Stripe Pooling and Vessel-Constraint Network for Fundus Image Artery/Vein Classification
The ratio of retinal artery to vein diameter is a prerequisite for quantitative analysis of chronic diseases,such as diabetes and hypertension,and is an important risk indicator for many cardiovascular diseases.With the development of deep learning technology,many methods based on convolution neural network have made great progress in the classification of fundus images based on their ability to capture high-level semantics.However,most of the methods are based on superimposed local convolution and pooling operation,which is difficult to be well applied to striped retinal blood vessel segmentation.In this paper,in order to extract the features of retinal blood vessels in the shape of stripes more effectively,we introduced stripe pooling to capture the long-distance dependence of spatial pixels.Taking into account the complex characteristics of arteriovenous interleaving and further combining with spatial pyramid pooling,a new mixed pooling technology was proposed to expand the receptive field and learning context information of the neural network.On the other hand,considering that the proportion of blood vessel and non-blood vessel distribution in the fundus image is extremely unbalanced,this paper introduced a blood vessel enhancement module,which used the information of blood vessel distribution and the information of blood vessel edge constrained by Gaussian kernel function as weights to correct the arteriovenous features and suppress the background features,thus solving the problem of the imbalance between blood vessel and background distribution.Experiments on three internationally available datasets,DRIVE,LES,and HRF,containing 40,22,and 45 color fundus images respectively,showed that the proposed algorithm achieved results of 0.955,0.946,and 0.967 in term of BACC scores,which verified that the method combining strip pooling and vascular enhancement effectively solved the problems of complex arteriovenous interlacing and category imbalance in fundus images,achieving accurate classification of retinal arteriovenous malformations,holding a high application value.
NETL
NSTL
万方数据
