Enhanced glaucoma detection using U-Net and U-Net+ architectures using deep learning techniques.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma diagnosis relies heavily on accurate optic disc and retinal nerve fiber layer analysis.
• Current diagnostic methods can be limited by image noise and segmentation challenges.
• Deep learning offers potential for improved automated glaucoma detection.

Purpose of the Study:

• To develop and evaluate an enhanced approach for glaucoma diagnosis using image processing and deep learning.
• To compare the efficacy of U-Net and U-Net+ for optic disc segmentation.
• To assess the overall diagnostic performance of the proposed integrated model.

Main Methods:

• Image noise reduction using median filtering.
• Optic disc segmentation employing U-Net and U-Net+ architectures.
• Feature extraction with Capsule Networks and classification via Extreme Learning Machines (ELM).

Main Results:

• Median filtering achieved 97.88% noise reduction (PSNR: 44.99).
• U-Net outperformed U-Net+ in segmentation (Dice: 0.8557, Jaccard: 0.7307).
• The integrated model demonstrated high diagnostic accuracy: 99% (DRISHTI-GS), 99.5% (DRIONS-DB), 98.5% (HRF).

Conclusions:

• The proposed deep learning approach significantly enhances glaucoma diagnosis accuracy.
• Median filtering and U-Net segmentation are effective preprocessing steps.
• This method holds promise for improving patient outcomes in glaucoma management.