Glauco-Net: A novel deep learning framework for early glaucoma diagnosis

Glaucoma, a leading cause of irreversible blindness, poses significant diagnostic challenges due to its subtle onset and progressive nature. Current methods using color fundus images often fail to detect early-stage glaucomatous changes. To address this, we introduce Glauco-Net, a novel algorithmic framework integrating four transformative innovations. First, the spectral-adaptive retinal reconstruction engine (SARRE) reconstructs hyperspectral details from RGB fundus images, uncovering latent spectral biomarkers of early glaucomatous damage. Second, the vascular topology signature embedding (VTSE) employs graph neural networks to analyze retinal vascular topology, capturing structural irregularities linked to disease progression. Third, the temporal texture evolution transformer (TTET) models spatiotemporal texture dynamics in sequential fundus images, detecting subtle textural shifts indicative of glaucoma. Finally, the peripapillary light scattering dynamics profiler (PLSDP) simulates light–tissue interactions to identify peripapillary structural anomalies linked to glaucomatous cupping and atrophy. These innovations are fused into a unified deep learning pipeline, achieving SOTA performance in glaucoma diagnosis. Evaluated on large-scale clinical datasets, Glauco-Net achieved an AUC exceeding 0.97, demonstrating superior sensitivity and specificity compared to existing methods. By leveraging advanced spectral, topological, and dynamic analyses, this framework not only enhances early detection but also provides deeper insights into glaucoma pathogenesis. Our work represents a paradigm shift in computational ophthalmology, offering a robust, multimodal approach to glaucoma diagnosis and paving the way for personalized monitoring and intervention strategies. The source code of the proposed Glauco-Net is available at https://github.com/livingjesus/Glauco-Net.