Retial Vessel Segmentation Using Generative Adversarial Learning With A Large Receptive Field

Morphological changes in retinal blood vessels are closely related to the occurrence of some diseases.The ophthalmologist can diagnose some diseases by observing the morphological structure of retinal vessels.Therefore,the automatic segmentation of retinal vessel is of great practical significance.There have been many retinal vessel segmentation algorithms,and the overall accuracies of these algorithms can exceed the accuracies of human observers.Due to the high complexity of retinal image,most retinal vessel segmentation methods are prone to some shortcomings such as oversegmentation or under-segmentation.This paper proposes a new retinal vessel segmentation using generative adversarial learning with a large receptive field,which can achieve a high segmentation accuracy.The proposed method trains the generative and discriminative networks based on the idea of two-player game,and finally the generative network can generate an more accurate vessel segmented map.As compared with existing methods,the proposed method makes improvements in the following three aspects.Firstly,a new generative network is proposed and the dilated residual block(DRB)is proposed based on the atrous convolution and residual block.The DRB structure can effectively increase the receptive field of the network while maintaining the overall information of feature maps,and thus the network can better capture the global information of the input image.Secondly,a new deep discriminative network based on the semantic information of the input image pair is proposed.The residual block is introduced into the discriminative network to prevent the degradation problem.Finally,a scheme for image filling and cropping is proposed to ensure the feature concatenation between different resolution under the premise of introducing the least pixels.In order to verify the performance of the proposed method,we have conducted many some experimental comparisons with existing methods on two public datasets(DRIVE and STARE).The accuracies are 95.63% and 96.84%,and ROC curves are 98.12% and 98.53%.The overall performance is better than existing methods.Experimental results show that the proposed method can not only obtain high segmentation accuracy and sensitivity,but also avoid some over-segmentation phenomenon with better performance.