Research On Segmentation And Detection Algorithm Of Vascular Image

Blood vessels are an important part of the human circulatory system.The segmentation and detection of vasular images are crucial to the diagnosis of vascular diseases.However,the structural complexity of blood vessels,low image contrast,and uneven filling of contrast agents pose great challenges to the automatic segmentation and detection of vascular images.It is urgent to solve the following problems:(1)It is difficult to segment bifurcated vessels and lesion vessels accurately.(2)The three-dimensional(3D)vessel segmentation is not applicable to clinical diagnosis,owing to its slow computation,and poor reconstruction accuracy.(3)The small data volume and uneven sample distribution of vascular images drag down the accuracy of the detection task.From two-dimensional(2D)images to 3D images,from accuracy to computing speed,and from blood vessel segmentation to detection,this paper progressively analyzes and studies various vascular images.The main research results cover the following aspects:(1)Two attention-based methods were proposed for retinal vessel segmentation.Firstly,Att-ResGAN was designed with residual module and attention module,and evaluated on DRIVE and STARE datasets.The accuracy and Dice similarity coefficient of the network were 0.9564 and 0.829 on the former dataset and 0.9690 and 0.841 on the latter dataset,respectively.To overcome the low segmentation accuracy facing abnormal vessels and bifurcated vessels,the authors further proposed an iterative U-Net guided by dual attention mechanism,namely,DAI-UNet The channel attention and spatial attention were fused,and the auxiliary network was added to complete the fine segmentation.In this way,the missing of pixels,a common problem of existing blood vessel segmentation models,was solved perfectly.Experiments prove that this network achieves better segmentation results than the previously proposed Att-ResGAN and other mainstream networks.(2)A 3D automatic segmentation method was developed for coronary arteries.The accurate extraction of coronary arteries from computed tomography angiography(CTA)images is the basis for 3D visualization of coronary arteries,and the diagnosis of coronary heart disease.This paper puts forward a coronary artery segmentation method based on the Hessian matrix and sphere model.Firstly,the vessel structure was enhanced by an improved Hessian filter Then,the voxels were classified into vessels and the background through maximum a posteriori estimation.Finally,a sphere model was adopted to extract the coronary arteries,and the branches were detected through clustering.Experimental results show that the model can automatically extract coronary arteries,reduce blood vessel breakage,and avoid the formation of pseudo-vascular structures.(3)A segmentation method of the aorta dissection via morphology-constrained stepwise deep mesh regression was proposed.Aortic dissection is an acute,severe vascular disease.The current segmentation methods for aortic dissection face multiple problems:rough edges,low resolution,slow speed,and unsuitability for clinical diagnosis of aortic dissection.To address these problems,this paper designs a structure of convolutional encoding and graph convolutional decoding to predict the deviation of the 3D mesh from the true lumen surface.The morphology of the true lumen of aortic dissection was used to constrain the initial mesh and guide the deformation.Moreover,a stepwise regression strategy was introduced to solve the mesh folding problem and improve the uniformity of the mesh points.The proposed method was verified on the aortic dissection dataset The verification results(Dice similarity coefficient:94.12%;95%Hausdorff Distance:2.85mm)of our method were better than the five state-of-the-art segmentation methods.In addition,our method only consumed 16.6s.The high accuracy and short time-consuming perfectly fit in with clinical requirements.(4)A coronary stenosis detection method was developed based on transfer learning.In order to solve the common problems of insufficient training data and low classification accuracy in vascular image detection this paper carries out pretraining on ImageNet,which contains tens of millions of images,and performs residual network combined with the attention mechanism training by transfer learning strategy,with the aim to detect coronary stenosis.Our detection method was verified on a coronary angiography dataset The detection results(accuracy:0.92;sensitivity:0.94;specificity:0.90)were much better than the results of direct training.