A Lightweight Diabetic Retinopathy Detection Model Using a Deep-Learning Technique.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a severe diabetes complication prevalent in Saudi Arabia.
• Current DR detection systems are computationally expensive and prone to false positives due to imbalanced datasets.
• There is a need for efficient, low-resource DR grading systems.

Purpose of the Study:

• To develop a lightweight deep learning (DL) model for DR severity grading.
• To address high computational costs and dataset imbalance issues in existing systems.
• To create a system suitable for limited computational resources and potential mobile application deployment.

Main Methods:

• Image pre-processing to handle noise and artifacts in fundus images.
• Feature extraction using You Only Look Once (Yolo) V7.
• Feature selection via a tailored Quantum Marine Predator Algorithm (QMPA).
• Severity prediction using a hyperparameter-optimized MobileNet V3 model.

Main Results:

• The proposed model achieved high accuracy (98.0% on APTOS, 98.4% on EyePacs) and F1 Scores (93.7% and 93.1%, respectively).
• Demonstrated reduced computational complexity: fewer parameters, FLOPs, lower learning rate, and less training time.
• Validated on large datasets (APTOS: 5590 images, EyePacs: 35,100 images).

Conclusions:

• The lightweight DL model offers an efficient solution for DR severity grading.
• The system's low computational requirements enable its use in remote areas and mobile applications.
• Future work will focus on enhancing DR detection from low-quality fundus images.