Research On Vessel Segmentation Algorithm Guided By Geometric Model

Vessel segmentation technology is a fundamental work for vessel centerline extraction,vessel topology estimation and vessel stenosis quantification.Accurate vessel segmentation has important research significance for the auxiliary diagnosis of coronary heart disease,glaucoma and other vascular diseases.In recent years,artificial intelligence technology has made rapid progress in the fields of computer vision and medical image analysis.How to effectively integrate machine learning with medical prior knowledge to further improve the performance of blood vessel segmentation is a research hotspot.In this paper,geometric features are used as a medium for the fusion of medical prior knowledge and machine learning,where the doctor’s subjective experiences,rules or perceptions are transformed into geometric models.The machine learning is guided by geometric information to enhance the ability of image processing tasks.In this paper,we studies vessel segmentation based on geometric model guidance,including coronary artery segmentation,retinal vessel segmentation and retinal artery/vein separation.The existing challenges of vessel segmentation include:1)The complex noise types of coronary CTA slices and the low contrast between coronary artery and adjacent tissues affect the accuracy of coronary artery segmentation;2)The complex shape and distribution trend of 3D coronary trees and the non-rigid geometric deformation at vessel bifurcations make it difficult to segment the complete vessel tree;3)The low contrast of tiny vessels in fundus images leads to the fracture of thin blood vessels,making it difficult to effectively maintain vascular structural connectivity;4)The prior knowledge in retinal artery/vein separation task is difficult to utilize effectively.To address the above challenges,the research scheme of this paper is as follows:1)Aiming at the problem that complex noise interference and low contrast affect the accuracy of coronary artery segmentation in coronary CTA slices,this paper proposes a 2D coronary artery slice segmentation method combining image denoising and multi-objective clustering.First,we propose a non-local mean denoising algorithm based on the local fractal dimension,which preserves the detailed information of the image while suppressing the noise.Then,we design an adaptive multi-objective clustering method based on fuzzy constraint histogram,which can extract fine vessel structures from complex background regions and various image artifacts.Experimental results show that this method can effectively suppress noise and improve the accuracy of vessel segmentation.2)Aiming at the problem that the complex distribution trend of 3D coronary trees and the non-rigid geometric deformation at bifurcations make it difficult to segment the complete vessel tree,this paper proposes a 3D coronary tree segmentation method based on toroidal model guidance.Assuming that the blood vessel is a freely curved cylinder in 3D space,the geometric features of coronary cross-sections between adjacent slices can always be found and matched on two cross-sections of a toroidal model.Based on this theoretical assumption,we first extract eight types of geometric features of blood vessels on coronary slices,and then analyze the geometric feature relationship of vessels between adjacent slices based on the toroidal model to estimate the coronary artery.Finally,we use the depth-first search algorithm to traverse all slices to obtain a complete vessel tree.Experimental results show that this method can effectively avoid vessel branch loss during coronary artery tracking.3)Aiming at the problem that it is difficult to effectively maintain the connectivity of vascular structures in retinal image segmentation results,this paper proposes a retinal vessel segmentation method based on geometric skeleton reconnection model.Firstly,we propose a multi-scale linear structure detection network to extract fine vessel structures.Secondly,combining the advantages of dual thresholds,we design an adaptive hysteresis threshold algorithm to extract vessel branches from vessel probability maps to maintain vessel connectivity and suppress background noise.Finally,we design a geometric skeleton-based vessel segment reconnection algorithm to bridge the interrupted vessel segments.Experimental results show that this method can ensure the accuracy of vessel segmentation while effectively maintaining the connectivity of vessel structures.4)Aiming at the problem that prior knowledge is difficult to be effectively utilized in retinal artery/vein separation task,this paper proposes a multi-level artery/vein separation method based on geometric decomposition and recombination model guidance.First,we use a vessel tree separation method based on convex decomposition to obtain a completely separated vessel segment.Then,a fine-grained vessel segment classification network is designed to evaluate the matching degree of vessel segments.According to the number of vessel segments at bifurcations/intersections,it is divided into two groups,and two different classification strategies are designed for vessel segment pairing.Finally,a multi-level artery/vein vessel labeling method is designed to correct the vessel categories from four levels:crossover tree,bifurcation tree,bifurcation subtree and vessel segment.Experimental results show that this method integrates prior knowledge with classification network,which can cope with the complex connection mode of blood vessels and effectively improve the accuracy of retinal artery/vein separation.