The Study Of Deep Convolutional Neural Network Based Retinal Vessel Segmentation

In recent years,retinal vessel segmentation technology has become an important component for disease screening and diagnosing in clinical medicine.The retinal vasculature is the only part of microvessel can be directly observed in the whole body's vascular system.The delineation of morphological attributes of retinal blood vessels can not only prevent ocular diseases effectively,but also has a certain connection with cardiovascular diseases such as diabetes and hypertension.Therefore,the research on retinal vessel segmentation is of great importance in clinical medicine and practice.Retinal vessel segmentation is a nontrivial task due to complex distribution of blood vessels,relatively low contrast between target and background,and potential presence of illumination and pathologies.We propose a new algorithm for retinal vessel segmentation based on deep convolutional neural network,and use fully convolutional networks to solve the segmentation task specifically.The main works and contributions of this dissertation are provided as follows:1.A retinal vessel segmentation algorithm based on HED-CRF is proposed.From the deep learning point of view,we train a segmentation network suitable for retinal images,the algorithm has no extra preprocesses,to retain the original inherent information as completely as possible.The validity of fully convolutional networks in retinal vessel segmentation is proved.Fully convolutional networks have generalization ability and representation capability for describing the feature of targets,combined with the spatial constraints of CRF,we put CRF into HED,which contains five convolution layers.Finally,we construct the HED-CRF network,an end-to-end model.Experimental results show that the algorithm based on HED-CRF has an advantage in both speed and accuracy,thus can get the retinal vessel segmentation results quickly and efficiently.2.A retinal vessel segmentation algorithm based on fully convolutional networks with side-output inter-connection is proposed.On the basis of HED-CRF network,the algorithm takes multi-scale segmentation features into consideration.To achieve the fine degree of manual segmentation results,we take the first side-output as the foundation,which contains the most abundant details,then fuse global information from deeper side-output into shallower side-output.Two networks using side-output inter-connection is proposed,the connection is from deep to shallow layer,and the first side-output is regarded as the output of the multi-scale fusion network.Conbined with spatial constraints of CRF,we take the first side-output as unary potential,and construst an end-to-end network finally.The experimental results show that the multi-scale fusion is helpful to get finer segmentation results,and the segmentation is more similar to human observer.When applied to standard benchmarks of fundus imaging,the DRIVE,STARE and CHASE_DB1 databases,the two methods proposed in this dissertation all achieve state-of-the-art segmentation performance,not only get fine morphological structures of blood vessels,but also is certain robust to illumination and pathologies.