Automated Classification of Physiologic, Glaucomatous, and Glaucoma-Suspected Optic Discs Using Machine Learning

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma diagnosis requires accurate analysis of optic disc images and patient data.
• Clinical annotation of datasets can be time-consuming and limit data volume, potentially impacting algorithm performance.
• Developing machine learning algorithms (MLAs) can assist ophthalmologists in diagnosing glaucoma.

Purpose of the Study:

• To evaluate an MLA for automated classification of physiological optic discs (PODs), glaucomatous optic discs (GODs), and glaucoma-suspected optic discs (GSODs).
• To assess the performance of MLA using color fundus photographs alone and in combination with metadata.
• To determine if MLA can achieve high diagnostic accuracy with limited data.

Main Methods:

• Utilized color fundus photographs and 14 types of metadata from 1168 eyes (584 patients).
• Performed machine learning (ML) in two steps: 1) images only, 2) images and metadata.
• Evaluated classification performance using Area Under the Curve (AUC) for GOD vs. GSOD and GOD vs. POD.

Main Results:

• Initial MLA classification using only images achieved AUCs of 0.84 (GOD vs. GSOD) and 0.88 (GOD vs. POD).
• Combining images with metadata significantly improved performance, yielding AUCs of 0.92 (GOD vs. GSOD) and 0.99 (GOD vs. POD).
• Excellent classification performance was achieved despite a relatively small dataset.

Conclusions:

• An MLA can effectively support ophthalmologists in diagnosing glaucoma.
• Integrating multimodal data (images and metadata) enhances MLA diagnostic capabilities.
• MLA shows promise for accurate glaucoma detection, even with limited annotated data.