Retinal Vessel Segmentation In Fundus Images Using A Convolutional Encoder-Decoder Architecture

Fundus images are one of the useful diagnostic tools for different retinal diseases such as diabetic retinopathy and hypertensive retinopathy.Retinal vessel segmentation in the fundus image can assist the doctor in diagnosing the potential patient and greatly reduce the workload of doctor.And it is of great significance to the clinical analysis of the doctor.Due to the uniqueness of retinal blood vessels,how to achieve more efficient blood vessel segmentation of the retina is still a research difficulty.Recently,convolutional neural networks(CNNs)have been widely used and have shown good performance in medical image segmentation.Based on the convolutional encoder-decoder structure,this dissertation proposes two different convolutional neural networks to improve the segmentation results and evaluates the performance on a publicly accessed database called Digital Retinal Images for Vessel Extraction(DRIVE).The main research work in this dissertation is as follows:First,this dissertation adds two kinds of skip connections on the convolutional encoder-decoder structure.The long skip connection connects the feature maps of the encoding process to the decoder of the corresponding level,so that the up-sampling process can obtain high-level semantics information and shallow detail information at the same time.Short skip connections enable the network to achieve better convergence by learning residual mappings without increasing network computational complexity.The experimental results show that the model with skip connections can get better segmentation performance.Secondly,in order to make full use of the features learned by the network in the encoding process,the integration method of multi-path fusion is proposed based on the skip connections segmentation model to fine-tune the segmentation result.The feature maps learned at different levels are upsampled,and the results of different paths are integrated to achieve end-to-end segmentation.Through experimental comparison,the segmentation model of multipath aggregation improves the accuracy of segmentation results and the value of AUC.Finally,in order to reduce the influence of the useless features and noise in the decoding process,a segmentation model based on attention mechanism is proposed.By adding different spatial soft attention to the encoder and the decoder,the feature map is given a certain weight coefficient for parameter spatially re-scaling,so that the important information in the feature map is more emphasized,and the unimportant information is suppressed.The experimental results show that the sensitivity of the model to the blood vessel pixels is also effectively improved while improving the segmentation performance.