A Single Network with an Attention Mechanism for Glaucoma Detection

Glaucoma gradually deteriorates the optic nerve to total blindness, causing irreversible visual loss. When glaucoma patients have visual field abnormalities, they are essentially in the advanced stages of the disease. Consequently, early detection to prevent visual impairment is crucial. Due to their non-invasive nature and simplicity of observation, color fundus photographs have become an important screening tool for eye diseases. This paper proposes an attention-based network based on color fundus photographs for early screening of glaucoma. Specifically, in this study, the EfficientNetV2S network is selected as the base network, and the attention mechanism designed in this paper is introduced as the final model. Due to the issue with data imbalance, focal loss is introduced for training. Experiments conducted on two public databases demonstrate that our model attains an AUC of 0.984, an F1 index of 0.986, and an acc of 0.972. In the REFUGEE and GAMMA datasets, the proposed attention module can provide an F1 boost of 1%. After the implementation of focal loss, F1 has increased by approximately 1.5% on multiple networks.