Detection of retinal diseases using an accelerated reused convolutional network

Area of Science:

• Computer Science
• Artificial Intelligence
• Medical Imaging

Background:

• Convolutional neural networks (CNNs) are evolving for improved accuracy, speed, and accessibility.
• Enhanced CNN accessibility is crucial for tasks like early eye disease detection, preventing vision loss.
• Existing CNN models for eye disease detection are often computationally complex and inaccessible for mobile applications.

Purpose of the Study:

• To increase the accessibility of deep neural network models for eye disease detection.
• To develop a computationally efficient yet accurate CNN model suitable for mobile devices.

Main Methods:

• Redesigned and optimized convolutional layers to create a novel layer named ArConv.
• Developed a new general CNN model incorporating the ArConv layers.
• Trained and evaluated the model on the RfMiD dataset under identical conditions against MobileNetV2.

Main Results:

• The proposed model, featuring ArConv layers, contains only 1.3 million parameters, significantly fewer than MobileNetV2 (2.2 million).
• Achieved higher accuracy (0.9328) compared to MobileNetV2 (0.9266) on the RfMiD test set.
• Demonstrated suitable complexity for deployment on mobile phones while maintaining high diagnostic accuracy.

Conclusions:

• The novel ArConv layers significantly enhance the accessibility and efficiency of CNNs for medical image analysis.
• The developed model offers a computationally efficient and accurate solution for mobile-based eye disease diagnosis.
• This research contributes to broader applications of AI in ophthalmology, facilitating early detection and prevention of vision disorders.