Dual-stage deep-learning method for glaucoma severity classification based on multiscale feature fusion

Glaucoma represents a chronic eye disease caused by progressive optic neuropathies that lead to visual field loss. Appropriate treatment necessitates early detection and precise assessment of disease severity. Accordingly, recent studies have demonstrated substantial efforts in the development of automated glaucoma classification methods. However, the accurate identification of glaucoma stages remains challenging given that most methods rely on single-stage pathways and single-scale feature extraction, which limit their capability to capture overlapping anatomical features. This challenge is further compounded by the scarcity of reliable datasets that represent the stages of disease progression. In response, this work proposed the use of glaucoma multiscale feature fusion network (GMFF-Net), which represents a novel two-stage framework for the classification of glaucoma severity. The first stage of GMFF-Net employs two parallel encoder heads designed to extract structural and anatomical information. Each head integrates multiscale feature extraction and hybrid attention mechanisms to capture variations across receptive fields while emphasizing clinically relevant regions. The resulting feature maps are then adaptively combined using the proposed fusion modules, whose outputs are passed to the deep classification head in the second stage for disease severity classification. Systematic experiments demonstrated the high efficiency of GMFF-Net in the classification of glaucoma stages and its superiority over seven cutting-edge classification models. It achieved an accuracy of 92.822 %, a precision of 0.9326, a recall of 0.9174, and an F1 score of 0.9296 using the Ibn Al-Haitham dataset. These results demonstrate the capability of the dual-stage framework to extract fine-grained features and provides a suitable solution for screening numerous complex diseases.