Machine Learning And Deep Learning For Glaucoma Detection: A Systematic Narrative Review Of Methods, Datasets, Evaluation Practice, And Clinical Translation

Abstract

Glaucoma is a leading cause of irreversible blindness and is often asymptomatic until substantial visual field loss occurs. Clinicians integrate intraocular pressure, optic nerve head assessment, optical coherence tomography (OCT), visual field testing, and risk factors—but screening and longitudinal care remain resource-intensive and variable across settings. Machine learning (ML), and especially deep learning (DL), promise automated assistance from fundus photographs, OCT volumes, and increasingly electronic health record text and multimodal combinations. This systematic narrative review synthesizes peer-reviewed work on ML/DL for glaucoma detection and related screening tasks, with emphasis on what is compared, on which data, and under which evaluation designs. We summarize trends across fundus-based classification and segmentation-plus-classification pipelines, OCT-based modeling, multimodal fusion, privacy-preserving federation, and explainability. We highlight persistent threats to validity, including small or homogeneous public benchmarks, weak external validation, spectrum bias, optimistic metrics, and limited reporting of calibration and fairness, and translate these into concrete recommendations for more clinically credible benchmarking.