Automatic Identification of Glaucoma Using Deep Learning Methods

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible blindness worldwide.
• Accurate and early detection of glaucoma is crucial for effective treatment.
• Automated methods can improve the efficiency and accessibility of glaucoma screening.

Purpose of the Study:

• To develop and evaluate an automatic classification method for glaucoma detection using fundus images.
• To adapt the GoogLeNet neural network architecture for this specific task.
• To assess the method's performance, particularly with varying image quality.

Main Methods:

• A two-stage methodology involving region of interest (ROI) detection and image classification.
• Utilizing a sliding-window approach combined with the GoogLeNet network for initial training.
• Employing data augmentation techniques to mitigate overfitting on smaller datasets.
• Training a secondary GoogLeNet model on the results of the initial stage for final glaucoma classification.

Main Results:

• The proposed method demonstrated good accuracy in classifying glaucoma from fundus images.
• The system performed well even when analyzing images of poor quality.
• Data augmentation proved effective in enhancing model robustness.

Conclusions:

• The automated GoogLeNet-based classification method is a promising tool for glaucoma detection.
• The approach shows potential for reliable screening, even with challenging image datasets.
• Further development could lead to widespread clinical application for early glaucoma diagnosis.