Knowledge distillation-based lightweight MobileNet model for diabetic retinopathy classification.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Diabetic retinopathy (DR) is a leading cause of preventable blindness globally.
• Manual screening of retinal images is labor-intensive and faces resource limitations, particularly in low-income countries.
• Deep learning (DL) shows promise for DR detection but often requires substantial computational resources.

Purpose of the Study:

• To develop a lightweight deep learning model for efficient diabetic retinopathy screening.
• To address the limitations of large DL models on resource-constrained devices.
• To enable effective DR detection suitable for edge deployment.

Main Methods:

• Proposed a lightweight student model based on the MobileNet architecture.
• Utilized depthwise separable convolutions for efficient model design.
• Employed knowledge distillation to transfer performance from a larger model to the lightweight one.

Main Results:

• Achieved 98.38% accuracy, precision, and recall for binary classification on the APTOS 2019 dataset.
• Attained 93.03% accuracy for ternary classification on the APTOS 2019 dataset.
• The model's efficient design is suitable for deployment on edge devices.

Conclusions:

• The proposed lightweight DL model provides an efficient and accurate method for diabetic retinopathy screening.
• Knowledge distillation effectively creates compact models for medical image analysis.
• This technology can help bridge the gap in DR screening accessibility, especially in resource-limited settings.