The proposed hybrid model structures further boosted DR classification performance.
• Key metrics improved: accuracy from 85.44% to 86.34%, sensitivity from 98.48% to 98.77%, specificity from 71.82% to 74.76%, precision from 90.27% to 91.37%, and F1 score from 93.62% to 93.9%.

Conclusions:

• The enhanced cross-entropy loss function is effective for accelerating training and improving DR classification.
• The novel hybrid model architectures demonstrate significant improvements in DR classification accuracy and other key metrics.
• These AI-driven advancements hold promise for earlier and more effective detection of diabetic retinopathy, potentially preventing blindness.