Research On Fundus Vascular Semantic Segmentation Methods Based On Supervised Learning

Fundus vascular segmentation technology is of great significance in DR screening and glaucoma diagnosis.The precision of vascular segmentation may directly affect the hard exudate of the fundus,microaneurysm and soft exudate and blood vessel diameter and shape.The precision of medical testing therefore affects the diagnosis of related eye diseases.The thicker fundus vessels are relatively easy to segment,while the peripheral blood vessels of the fundus are relatively small and low,and are not easily segmented.Therefore,it is difficult to achieve high segmentation precision for fundus vessels.In the case,conditional random fields,random forest algorithm,U-net model,deep image matting network model,dilated convolution,deformable convolution and residual are involved in this thesis.Specifically,the main work of the thesis includes:(1)A fundus segmentation model based on depth feature fusion and random forest integration is proposed.For the traditional conditional random field model of vessel segmentation,the vascular problems such as vascular end points and unclear constriction may be lost.We use the method of deep feature fusion to fuse the relevant information into its feature vector,and learn the blood vessels and background through model training.The dependencies thus improve the accuracy of vessel segmentation.Aiming at the problem of class imbalance between blood vessels and background,the random forest classification algorithm is integrated into the conditional random field vessel segmentation model,which can effectively solve the problem,so that the segmentation model can effectively suppress noise,robustness and improve the classification accuracy of blood vessels.And improve the accuracy of the vessel segmentation model.(2)A blood vessel segmentation model based on U-net network and deep matting network is proposed.Based on the limitations of manual extraction of features,we used a U-net network to implement port-to-port blood vessel segmentation.In view of the small blood vessel loss caused by the loss of spatial detail during sampling in the U-net network,we inspire from the rough segmentation to the fine segmentation idea,and based on the principle of image semantic segmentation accuracy of the map accuracy,the deep image matting network combined with the U-net network and the U-net network output as a coarse segmentation,the deep image matting network is used to refine the coarse segmentation,thereby increasing the precision of the blood vessel segmentation.For the incompatibility of the interface between the U-net network and the deep image matting network,we are dating the trimap generation algorithm.Through experiments and comparison with advanced blood vessel segmentation algorithms,we can verify that the improved vessel segmentation model has high precision for segmentation of small blood vessels.(3)A dilated residual block fusion deformable convolution blood vessel segmentation model is proposed.Aiming at the coexistence of complex problems based on U-net network and deep image matting network,the internal structure of U-net network is improved.At the same time,we are small in the case of U-net network.Inspired by the idea of dilated convolution,the vascular loss problem is integrated into the U-net network block.The downsampling uses convolution,which increases the convolution receptive field while preserving the spatial details and improving the small blood vessel segmentation.In view of the difference in the size and shape of blood vessels in different parts,it is difficult to improve the accuracy of blood vessel segmentation.Inspired by the deformable convolution,the deformable convolution is integrated into the U-net network block,so that the blood vessel model can Adapt to different sizes and shapes,improve the segmentation accuracy of segmented blood vessels and small blood stasis.In order to further improve the precision of small blood vessel segmentation,inspired by the improved U-net network model,we dated the horizontal sampling convolution layer to reduce the barrier between the low-level semantic features of the coding network and the high-level semantic features of the decoding network.In order to prevent the degradation of the network model from occurring,we place the integrated residual function into the U-shaped network block.