An Improved Mask-Guided Glaucoma Screening Method Based on Convolution Neural Network

In recent years, computer intelligent technology has been widely used in clinical diagnosis. Glaucoma is concealed and difficult to cure, so the early diagnosis is vital important for prevention. The existing automatic screening methods still have many problems in practical applications, such as excessive dependence on the segmentation accuracy of fundus images, rely on a large amount of data, poor interpretability, and so on. To solve these limitations, we propose a mask-guided glaucoma screening method. First, we segment the color fundus image to obtain the corresponding mask image. Then, the mask image and the fundus image are fused and input into our proposed classification network MG-Efficient Net for glaucoma prediction. In particular, mask image is not directly used for glaucoma classification, so the segmentation accuracy has little influence on the results. At the same time, the masked image contains the spatial information of the optic cup (OC) and the optic disc (OD), which can help the classification network pay more attention to the pathological areas to improve the classification accuracy. Abundant experiments have been performed on the public glaucoma classification dataset, and results show that the proposed method is superior to the existing methods on ORIGA, LAG, RIM-ONE and REFUGE datasets. Furthermore, the network attention experiment also proves the effectiveness of the mask-guided method we proposed. Keywords: Glaucoma screening, mask-guided, ROI extraction, mask generation, network attention