CN115908297A - Topology knowledge-based blood vessel segmentation modeling method in medical image - Google Patents
Topology knowledge-based blood vessel segmentation modeling method in medical image Download PDFInfo
- Publication number
- CN115908297A CN115908297A CN202211408809.9A CN202211408809A CN115908297A CN 115908297 A CN115908297 A CN 115908297A CN 202211408809 A CN202211408809 A CN 202211408809A CN 115908297 A CN115908297 A CN 115908297A
- Authority
- CN
- China
- Prior art keywords
- blood vessel
- vessel
- algorithm
- voxel
- segmentation
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides a topology knowledge-based blood vessel segmentation modeling method in medical images, which comprises the steps of firstly establishing a segmented arteriovenous blood vessel model based on a U-Net neural network, training the model by utilizing a sample image and a labeling result thereof, and adding data of partial labels into data set by utilizing a partial supervision method during training; the graph cutting algorithm is used as post-processing, the prediction result of the neural network is optimized, and the accuracy of the blood vessel cutting result is improved; and constructing a topological structure of the blood vessel by utilizing algorithms such as Grow Cuts, sequence traversal and the like, wherein the topological structure comprises the steps of carrying out segmentation processing on the blood vessel, obtaining the radius of the blood vessel and extracting an arteriovenous subtree structure. The invention utilizes the data mixed training of the complete label and the incomplete label in the data set, introduces the optimization algorithms such as graph cutting and the like, improves the robustness of the algorithm while ensuring the cutting precision, realizes the construction of the blood vessel topological structure, and forms a set of complete algorithm flow for extracting the blood vessel from the medical image and constructing the blood vessel topological structure.
Description
Technical Field
The invention belongs to the technical field of medical image processing, and particularly relates to a blood vessel segmentation and modeling method in a medical image based on topological knowledge.
Background
As a medical diagnostic tool, tomographic scanning modes such as Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) have been widely used to reveal vascular diseases. It is of clinical importance to study the vascular structure of the lungs in a Volume of Interest (VOI). In Diagnosis, some organs such as lung, brain, etc. need to analyze the vascular tree in the organ when Computer Aided Diagnosis (CAD) is used. Because the amount of data generated by medical images is large and the tubular structure of blood vessels is complex, a doctor generates a large amount of workload and increases a lot of burden when performing manual analysis. Therefore, the clinician's demand for fully automatically extracting the blood vessel tree structure is gradually increased, and the importance of the automatic quantitative CAD in diagnosis of blood vessel diseases, surgical planning, and the like is also shown.
In medical blood vessel examination methods, another important step is to extract the middle line representation to simplify specific visualization and quantitative evaluation of the blood vessel, which contains information on the radius of the blood vessel, the wall thickness of the blood vessel, etc. Accurate vessel segmentation and centerline detection as a reliable pre-processing step facilitates accurate determination of vessel anatomy or pathology, thereby guiding pre-operative planning of vascular disease treatment. The hierarchical structure of the blood vessels in the organ is theoretically obvious, and the pathological position can be accurately determined through the hierarchical structure, so that the hierarchical structure of the blood vessels in the organ is necessary and effective to be correctly constructed.
In recent years, there are many Networks for medical segmentation in deep learning, such as U-Net, deep Convolutional Neural Networks (CNNs). In the field of vessel segmentation, there are some network models which are good in representation, and these network models utilize the anatomical characteristics of the vessel to design the network structure in a targeted manner, so as to finally obtain the segmented vessel result. However, due to the complexity of blood vessels, in the result, local adhesion, short score, error and the like of arteries and veins can occur, and the robustness of the algorithm needs to be improved.
Meanwhile, most of the blood vessel segmentation work only aims at the segmentation of the blood vessel, and the extraction of the blood vessel information at a deeper level, such as the number of segments of the blood vessel, the radius of the blood vessel, the length of the blood vessel, the bifurcation condition of the blood vessel and other information, is rarely involved, and the blood vessel model in the organ can be constructed by acquiring the information, so that a doctor can be helped to quickly and accurately position the disease, and therefore, the extraction of the topological information of the blood vessel has important significance.
Disclosure of Invention
In order to solve the problems, the invention provides a blood vessel segmentation modeling method in a medical image based on topological knowledge. The method firstly trains a neural network by using a partial supervision thought in the aspect of blood vessel extraction, and then optimizes the arteriovenous segmentation result by using a graph segmentation algorithm as the post-processing of the arteriovenous segmentation method of the neural network, thereby improving the robustness of the neural network. Secondly, in the aspect of constructing a blood vessel topological structure, a thinning algorithm is used for extracting a central line of a blood vessel image obtained after the graph cut algorithm is optimized, a network topological algorithm is used for extracting a local graph structure of the blood vessel, then algorithms such as Grow Cuts and the like are used for obtaining segmented blood vessels of the blood vessel, and the primary hierarchical construction of the blood vessel tree is completed. And then, obtaining the radius of the blood vessel by using methods such as a section and the like, and finally, extracting an arteriovenous tree-shaped structure by using methods such as sequence traversal and the like, thereby completing the construction of the topological structure of the blood vessel.
The technical scheme of the invention is as follows:
a blood vessel segmentation modeling method in medical images based on topology knowledge comprises the following steps:
step A: creating a segmentation arteriovenous blood vessel model based on a U-Net neural network, adopting a partial supervision method, and training the segmentation arteriovenous blood vessel model by using image data and a labeling result thereof
Step A1, establishing a segmentation arteriovenous blood vessel model
The arteriovenous vascular model is divided by adopting a 3D U-Net network as a basic model, the left side of the U-Net neural network is an encoder part, and the right side of the U-Net neural network is a decoder part; down-sampling by the left encoder for K times, up-sampling by the right decoder for K times correspondingly, and restoring the size of an output image to the size of an input image, wherein the size of the input image is a fixed value; jump connection is also used in the arteriovenous vessel model segmentation, and the feature graph of each level of coder and the feature graph of each level of decoder are spliced to ensure the fineness of the segmentation result;
step A2, training and segmenting arteriovenous blood vessel model
Training with an existing data set; in order to fully use the existing data as much as possible and improve the robustness of the network, partial tagged data are added into a data set based on a partial supervision method; modifying a loss function of a segmented arteriovenous blood vessel model (network model) during training; during testing, a post-processing step B is added to improve the accuracy of arteriovenous segmentation; obtaining the output probability of the artery and the vein and the result of the blood vessel (including the artery and the vein) through a network model;
during training, the loss function used by the network model is a combination of a cross entropy loss function and a Dice loss function:
cross entropy loss function:
wherein, y i For the vessel labeling result corresponding to sample i, p i Outputting a result for the network model corresponding to the sample i, wherein N is the number of samples;
dice loss function:
wherein, y i For the vessel labeling result corresponding to sample i, p i Outputting a result for the network model corresponding to the sample i, wherein N is the number of samples;
loss function used for segmenting arteriovenous vascular models:
L network ＝α*L 1 +β*L 2
wherein, alpha is the coefficient of the cross entropy loss function, and beta is the coefficient of the Dice loss function;
and B: carrying out post-processing optimization on the blood vessel result segmented by the arteriovenous blood vessel model in the step A by utilizing algorithms such as graph cutting and the like
Step B1, extracting the central line of the blood vessel by using a thinning algorithm
Firstly, extracting a central line from a blood vessel result obtained by segmenting the arteriovenous blood vessel model in the step A based on a refinement algorithm, wherein the main process is to repeatedly delete boundary points by using the refinement algorithm under the condition of meeting topology invariance and geometric constraint; simultaneously, parallelly rechecking eight sub-voxels with non-overlapping neighborhoods until a connected point set serving as a skeleton is obtained; the method has the greatest advantages that the obtained framework can be ensured to be continuous, and the main topological structure of the original object can be maintained. The existing thinning algorithm is modified to a certain extent, so that the medical image data can be processed, and the centerline extraction of the blood vessel is realized.
The detailed algorithm comprises the following specific processes:
in a segmented blood vessel image, firstly, a local cube of 3 x 3 is constructed for a certain blood vessel voxel to check the local connectivity of the blood vessel voxel, and the local connectivity is represented by 26 neighborhoods and 6 neighborhoods; meanwhile, a local 3 x 3 cube is divided into 8 mutually overlapped 2 x 2 cubes by a central voxel; then, the following definitions are thus given:
v,N 6 (v),N 26 (v),N 2 (v)
wherein v denotes a vessel voxel, N 6 (v) 6 neighborhoods, N, representing vessel voxels v 26 (v) 26 neighborhoods, N, representing vessel voxels v 2 (v) A2 × 2 × 2 cube representing a part including a blood vessel voxel v;
during the refinement algorithm processing, the boundary of the three-dimensional blood vessel image is obtained from six directions, and the blood vessel voxels of the three-dimensional blood vessel image are divided into three types: boundary voxel V B Euler invariant voxel V E And simple voxel V S ；
The definition of the three types of voxels is as follows:
given a vessel voxel v ∈ S, all vessel voxels have a value of 1, and the corresponding complement set is
Corresponding to background, value 0; if only one of its connected objects belongs to the background, the vessel voxel is considered as boundary voxel V B ；
Given a vessel voxel v ∈ S, if the Euler number does not change when removing the voxel v from S, the vessel voxel is considered as a Euler invariant voxel; the euler number is specifically defined as:
V E ＝{v∣δG 26 (S∩N 2 (v))＝0}
wherein δ represents the amount of change in the parameter values before and after the deletion of the body vessel factor v, G 26 (S∩N 2 (v) N for all vessel voxels S) 2 (v) The euler numbers of the 26 neighborhoods in the set,
represents all vessel voxels->
N of (2) 2 (v) Euler numbers of 6 neighborhoods in the set; for determining N in the voxel v containing blood vessels 2 (v) In the set, the ith N 2 (v) Whether all 8 vertices of (1), if yes, oct i Is set to 1, otherwise is set to 0, f i 、e i Denotes the ith N ２ (v) The number of faces and sides of the cube; delta oct i ，δf i ,δe i Representing the corresponding variation value of each parameter after removing the vessel voxel v;
given a vessel voxel V ∈ S, if the connectivity of its 26 neighborhoods remains unchanged when V is removed from S, then the vessel voxel is considered as a simple voxel V S ；
Performing skeletonization by iteratively deleting the three voxels from the vessel region, wherein the vessel voxels are not changed after deletion, namely the whole skeleton, namely the corresponding central line, is obtained; meanwhile, in order to meet the topological relation and geometric constraint of the result, a sequence rechecking program is used for keeping the integral connectivity when the point is deleted;
step B2, extracting the graph structure of the blood vessel
After obtaining the central line of the blood vessel in the step B1, performing bifurcation point identification and segmentation processing on the central line of the blood vessel through a network topology algorithm, thereby obtaining a graph structure of the blood vessel; the method comprises the following specific steps:
firstly, identifying blood vessel bifurcation points in a blood vessel central line by utilizing a 26-neighborhood region of a blood vessel central point voxel, searching adjacent bifurcation points from different directions for each bifurcation point according to the extracted bifurcation point, and then counting the adjacent bifurcation points and a central point on a search path to obtain a section of blood vessel central line; then, different branches are searched on the same bifurcation point, and the local blood vessel centerline connection relation is extracted; finally, traversing the central line of the whole blood vessel to obtain the graph structure of the blood vessel;
step B3, post-processing optimization is carried out by using a graph cut algorithm
Although the blood vessels are obtained by segmenting the arteriovenous blood vessel model in the step A, adhesion and wrong segmentation can occur between the blood vessels, so that a graph cutting algorithm is combined with the steps B1 and B2 to carry out post-processing optimization;
the specific processing method is that the length of each blood vessel section and the intersection angle of the adjacent blood vessel sections are calculated by utilizing the blood vessel middle line section extracted in the step B2 and the local connection relation; for a certain central point of a blood vessel section, cutting the section of the blood vessel in the step A2 according to the position of the central point, then obtaining a target section after screening, and then calculating the average value of the distance from the central point to each point of the section to be used as the radius value of the current central point; and calculating the average value of all the discrete central point radiuses of the blood vessel section as the radius of the current blood vessel section for the whole blood vessel section. In order to improve the accuracy of vessel segmentation, firstly, the centerline of each section of vessel is regarded as a graph node, then the arteriovenous probability output by the network in the step A2 is comprehensively considered in a graph segmentation algorithm, and information such as the radius, the length, the intersection angle and the like of the vessel section is extracted, and finally the vessel centerline is segmented by minimizing an energy optimization function of the graph segmentation algorithm, so that the optimized two-classification vessel centerline result is obtained; the energy optimization function of the graph cut algorithm is as follows:
E(A)＝λ*R(A)+B(A)
wherein A represents a vessel centerline as a graph node, E (A) represents a loss function, R (A) is a prior penalty term, B (A) is a region similarity penalty term, and λ is a balance factor;
step B4, using the centerline result classified in the step B3 to perform regional growth on the blood vessel by using a Grow Cuts algorithm; the Grow Cuts algorithm extracts the region of interest through local competitive region growth; in the step B3, obtaining two classified vessel centerline points, then performing region growth on the centerline points and the vessel results obtained in the step A2 by using a GrowCuts algorithm, and finally obtaining optimized arteriovenous segmentation results;
step C: constructing a topological structure of the blood vessel result optimized in the step B by utilizing a GrowCuts algorithm
Step C1, obtaining the segmented blood vessel by using the Grow Cuts algorithm
Firstly, for the segmentation result of the blood vessel obtained in the step B2, endowing the central line of each segment of the blood vessel with different label values, and then growing the whole blood vessel by using the label value of the central point of each segment as a seed point (removing a bifurcation point) by using a Grow Cuts algorithm; after the growth is completed, the blood vessels in the result are displayed in different colors due to different label values of different sections, so that different levels of the blood vessels are seen, namely, the preliminary level construction of the blood vessel tree is completed;
step C2, obtaining the radius of the blood vessel by using a section method
In order to obtain the radius information of the blood vessel, firstly, a section cutting operation is carried out on the blood vessel optimized in the step B4 by using a section algorithm, then, the section of the blood vessel is cut by identifying the normal vector of the plane where the interested central point is located, and as the sections of other points can be obtained in the cutting process, the obtained sections need to be screened, and after the screening, a target section is obtained; then obtaining coordinates of all points on the target section, and calculating the average value of the distances from the points to the central point, namely the radius of the blood vessel;
step C3, extracting the arteriovenous tree structure by using a sequence traversal method
In order to determine the vessel condition of each level in the artery or vein subtree more specifically, the subtree root node of the artery or vein is determined manually, then the vessels of each level are counted by using a sequence traversal, different label values are given to the vessels of each level according to the result of the step C1, then the vessels of each level are grown by using a Grow Cuts algorithm, and finally the hierarchical display of a certain subtree of the artery or vein is completed.
In conclusion, the invention is suitable for the segmentation and modeling work of the blood vessel in the medical image, and the blood vessel segmentation and modeling are efficiently and accurately carried out by mainly utilizing the high efficiency of the neural network and the graph cutting algorithm in combination with the anatomical knowledge.
The invention has the beneficial effects that: in the aspect of vessel segmentation, when the neural network is used for training, a partial supervision method is introduced, and the data of partial labels are used for training, so that the robustness of the algorithm is enhanced; and the geometric knowledge of the blood vessel is integrated into the graph cut algorithm by combining the thinning algorithm and the network topology algorithm, so that the result of the neural network is further optimized, and the accuracy of the cut result is improved. In the aspect of vessel modeling, the construction of a vessel topological structure is realized by utilizing algorithms such as growth, cross section and sequence traversal, and the like, and the whole structure is accurate and efficient. The two aspects are combined to form a set of complete algorithm flow for extracting the blood vessel from the medical image and constructing the topological structure of the blood vessel.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
Fig. 2 is a network structure view of a segmentation arteriovenous vascular model.
Fig. 3 is a flow chart of segmenting a vessel and optimizing the vessel segmentation result.
Fig. 4 is a schematic illustration of segmenting, optimizing blood vessels and constructing a topology of blood vessels.
Fig. 5 is a probability map of the radius distribution of a vessel segment.
Detailed Description
The present invention will be further explained with reference to the following embodiments by taking CT images of human lungs as an example. The flow of the modeling method for vessel segmentation in medical images based on topology knowledge is shown in fig. 1. The device mainly comprises three parts: establishing an artery and vein vessel segmentation model based on a U-Net neural network, and training by using a partial supervision method; carrying out post-processing work, and further optimizing the segmentation result of the blood vessel by algorithms such as result collocation refinement and graph segmentation of the neural network; and (3) constructing a topological structure of the blood vessel, and extracting the radius, segmentation, subtree and the like of the blood vessel by using algorithms such as GrowCuts, cross sections and the like. The method comprises the following specific steps:
step A: a model for segmenting the artery and vein is established based on a U-Net neural network, a partial supervision method is adopted, and the segmented artery and vein model is trained by utilizing image data and a labeling result thereof
Step A1, creating a network model
The network structure of the arteriovenous vessel segmentation model is shown in fig. 1, and the model adopts a structure of a U-Net network, wherein a left encoder performs down-sampling 4 times, and a right decoder performs up-sampling 4 times, so that an output image is restored to the size of an input image (the input image size of the network is 128 × 128 × 96). And (3) layer jump connection is also used in the model, the characteristic graph of the coder at each level and the characteristic graph of the decoder at each level are spliced, and finally, a segmentation result graph with the same size as the original image is output by the network through a Softmax function.
Step A2, model training
The network is trained using the existing dataset. And (3) adding partial labeled data in the data set, modifying a loss function of the network during training, and adding a post-processing step B during testing so as to improve the accuracy of arteriovenous segmentation. Through the network model, the output probability of the artery and the vein and the result of the vessel (including the artery and the vein) can be obtained.
In training, to train the data of the partial label correctly, batch _ Size is set to 1. During training, epochs are set to 200, and the loss function used by the network is a combination of a cross-entropy loss function and a Dice loss function:
cross entropy loss function:
wherein, y i For the vessel labeling result corresponding to sample i, p i Outputting a result for the network model corresponding to the sample i, wherein N is the number of samples;
dice loss function:
wherein, y i For the vessel labeling result corresponding to sample i, p i Outputting a result for the network model corresponding to the sample i, wherein N is the number of samples;
loss function used by the network:
L network ＝α*L 1 +β*L 2
wherein, alpha is the coefficient of the cross entropy loss function, and beta is the coefficient of the Dice loss function. The hyper-parameters alpha and beta are set to 1 and 1, respectively.
And B: carrying out post-processing optimization on the neural network result in the step A by utilizing algorithms such as graph cut and the like
Step B1, extracting the central line of the blood vessel by using a thinning algorithm
Firstly, extracting a central line from the blood vessel result in the step A, wherein the central line is mainly based on a thinning method, and the main process is to repeatedly delete boundary points by using a thinning algorithm under the condition of meeting topology invariance and geometric constraint conditions.
Three voxels, namely a simple voxel, an Euler invariant voxel, a boundary voxel and the like, are deleted in an iterative way to perform skeletonization on vessel voxels, and when the voxels are not changed any more after deletion, the integral skeleton, namely the corresponding central line, can be obtained. While maintaining overall connectivity using a sequential rechecking procedure when deleting points in order to result in topological relationships and geometric constraints.
Step B2, extracting a blood vessel map structure
After the vessel central line is obtained in the step B1, the network topology algorithm is used for the central line to obtain the graph structure of the vessel, including information of bifurcation points, segmentation points, local connection relations and the like.
Step B3, post-processing optimization is carried out by using a graph cut algorithm
And (3) performing post-processing optimization by combining the graph cut algorithm with the steps B1 and B2. The specific processing method is to regard each segmented midline as a graph node, and then calculate the length of the blood vessel, the local intersection angle and the radius by using the segmented point and local connection information in the step B2. And finally, comprehensively considering the arteriovenous output probability in the step A2 and the geometric characteristics (radius, length, cross angle and the like) of adjacent vessel sections in a graph cut algorithm, and realizing the segmentation of the vessel midline by minimizing an energy optimization function of the graph cut algorithm so as to obtain an optimized result of the two classified vessel midline. The energy optimization function of the graph cut algorithm is as follows:
E(A)＝λ*R(A)+B(A)
where A represents a vessel centerline as a graph node, E (A) represents a loss function, R (A) is a prior penalty term, B (A) is a region similarity penalty term, and λ is a balance factor. The hyper-parameter lambda is set to 1.
And step B4, performing regional growth on the blood vessels by using the centerline result of the two classified blood vessels obtained in the step B3 and using a Grow Cuts algorithm.
In the step B3, two classified midline points are obtained, and then the midline points, the blood vessel results in the step A2 and the like are subjected to region growth by utilizing a GrowCuts algorithm, and finally, the optimized arteriovenous segmentation result is obtained.
And C: constructing a topological structure of the blood vessel result optimized in the step B by utilizing a GrowCuts algorithm
Step C1, obtaining segmented blood vessels by using methods such as Grow Cuts algorithm and the like
After the centerline segmentation is performed in step B2, a different label value is assigned to each centerline, and then similar to step B4, the label value of the center point of each centerline is used as a seed point (excluding the bifurcation point) by using the Grow Cuts algorithm to Grow to the whole blood vessel. After the growth is completed, the blood vessels in the result are displayed in different colors due to different label values of different segments, so that different levels of the blood vessels can be seen, and the preliminary level construction of the blood vessel tree is completed.
Step C2, obtaining the radius of the blood vessel by using a cross section method and the like
In order to obtain the radius information of the blood vessel, firstly, a section cutting operation is carried out on the blood vessel optimized in the step B4 by using a section algorithm, then, the section cutting operation is carried out on the blood vessel by identifying the normal vector of the plane where the interested central point is located, then, the obtained sections are screened, and finally, the target section is obtained. Then obtaining the coordinates of all points on the section, and calculating the average value of the distances from the points to the central point, namely the radius of the blood vessel.
Step C3, extracting the arteriovenous tree structure by using methods such as sequence traversal and the like
Determining subtree root nodes of the artery or vein of the blood vessel result obtained in the step B4 in a manual selection mode, then carrying out traversal statistics on the blood vessels of each level by using a sequence, endowing different label values to the blood vessels of each level according to the result obtained in the step C1, then growing by using a Grow Cuts algorithm, and finally completing the hierarchical display of a certain subtree of the artery or vein.
The blood vessel segmentation modeling of the human lung CT image is taken as an example for illustration, and although the shape and the gray scale change pattern of different human organs are different, the idea and the method of the present invention are also applicable to other organ subjects, such as human brain blood vessels. Accordingly, several modifications made without departing from the principle and basic idea of the invention should be construed as the protection scope of the invention.
Claims (1)
1. A blood vessel segmentation modeling method in medical images based on topology knowledge is characterized by comprising the following steps:
step A: establishing a segmentation arteriovenous blood vessel model, namely a network model, based on a U-Net neural network, and training the segmentation arteriovenous blood vessel model by using image data and a labeling result thereof by adopting a partial supervision method;
step A1, establishing a model for segmenting arteriovenous blood vessels
The arteriovenous vascular model is divided by adopting a 3D U-Net network as a basic model, the left side of the U-Net neural network is an encoder part, and the right side of the U-Net neural network is a decoder part; down-sampling by a left encoder for K times, and up-sampling by a right decoder for K times correspondingly, and recovering an output image to the size of an input image, wherein the size of the input image is a fixed value; jump connection is also used in the arteriovenous blood vessel segmentation model, and the characteristic diagram of each level of coder and the characteristic diagram of each level of decoder are spliced to ensure the fineness of the segmentation result;
step A2, training and segmenting arteriovenous blood vessel model
Training with an existing data set; in order to fully use the existing data as much as possible and improve the robustness of the network, partial tagged data are added into a data set based on a partial supervision method; during training, modifying a loss function for segmenting the arteriovenous vascular model; during testing, a post-processing step B is added to improve the accuracy of arteriovenous segmentation; obtaining the output probability and the blood vessel result of the artery and the vein through a network model;
during training, the loss function used by the network model is a combination of a cross entropy loss function and a Dice loss function:
cross entropy loss function:
wherein, y i For the vessel labeling result corresponding to sample i, p i Mesh for sample iOutputting a result by the complex model, wherein N is the number of samples;
dice loss function:
wherein, y i For the vessel labeling result corresponding to sample i, p i Outputting a result for the network model corresponding to the sample i, wherein N is the number of samples;
loss function used for segmenting arteriovenous vascular models:
L network ＝α*L 1 +β*L 2
wherein, alpha is the coefficient of the cross entropy loss function, and beta is the coefficient of the Dice loss function;
and B: carrying out post-processing optimization on the blood vessel result segmented by the arteriovenous blood vessel model in the step A by utilizing algorithms such as graph cutting and the like
Step B1, extracting the central line of the blood vessel by using a thinning algorithm
Firstly, extracting a central line from a blood vessel result obtained by segmenting the arteriovenous blood vessel model in the step A based on a refinement algorithm, wherein the main process is to repeatedly delete boundary points by using the refinement algorithm under the condition of meeting topology invariance and geometric constraint; simultaneously, parallelly rechecking eight sub-voxels with non-overlapping neighborhoods until a connected point set serving as a skeleton is obtained;
the detailed process of the refined algorithm is as follows:
in a three-dimensional blood vessel image obtained by segmentation, a local cube of 3 multiplied by 3 is firstly constructed for a certain blood vessel voxel to check the local connectivity of the blood vessel voxel and is represented by a 26 neighborhood and a 6 neighborhood; meanwhile, a local 3 x 3 cube is divided into 8 mutually overlapped 2 x 2 cubes by a central voxel; then, the following definitions are thus given:
v,N 6 (v),N 26 (v),N 2 (v)
wherein v denotes a vessel voxel, N 6 (v) 6 neighborhoods, N, representing vessel voxels v 26 (v) 2 representing a vessel voxel v6 neighborhoods, N 2 (v) A2 × 2 × 2 cube representing a part including a blood vessel voxel v;
during the refinement algorithm processing, the boundary of the three-dimensional blood vessel image is obtained from six directions, and the blood vessel voxels of the three-dimensional blood vessel image are divided into three types: boundary voxel V B Euler invariant voxel V E And simple voxel V S ；
The definition of the three types of voxels is as follows:
given a vessel voxel v ∈ S, all vessel voxels have a value of 1, and the corresponding complement set is
Corresponding to background, value 0; if only one of its connected objects belongs to the background, the vessel voxel is considered as boundary voxel V B ；
Given a vessel voxel v ∈ S, if the Euler number does not change when removing the voxel v from S, the vessel voxel is considered as a Euler invariant voxel; the euler number is specifically defined as:
V E ＝{v∣δG 26 (S∩N 2 (v))＝0}
wherein δ represents the amount of change in the parameter values before and after the deletion of the body vessel factor v, G 26 (S∩N 2 (v) N for all vessel voxels S) 2 (v) The euler numbers of the 26 neighborhoods in the set,
represents all vessel voxels->
N of (A) 2 (v) Euler numbers of 6 neighborhoods in the set; for determining N in the voxel v containing blood vessels 2 (v) In the set, the ith N 2 (v) Whether all 8 vertices of (1), if yes, oct i Setting to 1, otherwise setting to 0; f. of i 、e i Denotes the ith N 2 (v) The number of faces and sides of the cube; delta oct i ，δf i ,δe i Representing the corresponding variation value of each parameter after removing the vessel voxel v;
given a vessel voxel V ∈ S, if the connectivity of its 26 neighborhoods remains unchanged when V is removed from S, then the vessel voxel is considered as a simple voxel V S ；
Performing skeletonization by iteratively deleting the three voxels from the vessel region, wherein the vessel voxels are not changed after deletion, namely the whole skeleton, namely the corresponding central line, is obtained; meanwhile, in order to meet the topological relation and geometric constraint of the result, a sequence rechecking program is used for keeping the integral connectivity when the point is deleted;
step B2, extracting the graph structure of the blood vessel
After obtaining the midline of the blood vessel in the step B1, performing bifurcation point identification and segmentation processing on the midline of the blood vessel through a network topology algorithm, thereby obtaining a graph structure of the blood vessel; the method comprises the following specific steps:
firstly, identifying a vessel bifurcation point by using a 26 neighborhood of a vessel central point voxel in a vessel midline, searching adjacent bifurcation points from different directions for each bifurcation point according to the extracted bifurcation point, and then counting the adjacent bifurcation points and a central point on a search path to obtain a section of vessel midline; then, different branches are searched on the same bifurcation point, and the local blood vessel midline connection relation is extracted; finally, traversing the central line of the whole blood vessel to obtain the graph structure of the blood vessel;
step B3, post-processing optimization is carried out by using a graph cut algorithm
Although the blood vessels are obtained by segmenting the arteriovenous blood vessel segmentation model in the step A, the blood vessels are adhered and wrongly segmented, so that the post-processing optimization is carried out by combining the graph segmentation algorithm with the steps B1 and B2;
the specific processing method is that the length of each blood vessel section and the intersection angle of the adjacent blood vessel sections are calculated by utilizing the blood vessel middle line section extracted in the step B2 and the local connection relation; for a certain central point of a blood vessel section, cutting the section of the blood vessel in the step A2 according to the position of the central point, and then obtaining a target section after screening; then calculating the average value of the distance from the central point to each point of the cross section as the radius value of the current central point; calculating the average value of the radiuses of all the discrete central points of the blood vessel section as the radius of the current blood vessel section; in order to improve the accuracy of vessel segmentation, firstly, the centerline of each section of vessel is regarded as a graph node, then the arteriovenous probability output by the network in the step A2 and the radius, length and crossing angle of the extracted vessel section are comprehensively considered in a graph segmentation algorithm, and finally the vessel centerline is segmented by minimizing the energy optimization function of the graph segmentation algorithm, so that the optimized two-classification vessel centerline result is obtained; the energy optimization function of the graph cut algorithm is as follows:
E(A)＝λ*R(A)+B(A)
wherein A represents a vessel centerline as a graph node, E (A) represents a loss function, R (A) is a prior penalty term, B (A) is a region similarity penalty term, and λ is a balance factor;
step B4, using the centerline result classified in the step B3 to perform regional growth on the blood vessel by using a Grow Cuts algorithm;
the Grow Cuts algorithm extracts the region of interest through local competitive region growth; in the step B3, obtaining two classified vessel centerline points, then performing region growth on the centerline points and the vessel results in the step A2 by using a GrowCuts algorithm, and finally obtaining optimized arteriovenous segmentation results;
step C: constructing a topological structure of the blood vessel result optimized in the step B by utilizing a GrowCuts algorithm
Step C1, obtaining the segmented blood vessel by using the Grow Cuts algorithm
Firstly, for the segmentation result of the blood vessel obtained in the step B2, endowing the central line of each segment of the blood vessel with different label values, and then growing the blood vessel by using the label value of the central point of each segment as a seed point by using a Grow Cuts algorithm; after the growth is completed, the blood vessels in the result are displayed in different colors due to different label values of different sections, so that different levels of the blood vessels are seen, and the preliminary level construction of the blood vessel tree is completed;
step C2, obtaining the radius of the blood vessel by using a section method
In order to obtain the radius information of the blood vessel, firstly, a section cutting operation is carried out on the blood vessel optimized in the step B4 by using a section algorithm, then, the section of the blood vessel is cut by identifying the normal vector of the plane where the interested central point is located, and as the sections of other points can be obtained in the cutting process, the obtained sections need to be screened, and after the screening, a target section is obtained; then obtaining coordinates of all points on the target section, and calculating an average value of distances from the points to a central point, namely the radius of the blood vessel;
step C3, extracting the arteriovenous tree structure by using a sequence traversal method
In order to determine the vessel condition of each level in the artery or vein subtree more specifically, the subtree root node of the artery or vein is determined manually, then the vessels of each level are counted by using a sequence traversal, different label values are given to the vessels of each level according to the result of the step C1, then the vessels of each level are grown by using a Grow Cuts algorithm, and finally the hierarchical display of a certain subtree of the artery or vein is completed.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211408809.9A CN115908297A (en) | 2022-11-11 | 2022-11-11 | Topology knowledge-based blood vessel segmentation modeling method in medical image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211408809.9A CN115908297A (en) | 2022-11-11 | 2022-11-11 | Topology knowledge-based blood vessel segmentation modeling method in medical image |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115908297A true CN115908297A (en) | 2023-04-04 |
Family
ID=86479874
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211408809.9A Pending CN115908297A (en) | 2022-11-11 | 2022-11-11 | Topology knowledge-based blood vessel segmentation modeling method in medical image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115908297A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116342608A (en) * | 2023-05-30 | 2023-06-27 | 首都医科大学宣武医院 | Medical image-based stent adherence measurement method, device, equipment and medium |
| CN116503605A (en) * | 2023-06-01 | 2023-07-28 | 南京大学 | Pancreatic peripheral blood vessel segmentation marking method based on iterative trunk growth and weak supervision learning |
| CN117036253A (en) * | 2023-07-18 | 2023-11-10 | 强联智创（北京）科技有限公司 | Method for training a segmentation model for segmenting cerebral vessels and related products |
-
2022
- 2022-11-11 CN CN202211408809.9A patent/CN115908297A/en active Pending
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116342608A (en) * | 2023-05-30 | 2023-06-27 | 首都医科大学宣武医院 | Medical image-based stent adherence measurement method, device, equipment and medium |
| CN116342608B (en) * | 2023-05-30 | 2023-08-15 | 首都医科大学宣武医院 | Medical image-based stent adherence measurement method, device, equipment and medium |
| CN116503605A (en) * | 2023-06-01 | 2023-07-28 | 南京大学 | Pancreatic peripheral blood vessel segmentation marking method based on iterative trunk growth and weak supervision learning |
| CN116503605B (en) * | 2023-06-01 | 2023-10-13 | 南京大学 | Pancreatic peripheral blood vessel segmentation marking method based on iterative trunk growth and weak supervision learning |
| CN117036253A (en) * | 2023-07-18 | 2023-11-10 | 强联智创（北京）科技有限公司 | Method for training a segmentation model for segmenting cerebral vessels and related products |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107563983B (en) | Image processing method and medical imaging device | |
| CN111709953B (en) | Output method and device in lung lobe segment segmentation of CT (computed tomography) image | |
| JP5478171B2 (en) | Method and apparatus for classification of coronary image data | |
| Carvalho et al. | 3D segmentation algorithms for computerized tomographic imaging: a systematic literature review | |
| CN115908297A (en) | Topology knowledge-based blood vessel segmentation modeling method in medical image | |
| CN113674281A (en) | Liver CT automatic segmentation method based on deep shape learning | |
| Chen et al. | Tree-branch-searching multiresolution approach to skeletonization for virtual endoscopy | |
| CN110288611A (en) | Coronary vessel segmentation method based on attention mechanism and full convolutional neural networks | |
| CN114693933A (en) | Medical image segmentation device based on generation of confrontation network and multi-scale feature fusion | |
| CN113160120A (en) | Liver blood vessel segmentation method and system based on multi-mode fusion and deep learning | |
| KR101625955B1 (en) | Method of classifying artery and vein of organ | |
| Nardelli et al. | Deep-learning strategy for pulmonary artery-vein classification of non-contrast CT images | |
| Yao et al. | Tag-net: topology-aware graph network for centerline-based vessel labeling | |
| Liu et al. | Robust optimization-based coronary artery labeling from X-ray angiograms | |
| Zhang et al. | Pathological airway segmentation with cascaded neural networks for bronchoscopic navigation | |
| Wang et al. | Naviairway: a bronchiole-sensitive deep learning-based airway segmentation pipeline for planning of navigation bronchoscopy | |
| CN112638262A (en) | Similarity determination device, method, and program | |
| Wang et al. | Naviairway: a bronchiole-sensitive deep learning-based airway segmentation pipeline | |
| Ramos et al. | Fast and smart segmentation of paraspinal muscles in magnetic resonance imaging with CleverSeg | |
| CN111986216A (en) | RSG liver CT image interactive segmentation algorithm based on neural network improvement | |
| CN116958705A (en) | Medical image classifying system based on graph neural network | |
| Zhao et al. | Automatic aortic dissection centerline extraction via morphology-guided CRN tracker | |
| Shao et al. | A segmentation method of airway from chest ct image based on vgg-unet neural network | |
| CN116310114A (en) | Coronary artery point cloud data processing optimization method of multitask label model | |
| Affane et al. | Robust deep 3-d architectures based on vascular patterns for liver vessel segmentation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |