Circumpapillary OCT-focused hybrid learning for glaucoma grading using tailored prototypical neural networks

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of global blindness.
• Optical Coherence Tomography (OCT) is crucial for glaucoma detection.
• Current methods often focus on detection rather than grading.

Purpose of the Study:

• To propose a novel framework for glaucoma grading using raw circumpapillary OCT B-scans.
• To develop a hybrid network combining traditional and deep learning approaches.
• To improve the discrimination between healthy, early, and advanced glaucoma.

Main Methods:

• A hybrid OCT-based network integrating hand-crafted and deep learning features.
• An OCT-specific descriptor for retinal nerve fibre layer (RNFL) feature extraction.
• A novel CNN with residual and attention modules for feature refinement.
• Few-shot learning using static and dynamic prototypical networks on fused OCT databases.

Main Results:

• Achieved categorical accuracy of 0.9459 for validation and 0.8788 for testing in glaucoma grading.
• Demonstrated substantial improvements in discriminating between glaucoma stages.
• Class activation maps confirmed the RNFL as the most relevant structure for diagnosis, aligning with clinical opinion.

Conclusions:

• The proposed hybrid network offers a powerful new approach for glaucoma grading using OCT.
• Few-shot learning with prototypical networks enhances diagnostic capabilities.
• The model's focus on RNFL aligns with established clinical understanding of glaucoma progression.