The Research On The Key Technology Of Super-resolution Reconstruction And Vessel Segmentation For Fundus Image

Fundus images play an important role in medical diagnosis and can be used to diagnose and screen for a variety of ocular and systemic diseases.Fundus images contain a variety of tissue structures such as microvascular tumors,macula and blood vessels,among which the morphological characteristics of blood vessels are an important basis for diagnosis.In clinical practice,physicians usually need to manually label blood vessels in fundus images,however,this practice has some degree of limitations.First,due to the limitation of imaging effect and the professional level of doctors,manual labeling is subjective and uncertain,especially in the labeling of microscopic vessels,which may be missed.Secondly,the manual annotation process takes a lot of time,resulting in inefficient diagnosis and possible delays for patients with acute illnesses.Currently,super-resolution reconstruction and image segmentation techniques in the field of computer vision are considered to be the key to solve the problems of imaging effects and manual annotation.In order to solve the problems of insufficient imaging effect of fundus images and low data volume and resolution of datasets,this paper firstly proposes a ZSSR-based image super-resolution reconstruction method.The significance of this method is to perform super-resolution reconstruction on the dataset used by the segmentation network,and to solve the problems of blurred imaging images and insignificant regions of interest while enabling the image blocks after pre-processing and data enhancement to be fed into the segmentation network with larger size and higher resolution to improve the segmentation accuracy.The residual information distillation module proposed in this paper is introduced in the model to extract features at different levels and retain shallow features without increasing the network depth,making the model more lightweight.The multi-significant kernel attention blocks proposed in this paper are combined by feature fusion so that each depth feature is fused and extracted to achieve feature refinement and enhancement.The network is trained using the internal image training method proposed by ZSSR,which further improves the performance performance for the super-resolution task of fundus images with less data volume while achieving unsupervised learning.By conducting experiments on fundus image datasets,it is demonstrated that the method has better super-resolution reconstruction results on fundus images than other classical methods.Based on the careful study of medical features of fundus images and existing semantic segmentation methods,this paper proposes a U-Net-based vascular segmentation network DSENet for fundus images,in which the dataset is grayed out and CLAHE preprocessed to enhance the contrast of the image to make the vascular detail part more prominent than the background.The Large Kernel Space Pyramid(LKSP)module is proposed to obtain multi-scale information of images by convolutional kernels with different sizes of sensory fields.Multi-view attention(MVA)block is introduced to capture the interactions of multi-view cross dimensions to enhance the features of the encoder while assisting the decoder to generate more accurate mask maps.The effectiveness of the proposed module is demonstrated using ablation experiments on several fundus image datasets,and the quantitative and qualitative performance of the DSENet is also compared with the classical network for the vascular segmentation task in a comprehensive manner,demonstrating that the proposed network has higher segmentation accuracy.Finally,the data set was further processed using the super-resolution method proposed in Chapter 3,and it was found that the segmentation accuracy was significantly improved compared with that without the super-resolution method,demonstrating the advantages of combining the super-resolution reconstruction method proposed in this paper with the vessel segmentation method.