SODAS: Second-order optimization differential architecture search for diabetic retinopathy prediction

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy is the leading cause of blindness globally.
• Early diagnosis of diabetic retinopathy is critical for vision preservation.
• Existing diagnostic methods struggle with scale variations and real-world accuracy.

Purpose of the Study:

• To introduce a novel method, Second-Order Optimization Differential Architecture Search (SODAS), for predicting diabetic retinopathy.
• To overcome limitations of manual network design in detecting retinopathy at various scales.
• To improve accuracy and convergence speed in diabetic retinopathy diagnosis.

Main Methods:

• Utilized neural architecture search (NAS) to automatically design networks for grading diabetic retinopathy.
• Integrated Gumbel-Softmax sampling into NAS to minimize gradient information loss.
• Employed second-order optimization to accelerate the convergence of NAS.

Main Results:

• SODAS demonstrated significant improvements across multiple metrics.
• Average improvements included 12.1% in accuracy, 21.6% in Cohen's kappa, 28.8% in AUC, 24.1% in IBA, and 22.5% in F1-score.
• The method showed rapid convergence and effectiveness on benchmark datasets.

Conclusions:

• SODAS effectively addresses the limitations of current diabetic retinopathy detection methods.
• The proposed approach enhances diagnostic accuracy and efficiency.
• SODAS represents a significant advancement in AI-driven diabetic retinopathy screening.