A method and system for three-dimensional reconstruction of coronary angiograms

Abstract

The application discloses a kind of coronary angiography three-dimensional reconstruction method and system, it is related to medical image processing technical field, including: obtaining multi-angle coronary DSA angiography projection image and executing pre-processing, extract blood vessel initial center line and both sides initial contour, calculate each section contrast agent relative concentration, combined with blood vessel local curvature, concentration difference sign change and anatomical feature adaptive segmentation blood vessel section, filter reliable reference side and generate mirror virtual contour, calculate three kinds of contour deviation, construct multi-constraint objective function and correct center line section by section, integrate corrected center line and fit lumen parameter, solve blood vessel feature point three-dimensional coordinate and complete three-dimensional reconstruction.The application has the advantages that: effectively suppress the center line deviation caused by contrast agent concentration uneven, consider correction accuracy and blood vessel form continuity, the size of three-dimensional model output is accurate, topological correct, can be directly used for clinical quantitative diagnosis and interventional operation planning.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a three-dimensional reconstruction method and system for coronary angiography. Background Technology

[0002] Digital subtraction angiography (DSA) of the coronary arteries is currently the gold standard for diagnosing coronary artery disease. However, traditional two-dimensional DSA images can only provide projection information of blood vessels and cannot intuitively display the three-dimensional anatomical structure of the vessels, which can easily lead to missed diagnoses and misjudgments of the degree of stenosis. High-precision three-dimensional reconstruction technology of coronary arteries can restore the true morphology of blood vessels, providing key support for accurate clinical diagnosis and interventional surgery planning.

[0003] Existing three-dimensional reconstruction methods for coronary angiography can cause blurring and distortion at the edge of a vessel due to contrast agent dilution when the contrast agent concentration is uneven. This can lead to initial centerline shift and increased measurement error in vessel diameter. Uniform correction of the entire segment can easily cause reverse errors in the concentration inversion region, while segmented correction can produce cumulative errors and topological misalignment. Therefore, a three-dimensional reconstruction method and system for coronary angiography is needed to solve the above-mentioned problems. Summary of the Invention

[0004] To address the aforementioned technical problems, a three-dimensional reconstruction method and system for coronary angiography is provided. This technical solution solves the significant limitations of existing three-dimensional reconstruction methods for coronary angiography mentioned in the background art in scenarios with uneven contrast agent concentrations. Contrast agent dilution can lead to blurring and distortion of the edge of one side of the vessel, which in turn causes initial centerline offset and increased error in vessel diameter measurement. At the same time, using a uniform correction method for the entire segment is prone to generating reverse errors in the concentration inversion region, while using a segmented correction method will result in cumulative errors and topological misalignment problems.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A three-dimensional reconstruction method for coronary angiography includes: S1. Acquire multi-angle coronary DSA angiography projection images and perform preprocessing to obtain preprocessed coronary vessel grayscale images; S2. Perform multi-scale tubular structure enhancement and edge detection on the preprocessed coronary vessel grayscale image, extract the initial center line and bilateral initial contours of the vessel, sample the bilateral pixel grayscale values ​​cross-section by cross-section along the normal of the initial center line, and calculate the relative concentration of contrast agent in each cross-section. S3. Obtain the local curvature of the blood vessel, the sign change of the contrast agent concentration difference, and anatomical feature points, and adaptively segment a single continuous blood vessel into g independent blood vessel segments; S4. For each independent blood vessel segment, the mean concentration of contrast agent on both sides and the stability of the edge gradient within segment Xg are statistically analyzed. The side with higher concentration and more stable edge is selected as the reliable reference side, and the other side is selected as the weak distortion side. With the initial center line of segment Xg as the axis of symmetry, an axisymmetric mirror transformation is performed on the contour of the reliable reference side to generate a mirror virtual contour. S5. Perform pixel-level registration between the mirrored virtual contour of segmented Xg and the original contour of the weak distortion side, and calculate the contour spatial distance deviation, gray-scale mean deviation and edge gradient magnitude deviation. S6. Construct independent objective functions for centerline offset correction for each segment, use iterative gradient descent algorithm to solve for the optimal offset correction amount, and complete the correction and recalibration of the initial centerline segment by segment; S7. Smoothly stitch and perform global topology calibration on all segmented corrected centerlines to obtain complete and accurate vascular centerlines, and fit the geometric parameters of the lumen cross-section by combining bilateral contour information. S8. Solve for the three-dimensional coordinates of key feature points of blood vessels, construct a three-dimensional skeleton model of coronary arteries and complete surface reconstruction, and output a high-precision three-dimensional reconstruction model of coronary arteries.

[0006] In an optional embodiment, the step of performing multi-scale tubular structure enhancement and edge detection on the preprocessed vascular image, and extracting the initial centerline and bilateral initial contours of the blood vessels, specifically includes: S201. Obtain the pixel grayscale distribution features of the preprocessed coronary artery grayscale image, construct a multi-scale Hessian matrix, set the scale traversal interval to [1,8], and strengthen the linear tubular vessel target area and suppress point noise and sheet tissue interference areas by solving the eigenvalues ​​and eigenvectors of the Hessian matrix at different scales. S202. Set a low gradient threshold T1 and a high gradient threshold T2 to satisfy the quantitative constraint relationship T2=2.5T1. Solve for the pixel gradient magnitude and gradient direction, filter candidate edge pixels with gradient magnitude higher than the low gradient threshold, retain the effective edge pixels that are continuous and connected and satisfy the high gradient threshold constraint, and obtain the complete initial contour of both sides of the blood vessel. S203. Perform pixel-by-pixel erosion refinement on the lumen region surrounded by the edge of the connected blood vessel. Based on the connectivity constraint of eight neighboring pixels, remove redundant interference pixels and iterate continuously until a single-pixel connected skeleton curve is obtained, which is used as the initial center line of the blood vessel.

[0007] In an optional embodiment, the step of sampling the grayscale values ​​of pixels on both sides cross-section along the initial centerline normal and calculating the relative concentration of contrast agent at each cross-section specifically includes: S204. Obtain the discrete pixel coordinate sequence of the initial centerline of the blood vessel, solve for the local tangent vector and normal vector of the centerline, calculate the unit normal vector of the centerline point by point, and determine the bilateral vertical sampling direction of each blood vessel section. S205. Obtain the pointing range of the normal unit vector. Taking the single sampling point of the initial center line as the center, set the single-sided sampling pixel length as d, select a continuous set of pixel sampling points at equal intervals along the double-sided normal direction, and batch collect the original pixel gray values ​​at corresponding positions on both sides. S206. Obtain physical law data of DSA imaging, establish a linear correlation model between contrast agent iodine concentration and X linear attenuation coefficient, satisfying the linear equation: C=α·I+β, where α is the gray-to-concentration linear conversion coefficient, β is the baseline concentration compensation constant, C is the contrast agent iodine mass concentration in the blood vessel lumen, and I is the pixel gray value at this position after digital subtraction. S207. Perform mean calculation on the grayscale data of the sampled pixels on the left and right sides of the blood vessel, respectively, and substitute them into the linear correlation model to calculate the average contrast agent concentration value on each side point by point, so as to obtain the contrast agent relative concentration sequence continuously distributed on both sides of the blood vessel.

[0008] In an optional embodiment, the step of acquiring local vascular curvature, contrast agent concentration gradient sign change, and anatomical feature points, and adaptively segmenting a single continuous blood vessel into several independent vascular segments, specifically includes: S301. Obtain the discrete pixel coordinate sequence of the initial centerline of the blood vessel, and solve for the local curvature value of the blood vessel point by point according to the following formula. : ,in Curve function The first derivative, Curve function The second derivative of the curve function In and These are the horizontal and vertical pixel coordinates of the discrete pixel coordinate sequence of the initial centerline of the blood vessel, respectively. S302. Identify curvature abrupt change points whose curvature values ​​exceed a preset curvature threshold and whose continuous length is greater than or equal to 5 pixels, and use them as the first type of candidate segmentation points; S303. Obtain the relative concentration sequence of contrast agent on both sides of the blood vessel, and calculate the difference in contrast agent concentration between the two sides cross-section along the initial centerline using the following formula. : ,in This represents the average concentration of contrast agent in the sampling area on the left side of the cross-section. The average concentration of contrast agent in the sampling area on the right side of the cross section is recorded. The sign change of the concentration difference is recorded. The location points where the sign of the concentration difference is reversed and the continuous length is greater than or equal to 3 pixels are identified as the second type of candidate segmentation points. S304. Extract the bifurcation point of the blood vessel, the endpoint of the blood vessel, the stenosis lesion point with a diameter stenosis rate of ≥50%, and the center point of the calcified plaque as forced segmentation points, and integrate the first type of candidate segmentation points, the second type of candidate segmentation points, and the forced segmentation points to form an initial segmentation point set; S305. Preset minimum and maximum segment lengths, remove redundant segment points in the initial segment point set whose distance is less than the minimum segment length, merge segments whose length is less than the minimum segment length, split segments whose length is greater than the maximum segment length, obtain the optimized segment point set, and then divide a single continuous blood vessel into g independent blood vessel segments. Each segment maintains single connectivity, no bifurcation, smooth curvature change, and consistent contrast agent concentration on the high and low sides. S306. Assign a unique identifier to each independent blood vessel segment, record the starting point coordinates, ending point coordinates, and segment length information of each segment, and establish a blood vessel segment index table.

[0009] In an optional embodiment, for each independent vascular segment, the mean concentration of contrast agent on both sides within segment Xg and the stability of the edge gradient are statistically analyzed. The side with a higher concentration and more stable edges is selected as the reliable baseline side, and the other side is selected as the weak distortion side. Specifically, this includes: S401. Obtain the blood vessel segment index table, and read the pixel data and feature information of each independent blood vessel segment in sequence; S402. For a single vessel segment Xg, calculate the left-side comprehensive confidence score S of segment Xg according to the following formula. L Compared with the comprehensive confidence score S on the right R : ,in This represents the average concentration on the left or right side of segment Xg. For the left or right gradient variability variance of all edge pixels in segment Xg, and All are weighting coefficients; S403. Compare the overall confidence scores on the left and right sides, and select the side with the higher score as the confidence benchmark side for that segment, and the side with the lower score as the weak distortion side for that segment.

[0010] In an optional embodiment, the step of performing an axisymmetric mirror transformation on the reliable reference side profile with the initial centerline of segment Xg as the axis of symmetry to generate a mirrored virtual profile specifically includes: S404. Extract all original contour pixel coordinates of the segmented Xg reliable reference side and establish a reliable reference side contour point set; S405. Obtain the initial centerline continuous curve equation of segment Xg, establish a two-dimensional plane rectangular coordinate system with the centerline as the axis of symmetry, and solve for the horizontal and vertical distances and vertical coordinate parameters of any contour pixel relative to the axis of symmetry. S406. Perform coordinate mapping transformation of all blood vessel contours on the reliable reference side pixel by pixel according to the following formula: ,in The coordinates of the reference point on the axis of symmetry For the original contour pixel coordinates of the reliable reference side, These are the mirror pixel coordinates; S407. Perform connectivity verification and breakpoint completion on the discrete contour pixels after mirror transformation, use local linear interpolation algorithm to fill the tiny blank pixel areas generated by mirror transformation, and simultaneously retain all pixel coordinates, grayscale information and edge features of the original contour on the weak distortion side of segmented Xg, and fix the spatial coordinate system and pixel scale specifications of the two types of contours.

[0011] In an optional embodiment, the step of performing pixel-level registration between the mirrored virtual contour of segment Xg and the original contour on the weakly distorted side, and calculating the contour spatial distance deviation, gray-level mean deviation, and edge gradient magnitude deviation, specifically includes: S501. Uniformly sample the mirrored virtual contour of segment Xg at preset pixel intervals to obtain a set of feature control points: An iterative nearest-point registration algorithm is used to find the nearest matching point for each mirror control point on the original contour on the weakly distorted side, resulting in the original contour control point set: The pixel distance D between each pair of matching points is calculated using the following formula: The mean of all matching points is used as a quantitative indicator of the spatial distance deviation of the contour. S502. Extract the local pixel grayscale matrix of the mirrored virtual contour and the local pixel grayscale matrix of the original contour on the weakly distorted side, respectively. Calculate the average grayscale value of the corresponding regions of the two contours according to the following formula. : ,in This represents the total number of pixels sampled locally. This represents the grayscale value of a single pixel. S503. Solve for the global average grayscale deviation to quantitatively characterize the regional grayscale distribution differences caused by uneven contrast agent concentration; S504. Obtain the set of pixel gradient magnitude values ​​of the mirrored virtual contour edge and the set of pixel gradient magnitude values ​​of the original contour edge on the weakly distorted side, and calculate the gradient feature residual edge by edge according to the following formula. : By combining the residual mean and the residual distribution interval, a gradient modulus deviation evaluation index is constructed, in which... The gradient magnitude of the mirrored virtual contour edge points. The gradient magnitude of the original contour edge points on the weakly distorted side; S505. Normalize the contour spatial distance deviation quantification index, global gray-scale average deviation, and gradient modulus deviation evaluation index to obtain three types of normalized difference characteristic parameters for a single blood vessel segment.

[0012] In an optional embodiment, the step of constructing an independent centerline offset correction objective function for each segment, using an iterative gradient descent algorithm to solve for the optimal offset correction amount, and completing the initial centerline correction and recalibration segment by segment specifically includes: S601. Obtain the three types of normalized differential feature parameters for a single vessel segment. Using the lateral and normal offsets of the initial centerline of segment Xg as optimization variables, and combining the local curvature smoothing constraint of the vessel, construct an independent centerline offset correction objective function for this segment. The core expression of the objective function is: In the formula, , , These are the normalized contour spatial distance deviation quantification index, the global gray-scale average deviation, and the gradient magnitude deviation evaluation index, respectively. The curvature constraint balance coefficient, This is a penalty for fluctuations in the local curvature of blood vessels. S602. Set the upper limit of the maximum offset of the centerline in a single iteration, use the gradient descent iterative optimization algorithm to solve the minimum value of the objective function, and calculate the gradient direction of the fastest descent of the objective function by taking the first-order partial derivative with respect to the optimization variables. S603. Set the iterative learning rate, and update the coordinate offset of each discrete sampling point on the segmented Xg centerline successively in combination with the gradient direction of the fastest descent of the objective function to complete the single-round coordinate fine-tuning correction. S604. After a single iteration, recalculate the difference feature parameters of the corrected bilateral contours, substitute them into the objective function to solve for the function value again, and compare the changes in the objective function value before and after the two iterations. S605. Set an iteration convergence threshold. When the difference between the objective function of two consecutive iterations is less than the convergence threshold, the iteration operation is determined to have reached a stable state, and the loop update is stopped. S606. Extract the global optimal offset correction amount of the segment Xg finally solved after iterative convergence, and perform point-by-point translation correction on all discrete pixel coordinates of the initial center line of segment Xg according to the independent offset parameters corresponding to each point of the center line. S607. The corrected segmented Xg discrete centerline coordinate sequence is locally smoothed by using a cubic B-spline smoothing fitting algorithm to eliminate local micro-jitter and abrupt change in inflection point caused by point-by-point independent correction. S608. Repeat steps S601 to S607 to complete the centerline correction and recalibration of all independent vascular segments in sequence, and obtain the independent corrected centerline data of each segment.

[0013] In an optional embodiment, the step of smoothly stitching together all segmented corrected centerlines and performing global topology calibration to obtain a complete and accurate vessel centerline, and then fitting the geometric parameters of the lumen cross-section using bilateral contour information, specifically includes: S701. Obtain all segmented and corrected centerline data, extract the overlapping area of ​​two adjacent segments, calculate the coordinate difference of corresponding points of the two segmented and corrected centerlines in the overlapping area, perform smooth transition processing on the centerline coordinates of the overlapping area, and obtain the spliced ​​centerline. S702. Using fixed anatomical feature points as constraints, the centerline after splicing is finely adjusted globally to obtain a complete and continuous corrected vascular centerline; S703. Along the normal section direction of the corrected blood vessel centerline, simultaneously read the boundary pixel coordinates of the original contour of the reliable reference side and the contour of the corrected weak side, and extract the set of boundary points of the blood vessel wall on both sides of the single section. S704. Obtain the coordinate data of multiple boundary points of the cross section and substitute them into the elliptical cross section fitting model to solve the semi-major axis, semi-minor axis and cross section deflection angle of the elliptical cross section of the blood vessel lumen. S705. Calculate the parameters of the entire vessel segment's continuous cross-section by continuously traversing the direction of the vessel's centerline.

[0014] Furthermore, a three-dimensional reconstruction system for coronary angiography is proposed to implement the three-dimensional reconstruction method described in any of the preceding claims, comprising: The image preprocessing module is used to acquire multi-angle coronary DSA angiography projection images and perform preprocessing to obtain preprocessed coronary vessel grayscale images. The vascular feature extraction and segmentation module is used to perform multi-scale tubular structure enhancement and edge detection on the preprocessed image, extract the initial center line and bilateral initial contours of the blood vessel, calculate the relative concentration of contrast agent in each section, and combine the local curvature of the blood vessel, the sign change of the concentration difference and anatomical features to adaptively segment a single continuous blood vessel into multiple independent blood vessel segments. The mirror contour generation and difference calculation module is used to screen the credible reference side and the weak distortion side for each blood vessel segment, perform axisymmetric mirror transformation on the credible reference side contour to generate a mirror virtual contour, register the mirror virtual contour with the original contour of the weak distortion side, and calculate the contour spatial distance deviation, gray level mean deviation and edge gradient magnitude deviation. The centerline offset correction module is used to construct independent centerline offset correction objective functions for each segment, and to solve for the optimal offset correction amount using an iterative gradient descent algorithm, thereby completing the correction and recalibration of the initial centerline segment by segment. The centerline integration and lumen fitting module is used to smoothly stitch together and perform global topology calibration on all segmented corrected centerlines to obtain a complete and accurate vascular centerline, and to fit the geometric parameters of the lumen cross-section by combining bilateral contour information. The 3D model construction and output module is used to solve the 3D coordinates of key feature points of blood vessels, construct a 3D skeleton model of coronary arteries and complete surface reconstruction, and output a high-precision 3D reconstructed model of coronary arteries.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. The proposed three-dimensional reconstruction method for coronary angiography combines adaptive vessel segmentation based on local vessel curvature, contrast agent concentration difference sign change and anatomical features, as well as reliable baseline screening based on concentration and edge gradient and axisymmetric mirror transformation correction. This method achieves accurate correction of vessel centerline in scenarios with uneven contrast agent concentration, significantly reducing centerline offset and diameter measurement error caused by unilateral edge blurring distortion. 2. The proposed three-dimensional reconstruction method for coronary angiography constructs a multi-constrained centerline offset correction objective function that integrates contour spatial distance deviation, gray-level mean deviation, and edge gradient magnitude deviation. Combined with piecewise independent gradient descent optimization and topological calibration anchored by global anatomical features, it achieves a balance between correction accuracy and vascular morphological continuity, eliminates the reverse error of whole-segment correction and the cumulative error of segmented correction, and ensures the anatomical rationality of the corrected centerline. 3. The proposed method for three-dimensional reconstruction of coronary angiography achieves high-precision coronary three-dimensional model construction by using cross-sectional elliptical fitting of the lumen based on the corrected precise centerline, combined with dual-view triangulation three-dimensional reconstruction, volume-preserving Laplacian smoothing, and anatomically driven topology correction. This method suppresses vascular volume shrinkage and topological errors during mesh smoothing, and the output three-dimensional model has complete morphology and accurate dimensions, which can be directly used for clinical quantitative analysis and interventional surgery planning. Attached Figure Description

[0016] Figure 1 This is an overall flowchart of a three-dimensional reconstruction method for coronary angiography proposed in this invention; Figure 2 This is a system framework diagram of a three-dimensional reconstruction system for coronary angiography proposed in this invention. Detailed Implementation

[0017] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0018] Reference Figure 1 As shown, a three-dimensional reconstruction method for coronary angiography includes: S1. Acquire multi-angle coronary DSA angiography projection images and perform preprocessing to obtain preprocessed coronary vessel grayscale images; S2. Perform multi-scale tubular structure enhancement and edge detection on the preprocessed coronary vessel grayscale image, extract the initial center line and bilateral initial contours of the vessel, sample the bilateral pixel grayscale values ​​cross-section by cross-section along the normal of the initial center line, and calculate the relative concentration of contrast agent in each cross-section. S3. Obtain the local curvature of the blood vessel, the sign change of the contrast agent concentration difference, and anatomical feature points, and adaptively segment a single continuous blood vessel into g independent blood vessel segments; S4. For each independent blood vessel segment, the mean concentration of contrast agent on both sides and the stability of the edge gradient within segment Xg are statistically analyzed. The side with higher concentration and more stable edge is selected as the reliable reference side, and the other side is selected as the weak distortion side. With the initial center line of segment Xg as the axis of symmetry, an axisymmetric mirror transformation is performed on the contour of the reliable reference side to generate a mirror virtual contour. S5. Perform pixel-level registration between the mirrored virtual contour of segmented Xg and the original contour of the weak distortion side, and calculate the contour spatial distance deviation, gray-scale mean deviation and edge gradient magnitude deviation. S6. Construct independent objective functions for centerline offset correction for each segment, use iterative gradient descent algorithm to solve for the optimal offset correction amount, and complete the correction and recalibration of the initial centerline segment by segment; S7. Smoothly stitch and perform global topology calibration on all segmented corrected centerlines to obtain complete and accurate vascular centerlines, and fit the geometric parameters of the lumen cross-section by combining bilateral contour information. S8. Solve for the three-dimensional coordinates of key feature points of blood vessels, construct a three-dimensional skeleton model of coronary arteries and complete surface reconstruction, and output a high-precision three-dimensional reconstruction model of coronary arteries.

[0019] Furthermore, multi-angle coronary DSA angiography projection images are acquired and preprocessed to obtain preprocessed vascular images with uniform grayscale distribution, specifically including: S101. Acquire ECG monitoring signals and multi-angle raw coronary DSA angiography projection images; S102. Based on the frame time sequence relationship of the original contrast images, filter the blank mask image without contrast agent filling and the target imaging frame image filled with contrast agent, perform differential operation pixel by pixel to eliminate the static background interference of bones and soft tissues, and obtain the initial subtraction image. S103. Filter the initial subtraction image pixel by pixel to obtain the denoised blood vessel image; The specific parameters of the bilateral filtering algorithm are: the spatial filter kernel size is 5×5 pixels, and the standard deviation of the spatial Gaussian function σ. s =1.5, Gaussian function standard deviation σ in grayscale range r=20. This parameter combination can effectively suppress quantum noise while preserving the details of the blood vessel edges to the greatest extent.

[0020] S104. Divide the denoised blood vessel image into several local sub-image blocks, and adaptively adjust the dynamic range of local pixel grayscale to obtain the enhanced image; The local sub-image patch size is 128×128 pixels, and the contrast constraint factor for Contrast Adaptive Histogram Equalization (CLAHE) is 2.0. When the cumulative distribution function of the local histogram exceeds this threshold, the excess portion is evenly distributed across the entire grayscale range to avoid local over-enhancement.

[0021] S105. Obtain the RR interval time node of the ECG gating mark, select the key imaging frame corresponding to the 75% RR interval at the end of diastole as the reference frame, and then construct the inter-frame pixel motion vector field to calculate the horizontal and vertical offsets caused by the heartbeat. In this embodiment, the motion vector field is constructed using the pyramid Lucas-Kanade (LK) optical flow method. The pyramid has three layers, and the search window size for each layer is 15×15 pixels. For outliers with motion vector amplitudes greater than 5 pixels, the average motion vector of the neighboring eight points is used for replacement.

[0022] S106. Based on the displacement parameters of the inter-frame pixel motion vector field, perform motion compensation correction on the enhanced image; S107. Extract the global maximum gray value and global minimum gray value of a single image respectively, perform pixel gray-level normalization processing on the corrected image, and finally output the preprocessed coronary artery gray-level image. In this embodiment, extreme pixels with grayscale values ​​less than 5 or greater than 250 are identified as noise pixels, and their grayscale values ​​are truncated to 5 and 250 respectively to avoid extreme values ​​affecting subsequent processing.

[0023] Furthermore, multi-scale tubular structure enhancement and edge detection are performed on the preprocessed vascular images to extract the initial centerline and bilateral initial contours of the vessels, specifically including: S201. Obtain the pixel grayscale distribution features of the preprocessed coronary artery grayscale image, construct a multi-scale Hessian matrix, set the scale traversal interval to [1,8], and strengthen the linear tubular vessel target area and suppress point noise and sheet tissue interference areas by solving the eigenvalues ​​and eigenvectors of the Hessian matrix at different scales. Furthermore, step S201 specifically includes the following steps: S2011. Initialize the scaling parameters, set the scaling traversal interval to [1,8], the step size to 0.5, and generate 15 Gaussian second derivative kernels with different scales and kernel sizes of [missing information]. ; S2012. Convolve the preprocessed grayscale image with Gaussian second-order derivative kernels in the x-direction, y-direction, and xy-mixing direction respectively to obtain three second-order partial derivative images; S2013. For each pixel, construct a 2×2 Hessian matrix using the three second-order partial derivatives, solve for the two eigenvalues ​​λ1 and λ2 of the matrix, and sort them in ascending order of absolute value as |λ1|≤|λ2|. S2014. Based on the feature value distribution characteristics, determine the tubular structure. When λ2 < 0, use the Frangi response function to calculate the enhanced response value of the pixel at the current scale; otherwise, set the response value to 0. S2015. Iterate through all scales and take the maximum response value of each pixel at all scales as the original enhanced response; S2016. Perform 2% truncation normalization on the original enhanced response and map it to the gray range of 0 to 255 to obtain the final multi-scale tubular structure enhanced image.

[0024] The criteria for determining tubular structure enhancement are as follows: For the two eigenvalues ​​λ1 and λ2 of the Hessian matrix (|λ1|≤|λ2|), when λ2<0 and |λ2| / |λ1|>3 (large curvature in the direction perpendicular to the blood vessel and small curvature in the parallel direction), the pixel is determined to be a tubular structure pixel; otherwise, it is a background pixel. When λ1<0, λ2<0 and λ1≈λ2 (large curvature in both directions), the pixel is determined to be a dotted structure (noise). When λ1≈0 and λ2<0 (large curvature in one direction and small curvature in the other direction), it is determined to be a sheet-like structure (soft tissue).

[0025] S202. Set a low gradient threshold T1 and a high gradient threshold T2 to satisfy the quantitative constraint relationship T2=2.5T1. Solve for the pixel gradient magnitude and gradient direction, filter candidate edge pixels with gradient magnitude higher than the low gradient threshold, retain the effective edge pixels that are continuous and connected and satisfy the high gradient threshold constraint, and obtain the complete initial contour of both sides of the blood vessel. The low gradient threshold T1 ranges from 15 to 30 gray levels; in this embodiment, T1 = 20, corresponding to a high gradient threshold T2 = 50. Pixels with gradient magnitudes between T1 and T2 are only retained as edge pixels if they are connected to a pixel with a gradient magnitude greater than T2.

[0026] S203. Perform pixel-by-pixel erosion refinement on the lumen region surrounded by the edge of the connected blood vessel. Based on the connectivity constraint of eight neighboring pixels, remove redundant interference pixels and iterate continuously until a single-pixel connected skeleton curve is obtained, which is used as the initial center line of the blood vessel.

[0027] The iteration termination condition for the skeleton thinning algorithm is: the iteration stops when no pixels are deleted during a single iteration. After thinning, isolated skeleton branches with a length of less than 3 pixels are removed to avoid interference from pseudo-skeletons.

[0028] Furthermore, the grayscale values ​​of pixels on both sides are sampled cross-section by cross-section along the initial centerline normal, and the relative concentration of contrast agent at each cross-section is calculated, specifically including: S204. Obtain the discrete pixel coordinate sequence of the initial centerline of the blood vessel, solve for the local tangent vector and normal vector of the centerline, calculate the unit normal vector of the centerline point by point, and determine the bilateral vertical sampling direction of each blood vessel section. S205. Obtain the pointing range of the normal unit vector. Taking the single sampling point of the initial center line as the center, set the single-sided sampling pixel length as d, select a continuous set of pixel sampling points at equal intervals along the double-sided normal direction, and batch collect the original pixel gray values ​​at corresponding positions on both sides. In this embodiment, the unilateral sampling area is a range of 1-10 pixels away from the initial center line, avoiding background pixels outside the blood vessel edge and the central area near the center line, ensuring that the sampling area is located inside the blood vessel lumen.

[0029] S206. Obtain physical law data of DSA imaging, establish a linear correlation model between contrast agent iodine concentration and X linear attenuation coefficient, satisfying the linear equation: C=α·I+β, where α is the gray-to-concentration linear conversion coefficient, β is the baseline concentration compensation constant, C is the contrast agent iodine mass concentration in the blood vessel lumen, and I is the pixel gray value at this position after digital subtraction. Specifically, the parameters α and β of the linear correlation model were obtained in advance through a standard iodine concentration phantom calibration experiment. The specific calibration steps are as follows: S2061. Prepare a standard iodine concentration phantom containing seven concentration gradients of 0, 50, 100, 150, 200, 250, and 300 mgI / mL to simulate the actual contrast agent concentration range in the coronary arteries; S2062. Image the phantom using the same equipment parameters as clinical acquisition (tube voltage 80kVp, tube current 200mA, exposure time 10ms); S2063. For each concentration, acquire 10 frames of images and take the average gray value of the region of interest (ROI); S2064. Perform linear regression fitting with iodine concentration as the ordinate and average gray value as the abscissa; S2065. The parameters obtained by fitting in this embodiment are: α = -1.02 gray levels / (mgI / mL), β = 235 gray levels. It should be noted that the gray polarity may be opposite in different devices. If the contrast agent filling area is a high gray value, then α is a positive value.

[0030] S207. Perform mean calculation on the grayscale data of the sampled pixels on the left and right sides of the blood vessel, respectively, and substitute them into the linear correlation model to calculate the average contrast agent concentration value on each side point by point, so as to obtain the contrast agent relative concentration sequence continuously distributed on both sides of the blood vessel.

[0031] The specific implementation process of step S207 is as follows: S2071. For the i-th sampling point of the center line, along the positive and negative directions of its normal unit vector, extract pixels within a range of 1 to 10 pixels from the center line to form the left sampling region and the right sampling region. The sampling length selection is based on a sampling range of 1 to 10 pixels, corresponding to a vessel diameter of 2 to 20 pixels, which fully covers the clinically common main coronary artery (diameter 8 to 16 pixels) and distal branches (diameter 2 to 6 pixels). S2072. Filter background pixels and artifact pixels within the sampling area, and retain effective cavity pixels with gray values ​​in the range of 20 to 230; In this embodiment, the grayscale range of 20 to 230 excludes pure background (grayscale < 20) and device saturation artifacts (grayscale > 230), ensuring the validity of the sampled data; S2073. Use the 3σ criterion to remove outliers from the effective grayscale set. If the number of remaining effective pixels is less than 5, reuse the average grayscale value of the previous sampling point. S2074. Calculate the arithmetic mean of the effective grayscale sets on the left and right sides respectively to obtain the average grayscale value on one side; S2075. Substitute the one-sided average gray value into the pre-calibrated gray-concentration linear correlation model C=α·I+β to calculate the one-sided average contrast agent concentration. S2076. The original concentration sequence was smoothed using a 5-point moving average method to eliminate local random fluctuations; In this embodiment, the 5-point moving average window can effectively eliminate noise fluctuations without excessive smoothing that would mask concentration abrupt changes (such as concentration reversal points). S2077. Generate continuous left-side and right-side contrast agent relative concentration sequences for subsequent vessel segmentation and reliable baseline screening.

[0032] In this embodiment, only the concentration difference across the blood vessel needs to be compared; precise values ​​of absolute concentration are not required. Therefore, even minor calibration errors in α and β will not affect the relative magnitude of the concentrations on both sides, which is a significant robust advantage of this method.

[0033] Furthermore, by acquiring the local curvature of the blood vessel, the sign change of the contrast agent concentration gradient, and anatomical feature points, a single continuous blood vessel is adaptively segmented into several independent blood vessel segments, specifically including: S301. Obtain the discrete pixel coordinate sequence of the initial centerline of the blood vessel, and solve for the local curvature value of the blood vessel point by point according to the following formula. : ,in Curve function The first derivative, Curve function The second derivative of the curve function In and These are the horizontal and vertical pixel coordinates of the discrete pixel coordinate sequence of the initial centerline of the blood vessel, respectively. Specifically, for a discrete sequence of centerline pixel coordinates, the first and second derivatives are calculated using the central difference method: First derivative: Second derivative: For the first and last points of the centerline, the derivatives are calculated using forward and backward differences, respectively, and the second derivatives are reused from the second derivatives of adjacent points.

[0034] S302. Identify curvature abrupt change points whose curvature values ​​exceed a preset curvature threshold and whose continuous length is greater than or equal to 5 pixels, and use them as the first type of candidate segmentation points; In this embodiment, the preset curvature threshold κ0 = 0.03 pixels. -1 This corresponds to a curvature radius of approximately 33 pixels. When the curvature values ​​of five or more consecutive pixels exceed this threshold, it is determined to be a curvature abrupt change segment, and a segmentation point is set at the point of maximum curvature.

[0035] S303. Obtain the relative concentration sequence of contrast agent on both sides of the blood vessel, and calculate the difference in contrast agent concentration between the two sides cross-section along the initial centerline using the following formula. : ,in This represents the average concentration of contrast agent in the sampling area on the left side of the cross-section. The average concentration of contrast agent in the sampling area on the right side of the cross section is recorded. The sign change of the concentration difference is recorded. The location points where the sign of the concentration difference is reversed and the continuous length is greater than or equal to 3 pixels are identified as the second type of candidate segmentation points. S304. Extract the bifurcation point of the blood vessel, the endpoint of the blood vessel, the stenosis lesion point with a diameter stenosis rate of ≥50%, and the center point of the calcified plaque as forced segmentation points, and integrate the first type of candidate segmentation points, the second type of candidate segmentation points, and the forced segmentation points to form an initial segmentation point set; The specific implementation process of step S304 is as follows: S3041. Based on the skeleton topology of the initial center line of the blood vessel, count the number of connected pixels in the eight neighboring regions of each skeleton pixel, determine the point with a connection number of 1 as the blood vessel endpoint, and determine the point with a connection number ≥ 3 and all branch lengths ≥ 3 pixels as the blood vessel bifurcation point. Specifically, a 3-pixel branch length threshold can effectively filter out noisy pseudo-branches generated by skeleton thinning, while retaining real small side branches. S3042. Calculate the vessel diameter point by point along the initial centerline to generate a vessel diameter sequence. Use the sliding window method (a 21-pixel sliding window can effectively cover normal vessel segments and avoid reference diameter calculation errors caused by local fluctuations) to calculate the reference diameter of each point, and then calculate the diameter stenosis rate. The point with the most severe stenosis in a region with a stenosis rate of ≥50% for 3 or more consecutive pixels is identified as a significant stenosis lesion point. S3043. Compare the grayscale difference between the blank mask image and the imaging frame image, extract the regions with grayscale values ​​≥220 in the mask image and absolute grayscale values ​​≤20 in the difference image as calcified patches, and calculate the centroid of the connected domain of the calcified patch as the center point of the calcified patch. In this embodiment, a grayscale threshold of 220 corresponds to the typical brightness of calcified patches in an 8-bit DSA image, which can effectively distinguish between calcification and contrast agents. S3044. Integrate vascular endpoints, vascular bifurcation points, significant stenosis lesion points, and calcified plaque center points to form a set of forced segmentation points; S3045. Merge the first type of candidate segment points, the second type of candidate segment points, and the forced segment point set, and remove duplicate segment points with an index difference ≤ 3 pixels; S3046. Filter out abnormal candidate segmentation points that are less than 5 pixels away from the blood vessel endpoint and less than 10 pixels away from the bifurcation point, and retain all forced segmentation points; S3047. Sort the remaining segment points in order of the centerline axis to generate an initial set of segment points.

[0036] S305. Preset minimum and maximum segment lengths, remove redundant segment points in the initial segment point set whose distance is less than the minimum segment length, merge segments whose length is less than the minimum segment length, split segments whose length is greater than the maximum segment length, obtain the optimized segment point set, and then divide a single continuous blood vessel into g independent blood vessel segments. Each segment maintains single connectivity, no bifurcation, smooth curvature change, and consistent contrast agent concentration on the high and low sides. In this embodiment, a minimum segment length L is preset. min =20 pixels, maximum segment length L max =150 pixels. For segments that are too short (less than 20 pixels), merge them into adjacent longer segments; for segments that are too long (more than 150 pixels), split them at their midpoint.

[0037] S306. Assign a unique identifier to each independent blood vessel segment, record the starting point coordinates, ending point coordinates, and segment length information of each segment, and establish a blood vessel segment index table.

[0038] Furthermore, for each independent vessel segment, the mean contrast agent concentration and edge gradient stability within segment Xg were statistically analyzed bilaterally. The side with higher concentration and more stable edges was selected as the reliable baseline side, and the other side was selected as the weak distortion side. Specifically, this included: S401. Obtain the blood vessel segment index table, and read the pixel data and feature information of each independent blood vessel segment in sequence; S402. For a single vessel segment Xg, calculate the left-side comprehensive confidence score S of segment Xg according to the following formula. L Compared with the comprehensive confidence score S on the right R : ,in This represents the average concentration on the left or right side of segment Xg. For the left or right gradient variability variance of all edge pixels in segment Xg, and All are weighting coefficients; the smaller the gradient fluctuation variance, the more continuous and stable the edge contour. In this embodiment, the weighting coefficient , That is, the contrast agent concentration is weighted at 60%, and the marginal gradient stability is weighted at 40%. Gradient fluctuation variance The calculation formula is: ,in For the first The gradient magnitude of each edge pixel. This represents the average gradient magnitude of the edge pixels. When the difference between the combined confidence scores of the left and right sides is less than 5%, the two sides are considered to have comparable quality. In this case, the side with the smaller gradient fluctuation variance is selected as the confidence benchmark side.

[0039] S403. Compare the overall confidence scores on the left and right sides, and select the side with the higher score as the confidence benchmark side for that segment, and the side with the lower score as the weak distortion side for that segment.

[0040] Furthermore, using the initial centerline of segment Xg as the axis of symmetry, an axisymmetric mirror transformation is performed on the reliable reference side profile to generate a mirrored virtual profile, specifically including: S404. Extract all original contour pixel coordinates of the segmented Xg reliable reference side and establish a reliable reference side contour point set; S405. Obtain the initial centerline continuous curve equation of segment Xg, establish a two-dimensional plane rectangular coordinate system with the centerline as the axis of symmetry, and solve for the horizontal and vertical distances and vertical coordinate parameters of any contour pixel relative to the axis of symmetry. S406. Perform coordinate mapping transformation of all blood vessel contours on the reliable reference side pixel by pixel according to the following formula: ,in The coordinates of the reference point on the axis of symmetry For the original contour pixel coordinates of the reliable reference side, These are the mirror pixel coordinates; S407. Perform connectivity verification and breakpoint completion on the discrete contour pixels after mirror transformation, use local linear interpolation algorithm to fill the tiny blank pixel areas generated by mirror transformation, and simultaneously retain all pixel coordinates, grayscale information and edge features of the original contour on the weak distortion side of segmented Xg, and fix the spatial coordinate system and pixel scale specifications of the two types of contours.

[0041] For breakpoints longer than 3 pixels, a cubic spline interpolation algorithm is used for completion to ensure the continuity of the mirrored contour. The gradient magnitude of the completed contour pixels reuses the average gradient magnitude of adjacent contour pixels.

[0042] Furthermore, the mirrored virtual contours of segmented Xg are registered pixel-level with the original contours on the weakly distorted side, and the contour spatial distance deviation, gray-level mean deviation, and edge gradient magnitude deviation are calculated, specifically including: S501. Uniformly sample the mirrored virtual contour of segment Xg at preset pixel intervals to obtain a set of feature control points: An iterative nearest-point registration algorithm is used to find the nearest matching point for each mirror control point on the original contour on the weakly distorted side, resulting in the original contour control point set: The pixel distance D between each pair of matching points is calculated using the following formula: The mean of all matching points is used as a quantitative indicator of the spatial distance deviation of the contour. Before calculating the mean, the 3σ criterion is used to filter out abnormal matching points: matching points with a distance greater than 3 times the mean square error are removed to avoid distortion in the deviation calculation caused by local contour breaks or noise. In this embodiment, the preset pixel interval is 3 pixels to ensure that the control points uniformly cover the entire blood vessel contour. The iteration termination condition of the Iterative Closest Point (ICP) registration algorithm is: the change in mean square error between two consecutive iterations is less than 1 × 10⁻⁶. -4 Or the number of iterations reaches 10.

[0043] S502. Extract the local pixel grayscale matrix of the mirrored virtual contour and the local pixel grayscale matrix of the original contour on the weakly distorted side, respectively. Calculate the average grayscale value of the corresponding regions of the two contours according to the following formula. : ,in This represents the total number of pixels sampled locally. This represents the grayscale value of a single pixel. S503. Solve for the global average grayscale deviation to quantitatively characterize the regional grayscale distribution differences caused by uneven contrast agent concentration; Understandably, the global grayscale average deviation refers to the absolute difference in average grayscale value between the lumen region enclosed by the mirrored virtual contour and the lumen region enclosed by the original contour on the weakly distorted side within the same blood vessel segment.

[0044] S504. Obtain the set of pixel gradient magnitude values ​​of the mirrored virtual contour edge and the set of pixel gradient magnitude values ​​of the original contour edge on the weakly distorted side, and calculate the gradient feature residual edge by edge according to the following formula. : By combining the residual mean and the residual distribution interval, a gradient modulus deviation evaluation index is constructed, in which... The gradient magnitude of the mirrored virtual contour edge points. The gradient magnitude of the original contour edge points on the weakly distorted side; S505. Normalize the contour spatial distance deviation quantification index, global gray-scale average deviation, and gradient modulus deviation evaluation index to obtain three types of normalized difference characteristic parameters for a single blood vessel segment.

[0045] In this embodiment, the normalization method is max-min normalization, and the specific formula is: X nor = (XX min ) / (X max -X min Where X is the original deviation value, X min X is the minimum possible value of this type of deviation. max This represents the maximum possible value for this type of deviation. For contour space distance deviation, X... max Take half of the average diameter of the blood vessel segment; for the grayscale mean deviation, X max Take 255; for gradient magnitude deviation, X max Take 255.

[0046] Furthermore, an independent objective function for centerline offset correction is constructed for each segment, and the optimal offset correction amount is solved using an iterative gradient descent algorithm. The initial centerline is corrected and recalibrated segment by segment, specifically including: S601. Obtain the three types of normalized differential feature parameters for a single vessel segment. Using the lateral and normal offsets of the initial centerline of segment Xg as optimization variables, and combining the local curvature smoothing constraint of the vessel, construct an independent centerline offset correction objective function for this segment. The core expression of the objective function is: In the formula, , , These are the normalized contour spatial distance deviation quantification index, the global gray-scale average deviation, and the gradient magnitude deviation evaluation index, respectively. The curvature constraint balance coefficient, This is a penalty for fluctuations in the local curvature of blood vessels. In this embodiment, the specific values ​​of each parameter in the objective function are: , , , Among them, the contour space distance deviation has the highest weight because it most directly reflects the positional offset of the edge; the curvature constraint balance coefficient λ is used to prevent centerline distortion caused by over-correction. The penalty term for local curvature fluctuations in blood vessels... The calculation formula is: in The curvature value at the current point. and The curvature values ​​are those of adjacent points.

[0047] S602. Set the upper limit of the maximum offset of the centerline in a single iteration, use the gradient descent iterative optimization algorithm to solve the minimum value of the objective function, and calculate the gradient direction of the fastest descent of the objective function by taking the first-order partial derivative with respect to the optimization variables. The maximum offset per iteration is capped at 0.8 pixels to prevent abrupt centerline changes due to overcorrection in a single iteration. The learning rate η = 0.1 is used to balance convergence speed and stability. The convergence threshold is 1 × 10⁻⁶ pixels. -3 The iteration stops when the difference between the objective function values ​​of two consecutive iterations is less than the threshold. In this embodiment, the average number of iterations is 15.

[0048] S603. Set the iterative learning rate, and update the coordinate offset of each discrete sampling point on the segmented Xg centerline successively in combination with the gradient direction of the fastest descent of the objective function to complete the single-round coordinate fine-tuning correction. S604. After a single iteration, recalculate the difference feature parameters of the corrected bilateral contours, substitute them into the objective function to solve for the function value again, and compare the changes in the objective function value before and after the two iterations. S605. Set an iteration convergence threshold. When the difference between the objective function of two consecutive iterations is less than the convergence threshold, the iteration operation is determined to have reached a stable state, and the loop update is stopped. S606. Extract the global optimal offset correction amount of the segment Xg finally solved after iterative convergence, and perform point-by-point translation correction on all discrete pixel coordinates of the initial center line of segment Xg according to the independent offset parameters corresponding to each point of the center line. The specific implementation process of step S606 is as follows: S6061. Extract the optimal normal offset array of each centerline sampling point of segment Xg after iterative convergence, perform legality verification on the offset, and truncate abnormal offsets with absolute values ​​exceeding 2.0 pixels to ±2.0 pixels; S6062. For the i-th centerline sampling point, based on its initial coordinates... Normal unit vector and optimal normal offset According to the formula Calculate the corrected coordinates while preserving sub-pixel accuracy; S6063. The 3-point moving average method is used to perform local smoothing on the corrected coordinate sequence to eliminate sub-pixel-level micro-jitter caused by point-by-point independent correction. S6064. Lock the corrected coordinates and normal vectors of the starting and ending points of the segments to provide a reference for the smooth splicing of subsequent adjacent segments; S6065. Generate the final corrected discrete coordinate sequence of the centerline for this segment, and record the offset and correction error information for each point; S607. The corrected segmented Xg discrete centerline coordinate sequence is locally smoothed by using a cubic B-spline smoothing fitting algorithm to eliminate local micro-jitter and abrupt change in inflection point caused by point-by-point independent correction. The specific implementation method of step S607 is as follows: S6071. Obtain the discrete centerline coordinate sequence of segment Xg after point-by-point translation correction. This sequence retains sub-pixel accuracy and contains the axial order and coordinate information of each sampling point of the centerline. Lock the coordinates of the first and last endpoints of the segment to ensure that the splicing reference with adjacent segments remains unchanged.

[0049] S6072. Using the calibrated discrete coordinate points of the centerline as control points of the cubic B-spline, arranged in order of the centerline axis, when the segment length is less than 50 pixels, all coordinate points are used as control points, and when the segment length is greater than 50 pixels, one control point is selected for every three points, taking into account both fitting accuracy and computational efficiency.

[0050] S6073. Construct a cubic uniform B-spline basis function, set the spline order to 3, generate a uniform node vector that matches the control points, and ensure that the node interval is consistent with the sampling interval of the centerline to ensure that the spline curve fits the original direction of the centerline.

[0051] S6074. The least squares method is used to solve the fitting coefficients of the cubic B-spline curve. The control points are used as constraints to generate a smooth and continuous spline curve. During the fitting process, the curvature of the constraint curve changes gradually, avoiding unreasonable bends.

[0052] S6075. Along the original centerline sampling interval, resample on the fitted cubic B-spline curve to generate a smoothed discrete centerline coordinate sequence, and replace the corrected original coordinates point by point.

[0053] S6076. Verify the smoothed coordinate sequence, limit the coordinate smoothing offset to no more than 0.5 pixels, remove abnormal points of curvature change, and finally output a smooth segmented centerline coordinate sequence without local jitter and without inflection point change. S608. Repeat steps S601 to S607 to complete the centerline correction and recalibration of all independent vascular segments in sequence, and obtain the independent corrected centerline data of each segment.

[0054] Furthermore, all segmented corrected centerlines are smoothly stitched together and globally topologically calibrated to obtain a complete and accurate vessel centerline. The geometric parameters of the lumen cross-section are then fitted using bilateral contour information, specifically including: S701. Obtain all segmented and corrected centerline data, extract the overlapping area of ​​two adjacent segments, calculate the coordinate difference of corresponding points of the two segmented and corrected centerlines in the overlapping area, perform smooth transition processing on the centerline coordinates of the overlapping area, and obtain the spliced ​​centerline. Specifically, the length of the overlapping region is 5 pixels. A cubic spline interpolation algorithm is used to smoothly transition the centerline coordinates of the overlapping region. The transition weight changes linearly with the distance from the segment point: the closer to the segment point, the higher the weight of the adjacent segment.

[0055] S702. Using fixed anatomical feature points as constraints, the centerline after splicing is finely adjusted globally to obtain a complete and continuous corrected vascular centerline; Detailed implementation of step S702: S7021. Perform multi-frame consistency and bilateral contour consistency verification on the pre-extracted vascular endpoints, bifurcation points, significant stenosis lesion points and calcified plaque center points, and use the verified feature points as global anchor points and lock their coordinates; S7022. Construct a global deformation energy function based on thin plate spline interpolation. This function consists of a feature point fitting error term and a smoothness constraint term, with a smoothness weight coefficient of 0.01. The formula for the feature point fitting error term is as follows: In the formula, For the deformed first The coordinates of the feature points These are the original locked coordinates of the feature point. Since the feature point coordinates are locked, therefore... =0; The formula for the smoothness constraint term is: This formula measures the sum of squares of the second derivatives of the deformation function f(x,y), i.e., the bending energy. The smaller the bending energy, the smoother the deformation. For m anchor points, the deformation function of thin-plate spline interpolation can be expressed as: In the formula, , , For affine transformation parameters, The weighting coefficient for each anchor point, The total number of anchor points. These are radial basis functions; All deformation parameters can be obtained by solving the following system of equations: In the formula, It is an m×m matrix. , It is an m×3 matrix. , For the weight vector, For affine parameter vectors, y-coordinate vector of the anchor point (the same applies to the x-direction); S7023. Solve the linear equations to obtain the optimal deformation parameters, ensuring that the positions of all anchor points remain unchanged during the deformation process; S7024. Based on the deformation parameters, fine-tune the coordinates of all non-feature points on the centerline after splicing, while preserving sub-pixel accuracy; In this embodiment, for the i-th non-feature point Calculate its deformation in the x and y directions respectively. and The corrected coordinates are The same applies to the y-direction; the coordinates of all feature points remain unchanged. Thin-plate spline interpolation ensures that deformation is precisely matched at the anchor points and smoothly transitioned between anchor points without introducing new inflection points or abrupt changes, thus perfectly preserving the natural shape of blood vessels. S7025. Perform topological consistency, curvature rationality and contour distance checks on the fine-tuned centerline. If any abnormalities are found, adjust the smoothing weight coefficient and solve again. S7026. Generate the final, complete, and continuous corrected vessel centerline for subsequent fitting of the lumen cross-sectional geometry parameters; S703. Along the normal section direction of the corrected blood vessel centerline, simultaneously read the boundary pixel coordinates of the original contour of the reliable reference side and the contour of the corrected weak side, and extract the set of boundary points of the blood vessel wall on both sides of the single section. S704. Obtain the coordinate data of multiple boundary points of the cross section and substitute them into the elliptical cross section fitting model to solve the semi-major axis, semi-minor axis and cross section deflection angle of the elliptical cross section of the blood vessel lumen. In this embodiment, the Random Sampling Consensus (RANSAC) algorithm is used for ellipse fitting to eliminate outlier boundary points with large deviations, thereby improving fitting accuracy. When the number of boundary points involved in the fitting is less than 5, the ellipse parameters of adjacent cross sections are reused to avoid fitting failure.

[0056] S705. Calculate the parameters of the continuous distribution of the lumen cross sections along the direction of the vessel centerline. The specific implementation method of step S705 is as follows: S7051. Based on the corrected complete blood vessel centerline, the cross-section position is determined by 1-pixel equal-interval sampling, and the normal unit vector of each cross-section is recalculated using the center difference method; S7052. Search along the normal vector of each section to the left and right sides, and extract the vessel wall boundary points of the original contour of the reliable reference side and the corrected weak distortion side contour respectively. At least 8 uniformly distributed boundary points are extracted for each section. S7053. The Random Sampling Consensus (RANSAC) algorithm is used to perform robust ellipse fitting on the boundary points. After 100 iterations, the ellipse with the most interior points is selected as the optimal fitting result, and the semi-major axis, semi-minor axis and deflection angle of the ellipse are obtained. S7054. Verify the reasonableness of the dimensions, continuity, and area of ​​the fitted parameters, and mark any abnormal cross sections that do not meet the requirements; S7055. The parameters of abnormal sections are repaired by cubic spline interpolation. If there are more than 3 consecutive abnormal sections, linear interpolation is used. S7056. Use the 5-point moving average method to smooth the parameter sequence and eliminate local random fluctuations; S7057. Calculate the equivalent diameter and cross-sectional area of ​​each section to generate a sequence of lumen cross-sectional parameters for continuous distribution throughout the blood vessel.

[0057] Furthermore, the three-dimensional coordinates of key feature points of the blood vessels are solved, a three-dimensional skeleton model of the coronary arteries is constructed, and surface reconstruction is completed to output a high-precision three-dimensional reconstructed model of the coronary arteries, specifically including: S801. Obtain the calibration parameters of the coronary angiography equipment, extract the imaging camera intrinsic parameter matrix and multi-view extrinsic parameter matrix, and construct the projection matrix; In this embodiment, the projection matrix P is a 3×4 matrix, expressed as: P=K[R|t] where K is a 3×3 intrinsic parameter matrix, R is a 3×3 rotation matrix, and t is a 3×1 translation vector. For dual-view reconstruction, the projection matrices P1 and P2 of the two views are obtained respectively.

[0058] S802. Transform the blood vessel feature points matched from dual perspectives into homogeneous coordinates, and solve the three-dimensional spatial coordinates of key feature points through triangulation; In this embodiment, for the matching points x1 and x2 under dual perspectives, their three-dimensional coordinates X satisfy: P1X = x1, P2X = x2; the above equations are transformed into an overdetermined system of equations AX=0, and the least squares solution of the three-dimensional coordinates X is obtained by solving the system of equations through singular value decomposition (SVD).

[0059] S803. Perform neighborhood distance threshold filtering on 3D feature points to remove outlier noise; S804. Using the cleaned 3D feature points as control points, a 3D skeleton model of the coronary artery is fitted and generated. In this embodiment, a 3D skeleton is fitted using a 3rd-order non-uniform rational B-spline (NURBS) curve, with the number of control points being 1 / 3 of the original number of feature points, to ensure that the curve is smooth and retains key morphological features.

[0060] S805. Combining the set parameters of the lumen cross-section corresponding to the corrected blood vessel centerline, arrange the cross-sectional contours along the three-dimensional skeleton and generate a triangular mesh on the blood vessel surface; In this embodiment, the Delaunay triangulation algorithm is used for surface skinning. Each cross-section generates 16 evenly distributed contour points, and corresponding points of adjacent cross-sections are connected to form a closed triangular mesh. The Laplacian smoothing algorithm iterates 5 times with a smoothing coefficient of 0.5.

[0061] S806. Perform Laplacian smoothing and anatomical topology correction on the three-dimensional curved surface mesh, and finally output the three-dimensional reconstruction model of the coronary artery; The specific implementation method of step S806 is as follows: S8061. Preprocess the initial triangular mesh on the blood vessel surface to remove degenerate patches, isolated vertices, and duplicate vertices, and unify the normal direction of all patches; S8062. The Taubin volume-preserving Laplacian smoothing algorithm is used to smooth the mesh, with a smoothing coefficient λ=0.5, a shrinkage compensation coefficient μ=-0.53, and 5 iterations. During the smoothing process, the coordinates of the vessel bifurcation point, the significant stenosis lesion point, and the center point of the calcified plaque are locked to avoid the real lesion morphology being smoothed. S8063. Based on the coronary tree topology prior, automatically detect four types of topological anomalies in the mesh: self-intersection, boundary holes, non-manifold edges, and bifurcation connection errors. S8064. Targeted repair methods are adopted for different types of topology errors: For self-intersecting regions, intersecting patches are deleted and re-subdivided and filled; for boundary holes, a filling surface conforming to the morphology of blood vessels is generated based on lumen parameters; for non-manifold edges, redundant patches are deleted or missing patches are filled; for bifurcation connection errors, a standard Y-shaped bifurcation structure is regenerated and smoothly spliced. S8065. Optimize the quality of the repaired mesh by achieving mesh uniformity through mesh subdivision and merging, and by performing weighted average smoothing on vertex normals; S8066. Perform anatomical consistency verification to confirm that the deviation between the grid diameter and lumen parameters does not exceed 5%, the topology is consistent with the three-dimensional skeleton, and the lesion morphology matches the two-dimensional measurement results; S8067. Convert the final optimized 3D model into clinical standard formats such as STL and OBJ, and embed the centerline, lumen parameters and lesion quantification information to output a high-precision coronary artery 3D reconstruction model that can be used for clinical diagnosis.

[0062] Reference Figure 2 As shown, this invention proposes a three-dimensional reconstruction system for coronary angiography, used to implement any of the three-dimensional reconstruction methods described above, including: The image preprocessing module is used to acquire multi-angle coronary DSA angiography projection images and perform preprocessing to obtain preprocessed coronary vessel grayscale images. The vascular feature extraction and segmentation module is used to perform multi-scale tubular structure enhancement and edge detection on the preprocessed image, extract the initial center line and bilateral initial contours of the blood vessel, calculate the relative concentration of contrast agent in each section, and combine the local curvature of the blood vessel, the sign change of the concentration difference and anatomical features to adaptively segment a single continuous blood vessel into multiple independent blood vessel segments. The mirror contour generation and difference calculation module is used to screen the credible reference side and the weak distortion side for each blood vessel segment, perform axisymmetric mirror transformation on the credible reference side contour to generate a mirror virtual contour, register the mirror virtual contour with the original contour of the weak distortion side, and calculate the contour spatial distance deviation, gray level mean deviation and edge gradient magnitude deviation. The centerline offset correction module is used to construct independent centerline offset correction objective functions for each segment, and to solve for the optimal offset correction amount using an iterative gradient descent algorithm, thereby completing the correction and recalibration of the initial centerline segment by segment. The centerline integration and lumen fitting module is used to smoothly stitch together and perform global topology calibration on all segmented corrected centerlines to obtain a complete and accurate vascular centerline, and to fit the geometric parameters of the lumen cross-section by combining bilateral contour information. The 3D model construction and output module is used to solve the 3D coordinates of key feature points of blood vessels, construct a 3D skeleton model of coronary arteries and complete surface reconstruction, and output a high-precision 3D reconstructed model of coronary arteries.

[0063] The advantages of this invention are as follows: Addressing the issue of insufficient reconstruction accuracy caused by uneven contrast agent concentration in coronary angiography, it precisely corrects centerline offset caused by unilateral edge blurring by combining adaptive vessel segmentation based on anatomical features with a reliable baseline mirror transformation, significantly reducing diameter measurement errors. A multi-dimensional constrained objective function is constructed to achieve segment-by-segment optimization, coupled with topology calibration anchored to global anatomical features, maintaining the continuity and smoothness of vessel morphology while ensuring correction accuracy. In the 3D reconstruction stage, volume-preserving mesh smoothing and anatomy-driven topology correction are employed to output a dimensionally accurate and topologically correct 3D model, directly supporting clinical quantitative diagnosis and interventional surgery planning.

[0064] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A three-dimensional reconstruction method for coronary angiography, characterized in that, include: S1. Acquire multi-angle coronary DSA angiography projection images and perform preprocessing to obtain preprocessed coronary vessel grayscale images; S2. Perform multi-scale tubular structure enhancement and edge detection on the preprocessed coronary vessel grayscale image, extract the initial center line and bilateral initial contours of the vessel, sample the bilateral pixel grayscale values ​​cross-section by cross-section along the normal of the initial center line, and calculate the relative concentration of contrast agent in each cross-section. S3. Obtain the local curvature of the blood vessel, the sign change of the contrast agent concentration difference, and anatomical feature points, and adaptively segment a single continuous blood vessel into g independent blood vessel segments; S4. For each independent blood vessel segment, the mean concentration of contrast agent on both sides and the stability of the edge gradient within segment Xg are statistically analyzed. The side with higher concentration and more stable edge is selected as the reliable reference side, and the other side is selected as the weak distortion side. With the initial center line of segment Xg as the axis of symmetry, an axisymmetric mirror transformation is performed on the contour of the reliable reference side to generate a mirror virtual contour. S5. Perform pixel-level registration between the mirrored virtual contour of segmented Xg and the original contour of the weak distortion side, and calculate the contour spatial distance deviation, gray-scale mean deviation and edge gradient magnitude deviation. S6. Construct independent objective functions for centerline offset correction for each segment, use iterative gradient descent algorithm to solve for the optimal offset correction amount, and complete the correction and recalibration of the initial centerline segment by segment; S7. Smoothly stitch and perform global topology calibration on all segmented corrected centerlines to obtain complete and accurate vascular centerlines, and fit the geometric parameters of the lumen cross-section by combining bilateral contour information. S8. Solve for the three-dimensional coordinates of key feature points of blood vessels, construct a three-dimensional skeleton model of coronary arteries and complete surface reconstruction, and output a high-precision three-dimensional reconstruction model of coronary arteries.

2. The three-dimensional reconstruction method for coronary angiography according to claim 1, characterized in that, The process of performing multi-scale tubular structure enhancement and edge detection on the preprocessed vascular image, and extracting the initial centerline and bilateral initial contours of the blood vessels, specifically includes: S201. Obtain the pixel grayscale distribution features of the preprocessed coronary artery grayscale image, construct a multi-scale Hessian matrix, set the scale traversal interval to [1,8], and strengthen the linear tubular vessel target area and suppress point noise and sheet tissue interference areas by solving the eigenvalues ​​and eigenvectors of the Hessian matrix at different scales. S202. Set a low gradient threshold T1 and a high gradient threshold T2 to satisfy the quantitative constraint relationship T2=2.5T1. Solve for the pixel gradient magnitude and gradient direction, filter candidate edge pixels with gradient magnitude higher than the low gradient threshold, retain the effective edge pixels that are continuous and connected and satisfy the high gradient threshold constraint, and obtain the complete initial contour of both sides of the blood vessel. S203. Perform pixel-by-pixel erosion refinement on the lumen region surrounded by the edge of the connected blood vessel. Based on the connectivity constraint of eight neighboring pixels, remove redundant interference pixels and iterate continuously until a single-pixel connected skeleton curve is obtained, which is used as the initial center line of the blood vessel.

3. The three-dimensional reconstruction method for coronary angiography according to claim 2, characterized in that, The step of sampling the grayscale values ​​of pixels on both sides cross-section along the normal of the initial centerline and calculating the relative concentration of contrast agent at each cross-section specifically includes: S204. Obtain the discrete pixel coordinate sequence of the initial centerline of the blood vessel, solve for the local tangent vector and normal vector of the centerline, calculate the unit normal vector of the centerline point by point, and determine the bilateral vertical sampling direction of each blood vessel section. S205. Obtain the pointing range of the normal unit vector. Taking the single sampling point of the initial center line as the center, set the single-sided sampling pixel length as d, select a continuous set of pixel sampling points at equal intervals along the double-sided normal direction, and batch collect the original pixel gray values ​​at corresponding positions on both sides. S206. Obtain physical law data of DSA imaging, establish a linear correlation model between contrast agent iodine concentration and X linear attenuation coefficient, satisfying the linear equation: C=α·I+β, where α is the gray-to-concentration linear conversion coefficient, β is the baseline concentration compensation constant, C is the contrast agent iodine mass concentration in the blood vessel lumen, and I is the pixel gray value at this position after digital subtraction. S207. Perform mean calculation on the grayscale data of the sampled pixels on the left and right sides of the blood vessel, respectively, and substitute them into the linear correlation model to calculate the average contrast agent concentration value on each side point by point, so as to obtain the contrast agent relative concentration sequence continuously distributed on both sides of the blood vessel.

4. The three-dimensional reconstruction method for coronary angiography according to claim 3, characterized in that, The process of acquiring local vascular curvature, contrast agent concentration gradient sign changes, and anatomical feature points, and adaptively segmenting a single continuous blood vessel into several independent vascular segments, specifically includes: S301. Obtain the discrete pixel coordinate sequence of the initial centerline of the blood vessel, and solve for the local curvature value of the blood vessel point by point according to the following formula. : ,in Curve function The first derivative, Curve function The second derivative of the curve function In and These are the horizontal and vertical pixel coordinates of the discrete pixel coordinate sequence of the initial centerline of the blood vessel, respectively. S302. Identify curvature abrupt change points whose curvature values ​​exceed a preset curvature threshold and whose continuous length is greater than or equal to 5 pixels, and use them as the first type of candidate segmentation points; S303. Obtain the relative concentration sequence of contrast agent on both sides of the blood vessel, and calculate the difference in contrast agent concentration between the two sides cross-section along the initial centerline using the following formula. : ,in This represents the average concentration of contrast agent in the sampling area on the left side of the cross-section. The average concentration of contrast agent in the sampling area on the right side of the cross section is recorded. The sign change of the concentration difference is recorded. The location points where the sign of the concentration difference is reversed and the continuous length is greater than or equal to 3 pixels are identified as the second type of candidate segmentation points. S304. Extract the bifurcation point of the blood vessel, the endpoint of the blood vessel, the stenosis lesion point with a diameter stenosis rate of ≥50%, and the center point of the calcified plaque as forced segmentation points, and integrate the first type of candidate segmentation points, the second type of candidate segmentation points, and the forced segmentation points to form an initial segmentation point set; S305. Preset minimum and maximum segment lengths, remove redundant segment points in the initial segment point set whose distance is less than the minimum segment length, merge segments whose length is less than the minimum segment length, split segments whose length is greater than the maximum segment length, obtain the optimized segment point set, and then divide a single continuous blood vessel into g independent blood vessel segments. Each segment maintains single connectivity, no bifurcation, smooth curvature change, and consistent contrast agent concentration on the high and low sides. S306. Assign a unique identifier to each independent blood vessel segment, record the starting point coordinates, ending point coordinates, and segment length information of each segment, and establish a blood vessel segment index table.

5. The three-dimensional reconstruction method for coronary angiography according to claim 4, characterized in that, For each independent vascular segment, the mean contrast agent concentration and edge gradient stability within segment Xg are statistically analyzed bilaterally. The side with higher concentration and more stable edges is selected as the reliable baseline side, and the other side is selected as the weak distortion side. Specifically, this includes: S401. Obtain the blood vessel segment index table, and read the pixel data and feature information of each independent blood vessel segment in sequence; S402. For a single vessel segment Xg, calculate the left-side comprehensive confidence score S of segment Xg according to the following formula. L Compared with the comprehensive confidence score S on the right R : ,in This represents the average concentration on the left or right side of segment Xg. For the left or right gradient variability variance of all edge pixels in segment Xg, and All are weighting coefficients; S403. Compare the overall confidence scores on the left and right sides, and select the side with the higher score as the confidence benchmark side for that segment, and the side with the lower score as the weak distortion side for that segment.

6. The three-dimensional reconstruction method for coronary angiography according to claim 5, characterized in that, The step of performing an axisymmetric mirror transformation on the reliable reference side profile, using the initial centerline of segmented Xg as the axis of symmetry, to generate a mirrored virtual profile, specifically includes: S404. Extract all original contour pixel coordinates of the segmented Xg reliable reference side and establish a reliable reference side contour point set; S405. Obtain the initial centerline continuous curve equation of segment Xg, establish a two-dimensional plane rectangular coordinate system with the centerline as the axis of symmetry, and solve for the horizontal and vertical distances and vertical coordinate parameters of any contour pixel relative to the axis of symmetry. S406. Perform coordinate mapping transformation of all blood vessel contours on the reliable reference side pixel by pixel according to the following formula: ,in The coordinates of the reference point on the axis of symmetry For the original contour pixel coordinates of the reliable reference side, These are the mirror pixel coordinates; S407. Perform connectivity verification and breakpoint completion on the discrete contour pixels after mirror transformation, use local linear interpolation algorithm to fill the tiny blank pixel areas generated by mirror transformation, and simultaneously retain all pixel coordinates, grayscale information and edge features of the original contour on the weak distortion side of segmented Xg, and fix the spatial coordinate system and pixel scale specifications of the two types of contours.

7. The three-dimensional reconstruction method for coronary angiography according to claim 6, characterized in that, The step of performing pixel-level registration between the mirrored virtual contour of segmented Xg and the original contour of the weakly distorted side, and calculating the contour spatial distance deviation, gray-level mean deviation, and edge gradient magnitude deviation, specifically includes: S501. Uniformly sample the mirrored virtual contour of segment Xg at preset pixel intervals to obtain a set of feature control points: An iterative nearest-point registration algorithm is used to find the nearest matching point for each mirror control point on the original contour on the weakly distorted side, resulting in the original contour control point set: The pixel distance D between each pair of matching points is calculated using the following formula: The mean of all matching points is used as a quantitative indicator of the spatial distance deviation of the contour. S502. Extract the local pixel grayscale matrix of the mirrored virtual contour and the local pixel grayscale matrix of the original contour on the weakly distorted side, respectively. Calculate the average grayscale value of the corresponding regions of the two contours according to the following formula. : ,in This represents the total number of pixels sampled locally. This represents the grayscale value of a single pixel. S503. Solve for the global average grayscale deviation to quantitatively characterize the regional grayscale distribution differences caused by uneven contrast agent concentration; S504. Obtain the set of pixel gradient magnitude values ​​of the mirrored virtual contour edge and the set of pixel gradient magnitude values ​​of the original contour edge on the weakly distorted side, and calculate the gradient feature residual edge by edge according to the following formula. : By combining the residual mean and the residual distribution interval, a gradient modulus deviation evaluation index is constructed, in which... The gradient magnitude of the edge points of the mirrored virtual contour. The gradient magnitude of the original contour edge points on the weakly distorted side; S505. Normalize the contour spatial distance deviation quantification index, global gray-scale average deviation, and gradient modulus deviation evaluation index to obtain three types of normalized difference characteristic parameters for a single blood vessel segment.

8. The three-dimensional reconstruction method for coronary angiography according to claim 7, characterized in that, The process involves constructing independent centerline offset correction objective functions for each segment, employing an iterative gradient descent algorithm to solve for the optimal offset correction amount, and completing the initial centerline correction and recalibration segment by segment. Specifically, this includes: S601. Obtain the three types of normalized differential feature parameters for a single vessel segment. Using the lateral and normal offsets of the initial centerline of segment Xg as optimization variables, and combining the local curvature smoothing constraint of the vessel, construct an independent centerline offset correction objective function for this segment. The core expression of the objective function is: In the formula, , , These are the normalized contour spatial distance deviation quantification index, the global gray-scale average deviation, and the gradient magnitude deviation evaluation index, respectively. The curvature constraint balance coefficient, This is a penalty term for fluctuations in the local curvature of blood vessels. , , All are weighting coefficients; S602. Set the upper limit of the maximum offset of the centerline in a single iteration, use the gradient descent iterative optimization algorithm to solve the minimum value of the objective function, and calculate the gradient direction of the fastest descent of the objective function by taking the first-order partial derivative with respect to the optimization variables. S603. Set the iterative learning rate, and update the coordinate offset of each discrete sampling point on the segmented Xg centerline successively in combination with the gradient direction of the fastest descent of the objective function to complete the single-round coordinate fine-tuning correction. S604. After a single iteration, recalculate the difference feature parameters of the corrected bilateral contours, substitute them into the objective function to solve for the function value again, and compare the changes in the objective function value before and after the two iterations. S605. Set an iteration convergence threshold. When the difference between the objective function of two consecutive iterations is less than the convergence threshold, the iteration operation is determined to have reached a stable state, and the loop update is stopped. S606. Extract the global optimal offset correction amount of the segment Xg finally solved after iterative convergence, and perform point-by-point translation correction on all discrete pixel coordinates of the initial center line of segment Xg according to the independent offset parameters corresponding to each point of the center line. S607. The corrected segmented Xg discrete centerline coordinate sequence is locally smoothed by using a cubic B-spline smoothing fitting algorithm to eliminate local micro-jitter and abrupt change in inflection point caused by point-by-point independent correction. S608. Repeat steps S601 to S607 to complete the centerline correction and recalibration of all independent vascular segments in sequence, and obtain the independent corrected centerline data of each segment.

9. The three-dimensional reconstruction method for coronary angiography according to claim 8, characterized in that, The process involves smoothing and globally calibrating all segmented corrected centerlines to obtain a complete and accurate vessel centerline. This is then combined with bilateral contour information to fit the geometric parameters of the lumen cross-section. Specifically, this includes: S701. Obtain all segmented and corrected centerline data, extract the overlapping area of ​​two adjacent segments, calculate the coordinate difference of corresponding points of the two segmented and corrected centerlines in the overlapping area, perform smooth transition processing on the centerline coordinates of the overlapping area, and obtain the spliced ​​centerline. S702. Using fixed anatomical feature points as constraints, the centerline after splicing is finely adjusted globally to obtain a complete and continuous corrected vascular centerline; S703. Along the normal section direction of the corrected blood vessel centerline, simultaneously read the boundary pixel coordinates of the original contour of the reliable reference side and the contour of the corrected weak side, and extract the set of boundary points of the blood vessel wall on both sides of the single section. S704. Obtain the coordinate data of multiple boundary points of the cross section and substitute them into the elliptical cross section fitting model to solve the semi-major axis, semi-minor axis and cross section deflection angle of the elliptical cross section of the blood vessel lumen. S705. Calculate the parameters of the entire vessel segment's continuous cross-section by continuously traversing the direction of the vessel's centerline.

10. A three-dimensional reconstruction system for coronary angiography, used to implement the three-dimensional reconstruction method as described in any one of claims 1-9, characterized in that, include: The image preprocessing module is used to acquire multi-angle coronary DSA angiography projection images and perform preprocessing to obtain preprocessed coronary vessel grayscale images. The vascular feature extraction and segmentation module is used to perform multi-scale tubular structure enhancement and edge detection on the preprocessed image, extract the initial center line and bilateral initial contours of the blood vessel, calculate the relative concentration of contrast agent in each section, and combine the local curvature of the blood vessel, the sign change of the concentration difference and anatomical features to adaptively segment a single continuous blood vessel into multiple independent blood vessel segments. The mirror contour generation and difference calculation module is used to screen the credible reference side and the weak distortion side for each blood vessel segment, perform axisymmetric mirror transformation on the credible reference side contour to generate a mirror virtual contour, register the mirror virtual contour with the original contour of the weak distortion side, and calculate the contour spatial distance deviation, gray level mean deviation and edge gradient magnitude deviation. The centerline offset correction module is used to construct independent centerline offset correction objective functions for each segment, and to solve for the optimal offset correction amount using an iterative gradient descent algorithm, thereby completing the correction and recalibration of the initial centerline segment by segment. The centerline integration and lumen fitting module is used to smoothly stitch together and perform global topology calibration on all segmented corrected centerlines to obtain a complete and accurate vascular centerline, and to fit the geometric parameters of the lumen cross-section by combining bilateral contour information. The 3D model construction and output module is used to solve the 3D coordinates of key feature points of blood vessels, construct a 3D skeleton model of coronary arteries and complete surface reconstruction, and output a high-precision 3D reconstructed model of coronary arteries.

Images