Toward Efficient Identification of Retinal Diseases: A Lightweight Convolutional Neural Network-Based Approach Using...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal diseases cause significant vision loss, necessitating early detection.
• Optical coherence tomography (OCT) is crucial for diagnosing retinal conditions.
• Existing deep learning models for OCT analysis are often too complex for clinical edge devices.

Purpose of the Study:

• To develop an efficient, lightweight deep learning framework for real-time retinal disease diagnosis.
• To enable practical deployment of AI-powered diagnostic tools on clinical edge devices.

Main Methods:

• Proposed a novel lightweight deep learning architecture for retinal disease detection.
• Integrated a lite convolution block with depthwise separable convolutions for efficiency.
• Incorporated a global-local fusion block and squeeze-and-excitation mechanism for feature capture and refinement.
• Model parameter count is only 0.27 million.

Main Results:

• Achieved high accuracies on benchmark datasets: 99.70% (OCT 2017), 95.00% (OCT C8), and 97.26% (OCTDL).
• Demonstrated strong and stable performance through confusion matrix and ROC analysis.
• Grad-CAM visualizations provided enhanced model interpretability.

Conclusions:

• The proposed lightweight deep learning model offers a practical and efficient solution for real-time retinal disease diagnosis.
• The architecture is suitable for deployment on edge devices, facilitating clinical application.
• The study highlights the potential of efficient AI in improving early detection and management of retinal diseases.