Three-dimensional full-eye retinal vessel map construction method

Abstract

The invention discloses a three-dimensional full-eye retinal blood vessel map construction method. The method comprises the steps that original retinal blood vessel data are acquired and preprocessed; key points on the surface of the blood vessel are extracted, and a smoothed and refined three-dimensional mesh model is constructed; the vertexes of the surface where the retina deep blood vessels are located are screened, the geodesic distance and orientation from the position where the optic nerve is located to each vertex are calculated, and a three-dimensional latitude and longitude coordinate model is constructed; projecting the three-dimensional latitude and longitude coordinates to a two-dimensional plane according to a projection method; and performing feature extraction on the retinal blood vessel at the vertex of the surface, and combining the extracted features with the three-dimensional latitude and longitude coordinate model to obtain a three-dimensional latitude and longitude feature map model. According to the three-dimensional full-eye retinal blood vessel atlas construction method, the problem of cross-eyeball sample matching is solved through the three-dimensional latitude and longitude coordinate model, and a standard three-dimensional retinal blood vessel atlas template is established; and the visualization effect and the convenience of characteristic spectrum calculation are improved by using a projection method.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a method for constructing a three-dimensional full-eye retinal vascular atlas. Background Art

[0002] Eye atlases not only provide basic research tools for understanding complex visual functions, but also offer new perspectives and research means for understanding the underlying mechanisms of eye diseases, discovering biological markers for early diagnosis and efficacy evaluation of eye diseases, and establishing personalized and precise clinical treatments.

[0003] Currently, there is no atlas template specifically for the construction of retinal blood vessels. In traditional retinal vascular imaging methods, OCT does not have specific labeling ability, and has low resolution and limited field of view; while confocal microscopy requires cutting the eyeball to be imaged into a "four-leaf clover" structure, which destroys the integrity of the retinal blood vessels. Therefore, only light sheet fluorescence microscopy can achieve complete full-eye retinal vascular imaging, and has the advantages of high resolution and fast large-volume imaging.

[0004] However, different eyeball shapes are different and are prone to deformation during sample processing, resulting in the inability to directly perform cross-sample matching according to the original standard hemispherical shape of the retina. Therefore, it is necessary to develop a mapping method to map the full-eye retinal blood vessels into a unified spatial template for alignment in order to achieve precise drawing of the retinal vascular atlas. Summary of the Invention

[0005] In order to solve the problems in the background art, the present invention provides a method for constructing a three-dimensional full-eye retinal vascular atlas, which solves the problem of difficult matching caused by reasons such as retinal surface deformation in cross-sample matching, establishes a standard three-dimensional retinal vascular atlas template, and uses the projection method to improve the visualization effect and the convenience of feature atlas calculation, providing a reliable solution for the precise drawing of the retinal vascular atlas.

[0006] The technical solution adopted by the present invention is as follows:

[0007] The method for constructing a three-dimensional full-eye retinal vascular atlas includes the following steps:

[0008] S1) Obtain three-dimensional images of retinal blood vessels.

[0009] S2) Preprocess the three-dimensional images of retinal blood vessels to obtain preprocessed three-dimensional images of retinal blood vessels. The preprocessing includes binaryzation, blood vessel cavity filling, skeletonization, and morphological processing in sequence. The morphological processing is specifically: morphological processing is performed by the method of voxel connected components, and the connected component with the largest volume is retained.

[0010] S3) Sequentially perform surface key point extraction and 3D point cloud anomaly detection on the preprocessed 3D retinal vessel image to obtain 3D retinal vessel point cloud data.

[0011] Preferably, the rolling ball method is used for surface key point extraction.

[0012] Preferably, an outlier point cloud detection method based on statistical distribution is used for 3D point cloud anomaly detection.

[0013] S4) Use the 3D meshing method to establish a 3D mesh model of the retinal vessels based on the 3D retinal vessel point cloud data obtained in step S3: Use the triangulation meshing method to convert the surface key points in the 3D retinal vessel point cloud data into a triangular mesh composed of vertices and triangular patches to obtain the original triangular mesh model, and then sequentially perform surface smoothing processing and refinement processing on the original triangular mesh model to obtain the 3D mesh model of the retinal vessels.

[0014] S5) First, extract surface mesh vertices from the 3D mesh model of the retinal vessels established in step S4 and construct a 3D surface model: Rotate the triangular mesh model so that the optic nerve is at the top, divide the XY plane into several xy sub-planes, and use the mesh vertex with the maximum z value on each xy sub-plane as the surface mesh vertex, and generate a 3D surface model based on all surface mesh vertices. Among them, the surface mesh vertex is the mesh vertex on the surface where the deep vessels are located.

[0015] Secondly, according to the 3D surface model, calculate the geodesic distance and vector angle from the position of the optic nerve to each surface mesh vertex, convert them into longitude and latitude, and construct a 3D longitude and latitude coordinate model: Use the fast marching algorithm to calculate the geodesic distance from the position of the optic nerve to each surface mesh vertex;

[0016] Obtain the maximum geodesic distance and process it using the following formula to obtain the equatorial radius:

[0017] R = 2L max / π

[0018] In the formula, R represents the equatorial radius, and L max represents the maximum geodesic distance;

[0019] Process the geodesic distance of each surface mesh vertex using the following formula to obtain the latitude of the surface mesh vertex:

[0020] lon i = π / 2 - L i / R

[0021] In the formula, lon i represents the latitude of the i-th surface mesh vertex, and L irepresents the geodesic distance of the i-th surface grid vertex, and R represents the equatorial radius;

[0022] Meanwhile, for the projection point of each surface grid vertex on the XY plane, obtain the vector angle from the position of the optic nerve to the surface grid vertex, and convert the vector angle into longitude.

[0023] S6) Perform feature extraction on the three-dimensional surface model, and combine the extracted features and the three-dimensional longitude and latitude coordinate model to obtain a three-dimensional longitude and latitude feature map model: Perform feature extraction on the three-dimensional surface model to obtain the feature value at each surface grid vertex, and use the combination of the feature value at each surface grid vertex, the longitude and latitude of the surface grid vertex in the three-dimensional longitude and latitude coordinate model as a data point to construct a three-dimensional longitude and latitude feature map model.

[0024] In this step, the feature values at all surface grid vertices form the feature value distribution in the three-dimensional surface model and / or the three-dimensional longitude and latitude feature map model.

[0025] Preferably, the features include one or more of vascular density, vascular tortuosity, vascular branch length, and vascular branch point density.

[0026] S7) Process multiple different three-dimensional retinal vessel images according to steps S2 to S6, obtain the three-dimensional longitude and latitude feature map models corresponding to each three-dimensional retinal vessel image, map all the three-dimensional longitude and latitude feature map models to the same spatial template for alignment through spatial interpolation method, obtain the aligned three-dimensional longitude and latitude feature map models corresponding to each three-dimensional retinal vessel image, and then perform averaging processing on all the aligned three-dimensional longitude and latitude feature map models to obtain an integrated three-dimensional longitude and latitude feature map model.

[0027] Among them, the spatial template is the upper hemisphere of the unit sphere located by longitude and latitude grids.

[0028] The averaging process is carried out according to the following formula:

[0029]

[0030] In the formula, n is the total number of three-dimensional retinal vessel images, and m j is the feature value distribution in the aligned three-dimensional longitude and latitude feature map model corresponding to the j-th three-dimensional retinal vessel image, and M is the average feature value distribution in the integrated three-dimensional longitude and latitude feature map model.

[0031] Among them, the method of the present invention can achieve cross-eye sample matching. When the multiple different three-dimensional retinal vessel images are complete three-dimensional retinal vessel images from different retinal samples collected by a light sheet fluorescence microscope, an accurate integrated three-dimensional longitude and latitude feature map model can be constructed by the method of the present invention.

[0032] Furthermore, the construction method further includes the following steps: projecting the three-dimensional longitude and latitude coordinate model obtained in step S5 onto a two-dimensional plane for display using a projection method. This step can also replace the above-mentioned step S6.

[0033] Furthermore, the construction method further includes the following steps: projecting the three-dimensional longitude and latitude feature map model obtained in step S6 onto a two-dimensional plane for display using a projection method.

[0034] Furthermore, the construction method further includes the following steps: projecting the integrated three-dimensional longitude and latitude feature map model obtained in step S7 onto a two-dimensional plane for display using a projection method.

[0035] Preferably, the Lambert equal-area cylindrical projection method is used for projection.

[0036] The beneficial effects of the present invention are:

[0037] The three-dimensional full-eye retinal vessel map construction method of the present invention overcomes the problem of cross-eye sample matching through the three-dimensional longitude and latitude coordinate model, and establishes a standard three-dimensional retinal vessel map template; the visualization effect and the convenience of feature map calculation are improved by using the projection method. Brief Description of the Drawings

[0038] Figure 1 It is a flow chart of the retinal vessel map construction method of the present invention.

[0039] Figure 2 It is the three-dimensional original retinal vessel image data.

[0040] Figure 3 It is the three-dimensional retinal vessel map after binarization.

[0041] Figure 4 It is the three-dimensional retinal vessel map after skeletonization and morphological processing.

[0042] Figure 5 It is the key point map on the vessel surface after removing abnormal points.

[0043] Figure 6 It is the triangular mesh model map of the three-dimensional retinal vessels.

[0044] Figure 7 It is a schematic diagram of the vector angle.

[0045] Figure 8 It is a longitude and latitude coordinate model of three-dimensional retinal blood vessels.

[0046] Figure 9 It is a two-dimensional plan view after the Lambert equal-area cylindrical projection of the longitude and latitude coordinate model. Specific implementation manners

[0047] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0048] Specific embodiments of the present invention are as follows:

[0049] Embodiment 1

[0050] The process of the present invention is as Figure 1 shown. The steps of the method of the present invention will be further described below in conjunction with this embodiment:

[0051] S1) Obtain a three-dimensional image of the complete retinal blood vessels. Among them, the three-dimensional image of the retinal blood vessels is preferably a 3D grayscale image composed of a number of voxels.

[0052] Preferably, a three-dimensional image of the complete retinal blood vessels is obtained by a light sheet fluorescence microscope.

[0053] In the embodiment of the present invention, an image of a retinal sample is collected by a light sheet fluorescence microscope as a three-dimensional image of the retinal blood vessels. The original three-dimensional image of the retinal blood vessels is as Figure 2 shown, with a size of 2048×2048×1350, and the actual size corresponding to a single voxel is 1.625μm×1.625μm×2.00μm.

[0054] S2) Perform binaryzation, blood vessel cavity filling, skeletonization, and morphological processing on the three-dimensional image of the retinal blood vessels in sequence to obtain a preprocessed three-dimensional image of the retinal blood vessels.

[0055] Preferably, an adaptive threshold method is used for binaryzation processing to segment the retinal blood vessels and the background in the three-dimensional image data.

[0056] Preferably, a blood vessel cavity filling network based on deep learning is used for blood vessel cavity filling processing.

[0057] Preferably, a 3D skeletonization method of K3M is used for skeletonization processing.

[0058] Preferably, a voxel connected domain method is used for morphological processing. During the processing, the connected domain with the largest volume is retained to reduce the influence of noise points on the extraction of key points.

[0059] In the embodiments of the present invention, the preprocessing process is specifically as follows:

[0060] S2.1) Perform binarization processing on the three-dimensional image data of the retina by using the method of adaptive threshold, and segment the blood vessels and the background in the three-dimensional image data of the retinal blood vessels;

[0061] S2.2) Since large blood vessels only have relatively large signal intensities and gray values at the blood vessel wall positions and there are cavities inside the blood vessels. Therefore, after binarization, a deep learning network is used to fill the blood vessel cavities to ensure that complete binarized blood vessels can be extracted. After the blood vessel cavities are filled, the binarized image of the retinal blood vessels is as Figure 3 shown;

[0062] S2.3) For the data after the blood vessel cavity filling process, use the 3D skeletonization method of K3M to extract the skeleton lines of the blood vessels;

[0063] S2.4) The size of the retinal blood vessel skeleton image is 2048×2048×1350. Then, use the method of voxel connected components to perform morphological processing on the skeleton, retain the connected object with the largest volume, and obtain the three-dimensional image data of the retinal blood vessels after preprocessing, as Figure 4 shown.

[0064] S3) Perform surface key point extraction and three-dimensional point cloud anomaly detection on the three-dimensional image of the retinal blood vessels after preprocessing in sequence to obtain the three-dimensional point cloud data of the retinal blood vessels. By using the outlier point cloud detection method based on statistical distribution, the accuracy of surface extraction can be improved, and the influence of noise points on the construction of the surface three-dimensional model can be reduced.

[0065] Preferably, the rolling ball method is used to extract the surface key points of the three-dimensional image data of the retinal blood vessels after preprocessing.

[0066] Preferably, since the key points at the retinal edge positions are usually sparsely distributed, they are not included in the scope of outlier detection. Therefore, for the original three-dimensional point cloud data obtained after surface key points, only the non-edge regions are subjected to three-dimensional point cloud anomaly detection. The non-edge regions refer to the regions close to the position where the optic nerve is located and where the blood vessels are densely distributed.

[0067] Optionally, the non-edge regions can be set as follows: taking the position where the optic nerve is located as the highest point and the center position, setting the region with the Z-axis coordinate value higher than the preset Z-axis threshold as the non-edge region; or, taking the position where the optic nerve is located as the highest point and the center position, on the x-y plane, taking the circular region with the position where the optic nerve is located as the center and within the preset radius threshold as the non-edge region.

[0068] Preferably, the outlier point cloud detection method based on statistical distribution is used for three-dimensional point cloud anomaly detection.

[0069] In the embodiment of the present invention, this step is specifically as follows: Key points of the preprocessed three-dimensional retinal blood vessel image data are extracted by the rolling ball method and saved in the point cloud format. An outlier point cloud detection method based on statistical distribution is used to remove key points with abnormal spatial positions in the non-edge region, where the number of surrounding points is set to 35, the standard deviation multiple is set to 2, and among the surrounding key points, those with a distribution range exceeding twice the standard deviation are determined as outlier points. The non-edge region is the region where the Z-axis coordinate value is greater than 1200. The key points after removing outlier points are as Figure 5 shown.

[0070] S4) A three-dimensional meshing method is adopted to establish a three-dimensional grid model of the retinal blood vessels according to the three-dimensional retinal blood vessel point cloud data obtained in step S3.

[0071] Preferably, the three-dimensional grid model adopts a triangular grid model, which is generated by a triangulation method and mainly consists of vertices and triangular patches.

[0072] Furthermore, during this process, the original three-dimensional grid model is subjected to surface smoothing processing to reduce abnormal protrusions on the surface, and then the surface is refined. The original triangular patches are subdivided to increase the number of patches, and the smoothed and refined three-dimensional grid model is used for the next step of processing. By performing smoothing and refinement processing on the original triangular grid model, the accuracy of geodesic distance calculation can be improved.

[0073] In the embodiment of the present invention, this step is specifically as follows: For the surface key points after removing outlier points, they are converted into a triangular grid model consisting of vertices and triangular patches by a triangulation meshing method. Then, the triangular grid model is subjected to surface smoothing processing, and the number of smoothing iterations is set to 5. Then, the surface is refined, and the original triangular patches are subdivided to increase the number of patches, and the number of iterations is set to 2. The triangular grid model is as Figure 6 shown.

[0074] S5) Grid vertices on the surface where the deep blood vessels are located are extracted from the three-dimensional grid model of the retinal blood vessels established in step S4 as surface grid vertices, and a three-dimensional surface model is constructed. Then, according to the three-dimensional surface model, the geodesic distance and vector angle from the position of the optic nerve to each surface grid vertex are calculated, converted into longitude and latitude, and a three-dimensional longitude and latitude coordinate model is constructed.

[0075] This step is specifically as follows:

[0076] S5.1) Select the vertices of the 3D mesh model of the surface where the deep blood vessels are located from the 3D mesh model of the retinal blood vessels: Rotate the triangular mesh model so that the optic nerve is at the top. At this time, the upper surface of the triangular mesh model corresponds to the surface where the deep retinal blood vessels are located. Divide the XY plane into several xy sub-planes, and take the mesh vertex with the maximum z value on each xy sub-plane as the mesh vertex of the surface where the deep blood vessels are located, that is, the surface mesh vertex;

[0077] Construct a 3D surface model: Generate a 3D surface model based on all surface mesh vertices;

[0078] S5.2) Calculate the geodesic distance and vector angle of the mesh vertices, convert them to longitude and latitude, and construct a 3D longitude and latitude coordinate model based on the 3D surface model:

[0079] Use the fast marching method to calculate the geodesic distance from each surface mesh vertex to the location of the optic nerve;

[0080] Process the maximum geodesic distance using the following formula to obtain the equatorial radius:

[0081] R = 2L max / π

[0082] In the formula, R represents the equatorial radius, and L max represents the maximum geodesic distance;

[0083] Use the following formula for the geodesic distance of each surface mesh vertex to obtain the latitude:

[0084] lon i = π / 2 - L i / R

[0085] In the formula, lon i represents the latitude of the i-th surface mesh vertex, L i represents the geodesic distance of the i-th surface mesh vertex, and R represents the equatorial radius;

[0086] At the same time, for the projection point of each surface mesh vertex on the XY plane, obtain the vector angle from the location of the optic nerve to the surface mesh vertex, and convert the vector angle to longitude;

[0087] Construct a 3D longitude and latitude coordinate model: Generate a 3D longitude and latitude coordinate model based on all surface mesh vertices and their corresponding longitude and latitude.

[0088] In the embodiment of the present invention, this step is specifically:

[0089] S5.1) First, rotate the triangular mesh model so that the optic nerve is at the top. The entire retinal blood vessel is in the shape of an inverted bowl, and the surface where the deep retinal blood vessels are located is the upper surface. Divide the XY plane into several small squares with a size of 15×15 to cover the projection of the entire retinal blood vessel on the XY plane. Then, find the grid vertex with the highest Z - direction value within each square range as the surface grid vertex. Traverse all the small squares to obtain all the surface grid vertices;

[0090] S5.2) Consider the entire retina as the northern hemisphere of the earth, the position where the optic nerve is located as the "North Pole", and the position where the retinal edge is located as the equator. First, find the longest geodesic distance L from the "North Pole" to the retinal edge max , then the radius of the sphere is:

[0091] R = 2L max / π

[0092] In the formula, R represents the equatorial radius, and L max represents the maximum geodesic distance;

[0093] If the geodesic distance from the "North Pole" to any surface grid vertex is L i , then its corresponding latitude is:

[0094] lon i = π / 2 - L i / R

[0095] In the formula, lon i represents the latitude of the i - th surface grid vertex, L i represents the geodesic distance of the i - th surface grid vertex, and R represents the equatorial radius;

[0096] Obtain the azimuth angle of each surface grid vertex. The azimuth angle is specifically the vector angle on the XY plane from the "North Pole" to the surface grid vertex. In this embodiment, as Figure 7 shown, the vector angle is the angle between the vector from the position where the optic nerve is located to the surface grid vertex and the positive x - axis direction. When the surface grid vertex is above the x - axis, the vector angle takes a positive value, and the value range is from 0° to 180°; when the surface grid vertex is below the x - axis, the vector angle takes a negative value, and the value range is from 0° to - 180°. When converting to longitude and latitude, if the angle is in the range of 0° to 180°, it is east longitude (W), and if it is in the range of 0° to - 180°, it is west longitude (E).

[0097] The schematic diagram of the three - dimensional longitude - latitude coordinate model obtained in this embodiment is as Figure 8 shown.

[0098] S6) Project the three-dimensional longitude and latitude coordinate model obtained in step S5 onto a two-dimensional space using the Lambert equal-area cylindrical projection method to establish a position mapping relationship from the retinal blood vessel surface to the two-dimensional plane, so as to enhance the visualization effect and the convenience of feature map calculation. The three-dimensional longitude and latitude coordinate model obtained in this embodiment ( Figure 8 ) after being projected onto the two-dimensional plane is distributed as Figure 9 shown.

[0099] It should be noted that both Figure 8 and Figure 9 here are only schematic diagrams. In specific implementation, the distribution of surface grid vertices is relatively random and will not exactly be the shape of longitude and latitude grids.

[0100] It can be seen that the three-dimensional longitude and latitude coordinate model obtained by the method of the present invention can be projected onto a two-dimensional plane using the Lambert equal-area cylindrical projection, thereby enhancing the visualization effect and the convenience of feature map calculation.

[0101] Furthermore, the present invention also shows in the following embodiments a method for constructing a feature map for a single retinal sample (Embodiment 2) or multiple retinal samples (Embodiments 3 and 4) by combining the three-dimensional surface model and the three-dimensional longitude and latitude coordinate model obtained in step S5.

[0102] Embodiment 2

[0103] In this embodiment, a feature map is constructed for the three-dimensional image of blood vessels of a single retinal sample.

[0104] In this embodiment, steps S1 to S5 are the same as those in Embodiment 1. The difference is that step S6 is: extracting features from the three-dimensional surface model, and combining the extracted features with the three-dimensional longitude and latitude coordinate model to obtain a three-dimensional longitude and latitude feature map model.

[0105] Optionally, the extracted features are one or more of features such as blood vessel density, blood vessel tortuosity, blood vessel branch length, blood vessel branch point density, etc.

[0106] Taking the extracted feature as the blood vessel branch point density as an example in this embodiment, the specific process is as follows:

[0107] S6.1) For each surface grid vertex in the three-dimensional surface model obtained in step S5 of Embodiment 1, with any surface grid vertex as the center of the sphere and a radius r = 45, count the number N of blood vessel branch points within the radius r. Among them, the spatial positions of the blood vessel branch points are obtained from the preprocessed three-dimensional image of retinal blood vessels obtained in step S2.4 of Embodiment 1.

[0108] Then calculate the volume of the sphere Obtain the blood vessel bifurcation point density value N / V at the vertices of the surface mesh.

[0109] S6.2) Calculate the blood vessel bifurcation point density value for each vertex of the surface mesh according to step S6.1), and obtain the distribution of blood vessel bifurcation point density values at the positions of each vertex of the surface mesh for a single retinal sample.

[0110] S6.3) Take the combination of the blood vessel bifurcation point density value of each vertex of the surface mesh and its longitude and latitude in the three-dimensional longitude-latitude coordinate model as a data point, and generate a three-dimensional longitude-latitude feature map model based on the data points corresponding to all vertices of the surface mesh.

[0111] Furthermore, the method of the present invention can also be used to integrate the three-dimensional longitude-latitude feature map models of multiple retinal samples to obtain a unified integrated three-dimensional longitude-latitude feature map model. The integration process specifically refers to: obtaining the three-dimensional longitude-latitude feature map models of multiple retinal samples, that is, the distribution of blood vessel bifurcation point density values combined with the three-dimensional longitude-latitude coordinate model, mapping these three-dimensional longitude-latitude feature map models into the same spatial template and aligning them, and then calculating the average value of the blood vessel bifurcation point density at each longitude-latitude position in the spatial template through averaging processing to obtain the feature map corresponding to the blood vessel bifurcation point density.

[0112] Next, the present invention demonstrates an implementation scheme for integrating the three-dimensional longitude-latitude feature map models of multiple retinal samples to obtain an integrated three-dimensional longitude-latitude feature map model through Example 3 and Example 4.

[0113] Example 3

[0114] In this embodiment, the three-dimensional longitude-latitude coordinate models corresponding to the three-dimensional blood vessel images of multiple different retinal samples are obtained by the same process as steps S1 to S5 of Example 1, and the three-dimensional longitude-latitude feature map models corresponding to each retinal sample are obtained by the same process as Example 2. Then, the three-dimensional longitude-latitude feature map models corresponding to all retinal samples are mapped into the same spatial template for alignment.

[0115] The spatial template refers to the upper hemisphere of a unit sphere located by longitude-latitude grids (all longitude-latitude combinations with a longitude range from -180° to 180° and an interval of 1 degree, and a latitude range from 0° to 90° and an interval of 1°).

[0116] The specific process of mapping and alignment is as follows:

[0117] According to Example 2, the distribution of blood vessel bifurcation point density values at the positions of each vertex of the three-dimensional retinal blood vessel image is obtained. Combining with step S5 of Example 1, it can be known that each vertex of the surface grid corresponds to a longitude and latitude coordinate. Through spatial interpolation, the distribution of blood vessel bifurcation point density values m at all grid points of the longitude and latitude grid (all combinations of longitudes ranging from -180° to 180° with an interval of 1 degree and latitudes ranging from 0° to 90° with an interval of 1°) is calculated. m is a 361x91 matrix, and m lon,lat represents the blood vessel bifurcation point density value at each longitude and latitude (lon, lat) position. Thus, the mapping alignment from the three-dimensional longitude and latitude coordinate model of a single retinal sample to the spatial template is realized.

[0118] The eigenvalue distributions in the three-dimensional longitude and latitude feature map models corresponding to each retinal sample are mapped and aligned in the same way, and finally, the aligned three-dimensional longitude and latitude feature map models corresponding to each retinal sample are obtained.

[0119] Example 4

[0120] Taking the eigenvalue distribution in the aligned three-dimensional longitude and latitude feature map model as the mapped and aligned eigenvalue distribution, in this example, by averaging the mapped and aligned eigenvalue distributions corresponding to all retinal samples, an integrated three-dimensional longitude and latitude feature map model is obtained.

[0121] The specific process is as follows:

[0122] For a single retinal sample, after feature extraction in Example 2 and mapping alignment in Example 3, the eigenvalue distribution at all grid points of the longitude and latitude grid (all combinations of longitudes ranging from -180° to 180° with an interval of 1 degree and latitudes ranging from 0° to 90° with an interval of 1°) is obtained, that is, the mapped and aligned eigenvalue distribution, denoted as m j , where j represents the serial number of the retinal sample.

[0123] The mapped and aligned eigenvalue distributions corresponding to multiple samples are averaged to obtain an integrated three-dimensional longitude and latitude feature map model. The eigenvalue distribution in the integrated three-dimensional longitude and latitude feature map model is expressed as:

[0124]

[0125] Among them, n is the number of retinal samples, and m j is the eigenvalue distribution at all longitude and latitude grid points in the aligned three-dimensional longitude and latitude feature map model corresponding to the jth retinal sample, and M is the average eigenvalue distribution at all longitude and latitude grid points in the integrated three-dimensional longitude and latitude feature map model.

[0126] It can be seen that the three-dimensional longitude and latitude coordinate model constructed by the method of the present invention can uniformly map retinal blood vessels with various shapes into a standardized three-dimensional space template, thereby solving the problem of cross-sample matching. In addition, by mapping the retinal blood vessels of the entire eye into a unified space template for alignment, the accurate drawing of the retinal blood vessel atlas can also be achieved.

[0127] In summary, the method for constructing a three-dimensional retinal blood vessel atlas of the entire eye according to the present invention overcomes the problem of cross-eye sample matching through a three-dimensional longitude and latitude coordinate model, and establishes a standard three-dimensional retinal blood vessel atlas template; the projection method is used to improve the visualization effect and the convenience of feature atlas calculation.

[0128] The above specific embodiments are used to explain and illustrate the present invention, rather than limiting the present invention. Any modification and change made to the present invention within the spirit and scope of the claims of the present invention fall within the protection scope of the present invention.

[0129] The above is only a preferred embodiment of the present invention. Therefore, any equivalent changes or modifications made according to the structure, features, and principles described in the scope of the present invention patent application are included in the scope of the present invention patent application.

Claims

1. A method for constructing a three-dimensional whole-eye retinal vascular atlas, characterized in that, Including the following steps: S1) Obtain a three-dimensional image of retinal blood vessels; S2) Preprocess the three-dimensional image of retinal blood vessels to obtain a preprocessed three-dimensional image of retinal blood vessels; S3) Successively perform surface key point extraction and three-dimensional point cloud anomaly detection on the preprocessed three-dimensional image of retinal blood vessels to obtain three-dimensional point cloud data of retinal blood vessels; S4) Adopt a three-dimensional meshing method to establish a three-dimensional grid model of retinal blood vessels according to the three-dimensional point cloud data of retinal blood vessels obtained in step S3; S5) Extract surface grid vertices from the three-dimensional grid model of retinal blood vessels established in step S4 and construct a three-dimensional surface model; According to the three-dimensional surface model, calculate the geodesic distance and vector angle from the position where the optic nerve is located to each surface grid vertex, convert them into longitude and latitude, and construct a three-dimensional longitude and latitude coordinate model; S6) Perform feature extraction on the three-dimensional surface model, and combine the extracted features and the three-dimensional longitude and latitude coordinate model to obtain a three-dimensional longitude and latitude feature map model.

2. The three-dimensional full-eye retinal vascular atlas construction method according to claim 1, characterized in that: In step S5, the surface grid vertices are the grid vertices on the surface where the deep blood vessels are located; the process of constructing a three-dimensional surface model according to the three-dimensional grid model is specifically as follows: Rotate the triangular grid model so that the optic nerve is at the topmost position, divide the XY plane into several xy sub-planes, take the grid vertex with the maximum z value on each xy sub-plane as the surface grid vertex, and generate a three-dimensional surface model according to all surface grid vertices.

3. The three-dimensional full-eye retinal vascular atlas construction method according to claim 1, characterized in that: In step S5, the process of constructing a three-dimensional longitude and latitude coordinate model according to the three-dimensional surface model is specifically as follows: Use the fast marching algorithm to calculate the geodesic distance from the position where the optic nerve is located to each surface grid vertex; Obtain the maximum geodesic distance and process it with the following formula to obtain the equatorial radius: R = 2L max / π where R represents the equatorial radius and L max represents the maximum geodesic distance; Process the geodesic distance of each surface grid vertex with the following formula to obtain the latitude of the surface grid vertex: lon i = π / 2 - L i / R where lon i represents the latitude of the i-th surface grid vertex, L i represents the geodesic distance of the i-th surface grid vertex, and R represents the equatorial radius; At the same time, for the projection point of each surface grid vertex on the XY plane, obtain the vector angle from the position where the optic nerve is located to the surface grid vertex, and convert the vector angle into longitude.

4. The method for constructing a three-dimensional full-eye retinal vascular atlas according to claim 1, wherein: Step S6 is specifically as follows: Perform feature extraction on the three-dimensional surface model to obtain the eigenvalue at each surface grid vertex, and use the combination of the eigenvalue at each surface grid vertex, the longitude and latitude of the surface grid vertex in the three-dimensional longitude and latitude coordinate model as a data point to construct a three-dimensional longitude and latitude feature map model.

5. The method for constructing a three-dimensional full-eye retinal vascular atlas according to claim 1, wherein: The construction method further includes: S7) Process multiple different three-dimensional images of retinal blood vessels according to steps S2 to S6, obtain the three-dimensional longitude and latitude feature map models corresponding to each three-dimensional image of retinal blood vessels, map all the three-dimensional longitude and latitude feature map models to the same spatial template for alignment through spatial interpolation to obtain the aligned three-dimensional longitude and latitude feature map models corresponding to each three-dimensional image of retinal blood vessels, and then perform averaging processing on all the aligned three-dimensional longitude and latitude feature map models to obtain an integrated three-dimensional longitude and latitude feature map model; Among them, the spatial template is the upper hemisphere of a unit sphere positioned with longitude and latitude grids; The averaging processing is performed according to the following formula: where n is the total number of three-dimensional images of retinal blood vessels, and m j is the eigenvalue distribution in the aligned three-dimensional longitude and latitude feature map model corresponding to the j-th three-dimensional image of retinal blood vessels, and M is the average eigenvalue distribution in the integrated three-dimensional longitude and latitude feature map model.

6. The method for constructing a three-dimensional full-eye retinal vascular atlas according to claim 1, 4 or 5, characterized in that: The features include one or more of blood vessel density, blood vessel tortuosity, blood vessel branch length, and blood vessel branch point density.

7. The three-dimensional full-eye retinal vascular atlas construction method according to claim 5, characterized in that: The construction method further includes: using the projection method to project the three-dimensional longitude and latitude coordinate model obtained in step S5, the three-dimensional longitude and latitude feature map model obtained in step S6, and / or the integrated three-dimensional longitude and latitude feature map model obtained in step S7 onto a two-dimensional plane for display.

8. The method for constructing a three-dimensional full-eye retinal vascular atlas according to claim 7, wherein: Use the Lambert equal-area cylindrical projection method for projection.

9. The three-dimensional full-eye retinal vascular atlas construction method according to claim 1, characterized in that: In step S3, the rolling ball method is used for surface key point extraction; and / or, the outlier cloud detection method based on statistical distribution is used for three-dimensional point cloud anomaly detection.

10. The method for constructing a three-dimensional full-eye retinal vascular map according to claim 1, wherein: Step S4 is specifically: adopting the triangulation meshing method to convert the surface key points in the three-dimensional point cloud data of the retinal blood vessels into a triangular mesh composed of vertices and triangular patches, obtaining the original triangular mesh model, and then sequentially performing surface smoothing processing and refinement processing on the original triangular mesh model to obtain the three-dimensional mesh model of the retinal blood vessels.

Images