Hybrid quantum-classical deep learning framework for balanced multiclass diabetic retinopathy classification.

Area of Science:

• Artificial Intelligence
• Quantum Computing
• Medical Imaging

Background:

• Diabetic Retinopathy (DR) is a leading cause of preventable blindness.
• Accurate DR classification is vital for timely treatment.
• Existing models face challenges like class imbalance and scalability.

Purpose of the Study:

• To develop a novel hybrid quantum-classical deep learning framework for five-class DR classification.
• To address limitations of existing models in accuracy, efficiency, and scalability.
• To create a scalable AI diagnostic approach for early disease screening.

Main Methods:

• Utilized a ResNet-50 feature extractor with quantum-ready compression.
• Employed an 8-qubit Variational Quantum Circuit (VQC) with specific gates and entanglement.
• Implemented stratified sampling and mixed-precision training for efficiency and balanced generalization.

Main Results:

• Achieved a balanced accuracy of 80.96% on the APTOS 2019 dataset.
• Outperformed several classical deep learning baselines across all DR severity stages.
• Demonstrated computational efficiency and class-balanced learning.

Conclusions:

• The hybrid quantum-classical framework offers a scalable solution for DR classification.
• The model is suitable for telemedicine and low-resource clinical environments.
• This work provides a replicable framework for AI in medical imaging and disease screening.