Glaucoma screening using an attention-guided stereo ensemble network.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible blindness, necessitating early detection for effective management.
• Accurate feature extraction from optic nerve head stereo images is crucial for reliable glaucoma classification.
• Current screening methods may lack the sensitivity required for early-stage diagnosis.

Purpose of the Study:

• To develop and validate a deep ensemble network with an attention mechanism for enhanced glaucoma detection using stereo optic nerve head images.
• To improve the sensitivity and reduce the false-negative rate in glaucoma screening.
• To enable weakly-supervised training, minimizing the need for expensive segmentation labels.

Main Methods:

• A deep ensemble network combining a Convolutional Neural Network (CNN) for global features and an Attention-Guided Network for localized optic disc analysis was designed.
• Stereo image pairs were processed through sub-components, with outputs fused for robust prediction, compensating for image quality variations.
• The attention mechanism was trained in a weakly-supervised manner using only image-level annotations.

Main Results:

• The proposed deep ensemble network achieved a recall (sensitivity) of 95.48%, a significant improvement from the state-of-the-art 88.89%.
• The model maintained high precision and demonstrated stable performance across real patient images.
• A substantial reduction in the false-negative rate was observed, enhancing early diagnosis potential.

Conclusions:

• The developed deep ensemble network with attention mechanism offers a powerful and efficient tool for early glaucoma detection.
• Weakly-supervised training of the attention module makes the approach more practical and cost-effective.
• Improved detection accuracy and reduced false negatives can lead to timely interventions and better preservation of vision.