Comparison of Different Machine Learning Classifiers for Glaucoma Diagnosis Based on Spectralis OCT

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Early glaucoma detection is crucial for management.
• Optical coherence tomography (OCT) provides detailed structural data.
• Conflicting OCT parameters can challenge clinical diagnosis.

Purpose of the Study:

• To compare the efficacy of various machine learning classifiers (MLCs) in diagnosing glaucoma.
• To evaluate MLCs using Spectralis OCT parameters: cRNFL, BMO-MRW, ETDRS macular thickness, and PPAA.
• To identify the most influential OCT parameters and anatomical locations for glaucoma detection.

Main Methods:

• Five MLCs (CIT, LMT, C5.0, RF, XGBoost) were trained and tested.
• Logistic regression (LGR) served as a benchmark for comparison.
• Analysis focused on discriminating between normal and glaucomatous eyes using Spectralis OCT data.

Main Results:

• Random Forest (RF) demonstrated superior performance among the tested MLCs.
• Ganglion cell layer measurements were most predictive for early glaucoma.
• Circumpapillary retinal nerve fiber layer (cRNFL) measurements became more critical with increased glaucoma severity.
• Global, temporal, inferior, superotemporal, and inferotemporal regions showed significant influence.

Conclusions:

• MLCs, particularly RF, offer a reliable method for glaucoma diagnosis using Spectralis OCT parameters.
• Specific OCT parameters and locations have varying importance depending on glaucoma stage.
• Clinicians should integrate OCT findings judiciously within the broader clinical context.