语义特征图引导的青光眼筛查方法  

Glaucoma Screening Method Based on Semantic Feature Map Guidance

作　　者：郭璠[1] 李伟清 赵鑫[2] 邹北骥[2] Guo Fan;Li Weiqing;Zhao Xin;Zou Beiji(School of Automation,Central South University,Changsha 410083;School of Computer,Central South University,Changsha 410083)

机构地区：[1]中南大学自动化学院,长沙410083 [2]中南大学计算机学院,长沙410083

出　　处：《计算机辅助设计与图形学学报》2021年第3期363-375,共13页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金青年科学基金(61502537);湖南省自然科学基金(2018JJ3681);中南大学研究生自由探索与创新项目(2020zzts567).

摘　　要：现有的端到端青光眼筛查模型往往忽略细微病变区域而导致过拟合问题,并且其可解释性区域尚不明确.针对上述问题,提出一种语义特征图引导的青光眼筛查方法.利用基于MobileNet v2作为特征提取网络的DeepLab v3+分割模型进行视盘区域的分割定位,并且根据定位结果提取用于青光眼筛查的重点感兴趣区域,再通过设计注意力模块将其嵌入VGG分类网络实现青光眼的准确筛查.注意力模块利用语义特征图融合浅层特征综合生成注意力图,通过强化特征分类相关区域而抑制非相关区域,达到增强模型分类鲁棒性的目的.所提方法基于LAG database数据集测试的敏感性、特异性以及AUC分别为0.970,0.983和0.996,优于已有方法,且通过可视化注意力激活热图得到的模型决策关注区域更为精细,能够获得与医生诊断参考区域相一致的结果.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.

关 键 词：青光眼筛查 可解释性 注意力机制 卷积神经网络 决策关注区域 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]