Research On Retinal Vessels Images Segmentation Based On Deep Learning

In recent years,medical image has made great progress in diagnosis and treatment of lesions depending on the image processing technology of computer.As we all know,if the relevant content that doctors want can be well segmented in medical images,the efficiency of disease diagnosis or treatment will be greatly improved.Therefore,the segmentation of medical images has always been one of the most concerned issues.A lot of diseases which are in early stages or in sicking process can be reflected on blood vessel of retina in human body,such as common hypertension,diabetic,etc.In addition,fundus retinal blood vessels are the most easily observed blood vessels in the human body,which can be directly observed in a non-invasive way.Therefore,the segmentation of blood vessels from various fundus retinal images will bring great convenience to doctors' diagnosis and treatment.Nowadays,there are many methods for blood vessels segmentation,but they are not enough in efficiency,flexibility and accuracy.Therefore,it is necessary to design and propose a more efficient,flexible and accurate automated retinal blood vessels segmentation algorithm.Deep learning has a good performance in artificial intelligence,computer vision and so on.Therefore,the research of image segmentation based on deep learning has certain research significance and application prospects.The following aspects are the main work of this thesis:(1)Firstly,Because the retinal blood vessel images have numerous details,the Unet encoder is replaced with a VGG16 which is removed the fully connected layer and has a more feature extraction capability.Secondly,considering that too many pooling layers in VGG16 will permanently lose more feature information,we replace the last two pooling layers in VGG16 with dilated convolutions,and solve contradictory issues in the original pooling layers which will expand the receptive field and reduce the image.In addition,the method of feature fusion of high and low layer networks in Unet is followed,but considering that the cropping in the fusion operation has no effect on the segmentation effect,all convolution operations in Unet are replaced with Same convolution,thereby simplify the fusion process.Finally,the width and length of the blood vessels in the retinal blood vessel images are different.In order to enable the network to perceive blood vessels of different scales and improve the segmentation precision of small blood vessels,the pyramid feature fusion module is joined to integrate image context information,and the feature fusion of multi-scale and multi-local regions is used to further enhance the segmentation effect.(2)The size of datasets is expanded by randomly segmenting the retinal blood vessel image in the DRIVE dataset within the FOV.During the network training process,the image blocks are randomly rotated,translated to further expand the dataset.In order to make the proposed network segmentation effect better,image preprocessing is performed on all images according to the characteristics of the datasets,including image size modification,image graying,image normalization,and CLAHE contrast enhancement.The K-fold cross-validation method was used during the training phase to prevent overfitting.In the model prediction stage,the segmentation effect is improved by using image blocks overlap averaging.(3)In order to verify whether the network proposed in this thesis has a better segmentation effect than Unet,the improved segmentation results are compared and analyzed.The experiment proves that the network has a better segmentation effect than Unet.In order to further verify the generalization performance of the proposed network,the network of this thesis is also applied to the STARE dataset,which still has a good segmentation effect.Compared with other blood vessel segmentation methods in recent years,the network of this thesis also has certain advantages.