Abstract
Early-stage glaucoma diagnosis is crucial for preventing permanent structural damage and irreversible vision loss. While various machine-learning approaches have been developed for glaucoma diagnosis, only a few specifically address early-stage detection. Moreover, existing early-stage detection methods rely on unimodal information and exclude subjects with high myopia, which contradicts clinical practice and overlooks the adverse effect of high myopia on prediction performance. To develop a clinically practical tool, this study proposes a deep-learning-based, end-to-end early-stage glaucoma detection framework designed for a cohort likely with high myopia. This framework uniquely integrates functional information from visual field (VF) parameters of standard automated perimetry (SAP) and Pulsar perimetry (PP) with structural information derived from optical coherence tomography (OCT) thickness maps. It comprises three key components: 3D OCT ganglion cell complex (GCC) layer segmentation, thickness map generation, and early-stage glaucoma detection. Evaluated on 394 subjects using five-time, 10-fold cross-validation, the proposed system achieved a mean area under the receiver operating characteristic (ROC) curve of 0.887 ± 0.006, outperforming the Asaoka method without transfer learning and nine models based solely on VF parameters. Results further confirmed that incorporating SAP and PP parameters was essential for mitigating the adverse effects of high myopia.