A Lightweight CNN for Multiclass Retinal Disease Screening with Explainable AI.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Current deep learning models for retinal disease screening face challenges with class imbalance, large model sizes, and lack of transparency.
• Accurate and timely detection of retinal diseases is crucial to prevent irreversible vision loss.

Purpose of the Study:

• To develop a lightweight, attention-augmented convolutional neural network (CNN) for accurate and transparent retinal disease detection.
• To address limitations of existing deep learning screeners, including class imbalance, model size, and opaque reasoning.

Main Methods:

• A novel CNN architecture combining depthwise separable convolutions, squeeze-and-excitation, and global-context attention was developed.
• Gradient-based class activation mapping (Grad-CAM and Grad-CAM++) were integrated for pixel-level explainability.
• A severely imbalanced ten-class color-fundus dataset was re-balanced using synthetic minority oversampling technique (SMOTE) and task-specific augmentations.

Main Results:

• The lightweight network achieved 87.9% test accuracy, outperforming Inception-V3 by 58% error reduction.
• High true-positive rates (>95%) were recorded for eight disorders, with notable F1-scores for macular scar (0.77) and central serous chorioretinopathy (0.89).
• Saliency maps successfully highlighted key pathological features, validating the model's decision-making process.

Conclusions:

• Targeted class re-balancing, lightweight attention mechanisms, and integrated explainability enable accurate, transparent, and deployable retinal screening.
• The developed model is suitable for point-of-care ophthalmic triage, even on resource-limited hardware.
• This approach offers a promising solution for improving early detection and management of retinal diseases globally.