A Brain CT and CTP Image Registration Method Based on Point Cloud Technology

Abstract

The present invention provides a method for registering brain CT and CTP images based on point cloud technology, comprising the following steps: S1. Perform spatial resolution normalization processing on the CT image and the CTP image respectively, and perform normalization processing on the gray values of the CT image and the CTP image to adjust the gray scale range to a unified scale; S2. Extract feature points from the CT image and the CTP image after mask masking and normalization processing; S3. Convert the feature points into three-dimensional point cloud data; S4. Based on the rigid body transformation model, perform preliminary alignment on the above three-dimensional point cloud data, and estimate the rotation and translation parameters of the image; S5. Combine the similarity of the point cloud gray information, fuse the spatial and gray features of the multi-modal images, and further optimize the point cloud alignment result; S6. Use the geometric error measurement method to evaluate the spatial alignment accuracy of the point cloud registration; solve the registration problem of brain CT and CTP images caused by differences in resolution, gray scale and modal characteristics through point cloud technology.

Description

Technical Field

[0001] The present invention belongs to the field of image registration, and particularly relates to a method for registering brain CT and CTP images based on point cloud technology. Background Art

[0002] Multi-modal medical image registration is an important research direction in medical image processing and is widely used in the diagnosis and treatment of brain diseases. As two commonly used imaging modalities in clinical practice, CT (Computed Tomography) and CTP (Cerebral Perfusion Imaging) respectively provide anatomical structure information and cerebral hemodynamic information. However, due to significant differences in imaging mechanisms, spatial resolutions, slice thicknesses, and gray value ranges between CT and CTP images, there are many technical challenges in directly registering the two images.

[0003] Traditional multi-modal registration methods mainly include global registration algorithms based on gray value similarity and local registration algorithms based on feature points. However, these methods have the following problems:

[0004] Global registration methods (such as registration based on mutual information) are difficult to capture stable corresponding relationships in the case of significant gray value differences in multi-modal images. Local feature methods (such as SURF, ORB) perform poorly in cross-modal images, resulting in a decrease in registration accuracy due to unstable feature point matching. Although the rigid body transformation model can achieve preliminary alignment, it lacks the ability to fuse multi-modal gray value characteristics and cannot meet the requirements of high-precision registration. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for registering brain CT and CTP images based on point cloud technology, which solves the registration problem caused by differences in resolution, gray value, and modal characteristics between brain CT and CTP images through point cloud technology, realizes high-precision multi-modal image fusion, provides reliable technical support for the diagnosis, treatment, and postoperative evaluation of brain diseases, and has broad clinical application value.

[0006] To achieve the above purpose, the present invention provides the following technical solution: A method for accurately registering brain CT and CTP images based on point cloud technology, comprising the following steps:

[0007] S1. Perform spatial resolution standardization processing on the CT image and the CTP image respectively to sample images with different resolutions on a unified spatial scale; and perform normalization processing on the gray values of the CT image and the CTP image to adjust the gray value range to a unified scale;

[0008] Specific steps of the spatial resolution standardization processing:

[0009] Execution method: Adjust the CT and CTP images to the same resolution through interpolation algorithms (such as cubic spline interpolation, bilinear interpolation, etc.). For example, adjust the slice thickness of the CT image from 5 mm to 1.25 mm to be consistent with the CTP image.

[0010] Operation objective: Make the voxel spacing of the CT and CTP images consistent in the spatial coordinate system, so as to achieve standardization in space.

[0011] Result (effect): The difference in spatial resolution is eliminated, and there is a one-to-one correspondence between the voxels of the CT and CTP images.

[0012] Specific steps for gray value normalization:

[0013] Execution method: Calculate the normalized gray value of each voxel, for example, through the linear normalization formula: , where I norm is the normalized voxel value (i.e., gray value), and max(I) and min(I) are the global maximum and minimum gray values in the image.

[0014] Operation objective: Adjust the gray value ranges of the CT image and the CTP image to a unified standard scale (such as 0 to 1).

[0015] Effect: The problem of inconsistent gray value ranges is solved, providing consistent gray value input for multi-modal image registration.

[0016] The spatial resolution standardization process and the gray value normalization process are executed through interpolation algorithms and normalization formulas respectively, which belong to specific technical operation steps. The effect of the standardization process: After executing these steps, the images achieve consistency in spatial resolution and gray value distribution, which is the result brought by the standardization process.

[0017] S2. Extract feature points with rotation invariance, scale invariance, and gray value change invariance from the CT image and CTP image after spatial resolution standardization and gray value normalization;

[0018] S3. Convert the feature points into three-dimensional point cloud data containing spatial coordinates and corresponding gray texture information;

[0019] S4. Perform preliminary registration on the three-dimensional point cloud data based on the rigid body transformation model, estimate the rotation and translation parameters between the CT image and the CTP image, so as to achieve rough registration and alignment of the two images;

[0020] S5. On the basis of the preliminary alignment, combine the similarity of the point cloud gray information, fuse the spatial features and gray features of the multi-modal images, and further optimize the point cloud alignment result; among them, the comprehensive objective function adopted introduces cross-modal gray consistency on the basis of geometric position alignment, and by considering the dual factors of space and texture intensity, the CT and CTP are fused simultaneously in terms of position and brightness distribution;

[0021] S6. Adopt a geometric error measurement method to evaluate the spatial alignment accuracy of point cloud registration, and use a distribution similarity measurement method to verify the fusion effect of image gray information.

[0022] Furthermore, in the process of spatial resolution normalization in the step S1, perform trilinear interpolation on the CT image I CT (x, y, z) and the CTP image I CTP (x, y, z) respectively to make them consistent in resolution and slice thickness. The gray value of the interpolation point (x, y, z) is calculated by the following formula:

[0023] Determine the eight nearest neighboring points (x i , y j , z k ) in the original grid for the interpolation point (x, y, z). Let (x i , y j , z k ) represent the i-th neighboring point, and use the interpolation weight to calculate the difference weight of the interpolation point, and calculate the gray value of the interpolation point through the trilinear interpolation formula

[0024]

[0025] ;

[0026] The trilinear interpolation formula is based on the three-dimensional linear interpolation theory, and ensures the continuity and smoothness of the image data through weighted summation, effectively eliminating the differences in spatial sampling between the CT and CTP, and providing a unified spatial scale for subsequent feature point extraction.

[0027] Furthermore, the normalization process of the gray values of the CT image and the CTP image includes: normalizing the CT values of the image obtained after interpolation according to the following formula:

[0028]

[0029] Among them, M(x, y, z) is based on the window width I WW and the window level I WLA mask that masks the CT value and CTP value, and only normalizes the unmasked part; IHU(x,y,z) represents the gray value of a point in the CT image, and I(x,y,z) is the gray value of a voxel in the CT image. Unifying the brightness ranges of the CT and CTP images to the interval [0,1] reduces the cross-modal matching problem caused by the different gray dynamic ranges of the CT and CTP images, and ensures the comparability of subsequent feature point extraction and point cloud generation in the gray dimension.

[0030] x, y, and z represent the three-dimensional coordinates of the interpolation point in the unified resolution (in the target coordinate system). These coordinates are usually located between the original voxel points, so the gray value needs to be calculated through interpolation.

[0031] Original coordinate system: x i ,x i+1 : Two original sampling points adjacent to the interpolation point in the x direction; y i ,y i+1 : Two original sampling points adjacent to the interpolation point in the y direction; z i ,z i+1 : Two original sampling points adjacent to the interpolation point in the z direction. The gray value of the interpolation point (x,y,z) is calculated by trilinear interpolation from the gray values of the eight original voxel points where it is located.

[0032] Through trilinear interpolation, the resolution and slice thickness of the CT image and CTP image can be adjusted to a unified scale, so that the two are corresponding and consistent in spatial position. This processing can effectively eliminate the influence of resolution differences on subsequent registration, and provide consistent input data for feature extraction and point cloud generation.

[0033] Summary: The gray value of the interpolation point (x,y,z) is calculated by the trilinear interpolation formula. The weights u, v, and w in the formula represent the relative positions of the interpolation point in each dimension direction, and the gray values I000 to I111 of the eight adjacent voxel points participate in the weighted calculation. This process ensures the unity of the CT and CTP images in resolution and slice thickness, laying a foundation for subsequent processing.

[0034] Furthermore, in step S2, the feature point extraction described includes:

[0035] Perform Gaussian blur convolution on the normalized image I norm (x,y,z) to construct a scale space;

[0036] ; where G(x,y,z,σ) is the Gaussian kernel with scale σ, and L(x,y,z,σ) represents the blurred image of the image filtered by the Gaussian kernel at a certain scale σ;

[0037] Then, through the scale difference formula Search for local extreme points to obtain the coordinates (x, y, z) of the feature points, so that key points with robustness to noise, rotation, and scale changes can be detected at different resolutions, laying a foundation for subsequent point cloud generation.

[0038] Furthermore, in step S3, when converting the feature points into three-dimensional point cloud data, the feature point coordinates (x, y, z) are combined with their corresponding normalized gray value and local descriptor information and represented as:

[0039] , where I normI is the normalized gray value of the feature point, and D is the local texture or gradient descriptor; this point cloud form can simultaneously represent the spatial position and modal texture information of the feature points in the same coordinate system, providing a unified data structure for multi-modal image fusion and improving the registration efficiency.

[0040] Furthermore, in step S4, when performing a preliminary alignment on the three-dimensional point cloud data, a rigid body transformation model is used, where P represents the spatial coordinate vector of a point in the point cloud, R is the rotation matrix, and t is the translation vector. By sequentially adjusting the rotation and translation parameters, the angular and position errors generated during the shooting of CT and CTP are eliminated, ensuring the preliminary alignment consistency of the images in the three-dimensional coordinate space.

[0041] Furthermore, the rigid body transformation parameters R and t are optimized through the Iterative Closest Point (ICP) algorithm, and the objective function is adopted; where, where p i and q i are the corresponding point pairs selected by the closest point strategy in the CT point cloud and the CTP point cloud, and N is the number of corresponding point pairs; the ICP algorithm is based on the least squares principle and can gradually reduce the distance error between corresponding points during the iteration process until it converges to a high-precision geometric registration result, laying a reliable spatial position foundation for cross-modal fusion.

[0042] Furthermore, in step S5, during the process of point cloud alignment, the following comprehensive objective function is adopted, , where E ICP is the geometric distance error, α and β are weight coefficients, σ represents the measure of image gray similarity, represents the gray similarity, I CT and I CTP respectively come from the information of the CT image and the CTP image carried by the point cloud; this comprehensive objective introduces cross-modal gray consistency on the basis of geometric position alignment, and by considering the dual factors of space and texture intensity, enables the simultaneous fusion of CT and CTP in terms of position and brightness distribution, improving the final multi-modal registration accuracy and robustness.

[0043] Through systematic differential analysis of brain CT (Computed Tomography) and CTP (Cerebral Perfusion Imaging) in terms of resolution, gray-scale range, texture features, etc., and by using multiple means such as unified sampling, cross-modal feature detection, and point cloud fusion, the present invention significantly enhances the accuracy and robustness of cross-modal image registration and has considerable application value in both clinical diagnosis and scientific research analysis. The beneficial effects brought by the present invention are described in detail from multiple dimensions below. The length is relatively long to ensure sufficient explanation and logical coherence.

[0044] First, at the level of resolution and gray-scale normalization, the present invention adopts two major strategies of trilinear interpolation and linear normalization at the beginning. For brain CT images, they usually have a relatively large thickness and relatively sparse pixel sampling. In order to capture real-time blood flow characteristics, CTP uses a smaller slice thickness and dynamically collects data in the time dimension. This results in significant differences in both spatial resolution and brightness distribution between the two. If the CT image and the CTP image are directly superimposed or features are extracted according to the original resolution, it is very likely that corresponding positions cannot be found on some tomograms, or because the brightness of CT is concentrated in the range of bones and soft tissues, while CTP shows brightness fluctuations in high-perfusion or low-perfusion areas, thus unable to achieve effective matching. The present invention carefully designs the trilinear interpolation formula so that the gray-scale value of each interpolation point is obtained by the weighted sum of its neighboring coordinates, fully ensuring the smoothness and consistency of the image in three-dimensional space; combined with linear normalization, the gray-scales of CT and CTP are stretched to the same range interval. This idea of "equalizing the starting line" greatly reduces the coordinate and brightness mismatches caused by different resolution and gray-scale scales, enabling any subsequent cross-modal feature detection and point cloud fusion to be carried out under a unified measurement standard. The benefit of this is that it fundamentally solves the contradiction between CT and CTP of "one large grid and one small grid; one with concentrated brightness and the other with a very large dynamic range of brightness", allowing subsequent steps to focus more on the texture differences themselves rather than having to deal with the chaos of basic sampling first.

[0045] Next, at the level of feature extraction and point cloud generation, in view of the inherent differences between CT and CTP modalities in terms of "anatomy vs. hemodynamics", the present invention adopts a multi-scale difference detection method to mine key points that are significant at multiple fuzzy scales. Specifically, by performing a series of Gaussian smoothings on the normalized images and then taking the differences, local extreme points are searched to ensure that the points that appear as significant edges of bones or soft tissues on CT and the points that show significant perfusion differences on CTP can be captured in multi-scale images. This multi-scale detection enables the formation of relatively stable cross-modal correspondence relationships at some vascular bifurcations, ventricular margins, or lesion locations even though the texture mechanisms of CT and CTP are different. Then, these key points are transformed into point clouds, and the coordinate and descriptor information are recorded together. The advantage is that there is no need to consider the synchronization of resolution or gray level again when performing subsequent rigid body transformations and alignments in a three-dimensional coordinate system, thus greatly improving the freedom and accuracy of feature matching. Compared with traditional methods based only on mutual information or only on ordinary corner points, the present invention directly corresponds the "co-occurrence regions" of anatomical-hemodynamic features in three-dimensional space in the form of point clouds, making the subsequent alignment more intuitive and efficient.

[0046] At the level of rigid body transformation and gray level fusion, the present invention first uses the iterative closest point (ICP) method to minimize the geometric distance, thereby quickly correcting the deviations in the shooting positions and angles between CT and CTP; subsequently, in order to ensure the correct correspondence of blood flow distribution in the CT space, a cross-modal gray level similarity term is introduced on the basis of geometric errors. The purpose of this is that only geometric alignment may result in a brightness mismatch in high-perfusion regions, affecting lesion localization; while if only focusing on gray level similarity, the overall spatial alignment may be sacrificed. The fusion objective function proposed by the present invention incorporates both into the same optimization framework, enabling a more perfect superposition of the anatomical structure of the CT image and the blood flow information of the CTP image under the dual conditions of "spatial accuracy" and "gray level consistency". This has significant implications for clinical brain diagnosis: doctors can clearly see the anatomical locations such as bones and brain parenchyma, and at the same time observe the true distribution of cerebral blood flow, which is conducive to discovering the location of cerebral infarction, evaluating vascular patency, and formulating personalized interventional treatment plans.

[0047] Furthermore, in terms of the objective measurement of the registration result, the present invention is also evaluated through two major dimensions: geometric error and gray-level distribution similarity. Geometric error can ensure high-quality overlap of the CT and CTP point clouds in three-dimensional coordinates. Once this error exceeds a certain threshold, it indicates that there are still deficiencies in the initial alignment relying on rigid body transformation or subsequent gray-level fusion. The gray-level distribution similarity (or mutual information, correlation coefficient, etc.) can detect the fusion degree of the images in terms of brightness or texture. Some doctors who mainly observe the blood perfusion area can thus quickly discover whether there are situations where local blood flow is too low or too high and does not match the anatomical position, so as to judge whether there are serious diseases such as vascular stenosis, insufficient blood supply, or local bleeding. Thus, it can be seen that at the evaluation level, the dual evaluation of "position + gray level" undoubtedly improves the registration reliability and can better compare the advantages and disadvantages of different registration schemes in scientific research offline tests.

[0048] On this basis, the beneficial effects of the present invention can be summarized as follows:

[0049] One-time solution to multiple inconsistencies: The present invention simultaneously considers three major factors: resolution difference, gray-level dynamic range difference, and modality texture difference. By steps S1 and S2, the grid and gray level are unified at the initial level, and then multi-scale feature point detection is used to select cross-modal comparable key points. Finally, fine matching is carried out by means of point cloud + rigid body + gray-level fusion, etc., avoiding simple repairs of single points or single links, and the system is more complete and stable.

[0050] Balancing spatial and brightness alignment: Compared with traditional global algorithms based on gray-level similarity (such as mutual information) that are prone to failure when the texture of CTP changes greatly, the present invention first uses point cloud ICP for geometric alignment and then uses gray-level similarity to supplement the microscopic brightness distribution, enabling the registration process to achieve a balance between "macroscopic position alignment" and "local texture fitting". The CT and CTP not only overlap in the skeleton but also achieve higher consistency in the brain parenchyma and even blood flow details, which is of great help to real clinical needs.

[0051] Stronger robustness: The present invention has obvious robustness advantages in details such as multi-scale feature extraction, iterative closest point, and fused gray-level optimization. For example, multi-scale features can avoid feature point mismatches caused by noise and overly fine textures to the greatest extent; ICP can still converge to a reasonable alignment result through iteration when the pose deviation is large; the gray-level fusion term can further correct the matching of the spatial point cloud when the perfusion value changes extremely. These characteristics enable the method to maintain a high registration accuracy under various actual situations (such as patient movement, scanning condition differences, dynamic blood flow fluctuations, etc.).

[0052] The results can be quantitatively evaluated: By means of dual measurement of geometric error and gray-scale distribution similarity, the present invention enables users to clearly know whether it is accurately aligned in position and whether it is well fused in brightness. In a certain case, if the spatial alignment is good but the gray-scale similarity is low, it implies that there is still a certain deviation between the blood flow and the anatomical structure, and fine-tuning can be continued on the gray-scale fusion parameters in S5; conversely, if the geometric error is large, it indicates that the number of ICP iterations can be appropriately increased or key point pairs can be better selected. In this way, pure subjective judgment is effectively reduced, and it is more repeatable and objective.

[0053] Suitable for both clinical applications and scientific research verification: On the one hand, in actual clinical applications such as thrombolysis for stroke, blood supply evaluation of brain tumors, and blood flow analysis of aneurysms, since the images after CT and CTP fusion can intuitively show the corresponding relationship between the anatomy and blood flow of the lesion site, doctors can take timely measures; on the other hand, researchers can also use the present invention for offline tests under different noise levels and different perfusion parameters, and objectively measure the performance of the registration algorithm in combination with dual evaluation indicators.

[0054] Therefore, the present invention has comprehensive beneficial effects such as "eliminating multiple modality differences at one time, taking into account macroscopic position and local texture, providing quantifiable objective evaluation, being highly robust and adapting to real clinical needs", and provides a systematic and efficient technical solution for brain CT and CTP image registration in aspects such as resolution synchronization, gray-scale normalization, cross-modal feature extraction, rigid body transformation, gray-scale fusion, and evaluation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is the schematic diagram of the method of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0056] The differences between CT and CTP images are reflected in the differences in spatial resolution and slice thickness and the differences in gray-scale value distribution.

[0057] Differences in spatial resolution and slice thickness: CT images usually use a larger slice thickness, while CTP images have a smaller slice thickness, resulting in different spatial resolutions of the two images and the inability to directly correspond in spatial position. Differences in gray-scale value distribution: CT images mainly reflect anatomical structures, and their gray-scale value ranges are concentrated and stable; CTP images reflect the dynamic perfusion state of the brain, and their gray-scale values are affected by blood flow parameters and vary greatly in range. The inconsistency of gray-scale scales increases the complexity of image registration.

[0058] The core of multi-modal image registration is how to accurately find the corresponding relationship between the two-modal images, and the main challenges include:

[0059] Inconsistencies in resolution and grayscale: Due to different resolutions and grayscale ranges between CT and CTP images, significant spatial and grayscale errors will occur during direct registration. It is necessary to unify the spatial resolution and grayscale scale to solve the problems of resolution mismatch and inconsistent grayscale information, providing a consistent basis for subsequent feature extraction.

[0060] Differences in modal features: CT images mainly provide anatomical structure information, while CTP images reflect the hemodynamic characteristics of the brain. The differences in their texture characteristics and spatial structures are significant. It is necessary to extract robust multi-modal features to ensure the matching accuracy of feature points in cross-modal images.

[0061] Complex spatial alignment: Since the spatial positions of CT and CTP images during acquisition may be different, the registration requires simultaneous estimation of rotation and translation parameters. It is difficult to achieve high-precision alignment relying solely on spatial information, and it is necessary to optimize by combining the grayscale and texture characteristics of the images.

[0062] Consistency of multi-modal fusion: After completing the spatial alignment, it is necessary to ensure the fusion consistency of the images in terms of grayscale information and texture information to guarantee the accuracy and reliability of the registration results.

[0063] Therefore, the present invention solves the registration problem of brain CT and CTP images caused by differences in resolution, grayscale, and modal characteristics through point cloud technology, realizes high-precision multi-modal image fusion, provides reliable technical support for the diagnosis, treatment, and postoperative evaluation of brain diseases, and has broad clinical application value.

[0064] The present invention discloses a method for registering brain CT and CTP images based on point cloud technology. Through innovative designs in six major steps (S1 - S6) including image preprocessing, feature extraction, three-dimensional point cloud generation, rigid body transformation alignment, cross-modal grayscale fusion, and result evaluation, it overcomes the differences between CT images and CTP images in terms of resolution, grayscale range, and modal texture characteristics, and finally realizes multi-modal fusion with highly consistent anatomical positions and blood flow distributions.

[0065] In step S1, the present invention first solves the problem of inconsistency in spatial resolution and grayscale dynamic range between CT and CTP images. The reason for this is that if the resolution and brightness range are not synchronized, subsequent feature detection and point cloud alignment will fall into the dilemma of coordinate mismatch or incomparable grayscale. The specific technical means include:

[0066] (1) Perform trilinear interpolation on CT images and CTP images respectively to make them consistent in resolution and slice thickness; the interpolation formula is , where, (x i , y j , z krepresents the original grid coordinates adjacent to the interpolation point (x, y, z). In this embodiment, the number of adjacent points is eight points around the interpolation point, and w ijk is the weight for weighted summation based on distance, ensuring the smooth continuity of the image data;

[0067] (2)Perform linear normalization on the interpolated image

[0068]

[0069] Among them, M(x, y, z) is based on the window width I WW and the window level I WL is a mask for masking according to the CT value and CTP value. The unmasked part is normalized, and the brightness is compressed into the [0, 1] interval. The meaning of doing this is that through these two steps, CT and CTP can be aligned subsequently on the "uniform resolution + uniform gray level", no longer restricted by the original sampling and brightness differences.

[0070] Subsequently, in step S2, considering that CT emphasizes more on the anatomical edges of the brain and CTP focuses more on hemodynamic changes, the present invention adopts a multi-scale feature point detection method to extract key points that are less sensitive to cross-modal differences. The core approach is that on the one hand, the normalized image Inorm(x, y, z) is generated at different Gaussian smoothing scales σ .

[0071] On the other hand, the scale difference formula is used to find local extreme points (x, y, z) in the difference image. The reason for doing this is that only the points that are significant under multi-scale blurring are more likely to correspond to the same physical position in CT and CTP (such as the edge of brain tissue or the region of blood flow mutation), thus laying a reliable feature point foundation for subsequent cross-modal matching.

[0072] After obtaining the feature points, in step S3, these key points are converted into three-dimensional point cloud data to record spatial coordinates and texture information simultaneously. Specifically, the present invention combines the coordinates (x, y, z), the normalized gray level I norm (x, y, z) and the gradient descriptor D (such as the histogram of gradient directions) and represents them as , so that both the position and the modal gray level characteristics can be saved in a unified data structure. The reason for converting both CT and CTP into point clouds is that it is more natural to perform feature point correspondence and geometric alignment in the three-dimensional coordinate space, without having to worry about the resolution alignment of the image matrix or the gray level difference additionally; in addition, the point cloud form can also significantly reduce the useless background information, focus on the truly matchable feature points, and improve the calculation efficiency.

[0073] However, during the multimodal registration of brain CT and CTP images, due to differences in shooting angles and positions, there are usually significant angular deviations and position errors between the two images in the spatial coordinate system. To solve this problem, step S4 uses a rigid body transformation model to preliminarily align the three-dimensional point cloud data, and eliminates such errors by adjusting the rotation matrix R and the translation vector t. The core formula of rigid body transformation is P' = R·P + t, where P represents the spatial coordinate vector of a certain point in the point cloud, R is a 3×3 orthogonal matrix used to adjust the direction of the point cloud, and t represents the translation vector used to correct the position of the point cloud. To make the preliminarily aligned point cloud as close as possible to the target point cloud, the rigid body transformation model optimizes through the distance error objective function between the matching points This process first finds the matching point pairs p i and q i between the point clouds through nearest neighbor matching, then uses singular value decomposition (SVD) to optimize the rotation matrix R, and optimizes the translation vector t according to the centroid difference, and finally obtains the preliminarily aligned point cloud data.

[0074] Although the preliminary alignment of the two point clouds in space is achieved through rigid body transformation, since CT images mainly reflect anatomical structures, while CTP images reflect the brain perfusion state, it is difficult to fully achieve the consistency of multimodal registration only by using spatial information. Therefore, on the basis of the preliminary alignment, step S5 combines the similarity of the point cloud gray information, and further optimizes the alignment result by fusing the spatial and gray features of the multimodal images. The optimization process is carried out through the objective function where w i represents the gray similarity weight of the matching points, and the weight calculation formula is , and h(p i ) and h(q i ) are the gray histograms of the matching points respectively. This step makes the alignment result of the point cloud not only more accurate in space, but also has higher consistency in texture and gray features through the dynamic weight adjustment of the gray information similarity.

[0075] After the accurate alignment of the point cloud data is completed, it is necessary to verify the registration result to ensure that it meets the clinical requirements of multimodal image fusion. Step S6 evaluates the registration result from two aspects of spatial alignment accuracy and gray consistency through geometric error measurement methods and distribution similarity measurement methods respectively. The geometric error measurement uses the mean square error (MSE) as the evaluation index, and the calculation formula is where p i and q iare the matching points in the registered point cloud, and MSE is used to measure the alignment error of the two point clouds in spatial position. The gray-scale consistency verification calculates the distribution similarity of the gray-scale histograms through the Jensen-Shannon divergence (JSD), and the formula is is the Kullback-Leibler divergence, indicating the difference between the two gray-scale distributions. Through the above geometric and gray-scale evaluation metrics, when the MSE is less than the set threshold and the JSD meets the consistency standard, the registration result is considered successful.

[0076] Through the above steps, the rigid body transformation model corrects the spatial error in the preliminary alignment stage, and the gray-scale similarity optimization integrates the spatial and texture information of the multimodal images in the precise registration stage. Finally, through geometric and gray-scale verification, the consistency of the registration result in spatial position and texture characteristics is ensured, providing accurate and reliable technical support for the multimodal fusion of CT and CTP images. Embodiment

[0077] Embodiment 1 (Clinical diagnosis of stroke): In the radiology department of the hospital, after uniformly processing the CT and CTP images of stroke patients with S1-S5 in terms of resolution and gray scale, good registration is obtained through rigid body transformation and gray scale fusion. At S6, if the geometric error Egeo < 1 and the gray scale correlation coefficient Ssimilarity > 0.90, the physician can directly use the fused image to determine the location of the lesion and the blood flow defect area, and the treatment plan is more accurate and efficient.

[0078] Embodiment 2 (Evaluation of scientific research offline algorithm): Researchers obtain simulated brain CT and CTP from a public database (such as BrainWeb) and add noise or rotation deviation. After uniformly processing, extracting key points and generating point clouds with S1-S2, alignment and fusion are performed through S3-S5. S6 is quantified by geometric error and similarity, and the convergence performance of the method of the present invention under different noise levels or initial offsets is statistically analyzed. Compared with the traditional mutual information (MI) or the ICP-only scheme, it is often found that the method of the present invention still shows better robustness under high-noise or large-offset conditions.

[0079] Embodiment 3 (Cloud-based telemedicine): Deploy the above full process on the cloud platform to perform parallel registration on the batch CT / CTP uploaded by multiple hospitals. If the geometric or gray scale fusion effect is detected to meet the standard at S6, the final fused image can be returned to each hospital for sharing, significantly improving the multi-hospital collaborative diagnosis efficiency and achieving the goal of cross-regional expert joint diagnosis of brain lesions.

[0080] As can be seen from the above six major steps and multiple application scenarios, the present invention particularly highlights the significance of eliminating resolution and gray-scale differences (S1), capturing key points of cross-modal adaptation (S2-S3), performing pose alignment first and then gray-scale fusion (S4-S5), and dual metrics (S6); and respectively presents how to achieve a consistency basis through interpolation and normalization, how to overcome modal texture differences using multi-scale differences, how to efficiently perform rigid body transformation in 3D point clouds and finally fuse intensity distributions, etc. Without violating the above ideas and technical effects, equivalent replacement or local improvement of each formula or implementation detail still falls within the protection scope of the present invention.

[0081] The above description is only an embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, is similarly included in the patent protection scope of the present invention.

Claims

1. A brain CT and CTP image registration method based on point cloud technology, characterized in that: The following steps are involved: S1, performing standardization processing on the spatial resolution of the CT image and the CTP image respectively, and normalizing the grayscale values ​​of the two, so as to adjust the CT image and the CTP image to a uniform spatial resolution and grayscale range; S2, extracting feature points with rotation invariance, scale invariance and grayscale change invariance from CT images and CTP images after spatial resolution standardization and grayscale value normalization; S3, converting the feature points into three-dimensional point cloud data including spatial coordinates and corresponding grayscale texture information; S4, performing preliminary alignment on the three-dimensional point cloud data based on a rigid body transformation model, and estimating rotation and translation parameters of the image; S5. Based on the preliminary alignment, the similarity of the grayscale information of the point cloud is combined to fuse the spatial features and grayscale features of the multimodal images to further optimize the point cloud alignment results. Among them, the comprehensive objective function used introduces cross-modal grayscale consistency on the basis of geometric position alignment, and by considering the dual factors of space and texture intensity, CT and CTP are simultaneously fused in position and brightness distribution. S6. The geometric error measurement method is used to evaluate the spatial alignment accuracy of point cloud registration, and the distribution similarity measurement method is used to verify the fusion effect of image grayscale information.

2. The method for brain CT and CTP image registration based on point cloud technology according to claim 1, characterized in that: In the spatial resolution standardization process of step S1, the CT image I CT (x,y,z) and CTP image I CTP (x, y, z) are subjected to trilinear interpolation processing respectively to make the two consistent in resolution and layer thickness. The grayscale value of the interpolation point (x, y, z) is calculated by the following formula: Select the neighboring points of the interpolation point according to Calculate the difference weight of the interpolation point based on Calculate the gray value of the interpolation point; where W jik Represents the neighboring point I(x i ,y j ,z k )’s difference weight.

3. The brain CT and CTP image registration method based on point cloud technology according to claim 2, characterized in that: Normalizing the grayscale values ​​of the CT image and the CTP image includes: The grayscale value of the interpolated image is extracted and the mask is normalized according to the following formula: ； Among them, M(x,y,z) represents the window width I WW and window level I WL The mask for masking the CT value and CTP value is normalized only for the unmasked part, I HU (x, y, z) represents the gray value of the midpoint of the CT image. I(x,y,z) is the grayscale value of the voxel in the CT image.

4. The method for brain CT and CTP image registration based on point cloud technology according to claim 3, characterized in that: In step S2, the extraction of the feature points includes: For the normalized image I norm (x, y, z) performs Gaussian blur convolution to construct the scale space; ; Where G(x,y,z,σ) is a Gaussian kernel with a scale of σ, and L(x,y,z,σ) represents the blurred image of the image after Gaussian kernel filtering at a certain scale σ; By using the scale difference formula Find the local extreme point and get the coordinates of the feature point (x, y, z).

5. The brain CT and CTP image registration method based on point cloud technology according to claim 1, characterized in that: Converting feature points into 3D point cloud data includes: The feature point coordinates (x, y, z) and their corresponding normalized grayscale values ​​and local descriptor information are combined and expressed as: , where I normI is the normalized gray value of the feature point, and D is the local texture.

6. The brain CT and CTP image registration method based on point cloud technology according to claim 1, characterized in that: When performing preliminary alignment of 3D point cloud data, the rigid body transformation model is used , where P represents the spatial coordinate vector of a point in the point cloud, R is the rotation matrix, and t is the translation vector.

7. The brain CT and CTP image registration method based on point cloud technology according to claim 6, characterized in that: R and t are optimized by iterative closest point algorithm. The objective function during optimization is: ; Among them, p i With q i is the i-th pair of corresponding points selected by the closest point strategy in the CT point cloud and the CTP point cloud, and N is the number of corresponding point pairs.

8. The method for brain CT and CTP image registration based on point cloud technology according to claim 5 or 7, characterized in that: In step S5, during the point cloud alignment process, the following comprehensive objective function is adopted: ; Among them, E ICP is the geometric distance error, α and β are different weight coefficients, Represents grayscale similarity, I CT with I CTP They come from the information of CT images and CTP images carried by the point cloud respectively.

Images