A method for measuring the thickness of a blood vessel wall based on coronary reconstruction

Abstract

The present application belongs to the technical field of medical image processing, and particularly relates to a blood vessel wall thickness measurement method based on coronary artery reconstruction. The technical scheme provided by the present application is based on multi-row single heartbeat CT imaging, uses Laplace smoothing mapping to select the best heart cycle, uses a U-net architecture with dilated convolution to eliminate image artifacts, reconstructs the coronary artery, and performs segmentation of the blood vessel wall and the lumen of the reconstructed coronary artery, and uses the maximum inscribed sphere method to measure the blood vessel wall thickness. The problems of image artifacts in coronary artery imaging caused by heart movement, discontinuity of adjacent imaging caused by blood vessel contraction, and inability to accurately reconstruct are solved, and the problem of selecting the best period for measuring the coronary blood vessel wall thickness is solved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of medical image processing, and in particular to a blood vessel wall thickness measurement method based on coronary artery reconstruction. BACKGROUND

[0002] Coronary artery wall thickness refers to the thickness of the intima, media and adventitia of the coronary artery, and is an important indicator for assessing the risk of cardiovascular diseases. Typical cardiovascular diseases include atherosclerosis and hypertension. Atherosclerosis is a chronic inflammatory disease characterized by the accumulation of lipid plaques in the arterial wall. These plaques may thicken and calcify over time, causing the vessel wall to thicken and eventually leading to vascular stenosis. Hypertension is a condition in which the pressure exerted by blood flow on the vessel wall is consistently higher than normal. The vessel wall is stretched and thinned under long-term high pressure. These diseases are characterized by being difficult to detect before onset and difficult to screen early, but they all cause micro-deformation of the vessel wall. Therefore, it is necessary to measure the thickness of the coronary artery vessel wall.

[0003] The prerequisite for measuring the thickness of the coronary artery vessel wall is to obtain coronary artery imaging, and then to segment the captured image. The existing coronary artery imaging methods include coronary CT, coronary angiography, and vascular ultrasound. At present, due to the significant impact of heartbeats on coronary artery imaging, causing image artifacts, regardless of the imaging method used, the subsequent processing approach for obtaining complete and artifact-free coronary imaging is generally consistent, divided into two steps of artifact elimination and image reconstruction.

[0004] The patent with publication number "CN117064446A Vascular dynamic three-dimensional reconstruction system based on intravascular ultrasound" provides a heart vessel dynamic reconstruction method based on intravascular ultrasound imaging. This method performs three-dimensional reconstruction on the two-dimensional images obtained by intravascular ultrasound imaging, establishes an artifact elimination network to extract accurate blood vessel walls, and uses a generative adversarial network to compensate for frame missing after extracting key frames of the intravascular ultrasound sequence. Finally, based on the heart motion cycle, dynamic three-dimensional reconstruction of the heart vessel is achieved. However, the changes in blood vessel diameter and wall thickness caused by blood vessel contraction under heart motion result in different coronary vessel wall thicknesses at the same position captured at different times in imaging, making it extremely difficult to choose the timing of blood vessel wall thickness measurement. The selection of the best heart cycle and optimal phase is of great significance in blood vessel wall thickness measurement. This method includes multiple heart cycles in a single intravascular ultrasound imaging, aligns the lumen area of the blood vessel during the contraction and expansion periods under different heart cycles based on the gating frame, but ignores the consideration of blood vessel reconstruction in the best heart cycle. The blood vessel wall thickness obtained by segmentation after blood vessel reconstruction is dynamic and cannot accurately reflect the micro-deformation of the blood vessel. SUMMARY

[0005] To address the above issues, the present invention proposes a method for measuring vascular wall thickness based on coronary artery reconstruction, which can eliminate artifacts, solve the impact of lumen area caused by vascular contraction, reconstruct coronary arteries during the optimal cardiac cycle, and measure the coronary artery wall thickness, thereby improving the accuracy of the results.

[0006] The technical solution of the present invention is:

[0007] A method for measuring vessel wall thickness based on coronary artery reconstruction comprises the following steps:

[0008] S1. Use multi-slice single heartbeat CT imaging to acquire images of specified phases within multiple cardiac cycles within a set time. Define the number of images captured in a multi-slice single heartbeat CT imaging session as C, thereby obtaining an image dataset at different phases at various locations of the coronary artery. Define the image sequence obtained at coronary artery location c as V c ;

[0009] S2. Select the optimal adjacent phase single heartbeat CT imaging based on the collected data set, specifically:

[0010] The image sequence V c Mapping from high-dimensional space to low-dimensional space, and then selecting the cluster center through clustering algorithm, obtain the cluster center distance of CT imaging at the coronary artery position c in adjacent phases, which is defined as Then, the average value of the cluster center distance of each coronary artery position in adjacent phases is obtained according to the number of images taken in a multi-row single heartbeat CT imaging, which is expressed as Finally, the phase that achieves the best imaging effect in all sequences is obtained Expressed as K represents the number of cardiac cycles;

[0011] S3. Build an image artifact removal network to reconstruct the coronary arteries, specifically:

[0012] The image artifact removal network includes a deformable U-net, a scaling layer, and a spatial transformation network; the input of the deformable U-net is the forward sequence and the reverse sequence of the image, based on the obtained phase In chronological order from each V c Extraction in phase Images captured at the same time to obtain continuous frames Thus, we obtain the forward sequence And record it as continuous frame F, and get the reverse sequence at the same time And record it as continuous frame R; F and R are extracted by deformable U-net, and then the segmented image is obtained by 1x1x1 convolution and activation with a stride of 1. and The velocity field is obtained by convolution of 5x5x5 with a stride of 1 and Velocity field and After scaling the layer, the deformation field is obtained and The deformation field and segmentation results are input into the spatial transformation network to obtain a de-artifacted image sequence. Finally, the de-artifacted image sequence is passed through the reconstruction network to perform 3D reconstruction of the coronary arteries and obtain a 3D coronary artery map.

[0013] S4. Using the obtained three-dimensional image of the coronary artery, the maximum inscribed sphere method is used to measure the vessel wall thickness.

[0014] Furthermore, 128-slice single-heartbeat CT imaging was used, with a set time of 12-15 seconds, to capture 20 cardiac cycles, and a preset phase interval of 3% between each two single-heartbeat CT imagings to capture images within the phase range of 30%-60% of the systolic phase and 60%-90% of the diastolic phase.

[0015] Furthermore, the Laplace smoothing map is used to transform the image sequence V c Mapping from high-dimensional space to low-dimensional space, specifically the objective function optimized by the Laplace eigenmap is:

[0016]

[0017] Among them, V c As a set of data instances, y i 、y j is the vector representation of i and j in the data instance in the m-dimensional subspace, W ij The adjacency matrix constructed

[0018]

[0019] Where σ is a specified constant.

[0020] Furthermore, the method for obtaining the cluster center distance is:

[0021] For any phase n, the sum of its distances to the cluster centers of the adjacent phases n-1 and n+1 is defined as

[0022]

[0023] in It represents the distance between the cluster centers of the images at the same position c of the coronary artery captured at the 30%+(n-1)×3% phase and the 30%+n×3% phase.

[0024] Further, the deformable U-net network includes five layers of deformable U-net, the encoding stage is a shrinkage path, and the decoding stage is an expansion path; there is a series connection in the shrinkage path, and features extracted from normal convolution and dilated convolution are combined, in the expansion path, the feature map is up-sampled to the original resolution of the input image for inter-frame registration and segmentation; a series connection is used between the shrinkage path and the expansion path, and high-resolution features are fed into the up-sampled output.

[0025] Two sequence images are segmented in sequence by 1x1x1 convolution with a stride of 1 and an activation function to obtain a segmented image And In order to simultaneously retain global shape similarity and boundary details, the segmentation loss adopts a hybrid segmentation loss combining a class-weighted Tversky loss And a class-weighted cross-entropy loss

[0026]

[0027] Wherein And is a true segmentation containing partial frame annotations, gt represents the ground truth value, is calculated on the marked voxels;

[0028] A pixel velocity field is obtained by 5x5x5 convolution with a stride of 1 And That is, the displacement direction and motion velocity of each pixel are obtained. In the velocity field And , define And are the velocity fields extracted from the image frames obtained at the same position c when input in the forward order and in the reverse order respectively;

[0029] The velocity fields And obtained after the scaling layer are deformation fields And Define as the image frames obtained at position c in consecutive frames F and R, and record And respectively represent the deformation fields of to and to after de-artifacting; define the pixel smoothing loss as:

[0030]

[0031] Wherein N is the number of pixel points, i represents a pixel point, represents a deformation field The deformation field at pixel i in , where x and y represent the position of pixel i;

[0032] Furthermore, the deformation field and the segmented image are transformed through the spatial transformation network to obtain the image sequence after artifact removal, including the forward sequence and reverse sequence This image series has been segmented into membrane boundaries, including the inner and outer membranes;

[0033] In the spatial transformer network, deformation fields are used to segment images. and Perform spatial transformation:

[0034]

[0035] Where ◎ represents pixel multiplication, and Respectively and The segmented image at position c in , and Represents consecutive frames and The image frame obtained at position c in .

[0036] Define the loss function for reconstructing artifact-free image sequences

[0037]

[0038] Among them, u∈{0,1,2} represents three regions, namely the region outside the vessel wall, the region between the vessel wall and the intima boundary, and the region inside the intima. Ω(u≠0) represents all regions except the vessel wall, p represents the voxel belonging to the Ω(u≠0) region in the image frame, and R i (p) represents the image intensity at voxel p in the i-th frame in the continuous frame R, F i (p) represents the image intensity at voxel p in the i-th frame in the continuous frame F, Reconstructing continuous frames The image intensity at voxel p in the i-th frame, Reconstructing continuous frames The image intensity at voxel p in the i-th frame, and They represent the average image intensity of the i-th frame in the corresponding continuous frames;

[0039] Furthermore, the 3D coronary artery reconstruction method is as follows: the artifact-free forward sequence is passed through the reconstruction network, stacked along the membrane boundary along the coronary CT acquisition direction, linear interpolation is performed on the continuous 2D contours, and then the 3D surface is visualized by the marching cubes algorithm;

[0040] The total loss function of the network is:

[0041]

[0042] Among them, λ1, λ2 and λ3 are the parameters of various types of losses respectively. The network parameters are obtained by pre-training the network through the collected data based on the loss function.

[0043] Furthermore, the specific method of measuring the thickness of the blood vessel wall using the maximum inscribed sphere method is:

[0044] The maximum inscribed sphere method was used to draw the centerline of the coronary artery and lumen. tol and l q , calculate the thickness of the blood vessel wall at each position, define the same plane perpendicular to the entrance of the blood vessel in this section and the two midlines l tol and l q The intersection points are o tol and o q The radius of the intangent plane is r tol and r q , with o tol and o q The direction of the line is the x-axis, o q Take O as the origin and make a plane coordinate system. The thickness of the blood vessel wall at any point on the plane is m i for:

[0045]

[0046] where θ i is the angle between the line from the inscribed plane i to the origin and the x-axis, and r is the midline of the two lines l on the plane tol and l q Intersection o tol and o q straight-line distance.

[0047] The beneficial effects of the present invention are that, compared with existing technologies, the technical solution proposed in this invention is based on multi-slice single-heartbeat CT imaging, uses Laplace smoothing mapping to select the optimal cardiac cycle, uses a U-net architecture with dilated convolution to eliminate image artifacts, and performs segmentation of the vessel wall and lumen. The coronary arteries are reconstructed using a reconstruction network, and the maximum inscribed sphere method is used to measure vessel wall thickness. This solves the problems of coronary artery imaging artifacts caused by cardiac motion, discontinuity of adjacent images caused by vasoconstriction, and the inability to accurately reconstruct them, as well as the problem of selecting the optimal period for coronary vessel wall thickness measurement. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is the overall flow chart of the present invention.

[0049] Figure 2Schematic diagram of the structure of the coronary artery reconstruction model of the present invention.

[0050] Figure 3 Schematic diagram of the deformable U-net structure.

[0051] Figure 4 Schematic diagram of the maximum inscribed sphere method for calculating blood vessel wall thickness. DETAILED DESCRIPTION

[0052] The technical principles and solutions of the present invention are described in detail below with reference to the accompanying drawings:

[0053] like Figure 1 As shown, the present invention mainly includes the following steps:

[0054] 1) Original data set collection, specifically:

[0055] 128-slice single-beat CT imaging was used to capture 20 cardiac cycles, lasting approximately 12-15 seconds. The preset phase interval between each two single-beat CT images was 3%, and images were collected within the phase range of 30%-60% of the systolic phase and 60%-90% of the diastolic phase.

[0056] The number of images captured in a multi-slice single heartbeat CT scan is denoted as C. The images captured at different phases at different locations of the coronary artery are represented as Where V c is a sequence of images taken at the coronary artery position c in time sequence (i.e., at 20 different phases),

[0057] 2) Based on the collected data set, select the optimal adjacent phase single heartbeat CT imaging, specifically:

[0058] Input image sequence V c , Laplace smoothing mapping is used to map the real image from high-dimensional space to low-dimensional space.

[0059] The Laplacian eigenmap reconstructs the local structural features of the data manifold by constructing a graph with an adjacency matrix W. The main idea is that if two data instances i and j are very similar, then i and j should be as close as possible in the target subspace after dimensionality reduction. c As a set of data instances. Assume that the number of data instances is n, and the dimension of the target subspace, i.e. the final dimension reduction target, is m. Define a matrix Y of size n×m, where each row vector y i is the vector representation of data instance i in the target m-dimensional subspace (i.e., data instance i after dimensionality reduction). Our goal is to keep similar data instances i and j as close as possible in the target subspace after dimensionality reduction. Therefore, the objective function of the Laplace eigenmap optimization is as follows:

[0060] min∑ i ||y i -y j || 2 W ij (1)

[0061] Among them yi, y j is the vector representation of data instances i and j in the m-dimensional subspace, Wi j The adjacency matrix constructed is expressed as:

[0062]

[0063] Where σ is a specified constant.

[0064] After obtaining the low-dimensional feature vector, the AP clustering algorithm (Affinity propagation) is used to select the final cluster center.

[0065] For comparison, input image sequence V c The positions of the cluster centers of adjacent images in the input image sequence V c Medium Image Two adjacent phase images and The sum of the distances to the cluster centers is Where c∈[1,C], n∈[1,20].

[0066] For any phase n, take the adjacent phases n-1, n and n+1 (except when n=1 or n=20) before and after it, and record the cluster center distance sequence of the CT imaging corresponding to each coronary artery position in this group of adjacent phases as follows:

[0067]

[0068] in It represents the distance between the cluster centers of the images at the same position c of the coronary artery captured at the 30%+(n-1)×3% phase and the 30%+n×3% phase.

[0069] Calculate the average distance between the cluster centers of each coronary artery position in each adjacent phase

[0070]

[0071] Where C is the number of images captured in a multi-slice single heartbeat CT imaging session, is the cluster center distance of CT imaging at coronary artery position c in adjacent phases.

[0072] Comprehensively select the phase that achieves the best imaging effect (i.e., the least artifact effect) among all sequences Right now

[0073] 3) Build Figure 2 The image artifact removal network shown in the figure is used to reconstruct the coronary arteries, specifically:

[0074] (1) From each V c Extraction in phase Images captured at the same time to obtain continuous frames The forward sequence Recorded as continuous frame F (Forward Image), the reverse sequence It is recorded as continuous frame R (Reverse Image).

[0075] In the forward sequence In the c-th frame, the phase The image captured at position c is recorded as In reverse sequence In the c-th frame, the phase The image taken at position (C-c+1) is recorded as However, whether in F or R, The meaning is the same, both indicate the phase The image is taken at position c at time .

[0076] (2) The above continuous frames are input into a five-layer deformable U-net network, that is, dilated convolution is introduced into the U-net to obtain a larger receptive field, and the velocity field and segmentation image are finally obtained through inter-frame registration and segmentation.

[0077] like Figure 3 As shown in Figure 1, the deformable U-net architecture has two main paths: a contracting path during the encoding phase and an expanding path during the decoding phase. In the contracting path, concatenation is used to combine features extracted from normal and dilated convolutions to extract features at different scales. In the expanding path, feature maps are upsampled to the original resolution of the input image for inter-frame registration and segmentation. Furthermore, concatenation is used between the contracting and expanding paths to feed high-resolution features into the upsampled output, improving accuracy.

[0078] At the end of the deformable U-net architecture decoder, the pixel velocity field is obtained by a 5x5x5 convolution with a stride of 1. and That is to obtain the displacement direction and movement speed of each pixel. and middle, and They are the velocity fields extracted when the image frames obtained at the same position c are input in positive order and reverse order respectively.

[0079] At the same time, at the end of the deformable U-net architecture decoder, the two sequence images are segmented in sequence through a 1x1x1 convolution with a stride of 1 to obtain the segmented image and In order to preserve both global shape similarity and boundary details, a hybrid segmentation loss is adopted that combines class-weighted Tversky loss. and class-weighted cross entropy loss

[0080]

[0081] in and is the true segmentation containing partial frame annotations, and gt represents the ground truth. Computed over labeled voxels.

[0082] (3) Velocity field and After scaling the layer, the deformation field is obtained and definition is the image frame obtained at position c in the continuous frames F and R, and and Respectively represent the artifacts after arrive The deformation field and artifact removal arrive The deformation field of ; the pixel smoothing loss is defined as:

[0083]

[0084] Where N is the number of pixels, i represents the pixel, Representing the deformation field The deformation field at pixel i in , where x and y represent the position of pixel i.

[0085] (4) Use the deformation field and segmentation image to obtain the de-artifacted image sequence through the spatial transformation network.

[0086] The deformation field and the segmented image are input into the spatial transformation network to obtain the image sequence after artifact removal, including the forward sequence and reverse sequence The image series has been segmented into membrane boundaries, including the inner and outer membranes.

[0087] In the spatial transformer network, deformation fields are used to segment images. and Perform spatial transformation:

[0088]

[0089] Where ◎ represents pixel multiplication, and Respectively and The segmented image at position c in , and Represents consecutive frames and The image frame obtained at position c in .

[0090] In this method, the Normalized Cross Correlation (NCC) principle is used to measure the similarity between the images reconstructed from the forward sequence and the reverse sequence.

[0091]

[0092] Among them, u∈{0,1,2} represents three regions, namely the region outside the vessel wall, the region between the vessel wall and the intima boundary, and the region inside the intima. Ω(u≠0) represents all regions except the vessel wall, p represents the voxel belonging to the Ω(u≠0) region in the image frame, and R i (p) represents the image intensity at voxel p in the i-th frame in the continuous frame R, F i (p) represents the image intensity at voxel p in the i-th frame in the continuous frame F, Reconstructing continuous frames The image intensity at voxel p in the i-th frame, Reconstructing continuous frames The image intensity at voxel p in the i-th frame, and They represent the average image intensity of the i-th frame in the corresponding continuous frames.

[0093] Get the loss function of the reconstructed artifact-free image sequence

[0094]

[0095] (5) Use the reconstruction network to perform three-dimensional reconstruction of the coronary arteries based on the forward artifact removal sequence.

[0096] The reconstruction network is as follows: the membrane boundaries are stacked along the coronary CT acquisition direction, linear interpolation is performed on the consecutive 2D contours, and the 3D surface is then visualized using the marching cubes algorithm.

[0097] Finally, the total loss function for coronary artery reconstruction is:

[0098]

[0099] Where λ1, λ2 and λ3 are the parameters of each type of loss respectively.

[0100] 4) Measure vessel wall thickness using the maximum inscribed sphere method, specifically:

[0101] The maximum inscribed sphere method was used to draw the centerline of the coronary artery and lumen. tol and l q Calculate the thickness of the blood vessel wall at each location. The same plane perpendicular to the entrance of this section of blood vessels and the two midlines l tol and l q The intersection points are o tol and o q The radius of the intangent plane is r tol and r q , with o tol and o q The direction of the line is the x-axis, o q Take O as the origin and make a plane coordinate system. The thickness of the blood vessel wall at any point on the plane is m i for:

[0102]

[0103] where θ i is the angle between the line from the inscribed plane i to the origin and the x-axis, and r is the midline of the two lines l on the plane tol and l q Intersection o tol and o q straight-line distance.

[0104] In summary, the present invention has the following characteristics and advantages:

[0105] The present invention uses 128-row single-heartbeat CT imaging to collect coronary artery images of 20 cardiac cycles, uses Laplace smoothing mapping to select the optimal cardiac cycle, constructs an image artifact removal network, and performs segmentation of the vascular wall and lumen. The coronary artery is reconstructed through the reconstruction network, and the maximum inscribed sphere method is used to measure the vascular wall thickness.

[0106] Because cardiac motion causes vasoconstriction, the lumen area and vessel wall thickness at the same location vary at different phases. This results in uneven coronary artery images captured at adjacent locations during a single imaging session. Therefore, the use of 128-slice single-heartbeat CT imaging technology solves the problem of simultaneous imaging of various locations during coronary artery imaging and overcomes the acquisition difficulties associated with long imaging times.

[0107] Because cardiac motion can significantly affect coronary artery imaging during CT imaging, resulting in artifacts, a network is needed to eliminate these artifacts and select the optimal cardiac cycle for coronary artery reconstruction. Laplace smoothing is used to select the optimal phase, and a U-net architecture with dilated convolutions is used to extract detailed features, construct a deformation field, and reconstruct the coronary arteries at the optimal phase time.

[0108] For the segmented coronary arteries, the maximum inscribed sphere method is used to obtain the overall coronary artery and the midline of the lumen, and the maximum inscribed sphere plane is used to calculate the vessel wall thickness at any point in the coronary artery.

Claims

1. A method for measuring vascular wall thickness based on coronary artery reconstruction, characterized in that: The following steps are involved: S1. Use multi-slice single heartbeat CT imaging to acquire images of specified phases within multiple cardiac cycles within a set time. Define the number of images captured in a multi-slice single heartbeat CT imaging session as C, thereby obtaining an image dataset at different phases at various locations of the coronary artery. Define the image sequence obtained at coronary artery location c as V c ; S2. Select the optimal adjacent phase single heartbeat CT imaging based on the collected data set, specifically: The image sequence V c Mapping from high-dimensional space to low-dimensional space, and then selecting the cluster center through clustering algorithm, obtain the cluster center distance of CT imaging at the coronary artery position c in adjacent phases, which is defined as Then, the average value of the cluster center distance of each coronary artery position in adjacent phases is obtained according to the number of images taken in a multi-row single heartbeat CT imaging, which is expressed as Finally, the phase that achieves the best imaging effect in all sequences is obtained Expressed as K represents the number of cardiac cycles; S3. Build an image artifact removal network to reconstruct the coronary arteries, specifically: The image artifact removal network includes a deformable U-net, a scaling layer, and a spatial transformation network; the input of the deformable U-net is the forward sequence and the reverse sequence of the image, based on the obtained phase In chronological order from each V c Extraction in phase Images captured at the same time to obtain continuous frames Thus, we obtain the forward sequence And record it as continuous frame F, and get the reverse sequence at the same time And record it as continuous frame R; F and R are extracted by deformable U-net, and then the segmented image is obtained by 1x1x1 convolution and activation with a stride of 1. and The velocity field is obtained by convolution of 5x5x5 with a stride of 1 and Velocity field and After scaling the layer, the deformation field is obtained and The deformation field and the segmentation results are input into the spatial transformation network to obtain an artifact-free image sequence. Finally, the artifact-free image sequence is passed through the reconstruction network to perform three-dimensional reconstruction of the coronary arteries to obtain a three-dimensional image of the coronary arteries. S4. Using the obtained three-dimensional image of the coronary artery, the maximum inscribed sphere method is used to measure the thickness of the vascular wall.

2. The method for measuring vascular wall thickness based on coronary artery reconstruction according to claim 1, characterized in that: In S1, 128-slice single-beat CT imaging was used, with a set time of 12-15 seconds, capturing 20 cardiac cycles, and a preset phase interval of 3% between each two single-beat CT imagings, capturing images within the phase range of 30%-60% of the systolic phase and 60%-90% of the diastolic phase.

3. The method for measuring vascular wall thickness based on coronary artery reconstruction according to claim 2, characterized in that: In S2, Laplace smoothing mapping is used to transform the image sequence V c Mapping from high-dimensional space to low-dimensional space, specifically the objective function optimized by the Laplace eigenmap is: Among them, V c As a set of data instances, y i 、y j is the vector representation of i and j in the data instance in the m-dimensional subspace, W ij The adjacency matrix constructed Where σ is a specified constant.

4. The method for measuring vessel wall thickness based on coronary artery reconstruction according to claim 2, characterized in that: In S2, the cluster center distance is obtained as follows: For any phase n, the sum of its distances to the cluster centers of the adjacent phases n-1 and n+1 is defined as A c n : in It represents the distance between the cluster centers of the images at the same position c of the coronary artery captured at the 30%+(n-1)×3% phase and the 30%+n×3% phase.

5. The method for measuring vascular wall thickness based on coronary artery reconstruction according to claim 1, characterized in that: In S3, the deformable U-net network consists of five layers of deformable U-net, whose encoding stage is a contraction path and the decoding stage is an expansion path; there is a concatenation in the contraction path to merge the features extracted from the normal convolution and the dilated convolution, and in the expansion path the feature map is upsampled to the original resolution of the input image for inter-frame registration and segmentation; a concatenation is used between the contraction path and the expansion path to feed the high-resolution features into the upsampled output; The two sequence images are segmented in sequence by 1x1x1 convolution and activation function with a stride of 1 to obtain the segmented image and The segmentation loss adopts a hybrid segmentation loss combined with class-weighted Tversky loss and class-weighted cross entropy loss in and is the ground truth segmentation containing some frame annotations, gt represents the ground truth value, Computed on labeled voxels; Obtain the pixel velocity field through a 5x5x5 convolution with a stride of 1 and That is, obtain the displacement direction and movement speed of each pixel; in the velocity field and In the definition and are the velocity fields extracted from the image frames obtained at the same position c when input in positive and reverse order; Velocity field and After scaling the layer, the deformation field is obtained and definition is the image frame obtained at position c in the continuous frames F and R, and and Respectively represent the artifacts after arrive The deformation field and arrive The deformation field of ; the pixel smoothing loss is defined as: Where N is the number of pixels, i represents the pixel, Representing the deformation field The deformation field at pixel i in , where x and y represent the position of pixel i; In the spatial transformer network, deformation fields are used to segment images. and Perform spatial transformation to obtain the image sequence after artifact removal: Where ◎ represents pixel multiplication, and Respectively and The segmented image at position c in , and Represents consecutive frames and The image frame obtained at position c in ; Define the loss function for reconstructing artifact-free image sequences Among them, u∈{0,1,2} represents three regions, namely the region outside the vessel wall, the region between the vessel wall and the intima boundary, and the region inside the intima lumen. Ω(u≠0) represents all regions except the vessel wall. p represents the voxel belonging to the Ω(u≠0) region in the image frame. R i (p) represents the image intensity at voxel p in the i-th frame in the continuous frame R, F i (p) represents the image intensity at voxel p in the i-th frame in the continuous frame F, Reconstructing continuous frames The image intensity at voxel p in the i-th frame, Reconstructing continuous frames The image intensity at voxel p in the i-th frame, and They represent the average image intensity of the i-th frame in the corresponding continuous frames; The specific method for reconstructing 3D coronary arteries is as follows: the artifact-free forward sequence passes through the reconstruction network, stacks along the membrane boundary along the coronary CT acquisition direction, performs linear interpolation on continuous 2D contours, and then visualizes the 3D surface through the marching cubes algorithm; The final network total loss function is: Among them, λ1, λ2 and λ3 are the parameters of various types of losses respectively. The network parameters are obtained by pre-training the network through the collected data based on the loss function.

6. The method for measuring vessel wall thickness based on coronary artery reconstruction according to claim 1, characterized in that: In S4, the specific method for measuring vascular wall thickness using the maximum inscribed sphere method is: The maximum inscribed sphere method was used to draw the centerline of the coronary artery and lumen. tol and l q , calculate the thickness of the blood vessel wall at each position, define the same plane perpendicular to the entrance of the blood vessel in this section and the two midlines l tol and l q The intersection points are o tol and o q The radius of the intangent plane is r tol and r q , with o tol and o q The direction of the line is the x-axis, o q Take O as the origin and make a plane coordinate system. The thickness of the blood vessel wall at any point on the plane is m i for: where θ i is the angle between the line from the inscribed plane i to the origin and the x-axis, and r is the midline of the two lines l on the plane tol and l q Intersection o tol and o q straight-line distance.

Images