糖尿病视网膜病变(diabetic retinopathy,DR) 是糖尿病并发症最常见的疾病之一。由于视网膜图像病灶存在类间差异小、特征复杂的特点,导致传统深度学习网络对病灶点不聚焦、特征提取不充分等问题。针对上述问题,提出一种视网膜病变的自动分类模型D-SCNet。首先设计一种全局注意力模块SC,它以空间加通道的方式交替排列,相互促进,克服传统全局注意模块先通道后空间两维独立作用的注意方式；再使用高效激活函数EReLU进行非线性处理,加大对有效区域中黄色环、亮红色斑块和视盘区病变点数的关注度。最后,以轻量化DenseNet121为主干网络,将注意力模块SC放置于瓶颈结构的3×3卷积之后,得到新的特征提取瓶颈结构,促使网络提取有效信息,提高分类精度。测试结果表明,SC注意力模块具有一定的泛化作用,在传统典型网络Vgg、ResNet、Xception上对分类准确率分别提升3.04%、0.76%、2.48%；同时D-SCNet模型相比现有模型具有一定的优越性。若将该模型应用在临床上,可协助眼科医生对视网膜病变进行初级筛查。

Diabetic retinopathy (DR) is one of the most common complications of diabetes.Due to the small differences between inter classes and complex features of retinal image lesions,the traditional deep learning network has some problems,such as non-focusing on lesion points and inadequate feature extraction.In view of above problems,an automatic classification model for retinopathy,D-SCNet,is proposed.Firstly,a global attention module SC is designed,which alternately arranges in a spatial and channel manner and promotes each other,thereby overcoming the traditional two-dimensional independent attention mode where the global attention module first channels and then space.Then the high-efficiency activation function EReLU is used for nonlinear processing to increase the attention to the lesions numbers in the yellow rings,bright red patches and the optic disc areas.Finally,with lightweight DenseNet121 as the main backbone network,the attention module SC is placed after 3×3 convolution of the bottleneck structure to obtain a new feature extraction bottleneck structure,which promotes the network to extract effective information and improve classification accuracy.The test results show that the SC attention module has a certain generalization effect,and the classification accuracy on traditional typical networks Vgg,ResNet,and Xception is improved by 3.04%,0.76%,and 2.48% respectively.Meanwhile,the D-SCNet model has some advantages over the existing models.If this model is applied in clinical practice,it can assist ophthalmologists in conducting primary screening for retinopathy.

TP391.41

安徽理工大学医学 专项培育项目(YZ2023H2B006)、 安徽理工大学引进人才科研启动基金(2022yjrcc44)和国家自然科学基金(52174141)资助项目