Research On Retinal Image Vessels Segmentation Methods Based On Deep Transfer Learning

The shape and structure of retinal blood vessels are closely related to the development of diabetes,hypertension and other diseases in the human body,and retinal blood vessels themselves are also easy to observe,so the automatic segmentation of blood vessels in retinal images is of great significance to the medical field.In recent years,deep learning method has been widely applied in this issue,and after several years of development,many results have been achieved.Thererfore,in this paper,retinal image vessels segmentation is taken as the target task,and the application of deep learning and transfer learning in vessels segmentation is further studied.The main work is as follows:(1)The research status of retinal image vessels segmentation and machine learning,deep learning and transfer learning methods applied in vessels segmentation are researched and summarized.In addition,the problem that the number of labeled training samples cannot meet the requirement of deep learning is analyzed in depth,and the solution to this problem is discussed.(2)A retinal image vessels segmentation method based on network-based deep transfer learning is proposed.First,a deep neural network U-Net suitable for the vessels segmentation task is built,with DRIVE as the target domain dataset,and improving the vessels segmentation accuracy of the target domain dataset as the target task.Then,several relatively poor quality retinal image datasets such as STARE and the labeled neuron synaptic segmentation dataset CREMI are selected as the source domain of transfer learning to realize the pre-training of UNet network.Finally,the target domain training set is used to fine-tune the pre-training network.Experiment results show that this method can transfer the useful knowledge from the source task to the target task and improve the accuracy of vessels segmentation.(3)In order to solve the problem of negative transfer caused by the application of partial samples in source domain dataset to the target task,a retinal image vessels segmentation method based on instance-based deep transfer learning is proposed.First,according to the needs of feature extraction,the typical U-Net network structure is improved,which can appropriately reduce the dimension of the deep features of the image.Then the source domain samples whose features are similar to the target domain samples are selected by the semi-supervised clustering method.Finally,the selected samples are added into the training set to expand the scale of the training set.Experimental results show that this method can extract samples that are helpful to the target learning task from the source domain dataset,further improve the segmentation performance of the network model,and alleviate the negative transfer effect brought by the source domain.(4)To solve the problem of large difference between features of the target domain and the source domain,a retinal image vessels semantic segmentation method based on mapping-based deep transfer learning is proposed.This method takes the output of each convolutional layer of the pre-trained U-Net network as the feature representation space of the image,calculates and compares the difference of image feature distribution between the target domain and the source domain in the feature representation space,finds out the convolution layer with the smallest difference and determines it as the feature representation space,and freezes the network parameters before the feature representation layer during fine-tuning to map all the image to this space.Then the process of instance selection for transferring is combined to further reduce the differences.Experimental results show that this method can map all the samples to the new feature representation space,effectively reduce the difference between the data distributions of the target domain and the source domain,and make fuller use of the source domain information to improve the vessels segmentation accuracy of the target task.