A method and device for predicting blood flow reserve fraction based on coronary CT images

Abstract

An embodiment of the present invention relates to a method and apparatus for predicting blood flow reserve fraction based on coronary CT images. The method comprises: performing vascular network segmentation and inner wall edge and centerline identification on a CT image sequence; extracting blood flow velocity features based on a semantic graph sequence; recording the last image as a full coronary image and extracting image features from it; extracting vascular cross-section / stenosis / topology features based on the full coronary image and the corresponding semantic graph; fusing three types of features; searching for vascular paths on the fused feature graph; sampling path points and extracting features from the sampling points on each path; constructing a vascular tree based on the fused features; performing three-dimensional feature conversion on each path feature sequence; performing prediction based on each path feature sequence based on an FFR prediction model; and marking and displaying warnings on the vascular tree and full coronary image based on the prediction results. The present invention can avoid damage to the subject's blood vessels and improve the subject's comfort.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method and device for predicting blood flow reserve fraction based on coronary CT images. Background Art

[0002] Fractional Flow Reserve (FFR) refers to the ratio of the maximum blood flow that can be obtained in the myocardial area supplied by the coronary artery (CA) when there is a stenosis lesion in the coronary artery (CA), to the maximum blood flow that can be obtained in the same area under normal conditions in theory. The current conventional method of measuring FFR score is an invasive interventional test method. This invasive test method will cause certain vascular trauma to the subject each time it is tested, and a sufficiently long vascular recovery period must be separated between each two tests. In other words, this invasive test method cannot be provided to the subject as a sustainable monitoring method with a short time interval for FFR monitoring. Summary of the Invention

[0003] The purpose of the present invention is to address the deficiencies of the prior art and provide a method, device, electronic device and computer-readable storage medium for predicting blood flow reserve fraction based on coronary CT images. The present invention performs semantic segmentation of the vascular network on each CT image of a coronary electronic computed tomography (CT) image sequence to obtain a first semantic map, and identifies the inner wall edge and the center line of the vascular network on each first semantic map to obtain a second semantic map, and extracts blood flow velocity features based on the second semantic map sequence to obtain a first feature map; and records the last CT image in the CT image sequence as a full coronary image and extracts image features from it to obtain a second feature map; and extracts vascular cross-sectional features, vascular stenosis features and vascular topology features based on the full coronary image and the second semantic map to obtain a third feature map; and performs feature fusion on the first, second and third feature maps to obtain a first fused feature map; and extracts the first fused feature map. The present invention provides a method for performing a vascular path search based on a graph to obtain a first path sequence; sampling path points on each first path and extracting comprehensive features of each sampling point to obtain a first path feature sequence; constructing a three-dimensional vascular tree based on the first fusion feature graph to obtain a first vascular tree; performing three-dimensional feature conversion on each first path feature sequence based on the first vascular tree to obtain a second path feature sequence; performing FFR prediction based on each second path feature sequence using an FFR prediction model implemented based on a reference Bi-LSTM model structure to obtain a first path prediction sequence; marking the first vascular tree and the whole coronary image with FFR alarms based on all the obtained first path prediction sequences; and displaying the marked vascular tree and the marked image graph. The non-invasive blood flow reserve fraction prediction mechanism provided by the present invention can avoid damage to the subject's blood vessels, improve the subject's comfort, and serve as a sustainable monitoring method with a short time interval.

[0004] To achieve the above objectives, a first aspect of an embodiment of the present invention provides a method for predicting fractional flow reserve based on coronary CT images, the method comprising:

[0005] A received coronary CT image sequence is recorded as a corresponding first CT image sequence; the first image sequence includes a plurality of first CT images; each first CT image corresponds to a first image time; all first CT images have the same two-dimensional size; in the first CT image sequence, the first first CT image is a CT image when a contrast agent initially enters a coronary artery entrance, and the last first CT image is a CT image when the contrast agent fills the entire coronary vascular network;

[0006] Performing vascular network semantic segmentation processing on each of the first CT images to obtain a corresponding first semantic map; performing vascular inner wall edge and vascular centerline identification processing on the vascular network on each of the first semantic maps to obtain a corresponding second semantic map; and chronologically sorting all of the obtained second semantic maps to form a corresponding second semantic map sequence; the first and second semantic maps correspond one-to-one to the first CT image and are consistent with the two-dimensional size of the corresponding first CT image;

[0007] performing blood flow velocity feature extraction processing on the second semantic graph sequence to obtain a corresponding first feature graph; the first feature graph corresponds one-to-one to the first CT image and is consistent with the two-dimensional size of the corresponding first CT image;

[0008] Recording the last first CT image in the first CT image sequence as the corresponding full coronary image; performing image feature extraction processing on the full coronary image to obtain a corresponding second feature map; and performing vascular cross-sectional feature, vascular stenosis feature, and vascular topology feature extraction processing based on the full coronary image and the second semantic map to obtain a corresponding third feature map;

[0009] performing feature fusion processing on the first, second, and third feature maps corresponding to the full coronary image to obtain a corresponding first fused feature map;

[0010] Performing a vascular path search process on the first fusion feature map to obtain a corresponding first path sequence; sampling path points on each first path and extracting comprehensive features of each sampling point to obtain a corresponding first path feature sequence; the first path sequence includes a plurality of first paths; the first path feature sequence is composed of a plurality of first point two-dimensional features sequentially arranged;

[0011] Constructing a three-dimensional vascular tree based on the first fusion feature map to obtain a corresponding first vascular tree; and performing three-dimensional feature conversion on each of the first path feature sequences based on the first vascular tree to obtain a corresponding second path feature sequence; the second path feature sequence is composed of a plurality of first point three-dimensional features arranged in sequence;

[0012] Performing FFR prediction processing on each of the second path feature sequences based on an FFR prediction model to obtain a corresponding first path prediction sequence; the FFR prediction model is implemented based on a Bi-LSTM model structure; the first path prediction sequence includes multiple first-point FFR prediction values;

[0013] Perform FFR warning marking processing on the first vascular tree and the whole coronary image according to all the obtained first path prediction sequences to obtain corresponding first marked vascular tree and first marked image map; and display the first marked vascular tree and the first marked image map.

[0014] Preferably, the shape of the first CT image is H CT ×W CT ×C CT ;H CT 、W CT 、C CT are all integers greater than 0, H CT 、W CT are the height and width of the two-dimensional size of the first CT image, C CT is the characteristic dimension of the first CT image; the first CT image includes H CT ×W CT First pixel points; each of the first pixel points corresponds to a vector with a length of C CT The first pixel feature vector includes at least a first pixel coordinate and a first pixel value;

[0015] The shape of the first semantic graph is H SEG1 ×W SEG1 ×C SEG1 ;H SEG1 、W SEG1 、C SEG1 are all integers greater than 0, H SEG1 、W SEG1 are the height and width of the two-dimensional size of the first semantic graph, C SEG1 is the feature dimension of the first semantic graph, H SEG1 =H CT 、W SEG1 =W CT ; The first semantic graph includes H SEG1 ×W SEG1 second pixel points; each of the second pixel points corresponds to a vector with a length of C SEG1 The second pixel feature vector includes at least a second pixel coordinate and a first semantic type; the first semantic type includes a blood vessel point type and a background point type;

[0016] The shape of the second semantic graph is H SEG2 ×W SEG2 ×C SEG2 ;H SEG2 、W SEG2 、C SEG2 are all integers greater than 0, H SEG2 、W SEG2 are the height and width of the two-dimensional size of the second semantic graph, C SEG2 is the feature dimension of the second semantic graph, H SEG2 =H CT 、W SEG2 =WCT ; The second semantic graph includes H SEG2 ×W SEG2 third pixel points; each of the third pixel points corresponds to a vector with a length of C SEG2 The third pixel feature vector includes at least a third pixel coordinate and a second semantic type; the second semantic type includes a blood vessel edge point type, a blood vessel center point type and a background point type;

[0017] The shape of the first feature map is H F1 ×W F1 ×C F1 ;H F1 、W F1 、C F1 are all integers greater than 0, H F1 、W F1 are the height and width of the two-dimensional size of the first feature map, C F1 is the feature dimension of the first feature map, H F1 =H CT 、W F1 =W CT ; The first feature map includes H F1 ×W F1 fourth pixel points; each of the fourth pixel points corresponds to a vector with a length of C F1 The fourth pixel feature vector includes at least the fourth pixel coordinates and the blood flow velocity feature;

[0018] The shape of the second feature map is H F2 ×W F2 ×C F2 ;H F2 、W F2 、C F2 are all integers greater than 0, H F2 、W F2 are the height and width of the two-dimensional size of the second feature map, C F2 is the feature dimension of the second feature map, H F2 =H CT 、W F2 =W CT ; The second feature map includes H F2 ×W F2 fifth pixel points; each of the fifth pixel points corresponds to a vector with a length of C F2The fifth pixel feature vector includes at least the fifth pixel coordinate, the image gradient feature and the image texture feature vector; the image texture feature vector includes at least the image GLCM feature, the image GLSZM feature, the image GLRLM feature, the image GLDM feature and the image NGTDM feature;

[0019] The shape of the third characteristic map is H F3 ×W F3 ×C F3 ;H F3 、W F3 、C F3 are all integers greater than 0, H F3 、W F3 are the height and width of the two-dimensional size of the third feature map, C F3 is the feature dimension of the third feature map, H F3 =H CT 、W F3 =W CT ; The third characteristic map includes H F3 ×W F3 Sixth pixel points; each of the sixth pixel points corresponds to a vector with a length of C F3 The sixth pixel feature vector includes at least the sixth pixel coordinate, the blood vessel cross-section feature vector, the blood vessel stenosis feature vector and the blood vessel topology feature vector;

[0020] When the sixth pixel point does not correspond to a blood vessel center point, the blood vessel cross-section feature vector, the blood vessel stenosis feature vector, and the non-empty blood vessel topology feature vector are empty;

[0021] When the sixth pixel point corresponds to a blood vessel center point, the blood vessel cross-section feature vector includes at least the cross-section area and the cross-section brightness, the blood vessel stenosis feature vector includes at least the proximal radius, the minimum radius and the distal radius, and the blood vessel topology feature vector includes at least the topology point type, the front point pixel coordinates, the back point pixel coordinates and the blood vessel curvature; the topology point types include the coronary artery entrance center point, the middle straight center point, the middle bifurcation center point and the terminal center point; when the topology point type is the coronary artery entrance center point, the corresponding front point pixel coordinates are empty, and the corresponding back point pixel coordinates are the pixel coordinates of the next blood vessel center point connected to the blood vessel center point corresponding to the current pixel point; when the topology point type is the middle straight center point, the corresponding front and back point pixel coordinates are empty, and the corresponding back point pixel coordinates are the pixel coordinates of the next blood vessel center point connected to the blood vessel center point corresponding to the current pixel point; when the topology point type is the middle straight center point, the corresponding front and back point pixel coordinates are empty, and the corresponding back point pixel coordinates are the pixel coordinates of the next blood vessel center point connected to the blood vessel center point corresponding to the current pixel point. The coordinates are the pixel coordinates of the previous and next blood vessel center points connected to the blood vessel center point corresponding to the current pixel point; when the topological point type is an intermediate bifurcation center point, the corresponding front point pixel coordinates are the pixel coordinates of the previous blood vessel center point connected to the blood vessel center point corresponding to the current pixel point, and the corresponding back point pixel coordinates are composed of the pixel coordinates of the next blood vessel center point connected to the blood vessel center point corresponding to the current pixel point in the multiple rear bifurcated branches; when the topological point type is an end center point, the corresponding front point pixel coordinates are the pixel coordinates of the previous blood vessel center point connected to the blood vessel center point corresponding to the current pixel point, and the corresponding back point pixel coordinates are empty; the blood vessel curvature is the blood vessel centerline curvature of the blood vessel center point corresponding to the current pixel point;

[0022] The shape of the first fusion feature map is H FM ×W FM ×C FM ;H FM 、W FM 、C FM are all integers greater than 0, H FM 、W FM are the height and width of the two-dimensional size of the first fusion feature map, C FM is the feature dimension of the first fusion feature map, H FM =H CT 、W FM =W CT ; The first fusion feature map includes H FM ×W FM Seventh pixel points; each of the seventh pixel points corresponds to a vector with a length of C FM the seventh pixel feature vector; the seventh pixel feature vector includes at least the seventh pixel coordinate, the blood flow velocity feature, the image gradient feature, the image texture feature vector, the blood vessel cross-section feature vector, the blood vessel stenosis feature vector and the blood vessel topology feature vector.

[0023] Preferably, performing vascular network semantic segmentation processing on each of the first CT images to obtain a corresponding first semantic map specifically includes:

[0024] Set a shape of H for the current first CT image SEG1 ×W SEG1 ×C SEG1 The first semantic map; and initializing the first semantic type corresponding to each second pixel point in the current first semantic map to a background point type;

[0025] Each of the first pixel points of the current first CT image is taken as the corresponding current image pixel point; and when the first pixel value of the current image pixel point meets the preset contrast agent pixel value range, the first semantic type of the second pixel point corresponding to the second pixel coordinate on the current first semantic map and the first pixel coordinate of the current image pixel point is reset to a blood vessel point type.

[0026] Preferably, the step of performing vascular inner wall edge and vascular centerline recognition processing on each vascular network on the first semantic graph to obtain the corresponding second semantic graph specifically includes:

[0027] Set a shape H for the current first semantic graph SEG2 ×W SEG2 ×C SEG2 The second semantic map; and initializing the second semantic type corresponding to each of the third pixel points in the current second semantic map to a background point type;

[0028] taking each second pixel point of the first semantic type in the current first semantic map as a blood vessel point type as a corresponding lumen pixel point; and recording a pixel point connected image region consisting of all the lumen pixels as a corresponding blood vessel network region;

[0029] Performing regional edge recognition on the vascular network region, and taking the lumen pixel points on any edge line of the vascular network region as corresponding inner wall edge pixel points; and resetting the second semantic type of the third pixel points corresponding to the third pixel coordinates on the current second semantic map and the second pixel coordinates of each inner wall edge pixel point to a vascular edge point type;

[0030] The vascular centerline of the vascular network area is identified, and the lumen pixel points on any centerline of the vascular network area are used as corresponding centerline pixel points; and the second semantic type of the third pixel point corresponding to the third pixel coordinates on the current second semantic map and the second pixel coordinates of each centerline pixel point is reset to the vascular center point type.

[0031] Preferably, the step of performing blood flow velocity feature extraction processing according to the second semantic graph sequence to obtain the corresponding first feature graph specifically includes:

[0032] The total number of the second semantic graphs in the second semantic graph sequence is counted to obtain the corresponding first total number N; and each second semantic graph is recorded as a corresponding semantic graph P i ; and the first image time corresponding to each second semantic graph is recorded as the corresponding t i ; 1≤index i≤N;

[0033] And in each semantic graph P i Based on all the third pixel points whose second semantic type is the blood vessel center point type, the centerline skeleton is extracted to obtain the corresponding centerline skeleton B i ; and press the center line skeleton B i＝N The direction to achieve maximum overlap is for each centerline skeleton B 1≤i<N With the centerline skeleton B i＝N Perform registration;

[0034] And the semantic graph P i＝1 Each end point on the centerline skeleton B i＝N The corresponding point on is recorded as the corresponding time point d i＝1 ; and the semantic graph P i＝2 Each end point on the centerline skeleton B i＝N The corresponding point on is recorded as the corresponding time point d i＝2 ; And so on, finally the semantic graph P i＝N-1 Each end point on the centerline skeleton B i＝N The corresponding point on is recorded as the corresponding time point d i＝N-1 Then the center line skeleton B i＝N Each end point on the i＝N ;

[0035] And the centerline skeleton B i＝N From the skeleton starting point to each of the time points d i＝N The unique path is used as a corresponding first skeleton path; and the N time points d passed by each of the first skeleton paths are i Form a corresponding first moment point sequence;

[0036] and taking each of the first time point sequences as the corresponding current time point sequence; and taking the time point d in the current time point sequence as the corresponding current time point sequence; i＝N The corresponding speed v i＝N Initialized to 0, and at the time point d i＝N At the time point di＝N-1 The center line length is calculated to obtain the corresponding center line length s N-1 , and the time point d i＝N With the time point d i＝N-1 The time difference between them is recorded as the corresponding time difference △t N-1 =t i＝N -t i＝N-1 , and according to the time difference △t N-1 , the centerline length s N-1 and the speed v i＝N Calculate the time point d i＝N-1 The corresponding speed v i＝N-1 , And for the time point d i＝N-1 At time d i＝N-2 The center line length is calculated to obtain the corresponding center line length s N-2 , and the time point d i＝N-1 With the time point d i＝N-2 The time difference between them is recorded as the corresponding time difference △t N-2 =t i＝N-1 -t i＝N-2 , and according to the time difference △t N-2 , the centerline length s N-2 and the speed v i＝N-1 Calculate the time point d i＝N-2 The corresponding speed v i＝N-2 , And so on, until the time point d in the current time point sequence is calculated. i＝1 The corresponding speed v i＝1 until;

[0037] And the center line skeleton B i＝N All the time points d i Traverse; and when traversing, the current traversal time point d i Recorded as the corresponding current moment, and the third pixel point corresponding to the current moment as the corresponding current pixel point; and whether the current moment corresponds to two or more speeds v i Identify; if so, all the speeds v corresponding to the current moment i The average value of is taken as the first velocity feature corresponding to the current pixel point; if not, the velocity v uniquely corresponding to the current moment point is taken as i as the first velocity feature corresponding to the current pixel;

[0038] And at the center line skeleton B i＝NBased on the linear interpolation method, the time points d of two adjacent moments connected to each other are calculated. i-1 d i estimating the first velocity features of other third pixel points between the two points on the skeleton using the corresponding two first velocity features;

[0039] And the semantic graph P i＝N Set a shape to H F1 ×W F1 ×C F1 Initialize the blood flow velocity feature corresponding to each of the fourth pixel points in the current first feature map to 0; and compare the fourth pixel coordinates on the current first feature map with the centerline skeleton B i＝N The blood flow velocity feature of the fourth pixel point corresponding to the third pixel coordinate of each of the third pixel points is reset to the corresponding first velocity feature.

[0040] Preferably, performing image feature extraction processing on the full coronary image to obtain the corresponding second feature map specifically includes:

[0041] Set a shape H for the whole coronary image F2 ×W F2 ×C F2 The second feature map; and initializing the image gradient feature and the image texture feature vector corresponding to each fourth pixel point in the current second feature map to empty;

[0042] and performing grayscale conversion on the full coronary image to obtain a corresponding first grayscale image; the first grayscale image includes a plurality of first grayscale pixels; the first grayscale pixels correspond one-to-one to the first pixels of the first CT image; each of the first grayscale pixels corresponds to a first grayscale pixel feature vector; the first grayscale pixel feature vector includes at least grayscale pixel coordinates and grayscale values;

[0043] and calculating the two-dimensional grayscale gradient of each of the first grayscale pixels on the first grayscale image to obtain a corresponding first gradient; and resetting the image gradient feature of the fifth pixel corresponding to the fifth pixel coordinates on the current second feature map and the grayscale pixel coordinates of each of the first grayscale pixels to the corresponding first gradient;

[0044] The grayscale co-occurrence matrix, grayscale area size matrix, grayscale run matrix, grayscale dependency matrix and adjacent grayscale difference matrix of the first grayscale image are calculated to obtain the corresponding first, second, third, fourth and fifth matrices; and the matrix unit data corresponding to each first grayscale pixel in the first, second, third, fourth and fifth matrices are extracted as the corresponding first GLCM data, first GLSZM data, first GLRLM data, first GLDM data and first NGTDM data; and the image GLCM features, image GLSZM features, image GLRLM features, image GLDM features and image NGTDM features of the image texture feature vector of the fifth pixel corresponding to the fifth pixel coordinates on the current second feature map and the grayscale pixel coordinates of each first grayscale pixel are reset to the corresponding first GLCM data, the first GLSZM data, the first GLRLM data, the first GLDM data and the first NGTDM data.

[0045] Preferably, the extracting of vascular cross-sectional features, vascular stenosis features, and vascular topological features based on the full coronary image and the second semantic map to obtain a corresponding third feature map specifically includes:

[0046] Drawing a blood vessel edge line based on all the third pixel points whose second semantic type is a blood vessel edge point type on the second semantic map; and drawing a blood vessel centerline based on all the third pixel points whose second semantic type is a blood vessel center point type;

[0047] And traverse all the third pixel points of the second semantic type that are the blood vessel center point type on the second semantic map; and during the traversal, use the currently traversed third pixel point as the corresponding current center pixel point; and draw a perpendicular line to the blood vessel centerline at the location of the current center pixel point, and use the straight line segment between the two intersection points of the perpendicular line and the blood vessel edge as the corresponding first line segment, and use half the length of the first line segment as the corresponding minimum radius; and calculate the corresponding cross-sectional area based on the minimum radius; and take the average value of the first pixel values ​​of all the first pixel points that the first line segment passes through on the full coronary image. as the corresponding second pixel value, and based on the conversion relationship between pixel value and brightness value, perform brightness value conversion on the second pixel value and use the conversion result as the corresponding cross-sectional brightness; and use the Mth third pixel point on the blood vessel centerline where the current center pixel point is located, which is closer to the center point of the coronary artery entrance before the current center pixel point, as the corresponding proximal center point, and the Mth third pixel point which is farther away from the center point of the coronary artery entrance after the current center pixel point is as the corresponding distal center point, where M is an integer greater than 0; and draw perpendicular lines to the blood vessel centerline at the positions of the proximal and distal center points, and connect the perpendicular lines to the two sides of the blood vessel edge. half the length of the straight line segment between the intersection points as the corresponding proximal radius and the distal radius; and when the current center pixel point is the coronary entrance center point, the corresponding topological point type is set to the coronary entrance center point, and when the current center pixel point is an end point of a blood vessel center line, the corresponding topological point type is set to the end center point, and when the current center pixel point intersects only one blood vessel center line and is neither the coronary entrance center point nor an end point, the corresponding topological point type is set to the middle straight center point, and when the current center pixel point intersects with two or more blood vessel center lines, the corresponding The topological point type is a middle bifurcation center point; and the corresponding front point pixel coordinates and the back point pixel coordinates are set according to the topological point type; and the curvature of the blood vessel centerline at the location of the current center pixel point is estimated and the estimation result is used as the corresponding blood vessel curvature; and the obtained cross-sectional area and the cross-sectional brightness form a corresponding first eigenvector, and the obtained proximal radius, the minimum radius and the distal radius form a corresponding second eigenvector, and the obtained topological point type, the front point pixel coordinates, the back point pixel coordinates and the blood vessel curvature form a corresponding third eigenvector;

[0048] And set a shape H for the second semantic graph F3 ×W F3 ×C F3The third feature map of the present invention; and the blood vessel cross-section feature vector, the blood vessel stenosis feature vector and the blood vessel topology feature vector corresponding to each sixth pixel point in the current third feature map are initialized to empty; and the blood vessel cross-section feature vector, the blood vessel stenosis feature vector and the blood vessel topology feature vector of the sixth pixel point corresponding to the sixth pixel coordinate on the current third feature map and the third pixel point of each second semantic type being the blood vessel center point type are reset to the corresponding first, second and third feature vectors.

[0049] Preferably, performing feature fusion processing on the first, second, and third feature maps corresponding to the full coronary image to obtain the corresponding first fused feature map specifically includes:

[0050] Set a shape H for the whole coronary image FM ×W FM ×C FM The first fusion feature map; and initializing the blood flow velocity feature, the image gradient feature, the image texture feature vector, the blood vessel cross-section feature vector, the blood vessel stenosis feature vector, and the blood vessel topology feature vector corresponding to each of the seventh pixel points in the current first fusion feature map to corresponding zero-value features or zero-value feature vectors;

[0051] All the seventh pixel points of the current first fusion feature map are traversed; and the currently traversed seventh pixel point is used as the corresponding current pixel point during the traversal; and the fourth, fifth and sixth pixel points corresponding to the fourth, fifth and sixth pixel coordinates in the first, second and third feature maps and the seventh pixel coordinate of the current pixel point are recorded as the corresponding first, second and third matching pixel points; and the blood flow velocity feature of the current pixel point is set to the blood flow velocity feature of the corresponding first matching pixel point; and the image gradient feature and the image texture feature vector of the current pixel point are set to the image gradient feature and the image texture feature vector of the corresponding second matching pixel point; and the blood vessel cross-section feature vector, the blood vessel stenosis feature vector and the blood vessel topology feature vector of the current pixel point are set to the blood vessel cross-section feature vector, the blood vessel stenosis feature vector and the blood vessel topology feature vector of the corresponding third matching pixel point.

[0052] Preferably, performing a vascular path search process on the first fusion feature map to obtain a corresponding first path sequence specifically includes:

[0053] On the first fusion feature map, a vascular skeleton is extracted based on the vascular topology feature vectors of all non-zero feature vectors to obtain a corresponding first vascular skeleton; and the skeleton point on the first vascular skeleton corresponding to the seventh pixel point whose topology point type is the coronary entrance center point is recorded as a skeleton starting point, and the skeleton point corresponding to the seventh pixel point whose topology point type is the end center point is recorded as a skeleton end point; and the unique path from the skeleton starting point to each of the skeleton end points is taken as a corresponding first path; and all the obtained first paths form a corresponding first path sequence.

[0054] Preferably, sampling path points on each of the first paths and extracting comprehensive features of each sampling point to obtain a corresponding first path feature sequence specifically includes:

[0055] On the first fusion feature map, starting from the path starting point of each first path and downward along the current path, pixel sampling is performed every preset first interval distance, and the seventh pixel coordinates, the blood flow velocity feature, the image gradient feature, the image texture feature vector, the blood vessel cross-section feature vector, the blood vessel stenosis feature vector and the blood vessel topology feature vector of the currently sampled seventh pixel point are extracted to form a corresponding first-point two-dimensional feature; and all the first-point two-dimensional features corresponding to the current path are sorted in the sampling order to form a corresponding first path feature sequence.

[0056] Preferably, performing three-dimensional feature conversion on each of the first path feature sequences according to the first vascular tree to obtain a corresponding second path feature sequence specifically includes:

[0057] Performing a three-dimensional vascular centerline skeleton extraction process on the first vascular tree to obtain a corresponding second vascular skeleton; recording a voxel point on the second vascular skeleton corresponding to a center point of a coronary artery entrance as a first starting point, and recording a voxel point corresponding to an end point of each vascular tree as a second end point; and recording a unique path from the first starting point to each second end point as a corresponding first centerline path; the first centerline path and the first path have a one-to-one correspondence;

[0058] Starting from the first starting point of each first centerline path and downward along the current centerline path, voxel point sampling is performed every preset first interval distance and the first point two-dimensional feature corresponding to the current sampled voxel point is used as the corresponding current point feature; the seventh pixel coordinate of the current point feature is updated to the corresponding three-dimensional voxel coordinate; the image gradient feature of the current point feature is updated to the corresponding three-dimensional gradient feature; the front point pixel coordinate, the back point pixel coordinate and the blood vessel curvature of the blood vessel topology feature vector of the current point feature are updated to the corresponding three-dimensional front point voxel coordinate, three-dimensional back point voxel coordinate and three-dimensional blood vessel curvature; the updated current point feature is used as a corresponding first point three-dimensional feature; and all the first point three-dimensional features corresponding to the current path are sorted in the sampling order to form a corresponding second path feature sequence.

[0059] Preferably, performing FFR warning marking processing on the first vessel tree and the whole coronary image according to all the obtained first path prediction sequences to obtain the corresponding first marked vessel tree and first marked image map specifically includes:

[0060] A corresponding first prediction value set is formed by all first-point FFR prediction values ​​of all first-path prediction sequences; the three-dimensional voxel coordinates corresponding to each first-point FFR prediction value in the first prediction value set are marked as corresponding first voxel coordinates; one or more first-point FFR prediction values ​​corresponding to the same first voxel coordinate in the first prediction value set are grouped into a class to form a corresponding first voxel prediction value set; all first-point FFR prediction values ​​of each first voxel prediction value set are averaged to obtain a corresponding first voxel FFR average; the first voxel FFR average that is lower than a preset FFR threshold is recorded as the corresponding first alarm point FFR average; and the first voxel coordinates corresponding to each first alarm point FFR average are used as the corresponding first alarm point coordinates;

[0061] And identify whether the total number of the first alarm point coordinates is zero; if the total number of the first alarm point coordinates is zero, the first vascular tree and the full coronary image are used as the corresponding first marked vascular tree and the first marked image map; if the total number of the first alarm point coordinates is greater than zero, the first alarm point FFR mean corresponding to the preset first alarm color mark is calculated at the voxel point corresponding to each first alarm point coordinate on the first vascular tree, and the first alarm point FFR mean corresponding to the first alarm color mark is calculated at the first pixel point corresponding to each first alarm point coordinate on the full coronary image, and the first vascular tree and the full coronary image that have completed all the first alarm point FFR mean mark markings are used as the corresponding first marked vascular tree and the first marked image map.

[0062] Preferably, the method further comprises: performing model training on the FFR prediction model before using the FFR prediction model;

[0063] The performing model training on the FFR prediction model specifically includes:

[0064] Step 1301: recruit multiple volunteers without coronary artery stenosis symptoms and multiple volunteers with coronary artery stenosis symptoms of varying degrees to form a first volunteer group;

[0065] Wherein, the first volunteer set includes a plurality of first volunteers;

[0066] Step 1302: Perform multiple coronary artery path tests on each of the first volunteers using an invasive testing method. On each test path, feature data is collected from test points along the test path according to the data format of the first point's two-dimensional feature to obtain a corresponding test path two-dimensional feature sequence. The blood flow reserve fraction corresponding to each test point along the test path is measured to obtain a corresponding test path measurement sequence. Three-dimensional vascular tree modeling is performed based on the two-dimensional feature sequence to obtain a corresponding test path vascular tree. Based on the test path vascular tree, three-dimensional data conversion is performed on the two-dimensional feature sequence according to the data format of the first point's three-dimensional feature to obtain a corresponding test path three-dimensional feature sequence. The three-dimensional feature sequence corresponding to each test path and the test path measurement sequence are combined into a corresponding first data record. All of the obtained first data records are combined into a corresponding first data set.

[0067] The detection path two-dimensional feature sequence includes a plurality of detection point two-dimensional features, and the data format of the detection point two-dimensional features matches the data format of the first point two-dimensional features; the detection path three-dimensional feature sequence includes a plurality of detection point three-dimensional features, and the data format of the detection point three-dimensional features matches the data format of the first point three-dimensional features; the detection path measurement sequence includes a plurality of detection point FFR measurement values;

[0068] Step 1303: Set an integer greater than zero as the corresponding continuous pass count threshold; and set a pass count counter initialized to 0;

[0069] Step 1304: randomly select a plurality of first data records from the first data set according to a preset test set ratio to form a corresponding current test set, and use all remaining first data records to form a corresponding current validation set;

[0070] The current test set and the current validation set are both composed of a plurality of the first data records, and a ratio of the total number of the first data records in the current test set to the total number of the first data records in the first data set satisfies the test set ratio;

[0071] Step 1305: extract the first first data record of the current test set as the corresponding current data record;

[0072] Step 1306: Input the three-dimensional characteristic sequence of the detection path of the current data record into the FFR prediction model, and the FFR prediction model performs FFR prediction processing based on the three-dimensional characteristic sequence of the detection path input this time to obtain a corresponding detection path prediction sequence;

[0073] Wherein, the detection path prediction sequence includes multiple detection point FFR prediction values;

[0074] Step 1307 , subjecting the detection path prediction sequence and the detection path measurement sequence of the current data record to a preset first model loss function to calculate and obtain a corresponding first loss value;

[0075] Wherein, the first model loss function is implemented based on the L1 loss function or the L2 loss function;

[0076] Step 1308: Identify whether the first loss value satisfies a preset first loss value range. If the first loss value satisfies the first loss value range, identify whether the current data record is the last first data record in the current test set. If so, proceed to step 1309. If not, extract the next first data record in the current test set as the new current data record and return to step 1306 to continue training. If the first loss value does not satisfy the first loss value range, perform a round of model parameter optimization on the FFR prediction model based on a preset first model parameter optimizer in a direction to minimize the first model loss function. After this round of parameter optimization is completed, return to step 1306 to continue training.

[0077] Wherein, the first model parameter optimizer includes at least an SDG optimizer;

[0078] Step 1309: Traverse all first data records in the current validation set. During the traversal, the currently traversed first data record is used as the corresponding current validation record. The FFR prediction model performs FFR prediction processing based on the detection path feature sequence of the current validation record to obtain a corresponding current detection path prediction sequence. The current detection path prediction sequence and the detection path measurement sequence of the current validation record form a corresponding prediction-measurement sequence pair. At the end of the traversal, all the obtained prediction-measurement sequence pairs are substituted into preset first, second, and third error functions to calculate corresponding first, second, and third error values.

[0079] The first error function is implemented based on the MSE error function, the second error function is implemented based on the RMSE error function, and the third error function is implemented based on the MAPE error function;

[0080] Step 1310: Determine whether the first, second, and third error values ​​do not exceed the preset first, second, and third error thresholds. If none of the first, second, and third error values ​​exceed the corresponding first, second, and third error thresholds, increment the pass count counter by 1, and determine whether the incremented pass count counter is greater than the continuous pass count threshold. If so, proceed to step 1311; if not, return to step 1304 to continue training. If all of the first, second, or third error values ​​exceed the corresponding first, second, or third error thresholds, clear the pass count counter and return to step 1304 to continue training.

[0081] Step 1311: solidify the model parameters of the FFR prediction model and confirm that the model training is completed.

[0082] A second aspect of an embodiment of the present invention provides an apparatus for implementing the method for predicting blood flow reserve fraction based on coronary CT images described in the first aspect, the apparatus comprising: a data receiving module, a semantic recognition module, a blood flow velocity feature extraction module, an image and geometric feature extraction module, a feature fusion module, a path search module, a vascular tree processing module, an FFR prediction module, and a prediction result processing module;

[0083] The data receiving module is configured to receive a coronary CT image sequence recorded as a corresponding first CT image sequence; the first image sequence includes a plurality of first CT images; each first CT image corresponds to a first image time; all first CT images have the same two-dimensional size; in the first CT image sequence, the first first CT image is a CT image when a contrast agent initially enters a coronary artery entrance, and the last first CT image is a CT image when the contrast agent fills the entire coronary vascular network;

[0084] The semantic recognition module is configured to perform vascular network semantic segmentation on each of the first CT images to obtain a corresponding first semantic map; and to perform vascular inner wall edge and vascular centerline recognition on the vascular network on each of the first semantic maps to obtain a corresponding second semantic map; and to chronologically sort all of the obtained second semantic maps to form a corresponding second semantic map sequence; the first and second semantic maps correspond one-to-one to the first CT image and are consistent with the two-dimensional size of the corresponding first CT image;

[0085] The blood flow velocity feature extraction module is configured to perform blood flow velocity feature extraction processing based on the second semantic graph sequence to obtain a corresponding first feature graph; the first feature graph corresponds one-to-one to the first CT image and is consistent with a two-dimensional size of the corresponding first CT image;

[0086] The image and geometric feature extraction module is configured to record the last first CT image in the first CT image sequence as the corresponding full coronary image; perform image feature extraction on the full coronary image to obtain a corresponding second feature map; and perform vascular cross-sectional feature, vascular stenosis feature, and vascular topology feature extraction based on the full coronary image and the second semantic map to obtain a corresponding third feature map;

[0087] The feature fusion module is used to perform feature fusion processing on the first, second and third feature maps corresponding to the full coronary image to obtain a corresponding first fused feature map;

[0088] The path search module is configured to perform a vascular path search process on the first fusion feature map to obtain a corresponding first path sequence; and to sample path points on each first path and extract comprehensive features of each sampling point to obtain a corresponding first path feature sequence; the first path sequence includes a plurality of first paths; and the first path feature sequence is composed of a plurality of first point two-dimensional features sequentially arranged;

[0089] The vascular tree processing module is configured to construct a three-dimensional vascular tree based on the first fusion feature map to obtain a corresponding first vascular tree; and to perform three-dimensional feature conversion on each of the first path feature sequences based on the first vascular tree to obtain a corresponding second path feature sequence; the second path feature sequence is composed of a plurality of first point three-dimensional features arranged in sequence;

[0090] The FFR prediction module is used to perform FFR prediction processing on each second path feature sequence based on the FFR prediction model to obtain a corresponding first path prediction sequence; the FFR prediction model is implemented based on the Bi-LSTM model structure; the first path prediction sequence includes multiple first-point FFR prediction values;

[0091] The prediction result processing module is used to perform FFR alarm marking processing on the first vascular tree and the full coronary image according to all the first path prediction sequences obtained to obtain the corresponding first marked vascular tree and first marked image map; and display the first marked vascular tree and the first marked image map.

[0092] A third aspect of an embodiment of the present invention provides an electronic device, including: a memory, a processor, and a transceiver;

[0093] The processor is configured to be coupled to the memory, read and execute instructions in the memory, so as to implement the method steps described in the first aspect above;

[0094] The transceiver is coupled to the processor, and the processor controls the transceiver to send and receive messages.

[0095] A fourth aspect of an embodiment of the present invention provides a computer-readable storage medium, which stores computer instructions. When the computer instructions are executed by a computer, the computer executes the instructions of the method described in the first aspect above.

[0096] The embodiment of the present invention provides a method, device, electronic device and computer-readable storage medium for predicting blood flow reserve fraction based on coronary CT images. As can be seen from the above content, the embodiment of the present invention performs vascular network semantic segmentation on each CT image of the coronary CT image sequence to obtain a first semantic map, and identifies the inner wall edge and the center line of the vascular network on each first semantic map to obtain a second semantic map, and extracts blood flow velocity features based on the second semantic map sequence to obtain a first feature map; and records the last CT image in the CT image sequence as a full coronary image and extracts image features from it to obtain a second feature map; and extracts vascular cross-sectional features, vascular stenosis features and vascular topology features based on the full coronary image and the second semantic map to obtain a third feature map; and performs feature fusion on the first, second and third feature maps to obtain a first fused feature map; and performs feature fusion on the first fusion feature map. The first path sequence is obtained by searching the vascular path based on the feature map; the path points of each first path are sampled and the comprehensive features of each sampling point are extracted to obtain the first path feature sequence; a three-dimensional vascular tree is constructed based on the first fusion feature map to obtain the first vascular tree; the three-dimensional features of each first path feature sequence are converted based on the first vascular tree to obtain the second path feature sequence; the FFR prediction model implemented based on the reference Bi-LSTM model structure is used to perform FFR prediction based on each second path feature sequence to obtain the first path prediction sequence; the first vascular tree and the whole coronary image are marked with FFR alarms based on all the obtained first path prediction sequences; and the marked vascular tree and the marked image are displayed. The non-invasive blood flow reserve fraction prediction mechanism provided by the embodiment of the present invention avoids damage to the blood vessels of the subject, improves the comfort of the subject, and establishes a sustainable monitoring method with a short time interval. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 A schematic diagram of a method for predicting fractional flow reserve based on coronary CT images provided in Example 1 of the present invention;

[0098] Figure 2 This is a module structure diagram of a device for predicting fractional flow reserve based on coronary CT images provided in Example 2 of the present invention;

[0099] Figure 3 This is a structural diagram of an electronic device provided in Example 3 of the present invention. DETAILED DESCRIPTION

[0100] To make the objectives, technical solutions, and advantages of the present invention more apparent, the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the embodiments described herein are merely some, rather than all, of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0101] The first embodiment of the present invention provides a method for predicting blood flow reserve fraction based on coronary CT images, such as Figure 1 A schematic diagram of a method for predicting fractional flow reserve based on coronary CT images provided in Example 1 of the present invention is shown. The method mainly includes the following steps:

[0102] Step 1: receiving a coronary CT image sequence and recording it as a corresponding first CT image sequence;

[0103] Wherein, the first image sequence includes a plurality of first CT images;

[0104] The shape of the first CT image is H CT ×W CT ×C CT ;H CT 、W CT 、C CT are all integers greater than 0, H CT 、W CT are the height and width of the first CT image, respectively, CT is the characteristic dimension of the first CT image; the first CT image includes H CT ×W CT First pixel points; each first pixel point corresponds to a vector with a length of C CT The first pixel feature vector includes at least a first pixel coordinate and a first pixel value;

[0105] Each first CT image corresponds to a first image time; all first CT images have the same two-dimensional size; in the first CT image sequence, the first first CT image is the CT image when the contrast agent initially enters the coronary artery entrance, and the last first CT image is the CT image when the contrast agent fills the entire coronary vascular network.

[0106] Step 2: Performing vascular network semantic segmentation on each first CT image to obtain a corresponding first semantic map; then performing vascular wall edge and vascular centerline recognition on the vascular network in each first semantic map to obtain a corresponding second semantic map; and then chronologically sorting all the obtained second semantic maps to form a corresponding second semantic map sequence;

[0107] The first and second semantic maps correspond to the first CT image one-to-one and have the same two-dimensional size as the corresponding first CT image;

[0108] Specifically comprising: step 21, performing vascular network semantic segmentation processing on each first CT image to obtain a corresponding first semantic map;

[0109] Among them, the shape of the first semantic graph is H SEG1 ×W SEG1 ×C SEG1 ;H SEG1 、W SEG1 、C SEG1 are all integers greater than 0, H SEG1 、W SEG1 are the height and width of the two-dimensional size of the first semantic graph, C SEG1 is the feature dimension of the first semantic graph, H SEG1 =H CT 、W SEG1 =W CT ; The first semantic graph includes H SEG1 ×W SEG1 Second pixel points; each second pixel point corresponds to a vector with a length of C SEG1 The second pixel feature vector includes at least the second pixel coordinates and the first semantic type; the first semantic type includes a blood vessel point type and a background point type;

[0110] Specifically, step 211 sets a shape of H for the current first CT image. SEG1 ×W SEG1 ×C SEG1 and initializing the first semantic type corresponding to each second pixel point in the current first semantic map to the background point type;

[0111] Step 212: Each first pixel point of the current first CT image is used as a corresponding current image pixel point; and when the first pixel value of the current image pixel point meets a preset contrast agent pixel value range, the first semantic type of the second pixel point corresponding to the second pixel coordinate on the current first semantic map and the first pixel coordinate of the current image pixel point is reset to a blood vessel point type;

[0112] Here, the contrast agent pixel value range is a preset pixel value range;

[0113] Step 22: performing vascular wall edge and vascular centerline recognition processing on the vascular network on each first semantic graph to obtain a corresponding second semantic graph;

[0114] Among them, the shape of the second semantic graph is H SEG2 ×W SEG2 ×CSEG2 ;H SEG2 、W SEG2 、C SEG2 are all integers greater than 0, H SEG2 、W SEG2 are the height and width of the two-dimensional size of the second semantic graph, C SEG2 is the feature dimension of the second semantic graph, H SEG2 =H CT 、W SEG2 =W CT ; The second semantic graph includes H SEG2 ×W SEG2 third pixel points; each third pixel point corresponds to a vector with a length of C SEG2 The third pixel feature vector includes at least the third pixel coordinates and the second semantic type; the second semantic type includes a blood vessel edge point type, a blood vessel center point type and a background point type;

[0115] Specifically, step 221 is to set a shape H for the current first semantic graph. SEG2 ×W SEG2 ×C SEG2 and initializing the second semantic type corresponding to each third pixel point in the current second semantic map to the background point type;

[0116] Step 222: each second pixel point of the first semantic type being a blood vessel point type in the current first semantic map is used as a corresponding lumen pixel point; and a pixel point connected image region consisting of all the lumen pixels is recorded as a corresponding blood vessel network region.

[0117] Step 223: Perform regional edge recognition on the vascular network region, and use lumen pixels on any edge line in the vascular network region as corresponding inner wall edge pixels; and reset the second semantic type of the third pixel corresponding to the second pixel coordinate of each inner wall edge pixel in the current second semantic map to the vascular edge point type.

[0118] Step 224: Identify the vascular centerline of the vascular network region, and use the lumen pixel points on any centerline of the vascular network region as corresponding centerline pixel points; and reset the second semantic type of the third pixel point corresponding to the third pixel coordinate of the current second semantic map and the second pixel coordinate of each centerline pixel point to the vascular center point type.

[0119] In step 23, all the obtained second semantic graphs are sorted in chronological order to form a corresponding second semantic graph sequence.

[0120] Step 3: performing blood flow velocity feature extraction processing based on the second semantic graph sequence to obtain a corresponding first feature graph;

[0121] The first feature map corresponds to the first CT image one by one and is consistent with the two-dimensional size of the corresponding first CT image; the shape of the first feature map is H F1 ×W F1 ×C F1 ;H F1 、W F1 、C F1 are all integers greater than 0, H F1 、W F1 are the height and width of the two-dimensional size of the first feature map, C F1 is the feature dimension of the first feature map, H F1 =H CT 、W F1 =W CT ; The first feature map includes H F1 ×W F1 fourth pixel points; each fourth pixel point corresponds to a vector with a length of C F1 The fourth pixel feature vector includes at least the fourth pixel coordinates and the blood flow velocity feature;

[0122] Specifically, the step 31 includes: counting the total number of the second semantic graphs of the second semantic graph sequence to obtain a corresponding first total number N; and recording each second semantic graph as a corresponding semantic graph P i ; and the first image time corresponding to each second semantic graph is recorded as the corresponding t i ;

[0123] Where, 1≤index i≤N;

[0124] Step 32, and in each semantic graph P i Based on all the third pixel points whose second semantic type is the blood vessel center point type, the centerline skeleton is extracted to obtain the corresponding centerline skeleton B i ; and press the center line skeleton B i＝N The direction to achieve maximum overlap is for each centerline skeleton B 1≤i<N With centerline skeleton B i＝N Perform registration;

[0125] Step 33, and transform the semantic graph P i＝1 Each end point on the center line skeleton B i＝N The corresponding point on is recorded as the corresponding time point d i＝1 ; and the semantic graph P i＝2 Each end point on the center line skeleton B i＝N The corresponding point on is recorded as the corresponding time point d i＝2 ; And so on, finally the semantic graph P i＝N-1 Each end point on the center line skeleton Bi＝N The corresponding point on is recorded as the corresponding time point d i＝N-1 ; Then the center line skeleton B i＝N Each end point on the i＝N ;

[0126] Step 34, and the center line skeleton B i＝N From the starting point of the skeleton to each moment d i＝N The unique path of is taken as a corresponding first skeleton path; and the N time points d passed by each first skeleton path i Form a corresponding first moment point sequence;

[0127] Step 35, each first time point sequence is used as the corresponding current time point sequence; and the time point d in the current time point sequence is used as the corresponding current time point sequence. i＝N The corresponding speed v i＝N Initialized to 0, and at time point d i＝N At time d i＝N-1 The center line length is calculated to obtain the corresponding center line length s N-1 , and set time point d i＝N and time point d i＝N-1 The time difference between them is recorded as the corresponding time difference △t N-1 =t i＝N -t i＝N-1 , and according to the time difference △t N-1 , centerline length s N-1 and speed v i＝N Calculate the time point d i＝N-1 The corresponding speed v i＝N-1 , And for time point d i＝N-1 At time d i＝N-2 The center line length is calculated to obtain the corresponding center line length s N-2 , and set time point d i＝N-1 and time point d i＝N-2 The time difference between them is recorded as the corresponding time difference △t N-2 =t i＝N-1 -t i＝N-2 , and according to the time difference △t N-2 , centerline length s N-2 and speed v i＝N-1 Calculate the time point d i＝N-2 The corresponding speed v i＝N-2 , And so on, until the moment d in the current moment sequence is calculated i＝1 The corresponding speed v i＝1 until;

[0128] Step 36, and center line skeleton B i＝N All time points d i Traverse; and when traversing, point the current traversal time d i Record as the corresponding current moment, and take the third pixel point corresponding to the current moment as the corresponding current pixel point; and check whether the current moment corresponds to two or more speeds v i Identify; if so, all the speeds v corresponding to the current moment i The average value of is taken as the first velocity feature corresponding to the current pixel; if not, the velocity v corresponding to the current moment is taken as i As the first velocity feature corresponding to the current pixel;

[0129] Step 37, and at the center line skeleton B i＝N Based on the linear interpolation method, the time points d of two adjacent moments connected to each other are calculated. i-1 d i The corresponding two first velocity features are used to estimate the first velocity features of other third pixel points between the two points on the skeleton;

[0130] Step 38, and the semantic graph P i＝N Set a shape to H F1 ×W F1 ×C F1 and initialize the blood flow velocity feature corresponding to each fourth pixel point in the current first feature map to 0; and compare the fourth pixel coordinates on the current first feature map with the center line skeleton B i＝N The blood flow velocity feature of the fourth pixel point corresponding to the third pixel coordinate of each third pixel point is reset to the corresponding first velocity feature.

[0131] Step 4: Record the last first CT image in the first CT image sequence as the corresponding full coronary image; perform image feature extraction on the full coronary image to obtain a corresponding second feature map; and perform vascular cross-sectional feature extraction, vascular stenosis feature, and vascular topology feature extraction based on the full coronary image and the second semantic map to obtain a corresponding third feature map;

[0132] Specifically comprising: step 41, recording the last first CT image in the first CT image sequence as the corresponding whole coronary artery image;

[0133] Step 42: performing image feature extraction processing on the full coronary artery image to obtain a corresponding second feature map;

[0134] Among them, the shape of the second feature map is H F2 ×W F2 ×C F2 ;H F2 、W F2、C F2 are all integers greater than 0, H F2 、W F2 are the height and width of the two-dimensional size of the second feature map, C F2 is the feature dimension of the second feature map, H F2 =H CT 、W F2 =W CT ; The second feature map includes H F2 ×W F2 fifth pixel points; each fifth pixel point corresponds to a vector with a length of C F2 The fifth pixel feature vector includes at least the fifth pixel coordinate, the image gradient feature and the image texture feature vector; the image texture feature vector includes at least the image GLCM feature, the image GLSZM feature, the image GLRLM feature, the image GLDM feature and the image NGTDM feature;

[0135] Specifically, step 421 sets a shape H for the whole coronary image. F2 ×W F2 ×C F2 and initializing the image gradient feature and image texture feature vector corresponding to each fourth pixel point in the current second feature map to empty;

[0136] Step 422 , performing grayscale conversion on the entire coronary artery image to obtain a corresponding first grayscale image;

[0137] The first grayscale image includes a plurality of first grayscale pixels; the first grayscale pixels correspond one-to-one to the first pixels of the first CT image; each first grayscale pixel corresponds to a first grayscale pixel feature vector; the first grayscale pixel feature vector includes at least grayscale pixel coordinates and grayscale values;

[0138] Step 423, the two-dimensional grayscale gradient of each first grayscale pixel on the first grayscale image is calculated to obtain a corresponding first gradient; and the image gradient feature of the fifth pixel corresponding to the fifth pixel coordinate on the current second feature map and the grayscale pixel coordinate of each first grayscale pixel is reset to the corresponding first gradient;

[0139] Step 424: Gray-level co-occurrence matrix (GLC), gray-level size zone matrix (GLSZM), gray-level run length matrix (GLRLM), gray-level dependence matrix (GLC), and neighboring gray tone difference matrix (GLTD) of the first grayscale image are calculated. matrix, NGTDM) to obtain the corresponding first, second, third, fourth and fifth matrices; and extract the matrix unit data corresponding to each first grayscale pixel in the first, second, third, fourth and fifth matrices as the corresponding first GLCM data, first GLSZM data, first GLRLM data, first GLDM data and first NGTDM data; and reset the image GLCM features, image GLSZM features, image GLRLM features, image GLDM features and image NGTDM features of the image texture feature vector of the fifth pixel corresponding to the fifth pixel coordinates on the current second feature map and the grayscale pixel coordinates of each first grayscale pixel to the corresponding first GLCM data, first GLSZM data, first GLRLM data, first GLDM data and first NGTDM data;

[0140] Step 43: extracting blood vessel cross-sectional features, blood vessel stenosis features, and blood vessel topological features based on the full coronary artery image and the second semantic map to obtain a corresponding third feature map;

[0141] Among them, the shape of the third feature map is H F3 ×W F3 ×C F3 ;H F3 、W F3 、C F3 are all integers greater than 0, H F3 、W F3 are the height and width of the two-dimensional size of the third feature map, C F3 is the feature dimension of the third feature map, H F3 =H CT 、W F3 =W CT ; The third feature map includes H F3 ×W F3 Sixth pixel points; each sixth pixel point corresponds to a vector with a length of C F3 The sixth pixel feature vector includes at least the sixth pixel coordinate, the blood vessel cross-section feature vector, the blood vessel stenosis feature vector and the blood vessel topology feature vector;

[0142] When the sixth pixel point does not correspond to a blood vessel center point, the blood vessel cross-section feature vector, the blood vessel stenosis feature vector, and the non-empty blood vessel topology feature vector are empty;

[0143] When the sixth pixel point corresponds to a blood vessel center point, the blood vessel cross-section feature vector includes at least the cross-section area and the cross-section brightness, the blood vessel stenosis feature vector includes at least the proximal radius, the minimum radius and the distal radius, and the blood vessel topology feature vector includes at least the topology point type, the front point pixel coordinates, the back point pixel coordinates and the blood vessel curvature; the topology point types include the coronary artery entrance center point, the middle straight center point, the middle bifurcation center point and the terminal center point; when the topology point type is the coronary artery entrance center point, the corresponding front point pixel coordinates are empty, and the corresponding back point pixel coordinates are the pixel coordinates of the next blood vessel center point connected to the blood vessel center point corresponding to the current pixel point; when the topology point type is the middle straight center point, the corresponding front and back point pixel coordinates are The pixel coordinates of the previous and next blood vessel center points connected to the blood vessel center point corresponding to the current pixel point; when the topological point type is the middle bifurcation center point, the corresponding front point pixel coordinates are the pixel coordinates of the previous blood vessel center point connected to the blood vessel center point corresponding to the current pixel point, and the corresponding back point pixel coordinates are composed of the pixel coordinates of the next blood vessel center point connected to the blood vessel center point corresponding to the current pixel point in the multiple back bifurcated branches; when the topological point type is the terminal center point, the corresponding front point pixel coordinates are the pixel coordinates of the previous blood vessel center point connected to the blood vessel center point corresponding to the current pixel point, and the corresponding back point pixel coordinates are empty; the blood vessel curvature is the blood vessel centerline curvature of the blood vessel center point corresponding to the current pixel point;

[0144] The current step 43 specifically includes:

[0145] Step 431: Draw a blood vessel edge line on the second semantic map based on all third pixel points whose second semantic type is a blood vessel edge point type; and draw a blood vessel center line based on all third pixel points whose second semantic type is a blood vessel center point type.

[0146] Step 432: All third pixels of the second semantic type of the blood vessel center point on the second semantic map are traversed. During the traversal, the currently traversed third pixel is used as the corresponding current center pixel. A perpendicular line is drawn to the blood vessel centerline at the location of the current center pixel, and a straight line segment between the two intersection points of the perpendicular line and the blood vessel edge is used as the corresponding first line segment. Half the length of the first line segment is used as the corresponding minimum radius. The corresponding cross-sectional area is calculated based on the minimum radius. The average of the first pixel values ​​of all first pixels passed by the first line segment on the full coronary image is used as the corresponding second pixel value. The second pixel value is converted to a brightness value based on a conversion relationship between pixel value and brightness value, and the conversion result is used as the corresponding cross-sectional brightness. The Mth third pixel point on the blood vessel centerline where the current center pixel is located, which is closer to the coronary artery entrance center point and closer to the coronary artery entrance center point, is used as the corresponding proximal center point, and the Mth third pixel point after the current center pixel, which is farther away from the coronary artery entrance center point, is used as the corresponding distal center point, where M is an integer greater than 0. Perpendicular lines are drawn to the blood vessel centerline at the locations of the proximal and distal center points, and the perpendicular lines are used as the corresponding cross-sectional brightness. The half length of the straight line segment between the two intersection points with the blood vessel edge is used as the corresponding proximal radius and distal radius; and when the current center pixel point is the coronary artery entrance center point, the corresponding topological point type is set to the coronary artery entrance center point; and when the current center pixel point is an end point of a blood vessel center line, the corresponding topological point type is set to the end center point; and when the current center pixel point intersects only one blood vessel center line and is neither the coronary artery entrance center point nor an end point, the corresponding topological point type is set to the middle straight center point; and when the current center pixel point intersects with two or more blood vessel center lines, the corresponding topological point type is set to the middle bifurcation center point; and the corresponding front point pixel coordinates and back point pixel coordinates are set according to the topological point type; and the curvature of the blood vessel center line at the location of the current center pixel point is estimated and the estimated result is used as the corresponding blood vessel curvature; and a corresponding first eigenvector is formed by the obtained cross-sectional area and cross-sectional brightness, and a corresponding second eigenvector is formed by the obtained proximal radius, minimum radius and distal radius, and a corresponding third eigenvector is formed by the obtained topological point type, front point pixel coordinates, back point pixel coordinates and blood vessel curvature;

[0147] Step 433, set a shape H for the second semantic graph F3 ×W F3 ×C F3and initialize the vascular cross-section feature vector, vascular stenosis feature vector and vascular topology feature vector corresponding to each sixth pixel point in the current third feature map to empty; and reset the vascular cross-section feature vector, vascular stenosis feature vector and vascular topology feature vector of the sixth pixel point corresponding to the sixth pixel coordinate on the current third feature map and the third pixel point of each second semantic type being the vascular center point type to the corresponding first, second and third feature vectors.

[0148] Step 5: performing feature fusion processing on the first, second, and third feature maps corresponding to the full coronary image to obtain a corresponding first fused feature map;

[0149] Among them, the shape of the first fusion feature map is H FM ×W FM ×C FM ;H FM 、W FM 、C FM are all integers greater than 0, H FM 、W FM are the height and width of the two-dimensional size of the first fusion feature map, C FM is the feature dimension of the first fusion feature map, H FM =H CT 、W FM =W CT ; The first fusion feature map includes H FM ×W FM Seventh pixel points; each seventh pixel point corresponds to a vector with a length of C FM The seventh pixel feature vector includes at least the seventh pixel coordinate, blood flow velocity feature, image gradient feature, image texture feature vector, blood vessel cross-section feature vector, blood vessel stenosis feature vector and blood vessel topology feature vector;

[0150] Specifically, step 51 is to set a shape H for the whole coronary image. FM ×W FM ×C FM Initialize the blood flow velocity feature, image gradient feature, image texture feature vector, vascular cross-section feature vector, vascular stenosis feature vector, and vascular topology feature vector corresponding to each seventh pixel in the current first fusion feature map to the corresponding zero-value feature or zero-value feature vector;

[0151] Step 52, traverse all seventh pixel points of the current first fusion feature map; and during the traversal, the currently traversed seventh pixel point is used as the corresponding current pixel point; and the fourth, fifth and sixth pixel points corresponding to the fourth, fifth and sixth pixel coordinates in the first, second and third feature maps and the seventh pixel coordinate of the current pixel point are recorded as the corresponding first, second and third matching pixel points; and the blood flow velocity feature of the current pixel point is set to the blood flow velocity feature of the corresponding first matching pixel point; and the image gradient feature and image texture feature vector of the current pixel point are set to the image gradient feature and image texture feature vector of the corresponding second matching pixel point; and the vascular cross-section feature vector, vascular stenosis feature vector and vascular topology feature vector of the current pixel point are set to the vascular cross-section feature vector, vascular stenosis feature vector and vascular topology feature vector of the corresponding third matching pixel point.

[0152] Step 6: Performing a vascular path search process on the first fusion feature map to obtain a corresponding first path sequence; sampling path points on each first path and extracting comprehensive features of each sampling point to obtain a corresponding first path feature sequence;

[0153] The first path sequence includes a plurality of first paths; the first path feature sequence is composed of a plurality of first point two-dimensional features arranged in sequence;

[0154] Specifically, it includes: step 61, performing a blood vessel path search process on the first fusion feature map to obtain a corresponding first path sequence;

[0155] Specifically, the method includes: extracting a vascular skeleton based on all non-zero vascular topology feature vectors on the first fusion feature map to obtain a corresponding first vascular skeleton; recording the skeleton point corresponding to the seventh pixel point of the topology point type of the coronary artery entrance center point on the first vascular skeleton as a skeleton starting point, and recording the skeleton point corresponding to the seventh pixel point of each topology point type of the end center point as a skeleton end point; and taking a unique path from the skeleton starting point to each skeleton end point as a corresponding first path; and forming a corresponding first path sequence from all the obtained first paths;

[0156] Step 62: sampling path points on each first path and extracting comprehensive features of each sampling point to obtain a corresponding first path feature sequence;

[0157] Specifically, the method includes: starting from the starting point of each first path on the first fusion feature map and sampling pixels at every preset first interval distance along the current path downward, and extracting the seventh pixel coordinates, blood flow velocity features, image gradient features, image texture feature vectors, blood vessel cross-section feature vectors, blood vessel stenosis feature vectors and blood vessel topology feature vectors of the seventh pixel point currently sampled to form a corresponding first-point two-dimensional feature; and sorting all the first-point two-dimensional features corresponding to the current path in the order of sampling to form a corresponding first path feature sequence.

[0158] Here, the first interval distance is a preset distance parameter, which is not a straight-line distance parameter between two points, but the length of the blood vessel centerline between the two points.

[0159] Step 7: construct a three-dimensional vascular tree based on the first fusion feature map to obtain a corresponding first vascular tree; and perform three-dimensional feature conversion on each first path feature sequence based on the first vascular tree to obtain a corresponding second path feature sequence;

[0160] Specifically comprising: step 71, constructing a three-dimensional vascular tree according to the first fusion feature map to obtain a corresponding first vascular tree;

[0161] Here, the embodiment of the present invention uses a model or software for three-dimensional modeling to perform three-dimensional vascular tree modeling based on the positioning features of the first fusion feature map to obtain a corresponding first vascular tree, where the first vascular tree is composed of multiple three-dimensional voxel points;

[0162] Step 72 , performing three-dimensional feature conversion on each first path feature sequence according to the first blood vessel tree to obtain a corresponding second path feature sequence;

[0163] The second path feature sequence is composed of a plurality of first point three-dimensional features arranged sequentially;

[0164] Specifically, the process includes: step 721, performing a three-dimensional vascular centerline skeleton extraction process on the first vascular tree to obtain a corresponding second vascular skeleton; recording the voxel point corresponding to the center point of the coronary artery entrance on the second vascular skeleton as a first starting point, and recording the voxel points corresponding to the end points of each vascular tree as second end points; and recording the unique path from the first starting point to each second end point as a corresponding first centerline path;

[0165] wherein the first centerline path corresponds one-to-one to the first path;

[0166] Step 722: Starting from the first starting point of each first centerline path and downward along the current centerline path, voxel sampling is performed once every preset first interval distance, and the first point two-dimensional feature corresponding to the currently sampled voxel point is used as the corresponding current point feature; the seventh pixel coordinate of the current point feature is updated to the corresponding three-dimensional voxel coordinate; the image gradient feature of the current point feature is updated to the corresponding three-dimensional gradient feature; the front point pixel coordinate, the back point pixel coordinate, and the blood vessel curvature of the vascular topology feature vector of the current point feature are updated to the corresponding three-dimensional front point voxel coordinate, three-dimensional back point voxel coordinate, and three-dimensional blood vessel curvature; the updated current point feature is used as a corresponding first point three-dimensional feature; and all first point three-dimensional features corresponding to the current path are sorted in the sampling order to form a corresponding second path feature sequence.

[0167] Step 8: Perform FFR prediction processing on each second path feature sequence based on the FFR prediction model to obtain a corresponding first path prediction sequence;

[0168] Among them, the FFR prediction model is implemented based on the Bi-LSTM model structure; the first path prediction sequence includes multiple first-point FFR prediction values.

[0169] It should be noted that the FFR prediction model must be trained before it is used. The training of the FFR prediction model here specifically includes:

[0170] Step A01: recruit multiple volunteers without coronary artery stenosis symptoms and multiple volunteers with coronary artery stenosis symptoms of varying degrees to form a first volunteer group;

[0171] The first volunteer set includes multiple first volunteers;

[0172] Step A02: Multiple coronary artery path tests are performed on each first volunteer using an invasive testing method. Feature data is collected from detection points along each detection path according to the data format of the first-point two-dimensional feature to obtain a corresponding detection path two-dimensional feature sequence. The blood flow reserve fraction corresponding to each detection point along the path is measured to obtain a corresponding detection path measurement sequence. Three-dimensional vascular tree modeling is performed based on the detection path two-dimensional feature sequence to obtain a corresponding detection path vascular tree. Based on the detection path vascular tree, three-dimensional data conversion is performed on the detection path two-dimensional feature sequence according to the data format of the first-point three-dimensional feature to obtain a corresponding detection path three-dimensional feature sequence. The detection path three-dimensional feature sequence corresponding to each detection path and the detection path measurement sequence are combined into a corresponding first data record. All the obtained first data records are combined into a corresponding first data set.

[0173] The detection path two-dimensional feature sequence includes a plurality of detection point two-dimensional features, and the data format of the detection point two-dimensional features matches the data format of the first point two-dimensional features; the detection path three-dimensional feature sequence includes a plurality of detection point three-dimensional features, and the data format of the detection point three-dimensional features matches the data format of the first point three-dimensional features; the detection path measurement sequence includes a plurality of detection point FFR measurement values;

[0174] Step A03: setting an integer greater than zero as the corresponding continuous pass count threshold; and setting a pass count counter initialized to 0;

[0175] Step A04: randomly selecting a plurality of first data records from the first data set according to a preset test set ratio to form a corresponding current test set, and forming a corresponding current validation set from all remaining first data records;

[0176] The test set ratio is a preset ratio value, such as 80%; the current test set and the current validation set both consist of multiple first data records, and the ratio of the total number of first data records in the current test set to the total number of first data records in the first data set satisfies the test set ratio;

[0177] Step A05: extract the first data record of the current test set as the corresponding current data record;

[0178] Step A06: Input the three-dimensional characteristic sequence of the detection path of the current data record into the FFR prediction model, and the FFR prediction model performs FFR prediction processing based on the three-dimensional characteristic sequence of the detection path input at that time to obtain a corresponding detection path prediction sequence;

[0179] Among them, the detection path prediction sequence includes multiple detection point FFR prediction values;

[0180] Step A07: Substitute the detection path prediction sequence and the detection path measurement sequence of the current data record into a preset first model loss function to calculate and obtain a corresponding first loss value;

[0181] Wherein, the first model loss function is implemented based on the L1 loss function or the L2 loss function;

[0182] Step A08: Identify whether the first loss value satisfies a preset first loss value range; if the first loss value satisfies the first loss value range, identify whether the current data record is the last first data record of the current test set; if so, proceed to step A09; if not, extract the next first data record of the current test set as the new current data record and return to step A06 to continue training; if the first loss value does not satisfy the first loss value range, perform a round of model parameter optimization on the FFR prediction model in the direction of minimizing the first model loss function based on a preset first model parameter optimizer, and return to step A06 to continue training at the end of this round of parameter optimization; wherein the first loss value range is a preset loss value range; the first model parameter optimizer includes at least an SDG optimizer;

[0183] Step A09: Traverse all first data records of the current validation set; during the traversal, use the currently traversed first data record as the corresponding current validation record; perform FFR prediction processing on the detection path feature sequence of the current validation record by the FFR prediction model to obtain a corresponding current detection path prediction sequence; and form a corresponding prediction-measurement sequence pair by combining the current detection path prediction sequence and the detection path measurement sequence of the current validation record; and at the end of the traversal, subject all obtained prediction-measurement sequence pairs to preset first, second, and third error functions to calculate corresponding first, second, and third error values; wherein the first error function is implemented based on the mean square error (MSE) function, the second error function is implemented based on the root mean square error (RMSE) function, and the third error function is implemented based on the mean absolute percentage error (MAPE) function;

[0184] Step A10: Identify whether the first, second, and third error values ​​do not exceed the preset first, second, and third error thresholds; if the first, second, and third error values ​​do not exceed the corresponding first, second, and third error thresholds, then increment the pass count counter by 1, and determine whether the incremented pass count counter is greater than the continuous pass count threshold. If so, proceed to step A11; if not, return to step A04 to continue training; if the first, second, or third error values ​​all exceed the corresponding first, second, or third error thresholds, then clear the pass count counter and return to step A04 to continue training; here, the first, second, and third error thresholds are three preset error thresholds;

[0185] Step A11: solidify the model parameters of the FFR prediction model and confirm that the model training is completed.

[0186] Step 9: Perform FFR warning marking processing on the first vessel tree and the whole coronary image based on all the obtained first path prediction sequences to obtain corresponding first marked vessel tree and first marked image map; and display the first marked vessel tree and first marked image map;

[0187] Specifically, it includes: step 91, performing FFR warning marking processing on the first vessel tree and the whole coronary image according to all the obtained first path prediction sequences to obtain the corresponding first marked vessel tree and first marked image map;

[0188] Specifically, the method includes: step 911, forming a corresponding first prediction value set from all first-point FFR prediction values ​​of all first path prediction sequences; marking the three-dimensional voxel coordinates corresponding to each first-point FFR prediction value in the first prediction value set as a corresponding first voxel point coordinate; clustering one or more first-point FFR prediction values ​​corresponding to the same first voxel point coordinate in the first prediction value set into a class to form a corresponding first voxel point prediction value set; calculating the mean of all first-point FFR prediction values ​​in each first voxel point prediction value set to obtain a corresponding first voxel point FFR mean; recording the first voxel point FFR mean below a preset FFR threshold as the corresponding first alarm point FFR mean; and using the first voxel point coordinates corresponding to each first alarm point FFR mean as the corresponding first alarm point coordinate.

[0189] Here, the FFR threshold is a pre-set blood flow reserve fraction threshold;

[0190] Step 912: Whether the total number of first alarm point coordinates is zero is identified; if the total number of first alarm point coordinates is zero, the first vessel tree and the full coronary image are used as the corresponding first labeled vessel tree and first labeled image; if the total number of first alarm point coordinates is greater than zero, the first alarm point FFR mean corresponding to each first alarm point coordinate is determined based on a preset first alarm color at the voxel point on the first vessel tree, and the first alarm point FFR mean corresponding to each first alarm point coordinate is determined based on the first alarm color at the first pixel point on the full coronary image. The first vessel tree and the full coronary image, after all first alarm point FFR mean values ​​are marked, are used as the corresponding first labeled vessel tree and first labeled image.

[0191] Here, the first alarm color is a preset color;

[0192] Step 92: Display the first labeled blood vessel tree and the first labeled image.

[0193] Figure 2This is a module structure diagram of a device for predicting blood flow reserve fraction based on coronary CT images provided in the second embodiment of the present invention. The device is a terminal device or server that implements the aforementioned method embodiment, and can also be a device that enables the aforementioned terminal device or server to implement the aforementioned method embodiment. For example, the device can be a device or chip system of the aforementioned terminal device or server. Figure 2 As shown, the device includes: a data receiving module 201, a semantic recognition module 202, a blood flow velocity feature extraction module 203, an image and geometric feature extraction module 204, a feature fusion module 205, a path search module 206, a vascular tree processing module 207, an FFR prediction module 208 and a prediction result processing module 209.

[0194] The data receiving module 201 is used to receive a coronary CT image sequence recorded as a corresponding first CT image sequence; the first image sequence includes multiple first CT images; each first CT image corresponds to a first image time; the two-dimensional size of all first CT images is the same; in the first CT image sequence, the first first CT image is a CT image when the contrast agent initially enters the coronary entrance, and the last first CT image is a CT image when the contrast agent fills the entire coronary vascular network.

[0195] The semantic recognition module 202 is used to perform semantic segmentation of the vascular network on each first CT image to obtain a corresponding first semantic map; and to perform recognition processing of the vascular inner wall edge and vascular centerline of the vascular network on each first semantic map to obtain a corresponding second semantic map; and to sort all the obtained second semantic maps in chronological order to form a corresponding second semantic map sequence; the first and second semantic maps correspond one-to-one to the first CT image and are consistent with the two-dimensional size of the corresponding first CT image.

[0196] The blood flow velocity feature extraction module 203 is used to perform blood flow velocity feature extraction processing according to the second semantic graph sequence to obtain a corresponding first feature graph; the first feature graph corresponds one-to-one to the first CT image and is consistent with the two-dimensional size of the corresponding first CT image.

[0197] The image and geometric feature extraction module 204 is used to record the last first CT image in the first CT image sequence as the corresponding full coronary image; perform image feature extraction processing on the full coronary image to obtain a corresponding second feature map; and perform vascular cross-sectional feature, vascular stenosis feature, and vascular topology feature extraction processing based on the full coronary image and the second semantic map to obtain a corresponding third feature map.

[0198] The feature fusion module 205 is used to perform feature fusion processing on the first, second and third feature maps corresponding to the full coronary image to obtain a corresponding first fused feature map.

[0199] The path search module 206 is used to perform vascular path search processing on the first fusion feature map to obtain a corresponding first path sequence; and to sample path points on each first path and extract comprehensive features of each sampling point to obtain a corresponding first path feature sequence; the first path sequence includes multiple first paths; and the first path feature sequence is composed of multiple first point two-dimensional features arranged in sequence.

[0200] The vascular tree processing module 207 is used to construct a three-dimensional vascular tree based on the first fusion feature map to obtain a corresponding first vascular tree; and to perform three-dimensional feature conversion on each first path feature sequence based on the first vascular tree to obtain a corresponding second path feature sequence; the second path feature sequence is composed of multiple first point three-dimensional features arranged in sequence.

[0201] The FFR prediction module 208 is used to perform FFR prediction processing on each second path feature sequence based on the FFR prediction model to obtain a corresponding first path prediction sequence; the FFR prediction model is implemented based on the Bi-LSTM model structure; the first path prediction sequence includes multiple first-point FFR prediction values.

[0202] The prediction result processing module 209 is used to perform FFR warning marking processing on the first vessel tree and the whole coronary image according to all the obtained first path prediction sequences to obtain the corresponding first marked vessel tree and first marked image map; and display the first marked vessel tree and the first marked image map.

[0203] An embodiment of the present invention provides a device for predicting blood flow reserve fraction based on coronary CT images, which can execute the method steps in the above method embodiment. Its implementation principles and technical effects are similar and will not be repeated here.

[0204] It should be noted that the division of the various modules of the above device is merely a division of logical functions. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can all be implemented in the form of software called by a processing element; or all be implemented in the form of hardware; or some modules can be implemented in the form of software called by a processing element, and some modules can be implemented in the form of hardware. For example, the data receiving module can be a separate processing element, or it can be integrated into a chip of the above device. In addition, it can be stored in the form of program code in the memory of the above device, and called by a processing element of the above device to perform the functions of the above-mentioned module. The implementation of other modules is similar. In addition, these modules can all or partly be integrated together, or they can be implemented independently. The processing element described here can be an integrated circuit with signal processing capabilities. During implementation, each step of the above method or each of the above modules can be completed by hardware integrated logic circuits in the processor element or by software instructions.

[0205] For example, the above modules may be one or more integrated circuits configured to implement the above methods, such as one or more application-specific integrated circuits (ASICs), one or more digital signal processors (DSPs), or one or more field programmable gate arrays (FPGAs). For another example, when a module is implemented by scheduling program code through a processing element, the processing element may be a general-purpose processor, such as a central processing unit (CPU) or other processor that can call program code. For another example, these modules may be integrated together and implemented in the form of a system-on-a-chip (SOC).

[0206] In the above embodiments, all or part of the embodiments may be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments may be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the above method embodiments are generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The above-mentioned computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the above-mentioned computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center via wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, Bluetooth, microwave, etc.) means. The above-mentioned computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media. The above-mentioned available medium can be a magnetic medium (such as a floppy disk, hard disk, tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)), etc.

[0207] Figure 3This is a schematic diagram of the structure of an electronic device provided in the third embodiment of the present invention. The electronic device can be a terminal device or server that implements the method of the aforementioned embodiment, or it can be a terminal device or server that implements the method of the aforementioned embodiment connected to the aforementioned terminal device or server. Figure 3 As shown, the electronic device may include: a processor 301 (such as a CPU), a memory 302, and a transceiver 303; the transceiver 303 is coupled to the processor 301, and the processor 301 controls the transceiver 303's transceiver actions. Various instructions may be stored in the memory 302 for completing various processing functions and implementing the processing steps described in the aforementioned embodiment method. Preferably, the electronic device involved in the embodiment of the present invention further includes: a power supply 304, a system bus 305, and a communication port 306. The system bus 305 is used to realize communication connections between components. The above-mentioned communication port 306 is used for connecting and communicating between the electronic device and other peripherals.

[0208] exist Figure 3 The system bus 305 mentioned in the figure can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The system bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus. The communication interface is used to realize communication between the database access device and other devices (such as clients, read-write libraries, and read-only libraries). The memory may include random access memory (RAM) and may also include non-volatile memory (Non-Volatile Memory), such as at least one disk storage.

[0209] The above-mentioned processors can be general-purpose processors, including central processing units (CPUs), network processors (NPs), graphics processing units (GPUs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0210] It should be noted that an embodiment of the present invention further provides a computer-readable storage medium, which stores instructions. When the computer-readable storage medium is run on a computer, it enables the computer to execute the methods and processing procedures provided in the above embodiments.

[0211] An embodiment of the present invention further provides a chip for executing instructions, which is used to execute the processing steps described in the above method embodiment.

[0212] The embodiment of the present invention provides a method, device, electronic device and computer-readable storage medium for predicting blood flow reserve fraction based on coronary CT images. As can be seen from the above content, the embodiment of the present invention performs vascular network semantic segmentation on each CT image of the coronary CT image sequence to obtain a first semantic map, and identifies the inner wall edge and the center line of the vascular network on each first semantic map to obtain a second semantic map, and extracts blood flow velocity features based on the second semantic map sequence to obtain a first feature map; and records the last CT image in the CT image sequence as a full coronary image and extracts image features from it to obtain a second feature map; and extracts vascular cross-sectional features, vascular stenosis features and vascular topology features based on the full coronary image and the second semantic map to obtain a third feature map; and performs feature fusion on the first, second and third feature maps to obtain a first fused feature map; and performs feature fusion on the first fusion feature map. The first path sequence is obtained by searching the vascular path based on the feature map; the path points of each first path are sampled and the comprehensive features of each sampling point are extracted to obtain the first path feature sequence; a three-dimensional vascular tree is constructed based on the first fusion feature map to obtain the first vascular tree; the three-dimensional features of each first path feature sequence are converted based on the first vascular tree to obtain the second path feature sequence; the FFR prediction model implemented based on the reference Bi-LSTM model structure is used to perform FFR prediction based on each second path feature sequence to obtain the first path prediction sequence; the first vascular tree and the whole coronary image are marked with FFR alarms based on all the obtained first path prediction sequences; and the marked vascular tree and the marked image are displayed. The non-invasive blood flow reserve fraction prediction mechanism provided by the embodiment of the present invention avoids damage to the blood vessels of the subject, improves the comfort of the subject, and establishes a sustainable monitoring method with a short time interval.

[0213] Professionals should also be further aware that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0214] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be implemented using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0215] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting blood flow reserve fraction based on coronary CT images, characterized in that: The method comprises: A received coronary CT image sequence is recorded as a corresponding first CT image sequence; the first CT image sequence includes a plurality of first CT images; each first CT image corresponds to a first image time; all first CT images have the same two-dimensional size; in the first CT image sequence, the first first CT image is a CT image when a contrast agent initially enters a coronary artery entrance, and the last first CT image is a CT image when the contrast agent fills the entire coronary vascular network; Performing vascular network semantic segmentation processing on each of the first CT images to obtain a corresponding first semantic map; performing vascular inner wall edge and vascular centerline identification processing on the vascular network on each of the first semantic maps to obtain a corresponding second semantic map; and chronologically sorting all of the obtained second semantic maps to form a corresponding second semantic map sequence; the first and second semantic maps correspond one-to-one to the first CT image and are consistent with the two-dimensional size of the corresponding first CT image; performing blood flow velocity feature extraction processing on the second semantic graph sequence to obtain a corresponding first feature graph; the first feature graph corresponds one-to-one to the first CT image and is consistent with the two-dimensional size of the corresponding first CT image; Recording the last first CT image in the first CT image sequence as the corresponding full coronary image; performing image feature extraction processing on the full coronary image to obtain a corresponding second feature map; and performing vascular cross-sectional feature, vascular stenosis feature, and vascular topology feature extraction processing based on the full coronary image and the second semantic map to obtain a corresponding third feature map; performing feature fusion processing on the first, second, and third feature maps corresponding to the full coronary image to obtain a corresponding first fused feature map; Performing a vascular path search process on the first fusion feature map to obtain a corresponding first path sequence; sampling path points on each first path and extracting comprehensive features of each sampling point to obtain a corresponding first path feature sequence; the first path sequence includes a plurality of first paths; the first path feature sequence is composed of a plurality of first point two-dimensional features sequentially arranged; Constructing a three-dimensional vascular tree based on the first fusion feature map to obtain a corresponding first vascular tree; and performing three-dimensional feature conversion on each of the first path feature sequences based on the first vascular tree to obtain a corresponding second path feature sequence; the second path feature sequence is composed of a plurality of first point three-dimensional features arranged in sequence; Performing FFR prediction processing on each of the second path feature sequences based on an FFR prediction model to obtain a corresponding first path prediction sequence; the FFR prediction model is implemented based on a Bi-LSTM model structure; the first path prediction sequence includes multiple first-point FFR prediction values; Perform FFR warning marking processing on the first vascular tree and the whole coronary image according to all the obtained first path prediction sequences to obtain corresponding first marked vascular tree and first marked image map; and display the first marked vascular tree and the first marked image map.

2. The method for predicting fractional flow reserve based on coronary CT images according to claim 1, characterized in that: The shape of the first CT image is H CT ×W CT ×C CT ;H CT 、W CT 、C CT are all integers greater than 0, H CT 、W CT are the height and width of the two-dimensional size of the first CT image, C CT is the characteristic dimension of the first CT image; the first CT image includes H CT ×W CT First pixel points; each of the first pixel points corresponds to a vector with a length of C CT The first pixel feature vector includes at least a first pixel coordinate and a first pixel value; The shape of the first semantic graph is H SEG1 ×W SEG1 ×C SEG1 ;H SEG1 、W SEG1 、C SEG1 are all integers greater than 0, H SEG1 、W SEG1 are the height and width of the two-dimensional size of the first semantic graph, C SEG1 is the feature dimension of the first semantic graph, H SEG1 =H CT 、W SEG1 =W CT ; The first semantic graph includes H SEG1 ×W SEG1 second pixel points; each of the second pixel points corresponds to a vector with a length of C SEG1 The second pixel feature vector includes at least a second pixel coordinate and a first semantic type; the first semantic type includes a blood vessel point type and a background point type; The shape of the second semantic graph is H SEG2 ×W SEG2 ×C SEG2 ;H SEG2 、W SEG2 、C SEG2 are all integers greater than 0, H SEG2 、W SEG2 are the height and width of the two-dimensional size of the second semantic graph, C SEG2 is the feature dimension of the second semantic graph, H SEG2 =H CT 、W SEG2 =W CT ; The second semantic graph includes H SEG2 ×W SEG2 third pixel points; each of the third pixel points corresponds to a vector with a length of C SEG2 The third pixel feature vector includes at least a third pixel coordinate and a second semantic type; the second semantic type includes a blood vessel edge point type, a blood vessel center point type and a background point type; The shape of the first feature map is H F1 ×W F1 ×C F1 ;H F1 、W F1 、C F1 are all integers greater than 0, H F1 、W F1 are the height and width of the two-dimensional size of the first feature map, C F1 is the feature dimension of the first feature map, H F1 =H CT 、W F1 =W CT ; The first feature map includes H F1 ×W F1 fourth pixel points; each of the fourth pixel points corresponds to a vector with a length of C F1 The fourth pixel feature vector includes at least the fourth pixel coordinates and the blood flow velocity feature; The shape of the second feature map is H F2 ×W F2 ×C F2 ;H F2 、W F2 、C F2 are all integers greater than 0, H F2 、W F2 are the height and width of the two-dimensional size of the second feature map, C F2 is the feature dimension of the second feature map, H F2 =H CT 、W F2 =W CT ; The second feature map includes H F2 ×W F2 fifth pixel points; each of the fifth pixel points corresponds to a vector with a length of C F2 The fifth pixel feature vector includes at least the fifth pixel coordinate, the image gradient feature and the image texture feature vector; the image texture feature vector includes at least the image GLCM feature, the image GLSZM feature, the image GLRLM feature, the image GLDM feature and the image NGTDM feature; The shape of the third characteristic map is H F3 ×W F3 ×C F3 ;H F3 、W F3 、C F3 are all integers greater than 0, H F3 、W F3 are the height and width of the two-dimensional size of the third feature map, C F3 is the feature dimension of the third feature map, H F3 =H CT 、W F3 =W CT ; The third characteristic map includes H F3 ×W F3 Sixth pixel points; each of the sixth pixel points corresponds to a vector with a length of C F3 The sixth pixel feature vector includes at least the sixth pixel coordinate, the blood vessel cross-section feature vector, the blood vessel stenosis feature vector and the blood vessel topology feature vector; When the sixth pixel point does not correspond to a blood vessel center point, the blood vessel cross-section feature vector, the blood vessel stenosis feature vector, and the blood vessel topology feature vector are empty; When the sixth pixel point corresponds to a blood vessel center point, the blood vessel cross-section feature vector includes at least the cross-section area and the cross-section brightness, the blood vessel stenosis feature vector includes at least the proximal radius, the minimum radius and the distal radius, and the blood vessel topology feature vector includes at least the topology point type, the front point pixel coordinates, the back point pixel coordinates and the blood vessel curvature; the topology point types include the coronary artery entrance center point, the middle straight center point, the middle bifurcation center point and the terminal center point; when the topology point type is the coronary artery entrance center point, the corresponding front point pixel coordinates are empty, and the corresponding back point pixel coordinates are the pixel coordinates of the next blood vessel center point connected to the blood vessel center point corresponding to the current pixel point; when the topology point type is the middle straight center point, the corresponding front and back point pixel coordinates are empty, and the corresponding back point pixel coordinates are the pixel coordinates of the next blood vessel center point connected to the blood vessel center point corresponding to the current pixel point; when the topology point type is the middle straight center point, the corresponding front and back point pixel coordinates are empty, and the corresponding back point pixel coordinates are the pixel coordinates of the next blood vessel center point connected to the blood vessel center point corresponding to the current pixel point. The coordinates are the pixel coordinates of the previous and next blood vessel center points connected to the blood vessel center point corresponding to the current pixel point; when the topological point type is an intermediate bifurcation center point, the corresponding front point pixel coordinates are the pixel coordinates of the previous blood vessel center point connected to the blood vessel center point corresponding to the current pixel point, and the corresponding back point pixel coordinates are composed of the pixel coordinates of the next blood vessel center point connected to the blood vessel center point corresponding to the current pixel point in the multiple rear bifurcated branches; when the topological point type is an end center point, the corresponding front point pixel coordinates are the pixel coordinates of the previous blood vessel center point connected to the blood vessel center point corresponding to the current pixel point, and the corresponding back point pixel coordinates are empty; the blood vessel curvature is the blood vessel centerline curvature of the blood vessel center point corresponding to the current pixel point; The shape of the first fusion feature map is H FM ×W FM ×C FM ;H FM 、W FM 、C FM are all integers greater than 0, H FM 、W FM are the height and width of the two-dimensional size of the first fusion feature map, C FM is the feature dimension of the first fusion feature map, H FM =H CT 、W FM =W CT ; The first fusion feature map includes H FM ×W FM Seventh pixel points; each of the seventh pixel points corresponds to a vector with a length of C FM the seventh pixel feature vector; the seventh pixel feature vector includes at least the seventh pixel coordinate, the blood flow velocity feature, the image gradient feature, the image texture feature vector, the blood vessel cross-section feature vector, the blood vessel stenosis feature vector and the blood vessel topology feature vector.

3. The method for predicting fractional flow reserve based on coronary CT images according to claim 2, characterized in that: The performing vascular network semantic segmentation processing on each of the first CT images to obtain a corresponding first semantic map specifically includes: Set a shape of H for the current first CT image SEG1 ×W SEG1 ×C SEG1 The first semantic map; and initializing the first semantic type corresponding to each second pixel point in the current first semantic map to a background point type; Each of the first pixel points of the current first CT image is taken as the corresponding current image pixel point; and when the first pixel value of the current image pixel point meets the preset contrast agent pixel value range, the first semantic type of the second pixel point corresponding to the second pixel coordinate on the current first semantic map and the first pixel coordinate of the current image pixel point is reset to a blood vessel point type.

4. The method for predicting fractional flow reserve based on coronary CT images according to claim 2, characterized in that: The step of performing a vascular inner wall edge and vascular centerline recognition process on each vascular network on the first semantic graph to obtain a corresponding second semantic graph specifically includes: Set a shape H for the current first semantic graph SEG2 ×W SEG2 ×C SEG2 The second semantic map; and initializing the second semantic type corresponding to each of the third pixel points in the current second semantic map to a background point type; taking each second pixel point of the first semantic type in the current first semantic map as a blood vessel point type as a corresponding lumen pixel point; and recording a pixel point connected image region consisting of all the lumen pixels as a corresponding blood vessel network region; Performing regional edge recognition on the vascular network region, and taking the lumen pixel points on any edge line of the vascular network region as corresponding inner wall edge pixel points; and resetting the second semantic type of the third pixel points corresponding to the third pixel coordinates on the current second semantic map and the second pixel coordinates of each inner wall edge pixel point to a vascular edge point type; The vascular centerline of the vascular network area is identified, and the lumen pixel points on any centerline of the vascular network area are used as corresponding centerline pixel points; and the second semantic type of the third pixel point corresponding to the third pixel coordinates on the current second semantic map and the second pixel coordinates of each centerline pixel point is reset to the vascular center point type.

5. The method for predicting fractional flow reserve based on coronary CT images according to claim 2, characterized in that: The step of performing blood flow velocity feature extraction processing according to the second semantic graph sequence to obtain a corresponding first feature graph specifically includes: The total number of the second semantic graphs in the second semantic graph sequence is counted to obtain the corresponding first total number N; and each second semantic graph is recorded as a corresponding semantic graph P i ; and the first image time corresponding to each second semantic graph is recorded as the corresponding t i ; 1≤index i≤N; And in each semantic graph P i Based on all the third pixel points whose second semantic type is the blood vessel center point type, the centerline skeleton is extracted to obtain the corresponding centerline skeleton B i ; and press the center line skeleton B i=N The direction to achieve maximum overlap is for each centerline skeleton B 1≤i<N With the centerline skeleton B i=N Perform registration; And the semantic graph P i=1 Each end point on the centerline skeleton B i=N The corresponding point on is recorded as the corresponding time point d i=1 ; and the semantic graph P i=2 Each end point on the centerline skeleton B i=N The corresponding point on is recorded as the corresponding time point d i=2 ; And so on, finally the semantic graph P i=N-1 Each end point on the centerline skeleton B i=N The corresponding point on is recorded as the corresponding time point d i=N-1 Then the center line skeleton B i=N Each end point on the i=N ; And the centerline skeleton B i=N From the skeleton starting point to each of the time points d i=N The unique path is used as a corresponding first skeleton path; and the N time points d passed by each of the first skeleton paths are i Form a corresponding first moment point sequence; and taking each of the first time point sequences as the corresponding current time point sequence; and taking the time point d in the current time point sequence as the corresponding current time point sequence; i=N The corresponding speed v i=N Initialized to 0, and at the time point d i=N At the time point d i=N-1 The center line length is calculated to obtain the corresponding center line length s N-1 , and the time point d i=N With the time point d i=N-1 The time difference between them is recorded as the corresponding time difference △t N-1 =t i=N -t i=N-1 , and according to the time difference △t N-1 , the centerline length s N-1 and the speed v i=N Calculate the time point d i=N-1 The corresponding speed v i=N-1 , ; and for the time point d i=N-1 At time d i=N-2 The center line length is calculated to obtain the corresponding center line length s N-2 , and the time point d i=N-1 With the time point d i=N-2 The time difference between them is recorded as the corresponding time difference △t N-2 =t i=N-1 -t i=N-2 , and according to the time difference △t N-2 , the centerline length s N-2 and the speed v i=N-1 Calculate the time point d i=N-2 The corresponding speed v i=N-2 , ; and so on, until the time point d in the current time point sequence is calculated i=1 The corresponding speed v i=1 until; And the center line skeleton B i=N All the time points d i Traverse; and when traversing, the current traversal time point d i Recorded as the corresponding current moment, and the third pixel point corresponding to the current moment as the corresponding current pixel point; and whether the current moment corresponds to two or more speeds v i Identify; if so, all the speeds v corresponding to the current moment i The average value of is taken as the first velocity feature corresponding to the current pixel point; if not, the velocity v uniquely corresponding to the current moment point is taken as i as the first velocity feature corresponding to the current pixel; And at the center line skeleton B i=N Based on the linear interpolation method, the time points d of two adjacent moments connected to each other are calculated. i-1 d i estimating the first velocity features of other third pixel points between the two points on the skeleton using the corresponding two first velocity features; And the semantic graph P i=N Set a shape to H F1 ×W F1 ×C F1 Initialize the blood flow velocity feature corresponding to each of the fourth pixel points in the current first feature map to 0; and compare the fourth pixel coordinates on the current first feature map with the centerline skeleton B i=N The blood flow velocity feature of the fourth pixel point corresponding to the third pixel coordinate of each of the third pixel points is reset to the corresponding first velocity feature.

6. The method for predicting fractional flow reserve based on coronary CT images according to claim 2, characterized in that: The performing image feature extraction processing on the full coronary image to obtain a corresponding second feature map specifically includes: Set a shape H for the whole coronary image F2 ×W F2 ×C F2 and initializing the image gradient feature and the image texture feature vector corresponding to each of the fifth pixels in the current second feature map to be empty; and performing grayscale conversion on the full coronary image to obtain a corresponding first grayscale image; the first grayscale image includes a plurality of first grayscale pixels; the first grayscale pixels correspond one-to-one to the first pixels of the first CT image; each of the first grayscale pixels corresponds to a first grayscale pixel feature vector; the first grayscale pixel feature vector includes at least grayscale pixel coordinates and grayscale values; and calculating the two-dimensional grayscale gradient of each of the first grayscale pixels on the first grayscale image to obtain a corresponding first gradient; and resetting the image gradient feature of the fifth pixel corresponding to the fifth pixel coordinates on the current second feature map and the grayscale pixel coordinates of each of the first grayscale pixels to the corresponding first gradient; The grayscale co-occurrence matrix, grayscale area size matrix, grayscale run matrix, grayscale dependency matrix and adjacent grayscale difference matrix of the first grayscale image are calculated to obtain the corresponding first, second, third, fourth and fifth matrices; and the matrix unit data corresponding to each first grayscale pixel in the first, second, third, fourth and fifth matrices are extracted as the corresponding first GLCM data, first GLSZM data, first GLRLM data, first GLDM data and first NGTDM data; and the image GLCM features, image GLSZM features, image GLRLM features, image GLDM features and image NGTDM features of the image texture feature vector of the fifth pixel corresponding to the fifth pixel coordinates on the current second feature map and the grayscale pixel coordinates of each first grayscale pixel are reset to the corresponding first GLCM data, the first GLSZM data, the first GLRLM data, the first GLDM data and the first NGTDM data.

7. The method for predicting fractional flow reserve based on coronary CT images according to claim 2, characterized in that: The extracting of vascular cross-sectional features, vascular stenosis features, and vascular topological features based on the full coronary image and the second semantic map to obtain a corresponding third feature map specifically includes: Drawing a blood vessel edge line based on all the third pixel points whose second semantic type is a blood vessel edge point type on the second semantic map; and drawing a blood vessel centerline based on all the third pixel points whose second semantic type is a blood vessel center point type; And traverse all the third pixel points of the second semantic type that are the blood vessel center point type on the second semantic map; and during the traversal, use the currently traversed third pixel point as the corresponding current center pixel point; and draw a perpendicular line to the blood vessel centerline at the location of the current center pixel point, and use the straight line segment between the two intersection points of the perpendicular line and the blood vessel edge as the corresponding first line segment, and use half the length of the first line segment as the corresponding minimum radius; and calculate the corresponding cross-sectional area based on the minimum radius; and take the average value of the first pixel values ​​of all the first pixel points that the first line segment passes through on the full coronary image. As the corresponding second pixel value, and based on the conversion relationship between pixel value and brightness value, perform brightness value conversion on the second pixel value and use the conversion result as the corresponding cross-sectional brightness; and use the Mth third pixel point on the blood vessel centerline where the current center pixel point is located, which is closer to the center point of the coronary artery entrance before the current center pixel point, as the corresponding proximal center point, and the Mth third pixel point which is farther away from the center point of the coronary artery entrance after the current center pixel point is as the corresponding distal center point, where M is an integer greater than 0; and draw perpendicular lines to the blood vessel centerline at the positions of the proximal and distal center points, and align the perpendicular lines with the blood vessel edge. The half length of the straight line segment between the two intersection points is used as the corresponding proximal radius and the distal radius; and when the current center pixel point is the coronary entrance center point, the corresponding topological point type is set to the coronary entrance center point, and when the current center pixel point is an end point of a blood vessel center line, the corresponding topological point type is set to the end center point, and when the current center pixel point intersects only one blood vessel center line and is neither the coronary entrance center point nor an end point, the corresponding topological point type is set to the middle straight center point, and when the current center pixel point intersects with two or more blood vessel center lines, the corresponding The topological point type is a middle bifurcation center point; and the corresponding front point pixel coordinates and the back point pixel coordinates are set according to the topological point type; and the curvature of the blood vessel centerline at the location of the current center pixel point is estimated and the estimation result is used as the corresponding blood vessel curvature; and the obtained cross-sectional area and the cross-sectional brightness form a corresponding first eigenvector, and the obtained proximal radius, the minimum radius and the distal radius form a corresponding second eigenvector, and the obtained topological point type, the front point pixel coordinates, the back point pixel coordinates and the blood vessel curvature form a corresponding third eigenvector; And set a shape H for the second semantic graph F3 ×W F3 ×C F3 The third feature map of the present invention; and the blood vessel cross-section feature vector, the blood vessel stenosis feature vector and the blood vessel topology feature vector corresponding to each sixth pixel point in the current third feature map are initialized to empty; and the blood vessel cross-section feature vector, the blood vessel stenosis feature vector and the blood vessel topology feature vector of the sixth pixel point corresponding to the sixth pixel coordinate on the current third feature map and the third pixel point of each second semantic type being the blood vessel center point type are reset to the corresponding first, second and third feature vectors.

8. The method for predicting fractional flow reserve based on coronary CT images according to claim 2, characterized in that: The performing feature fusion processing on the first, second, and third feature maps corresponding to the full coronary image to obtain a corresponding first fused feature map specifically includes: Set a shape H for the whole coronary image FM ×W FM ×C FM The first fusion feature map; and initializing the blood flow velocity feature, the image gradient feature, the image texture feature vector, the blood vessel cross-section feature vector, the blood vessel stenosis feature vector, and the blood vessel topology feature vector corresponding to each of the seventh pixel points in the current first fusion feature map to corresponding zero-value features or zero-value feature vectors; All the seventh pixel points of the current first fusion feature map are traversed; and the currently traversed seventh pixel point is used as the corresponding current pixel point during the traversal; and the fourth, fifth and sixth pixel points corresponding to the fourth, fifth and sixth pixel coordinates in the first, second and third feature maps and the seventh pixel coordinate of the current pixel point are recorded as the corresponding first, second and third matching pixel points; and the blood flow velocity feature of the current pixel point is set to the blood flow velocity feature of the corresponding first matching pixel point; and the image gradient feature and the image texture feature vector of the current pixel point are set to the image gradient feature and the image texture feature vector of the corresponding second matching pixel point; and the blood vessel cross-section feature vector, the blood vessel stenosis feature vector and the blood vessel topology feature vector of the current pixel point are set to the blood vessel cross-section feature vector, the blood vessel stenosis feature vector and the blood vessel topology feature vector of the corresponding third matching pixel point.

9. The method for predicting fractional flow reserve based on coronary CT images according to claim 2, characterized in that: The performing of a vascular path search process on the first fusion feature map to obtain a corresponding first path sequence specifically includes: On the first fusion feature map, a vascular skeleton is extracted based on the vascular topology feature vectors of all non-zero feature vectors to obtain a corresponding first vascular skeleton; and the skeleton point on the first vascular skeleton corresponding to the seventh pixel point whose topology point type is the coronary entrance center point is recorded as a skeleton starting point, and the skeleton point corresponding to the seventh pixel point whose topology point type is the end center point is recorded as a skeleton end point; and the unique path from the skeleton starting point to each of the skeleton end points is taken as a corresponding first path; and all the obtained first paths form a corresponding first path sequence.

10. The method for predicting fractional flow reserve based on coronary CT images according to claim 2, characterized in that: Sampling path points on each of the first paths and extracting comprehensive features of each sampling point to obtain a corresponding first path feature sequence specifically includes: On the first fusion feature map, starting from the path starting point of each first path and downward along the current path, pixel sampling is performed every preset first interval distance, and the seventh pixel coordinates, the blood flow velocity feature, the image gradient feature, the image texture feature vector, the blood vessel cross-section feature vector, the blood vessel stenosis feature vector and the blood vessel topology feature vector of the currently sampled seventh pixel point are extracted to form a corresponding first-point two-dimensional feature; and all the first-point two-dimensional features corresponding to the current path are sorted in the sampling order to form a corresponding first path feature sequence.

11. The method for predicting fractional flow reserve based on coronary CT images according to claim 10, characterized in that: The performing three-dimensional feature conversion on each of the first path feature sequences according to the first vascular tree to obtain a corresponding second path feature sequence specifically includes: Performing a three-dimensional vascular centerline skeleton extraction process on the first vascular tree to obtain a corresponding second vascular skeleton; recording a voxel point on the second vascular skeleton corresponding to a center point of a coronary artery entrance as a first starting point, and recording a voxel point corresponding to an end point of each vascular tree as a second end point; and recording a unique path from the first starting point to each second end point as a corresponding first centerline path; the first centerline path and the first path have a one-to-one correspondence; Starting from the first starting point of each first centerline path and downward along the current centerline path, voxel point sampling is performed every preset first interval distance and the first point two-dimensional feature corresponding to the current sampled voxel point is used as the corresponding current point feature; the seventh pixel coordinate of the current point feature is updated to the corresponding three-dimensional voxel coordinate; the image gradient feature of the current point feature is updated to the corresponding three-dimensional gradient feature; the front point pixel coordinate, the back point pixel coordinate and the blood vessel curvature of the blood vessel topology feature vector of the current point feature are updated to the corresponding three-dimensional front point voxel coordinate, three-dimensional back point voxel coordinate and three-dimensional blood vessel curvature; the updated current point feature is used as a corresponding first point three-dimensional feature; and all the first point three-dimensional features corresponding to the current path are sorted in the sampling order to form a corresponding second path feature sequence.

12. The method for predicting fractional flow reserve based on coronary CT images according to claim 11, characterized in that: The performing FFR warning marking processing on the first vessel tree and the whole coronary image according to all the obtained first path prediction sequences to obtain the corresponding first marked vessel tree and first marked image map specifically includes: A corresponding first prediction value set is formed by all first-point FFR prediction values ​​of all first-path prediction sequences; the three-dimensional voxel coordinates corresponding to each first-point FFR prediction value in the first prediction value set are marked as corresponding first voxel coordinates; one or more first-point FFR prediction values ​​corresponding to the same first voxel coordinate in the first prediction value set are grouped into a class to form a corresponding first voxel prediction value set; all first-point FFR prediction values ​​of each first voxel prediction value set are averaged to obtain a corresponding first voxel FFR average; the first voxel FFR average that is lower than a preset FFR threshold is recorded as the corresponding first alarm point FFR average; and the first voxel coordinates corresponding to each first alarm point FFR average are used as the corresponding first alarm point coordinates; And identify whether the total number of the first alarm point coordinates is zero; if the total number of the first alarm point coordinates is zero, the first vascular tree and the full coronary image are used as the corresponding first marked vascular tree and the first marked image map; if the total number of the first alarm point coordinates is greater than zero, the first alarm point FFR mean corresponding to the preset first alarm color mark is calculated at the voxel point corresponding to each first alarm point coordinate on the first vascular tree, and the first alarm point FFR mean corresponding to the first alarm color mark is calculated at the first pixel point corresponding to each first alarm point coordinate on the full coronary image, and the first vascular tree and the full coronary image that have completed all the first alarm point FFR mean mark markings are used as the corresponding first marked vascular tree and the first marked image map.

13. The method for predicting fractional flow reserve based on coronary CT images according to claim 1, characterized in that: The method further includes: performing model training on the FFR prediction model before using the FFR prediction model; The performing model training on the FFR prediction model specifically includes: Step 1301: recruit multiple volunteers without coronary artery stenosis symptoms and multiple volunteers with coronary artery stenosis symptoms of varying degrees to form a first volunteer group; Wherein, the first volunteer set includes a plurality of first volunteers; Step 1302: Perform multiple coronary artery path tests on each of the first volunteers using an invasive testing method. On each test path, feature data is collected from test points along the test path according to the data format of the first point's two-dimensional feature to obtain a corresponding test path two-dimensional feature sequence. The blood flow reserve fraction corresponding to each test point along the test path is measured to obtain a corresponding test path measurement sequence. Three-dimensional vascular tree modeling is performed based on the two-dimensional feature sequence to obtain a corresponding test path vascular tree. Based on the test path vascular tree, three-dimensional data conversion is performed on the two-dimensional feature sequence according to the data format of the first point's three-dimensional feature to obtain a corresponding test path three-dimensional feature sequence. The three-dimensional feature sequence corresponding to each test path and the test path measurement sequence are combined into a corresponding first data record. All of the obtained first data records are combined into a corresponding first data set. The detection path two-dimensional feature sequence includes a plurality of detection point two-dimensional features, and the data format of the detection point two-dimensional features matches the data format of the first point two-dimensional features; the detection path three-dimensional feature sequence includes a plurality of detection point three-dimensional features, and the data format of the detection point three-dimensional features matches the data format of the first point three-dimensional features; the detection path measurement sequence includes a plurality of detection point FFR measurement values; Step 1303: Set an integer greater than zero as the corresponding continuous pass count threshold; and set a pass count counter initialized to 0; Step 1304: randomly select a plurality of first data records from the first data set according to a preset test set ratio to form a corresponding current test set, and use all remaining first data records to form a corresponding current validation set; The current test set and the current validation set are both composed of a plurality of the first data records, and a ratio of the total number of the first data records in the current test set to the total number of the first data records in the first data set satisfies the test set ratio; Step 1305: extract the first first data record of the current test set as the corresponding current data record; Step 1306: Input the three-dimensional characteristic sequence of the detection path of the current data record into the FFR prediction model, and the FFR prediction model performs FFR prediction processing based on the three-dimensional characteristic sequence of the detection path input this time to obtain a corresponding detection path prediction sequence; Wherein, the detection path prediction sequence includes multiple detection point FFR prediction values; Step 1307 , subjecting the detection path prediction sequence and the detection path measurement sequence of the current data record to a preset first model loss function to calculate and obtain a corresponding first loss value; Wherein, the first model loss function is implemented based on the L1 loss function or the L2 loss function; Step 1308: Identify whether the first loss value satisfies a preset first loss value range. If the first loss value satisfies the first loss value range, identify whether the current data record is the last first data record in the current test set. If so, proceed to step 1309. If not, extract the next first data record in the current test set as the new current data record and return to step 1306 to continue training. If the first loss value does not satisfy the first loss value range, perform a round of model parameter optimization on the FFR prediction model based on a preset first model parameter optimizer in a direction to minimize the first model loss function. After this round of parameter optimization is completed, return to step 1306 to continue training. Wherein, the first model parameter optimizer includes at least an SDG optimizer; Step 1309: Traverse all first data records in the current validation set. During the traversal, the currently traversed first data record is used as the corresponding current validation record. The FFR prediction model performs FFR prediction processing based on the detection path feature sequence of the current validation record to obtain a corresponding current detection path prediction sequence. The current detection path prediction sequence and the detection path measurement sequence of the current validation record form a corresponding prediction-measurement sequence pair. At the end of the traversal, all the obtained prediction-measurement sequence pairs are substituted into preset first, second, and third error functions to calculate corresponding first, second, and third error values. The first error function is implemented based on the MSE error function, the second error function is implemented based on the RMSE error function, and the third error function is implemented based on the MAPE error function; Step 1310: Determine whether the first, second, and third error values ​​do not exceed the preset first, second, and third error thresholds. If none of the first, second, and third error values ​​exceed the corresponding first, second, and third error thresholds, increment the pass count counter by 1, and determine whether the incremented pass count counter is greater than the continuous pass count threshold. If so, proceed to step 1311; if not, return to step 1304 to continue training. If all of the first, second, or third error values ​​exceed the corresponding first, second, or third error thresholds, clear the pass count counter and return to step 1304 to continue training. Step 1311: solidify the model parameters of the FFR prediction model and confirm that the model training is completed.

14. A device for executing the method for predicting fractional flow reserve based on coronary CT images according to any one of claims 1 to 13, characterized in that: The device includes: a data receiving module, a semantic recognition module, a blood flow velocity feature extraction module, an image and geometric feature extraction module, a feature fusion module, a path search module, a vascular tree processing module, an FFR prediction module and a prediction result processing module; The data receiving module is configured to receive a coronary CT image sequence recorded as a corresponding first CT image sequence; the first CT image sequence includes a plurality of first CT images; each first CT image corresponds to a first image time; all first CT images have the same two-dimensional size; in the first CT image sequence, the first first CT image is a CT image when a contrast agent initially enters a coronary artery entrance, and the last first CT image is a CT image when the contrast agent fills the entire coronary vascular network; The semantic recognition module is configured to perform vascular network semantic segmentation on each of the first CT images to obtain a corresponding first semantic map; and to perform vascular inner wall edge and vascular centerline recognition on the vascular network on each of the first semantic maps to obtain a corresponding second semantic map; and to chronologically sort all of the obtained second semantic maps to form a corresponding second semantic map sequence; the first and second semantic maps correspond one-to-one to the first CT image and are consistent with the two-dimensional size of the corresponding first CT image; The blood flow velocity feature extraction module is configured to perform blood flow velocity feature extraction processing based on the second semantic graph sequence to obtain a corresponding first feature graph; the first feature graph corresponds one-to-one to the first CT image and is consistent with a two-dimensional size of the corresponding first CT image; The image and geometric feature extraction module is configured to record the last first CT image in the first CT image sequence as the corresponding full coronary image; perform image feature extraction on the full coronary image to obtain a corresponding second feature map; and perform vascular cross-sectional feature, vascular stenosis feature, and vascular topology feature extraction based on the full coronary image and the second semantic map to obtain a corresponding third feature map; The feature fusion module is used to perform feature fusion processing on the first, second and third feature maps corresponding to the full coronary image to obtain a corresponding first fused feature map; The path search module is configured to perform a vascular path search process on the first fusion feature map to obtain a corresponding first path sequence; and to sample path points on each first path and extract comprehensive features of each sampling point to obtain a corresponding first path feature sequence; the first path sequence includes a plurality of first paths; and the first path feature sequence is composed of a plurality of first point two-dimensional features sequentially arranged; The vascular tree processing module is configured to construct a three-dimensional vascular tree based on the first fusion feature map to obtain a corresponding first vascular tree; and to perform three-dimensional feature conversion on each of the first path feature sequences based on the first vascular tree to obtain a corresponding second path feature sequence; the second path feature sequence is composed of a plurality of first point three-dimensional features arranged in sequence; The FFR prediction module is used to perform FFR prediction processing on each second path feature sequence based on the FFR prediction model to obtain a corresponding first path prediction sequence; the FFR prediction model is implemented based on the Bi-LSTM model structure; the first path prediction sequence includes multiple first-point FFR prediction values; The prediction result processing module is used to perform FFR alarm marking processing on the first vascular tree and the full coronary image according to all the first path prediction sequences obtained to obtain the corresponding first marked vascular tree and first marked image map; and display the first marked vascular tree and the first marked image map.

15. An electronic device, characterized in that: include: memory, processors, and transceivers; The processor is configured to be coupled to the memory, read and execute instructions in the memory, so as to implement the method according to any one of claims 1 to 13; The transceiver is coupled to the processor, and the processor controls the transceiver to send and receive messages.

16. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer instructions, and when the computer instructions are executed by a computer, the computer is caused to execute the method according to any one of claims 1 to 13.

Images