Glaucoma Screening Method Based on Semantic Feature Map Guidance

Existing end-to-end glaucoma screening models don’t pay enough attention to the region with subtle lesions,and this leads to over-fitting.Besides,the interpretability of the existing models is not clear.Therefore,a glaucoma screening method based on semantic feature map guidance is proposed in this paper.We first adopted the DeepLab v3+segmentation model using MobileNet v2 as the feature extraction network to locate the optic disc.Then we extracted the region of interest for glaucoma screening based on the positioning result.Finally,an attention module was designed and embedded in the VGG classification network to achieve accurate glaucoma screening.The attention block of the method uses the semantic discrimination of the semantic feature map and the shallow feature fusion to generate the attention map.Thus,the areas that related to the shallow feature map classification can be strengthened and the non-related areas will be suppressed to enhance the robustness of the model classification ability.The sensitivity,specificity and AUC of the proposed method tested on the LAG database data set are 0.970,0.983 and 0.996,respectively,indicating that the screening accuracy of the proposed method is relatively higher than that of the existing methods.Besides,the model decision attention areas obtained by visualizing the attention activation map are more accurate and consistent with the doctors’diagnosis reference regions.