Research Of Drusen Segmentation From Retinal Images Based On Supervised Feature Learning

The macula is an essential structure of the retina.Meanwhile,it is the most sensitive area of human vision.People would lose central vision when the macula has suffered damage through injury or disease.Age-related macular degeneration(AMD)is one of the most common causes of blindness in old people.Drusen is the leading cause of AMD and is considered as a specific physical sign of early AMD.Therefore,it is essential to segment drusen accurately for determining if and when a patient will suffer visual loss from AMD.Recent years,with the image processing techniques applied to medical images,computer aided diagnosis significantly improve the level of clinical diagnosis.As a result,the algorithms of drusen segmentation has important value for research.This paper analyzes the features of drusen in retinal images.Then we summarizes and analyzes the existing segmentation methods and the drusen segmentation algorithms.The hand-crafted features which are used in drusen segmentation have some shortcomings,such as indiscriminate,time-consuming,relying on experience and so on.According to the characteristics of drusen and the shortcomings of existing drusen segmentation methods,this paper proposes two segmentation methods based on supervised learning which can learn features automatically.We summarize the main work and innovations of this article as follows:1.This paper proposes a drusen segmentation method via supervised feature learning from retinal images.Different from traditional feature extraction methods,this method combines supervised learning with image label information to establish a new feature which is more discriminative and compact.It applies the generalized low rank approximation matrix to reduce dimensionality of images samples.Next,a supervised manifold regularization is obtained by combining the low rank matrix with label information of images.Then the new features are obtained by iterative optimization via an alternate optimization.Finally,these features are used to train SVM and new retinal images are tested by the SVM.The proposed method is tested on STARE and DRIVE dataset;experimental results demonstrate that the proposed method performs well.2.A drusen segmentation method based ResNet convolutional neural network is proposed.Convolutional neural network can extract discriminant features by automatic learning,but it has not been used in drusen segmentation.According to the characteristics of drusen,this paper modifies the structure and parameters of Residual Networks(ResNet)and applies it to segment drusen from retinal images.The ResNet can automatically optimize parameters and extract features,moreover,the extracted features are more discriminant.Experimental results on STARE dataset demonstrate the effectiveness of the proposed method.