Retinal Glaucoma Public Datasets: What Do We Have and What Is Missing?

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Public databases for glaucoma research contain retinal images, primarily of the optic papilla, for automated screening using deep learning.
• Current databases suffer from heterogeneity in image acquisition, lack of optic papilla segmentation, and insufficient data for early diagnosis.
• Existing datasets are not interchangeable and use varied classification criteria, hindering reliable deep learning model development.

Purpose of the Study:

• To propose improvements in the structure and presentation of public databases for automated screening of glaucomatous papillae.
• To enhance existing datasets with relevant medical information and standardized segmentation.
• To facilitate the development of robust deep learning models for early glaucoma detection.

Main Methods:

• Analysis of existing public glaucoma image databases and their limitations.
• Development of a structured approach for presenting retinal images with optic papilla and cupping segmentation.
• Inclusion of medical context and standardized classification criteria for glaucoma diagnosis.

Main Results:

• Identified critical shortcomings in current public glaucoma databases, including data heterogeneity and lack of standardization.
• Proposed a framework for enhanced database structure, incorporating expert-made segmentations and clear classification guidelines.
• Highlighted the need for simultaneous imaging of both eyes and early diagnostic markers.

Conclusions:

• Improving public glaucoma databases is crucial for advancing automated screening and early diagnosis.
• Standardized, well-annotated datasets are essential for training reliable deep learning models for glaucoma detection.
• Enhanced databases will support large-scale population screening, reduce diagnostic ambiguity, and promote timely treatment to prevent blindness.