Arterial aneurysm different-modal image registration method based on aneurysm wall thickness constraint

Abstract

The invention belongs to the technical field of medical image processing, and particularly relates to an aneurysm different-modal image registration method based on aneurysm wall thickness constraint. The method comprises the steps of screening key points in a target image point cloud to be registered and a source image point cloud to be registered according to an internal form descriptor algorithm, calculating curved surface normal vectors at the key points, estimating feature descriptors of the target and source point clouds by using a fast point feature histogram algorithm by taking the tumor wall thickness as a constraint, and obtaining the target and source image point clouds. Performing initial matching on the two point clouds by using a sampling consistency initial registration algorithm; and carrying out secondary optimization on the initial registration result of the point clouds by using an iterative nearest point algorithm to obtain a transformation matrix of the point clouds of the images with different modalities, and completing registration. According to the method, the technical problem of registration under the conditions of large non-rigid deformation and inconsistent scales is solved, the method can be used for three-dimensional registration of an aneurysm preoperative magnetic resonance image and a postoperative tissue transparency image, and a reference is provided for analyzing the risk of aneurysm rupture from the preoperative magnetic resonance image.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a method for registering images of different modalities of an aneurysm. Background Art

[0002] Intracranial aneurysm is a common vascular disease. Due to the high mortality and disability rate after rupture and bleeding, it is very important to assess the risk of aneurysm rupture. There are many factors that affect aneurysm rupture, among which the inflammation level of the aneurysm plays an important role. In order to use preoperative magnetic resonance imaging to evaluate the inflammation level of intracranial aneurysms, it is necessary to align the preoperative three-dimensional magnetic resonance imaging of the aneurysm with the postoperative three-dimensional tissue clearing staining image. Commonly used three-dimensional registration algorithms include iterative closest point (ICP) algorithm [1] , the transformation matrix is ​​found by sampling points iteratively, so that the distance between the two point clouds is minimized to achieve registration. Due to the difference in the morphology of aneurysms before and after surgery, there are scale changes, position changes and non-rigid deformations in the corresponding images, and the commonly used 3D registration algorithms cannot achieve the desired registration effect. Summary of the invention

[0003] The purpose of the present invention is to propose a method for registering images of different modalities of an aneurysm based on aneurysm wall thickness constraint to achieve a better registration effect.

[0004] The present invention adopts a fast point feature histogram (FPFH) algorithm based on aneurysm wall thickness constraint. [2] Perform a rough registration, and then use the iterative closest point algorithm for fine registration, which greatly improves the registration effect. The specific steps are as follows.

[0005] (I) Based on the internal shape signatures (ISS) algorithm [3] , filter the key points in the target point cloud and the source point cloud to be registered. Key points refer to points with significant geometric changes in the local area of ​​the point cloud. The specific screening method is as follows: For each point, calculate the covariance matrix of the points in its local neighborhood, perform eigenvalue decomposition on the covariance matrix, and obtain three eigenvalues, which are recorded as λ1, λ2, and λ3 from large to small. If the three eigenvalues ​​λ1, λ2, and λ3 corresponding to a point have the same value, the point is discarded. According to the number of key points N that need to be extracted, compare the smallest eigenvalue λ3 corresponding to all points, and filter out the top N with larger λ3 as candidate key points. Repeat the above process at different scales, and select points that are stable at multiple scales as the final key points. After obtaining the key points, calculate the surface normal vectors at all key points;

[0006] (ii) Taking the tumor wall thickness as a constraint, the fast point feature histogram algorithm is used to estimate the feature descriptors of the target and source point clouds, construct a point feature histogram, and use the sampling consistency initial registration algorithm [4] Perform an initial match between the two point clouds.

[0007] The specific process is as follows:

[0008] (1) First, the centroid of the aneurysm in different modality images is calculated, and a spherical coordinate system is established with the centroid as the center. The position of each point on the aneurysm wall is expressed as a spherical coordinate (r, θ, φ), where r is the radius, θ is the azimuth, and φ is the polar angle. For all θ and φ, the spherical space is divided into multiple concentric fan-cone intervals, and the aneurysm wall thickness of each area is calculated.

[0009] (2) For the key points in the target point cloud and the source point cloud to be registered, the feature descriptor of each point is estimated using the fast point feature histogram method with the thickness of the tumor wall as a constraint, and a point feature histogram is constructed;

[0010] (3) Then, using the sampling consistency initial registration algorithm, sampling points are first extracted from the point cloud to be registered, and a threshold is set to avoid the sampling points having the same features; a point set with similar feature descriptors as the sampling points is selected from the target point cloud, and then one point is randomly selected from the point set to correspond to the point to be registered, and the transformation matrix and registration error between the points are calculated. The transformation matrix with the smallest registration error is selected through iteration.

[0011] (III) The initial registration result of the point cloud is optimized again using the iterative closest point algorithm to obtain the transformation matrix of the point cloud of different modal images and complete the registration.

[0012] In the present invention, the mean square error RMSE and the linear correlation coefficient R of the two aneurysm wall thickness sequences are used to evaluate the registration results of the aneurysm three-dimensional image. The mean square error RMSE is used as an indicator to verify the accuracy of the algorithm, and its calculation formula is:

[0013]

[0014] In formula (1), n ​​is the number of corresponding point pairs in the target point cloud and the source point cloud, Represents the pairing error between corresponding point pairs.

[0015] Since the two images originate from the same aneurysm, the changes in the wall thickness of the same area on the different modal images are similar, that is, the wall thickness sequences are linearly correlated. The present invention uses a linear correlation coefficient R to reflect the linear correlation of the wall thickness of the same area on the two images. In the same φ section, the wall thickness sequence from small to large θ is used as the data set X and Y, respectively, and the linear correlation of X and Y is calculated. The calculation formula of the linear correlation coefficient R is:

[0016]

[0017] In formula (2), x i and i Represent the thickness values ​​at different positions of the two modal images, and Represent the mean thickness of the two modal images.

[0018] Compared with the prior art, the present invention adopts a two-step strategy of coarse registration and fine registration, and on the basis of the existing registration algorithm, adds the aneurysm wall thickness as a constraint, greatly improving the registration accuracy. The present invention solves the technical difficulties of registration in the presence of large non-rigid deformation and inconsistent scales, and can be used for the three-dimensional registration of aneurysm preoperative magnetic resonance images and postoperative tissue clearing images, providing a reference for analyzing the risk of aneurysm rupture from preoperative magnetic resonance images. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of the aneurysm different modality image registration method based on aneurysm wall thickness constraint of the present invention.

[0020] Figure 2 This is a schematic diagram of calculating the thickness of the tumor wall at different positions in spherical coordinates according to the present invention.

[0021] Figure 3 An example of a point cloud of the tumor wall to be registered. The left image is the target point cloud, and the right image is the source point cloud.

[0022] Figure 4 The left figure is the result of the present invention, and the right figure is the result of the ICP algorithm. DETAILED DESCRIPTION

[0023] In the specific implementation process, in order to evaluate the registration accuracy, it is necessary to know the gold standard of registration. First, the deformation parameters are customized, and the 3D graphics software Blender3.5 is used to use the magnetic resonance images of one modality to generate the deformed magnetic resonance images of another modality, which are used as the source image and target image to be registered. Figure 3An example of a point cloud corresponding to a tumor wall to be registered is shown. Then, the source image and the target image pair are registered using the method of the present invention and the ICP algorithm, respectively, and the results are compared. In order to calculate the thickness of the tumor wall more accurately, the present invention is divided into 40 orders in the θ and φ directions, respectively, and the entire spherical coordinate system is divided into 1600 blocks. The maximum value of the difference between the inner diameter and the outer diameter of the tumor wall contained in each interval is calculated as the thickness value of the tumor wall at that position. The registration experiment has two processes: coarse registration and fine registration. In the coarse registration stage, the number of key points N in the internal morphological descriptor algorithm is not more than one tenth of the number of all points in the point cloud. The point pair distance threshold in the sampling consistency initial registration algorithm is set to 0.05mm, the thickness threshold is 0.3mm, the maximum number of iterations is 100 times, and the point pair Euclidean distance threshold is 0.01mm. In comparison, the maximum number of iterations of the traditional ICP algorithm is also 100 times, and the point pair Euclidean distance threshold is also 0.01mm.

[0024] The hardware environment of the experiment is Intel Xeon W-2223CPU with 16GB of running memory; the system environment is 64-bit Windows system; the software environment is Visual Studio compiler and C++ development environment.

[0025] Figure 4 The table 1 shows the registration result of a data pair. It can be seen that the method of the present invention has a better registration effect than the ICP algorithm. The quantitative comparison results are shown in Table 1. It can be seen that the method of the present invention has a smaller registration error RMSE and a better linear correlation coefficient R, and the registration effect is better.

[0026] Table 1. Registration results of constructed aneurysm images in different modalities

[0027]

[0028] References

[0029] [1] et al.Robust generalized total least squares iterativeclosest point registration. / / Medical Image Computing and Computer-AssistedIntervention–MICCAI,2004:234-241.

[0030] [2]Zhong Y,et al.Intrinsic shape signatures:A shape descriptor for 3Dobject recognition. / / 2009IEEE 12th International Conference on ComputerVision Workshops,ICCV Workshops,Kyoto,Japan,2009:689-696.

[0031] [3]Rusu R,et al.Aligning point cloud views using persistent featurehistograms. / / 2008 IEEE / RSJ International Conference on Intelligent Robots andSystems,Nice,France,2008:3384–3391.

[0032] [4]Rusu R,et al.Fast point feature histograms(FPFH)for 3Dregistration. / / 2009IEEE International Conference on Robotics and Automation,2009:3212-3217。

Claims

1. A method for image registration of different modalities of aneurysms based on aneurysm wall thickness constraint, characterized in that: The specific steps are: (i) Select key points in the target point cloud and the source point cloud to be registered according to the internal morphological descriptor algorithm, and calculate the surface normal vector at the key points; key points refer to points with significant geometric changes in the local area of ​​the point cloud; (ii) Taking the tumor wall thickness as a constraint, the fast point feature histogram algorithm is used to estimate the feature descriptors of the target and source point clouds, a point feature histogram is constructed, and the two point clouds are initially matched using the sampling consistency initial registration algorithm; (III) The initial registration result of the point cloud is optimized again using the iterative closest point algorithm to obtain the transformation matrix of the point cloud of different modal images and complete the registration.

2. The aneurysm different modality image registration method based on aneurysm wall thickness constraint according to claim 1, characterized in that: The method for selecting key points in step (a) is as follows: for each point, calculate the covariance matrix of the points in its local neighborhood, perform eigenvalue decomposition on the covariance matrix, and obtain three eigenvalues, which are recorded as λ1, λ2, and λ3 from large to small; if the three eigenvalues ​​λ1, λ2, and λ3 corresponding to a point have the same value, then discard the point; according to the number of key points N that need to be extracted, compare the smallest eigenvalue λ3 corresponding to all points, and select the first N points with larger λ3 as candidate key points; repeat the above process at different scales, and select points that are stable at multiple scales as the final key points.

3. The aneurysm different modality image registration method based on aneurysm wall thickness constraint according to claim 1, characterized in that: The specific process of step (ii) is as follows: (1) First, the centroid of the aneurysm in different modality images is calculated, and a spherical coordinate system is established with the centroid as the center. The position of each point on the aneurysm wall is expressed as a spherical coordinate (r, θ, φ), where r is the radius, θ is the azimuth, and φ is the polar angle. For all θ and φ, the spherical space is divided into multiple concentric fan-cone intervals, and the aneurysm wall thickness of each area is calculated. (2) For the key points in the target point cloud and the source point cloud to be registered, the feature descriptor of each point is estimated using the fast point feature histogram method with the thickness of the tumor wall as a constraint, and a point feature histogram is constructed; (3) Then, using the sampling consistency initial registration algorithm, sampling points are first extracted from the point cloud to be registered, and a threshold is set to avoid the sampling points having the same features; a point set with similar feature descriptors as the sampling points is selected from the target point cloud, and then one point is randomly selected from the point set to correspond to the point to be registered, and the transformation matrix and registration error between the points are calculated. The transformation matrix with the smallest registration error is selected through iteration.

Images