Attention Mechanism-Based Glaucoma Classification Model Using Retinal Fundus Images.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate diagnosis of eye diseases like glaucoma is crucial for preventing vision loss.
• Automated classification using deep learning can aid in early detection and management.
• Feature extraction from retinal fundus images is key for robust diagnostic models.

Purpose of the Study:

• To develop and evaluate an attention-based classification model for diagnosing glaucoma using fundus images.
• To investigate the impact of incorporating extracted retinal structures (vessels and optic disc) into the classification process.
• To compare the proposed model's performance against existing research models.

Main Methods:

• A ResU-Net model and Hough Circle Transform were used for segmenting retinal vessels and the optic disc.
• A multi-input Convolutional Neural Network (CNN) model integrated preprocessed fundus images and extracted structures.
• Attention mechanisms were employed to enhance feature learning from the fundus structures.

Main Results:

• The proposed attention-based model achieved superior performance in glaucoma classification compared to other models, even with limited data.
• Ablation studies confirmed that attention mechanisms significantly improved performance by effectively utilizing fundus structures.
• Challenges were noted in classifying normal cases due to potential feature degradation during structure extraction.

Conclusions:

• Attention mechanisms integrated with structural feature extraction offer a promising approach for accurate glaucoma classification from fundus images.
• The model's effectiveness on smaller datasets suggests potential for practical clinical application.
• Future work should include more fundus structures and a broader range of eye diseases.