System for predicting risk of cerebral vasospasm after intracerebral hemorrhage based on magnetic resonance swi

Abstract

The application relates to the technical field of physiological data processing, in particular to a brain hemorrhage post-cerebral vasospasm risk prediction system based on magnetic resonance SWI. The system extracts a hematoma lesion area by using a magnetic resonance amplitude image and carries out theoretical calculation, linear regression is carried out on the ideal distribution map as an independent variable and the magnetic resonance phase image, and a microvessel phase image retaining local venous characteristics is extracted. By constructing a conduction velocity field and carrying out conduction simulation, the local characteristics are converted into a venous probability distribution map reflecting connectivity, and only the structure with a continuous path leading to a preset venous convergence area is counted. Finally, based on the risk prediction result of the venous visualization volume feedback, the cerebral tissue hypoxia load caused by vasospasm can be specifically quantified, the interference of artifacts on clinical evaluation is avoided, and the accuracy of cerebral vasospasm risk prediction is improved.

Description

Technical Field

[0001] This invention relates to the field of physiological data processing technology, specifically to a system for predicting the risk of cerebral vasospasm after cerebral hemorrhage based on magnetic resonance imaging (SWI). Background Technology

[0002] Cerebral vasospasm following intracerebral hemorrhage is a key pathological process leading to delayed ischemic neurological deterioration. Abnormal enhancement of deep medullary veins is an early imaging marker reflecting elevated brain tissue oxygen uptake fraction (OEF). On susceptibility-weighted imaging (SWI), hypoxic veins appear as significant low-signal tubular structures, and basic information can be determined using simple image processing methods such as threshold segmentation.

[0003] However, in SWI images of patients with cerebral hemorrhage, the strong magnetic susceptibility effect produced by hemosiderin within the hematoma lesion induces significant dipole magnetic field interference in the surrounding brain tissue. This interference forms radial or ring-shaped phase stripe artifacts around the hematoma. These artifacts are highly similar to real deep medullary veins in morphological characteristics and signal intensity, making it difficult for existing image processing algorithms to distinguish the interfered artifacts from real venous signals. This makes it difficult to accurately quantify the venous metabolic status of the blood supply area surrounding the hematoma using SWI images, thus hindering the timely and objective assessment of the risk of vasospasm in the relevant arteries. Summary of the Invention

[0004] To address the technical problem in existing SWI image processing technologies that fail to distinguish between interfering artifacts and genuine venous signals, resulting in indistinguishability, this invention aims to provide a SWI-based system for predicting the risk of cerebral vasospasm after intracerebral hemorrhage. The specific technical solution adopted is as follows: This invention proposes a risk prediction system for cerebral vasospasm after intracerebral hemorrhage based on magnetic resonance imaging (SWI), the system comprising: The magnetic resonance imaging acquisition module is used to acquire magnetic resonance amplitude images and magnetic resonance phase images of the patient's brain; the magnetic resonance amplitude image includes the hematoma lesion area; The microvascular phase image extraction module is used to perform theoretical calculations based on the location of the hematoma lesion area to obtain a unit hematoma magnetic field distribution map; the unit hematoma magnetic field distribution map is used as the independent variable and the magnetic resonance phase image is used as the dependent variable to perform linear regression, and the residual of each voxel point is used as the microvascular phase image. The vein probability distribution map acquisition module is used to extract the structure tensor from the microvascular phase image in a multi-scale Gaussian space using the Hessian matrix algorithm. It determines the tubular probability of each voxel point based on the eigenvalue distribution of the structure tensor at each scale and filters out the vessel axis vector from the feature vector. It obtains the orientation difference weight of each voxel position by utilizing the directional difference between the gradient vector of each voxel position on the unit hematoma magnetic field distribution map and the vessel axis vector. It constructs the conduction velocity field for all voxel positions based on the orientation difference weight and the tubular probability. Based on the conduction velocity field and a preset vein convergence region, it performs conduction simulation to obtain the vein probability distribution map. The risk prediction module determines the venous contrast volume based on the venous probability distribution map and feeds back the risk prediction result based on the size of the venous contrast volume.

[0005] Furthermore, the method for extracting the hematoma lesion area includes: The magnetic resonance amplitude image is segmented using the Otsu threshold based on the signal value. The voxel with the smallest signal value is selected as the voxel of the hematoma lesion. The region formed by the voxel of the hematoma lesion is subjected to a three-dimensional morphological closing operation to obtain the hematoma lesion region.

[0006] Furthermore, the calculation method for the unit hematoma magnetic field distribution map includes: In the magnetic resonance amplitude image, the magnetic susceptibility of the voxel point in the hematoma lesion area is set to 1, and the magnetic susceptibility of other voxel points is set to 0. The magnetic field distribution map of the unit hematoma is calculated in the frequency domain based on the dipole field theory.

[0007] Furthermore, the method for extracting the structure tensor includes: For each scale, the corresponding scale is used as the standard deviation in the Gaussian kernel function. The second derivative of the Gaussian kernel function is convolved with the normalized microvascular phase image to obtain the Hessian matrix as the tubular structure tensor at the corresponding scale.

[0008] Furthermore, the method for obtaining the tubular probability includes: For each tubular structure tensor, three eigenvalues ​​are obtained: a first eigenvalue, a second eigenvalue, and a third eigenvalue. The first eigenvalue has the smallest absolute value, and the third eigenvalue has the largest absolute value. The absolute value of the second eigenvalue is used as the numerator, and the absolute value of the third eigenvalue is used as the denominator to obtain the tubular roundness. The absolute value of the first eigenvalue is used as the numerator, and the absolute value of the second eigenvalue is used as the denominator to obtain the irregular structure probability. The irregular structure probability is negatively correlated and then multiplied by the tubular roundness to obtain the initial tubular probability. The maximum initial tubular probability at all scales is used as the tubular probability.

[0009] Furthermore, the method for filtering the vascular axial vector includes: In the tubular structure tensor at the scale corresponding to the tubular probability, the direction of the eigenvector corresponding to the first eigenvalue is selected as the vascular axis vector.

[0010] Furthermore, the method for obtaining the divergence weights includes: For each voxel point, the cosine similarity between the gradient vector and the blood vessel axis vector is obtained. The magnetic field strength in the unit hematoma magnetic field distribution map is normalized and used as the weight of the cosine similarity to obtain the weighted cosine similarity. The negative correlation mapping of the weighted cosine similarity is used to obtain the orientation difference weight.

[0011] Furthermore, the method for obtaining the conduction velocity field includes: For each voxel, the corresponding tubular probability and the orientation difference weight are multiplied together and then added to the preset minimum conduction velocity to obtain the final conduction velocity of each voxel, thus forming the conduction velocity field.

[0012] Furthermore, the method for obtaining the vein probability distribution map includes: Using the preset venous convergence area as the source point, an equation is constructed based on the conduction velocity field. The equation is solved using the fast travel method. The minimum geodesic distance from the source point to each voxel point is calculated. The minimum geodesic distance of each voxel point constitutes the venous return impedance diagram. The probability distribution map of veins is obtained by negatively correlated mapping and normalization of each pixel value in the venous return impedance map.

[0013] Furthermore, the method for obtaining the venous contrast volume includes: In the vein probability distribution map, voxel points with vein probabilities greater than a preset probability threshold are used as vein imaging voxel points, and the volume of the three-dimensional region formed by the vein imaging voxel points is used as the vein imaging volume.

[0014] The present invention has the following beneficial effects: This invention addresses the problem that the strong magnetic field dipole effect generated by cerebral hematoma can induce radial artifacts in SWI phase images that resemble deep vein morphology, making it difficult to accurately quantify the metabolic state of brain tissue surrounding the hematoma. This system utilizes magnetic resonance amplitude images to extract the hematoma lesion region and performs theoretical calculations to obtain a noise-free magnetic field distribution map of a unit hematoma. By using this ideal distribution map as an independent variable and performing linear regression with the magnetic resonance phase image, large-scale background magnetic fields can be removed, and microvascular phase images retaining local venous features can be extracted. Furthermore, this system uses the Hessian matrix algorithm to extract the structural tensor from the microvascular phase image and combines it with the gradient vector of the unit hematoma magnetic field distribution map to calculate the directional difference between the vessel axis vector and the magnetic field gradient vector. Using this difference to construct a direction difference weight, artifact signals conforming to the magnetic field gradient distribution can be effectively suppressed based on physical field characteristics, while retaining true venous signals following anatomical paths. Based on this, by constructing a conduction velocity field and performing conduction simulation, local features are transformed into a venous probability distribution map reflecting connectivity, ensuring that only structures with continuous paths leading to a predetermined venous convergence area are statistically analyzed. Ultimately, the risk prediction results based on the venous contrast volume feedback can specifically quantify the hypoxic load of brain tissue caused by vasospasm, avoid the interference of artifacts on clinical assessment, and improve the accuracy of cerebral vasospasm risk prediction. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 The diagram below shows a structural block diagram of a cerebral vasospasm risk prediction system based on magnetic resonance imaging (SWI) after cerebral hemorrhage, as provided in one embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a magnetic resonance imaging (SWI)-based system for predicting the risk of cerebral vasospasm after intracerebral hemorrhage, as proposed by the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] The following description, in conjunction with the accompanying drawings, details the specific scheme of the cerebral vasospasm risk prediction system based on magnetic resonance imaging (SWI) for cerebral hemorrhage provided by this invention.

[0020] Please see Figure 1 The diagram illustrates a structural block diagram of a cerebral vasospasm risk prediction system based on magnetic resonance imaging (SWI) after cerebral hemorrhage, according to an embodiment of the present invention. The system includes: a magnetic resonance image acquisition module 101, a microvascular phase image extraction module 102, a vein probability distribution map acquisition module 103, and a risk prediction module 104.

[0021] It should be noted that this system, as a software module, can be directly integrated into existing MRI image post-processing workstations or PACS systems. The system input directly reads amplitude and phase images in standard DICOM format without requiring changes to the underlying scanning hardware. In practice, the core computing module is encapsulated as a post-processing plugin or a background microservice, directly calling the workstation's existing computing resources to perform physical field calculations and tensor extraction algorithms. The output vein probability distribution map is generated as a pseudo-color layer and overlaid on the original image. The risk prediction results are converted into a standard structured report and pushed to the front-end image reading interface. The overall architecture does not change the doctor's original image reading habits and has extremely high engineering feasibility.

[0022] The magnetic resonance imaging acquisition module 101 is used to call up the magnetic resonance amplitude image and magnetic resonance phase image of the patient's brain from the existing magnetic resonance imaging system. The magnetic resonance amplitude image can provide anatomical structure information, and the T2 signal can reflect the relaxation characteristics of the tissue, which is represented as a standard grayscale image. It can be used for spatial localization and lesion identification. In this embodiment of the invention, the low signal value feature of the hematoma region on the magnetic resonance amplitude image can be used to accurately extract the hematoma region. The magnetic resonance phase image provides physical magnetic field information, which can reflect the degree of distortion of the local voxel to the main magnetic field. It is directly related to the magnetization of the material. This phase image can characterize the metabolic state and vascular visualization, because deoxyhemoglobin in the vein and hemosiderin in the hematoma are both magnetic and will generate characteristic signals in the phase image.

[0023] Because magnetic resonance amplitude imaging can characterize the location of hematoma lesions based on signal value magnitude, but cannot assess the degree of vascular hypoxia; while magnetic resonance phase imaging can reflect the degree of vascular hypoxia based on magnetic field information, it is affected by the large magnetic field generated by the hematoma lesion area, resulting in significant artifacts that obscure useful features. Therefore, this embodiment of the invention extracts the hematoma lesion area from the magnetic resonance amplitude image, first determining the location of the hematoma interference source in the magnetic resonance amplitude image, and then calculating the artifact distribution caused by the hematoma lesion area in the phase image through physical calculations in subsequent modules. Furthermore, true veins and false artifacts in the magnetic resonance phase image are highly similar in brightness, and the difference between them in the vector direction must be calculated by combining the location information of the hematoma lesion area provided by the magnetic resonance amplitude image, thereby achieving effective identification of the vein area.

[0024] It should be noted that both the magnetic resonance amplitude image and the magnetic resonance phase image originate from a single three-dimensional susceptibility-weighted imaging (SWI) scan performed on the same patient at the same time. Both are defined in the exact same discrete three-dimensional Cartesian coordinate space, meaning the voxel positions are identical, with no spatial registration error or temporal phase error. The subsequent descriptions of voxels in this embodiment refer to all voxel points, and the processing method for all voxel points is the same. Therefore, only one voxel point is used as an example, and the features obtained in the subsequent analysis also include the spatial attributes of the voxel point in the three-dimensional Cartesian coordinate space.

[0025] Preferably, in some implementations of the present invention, the method for extracting the hematoma lesion area includes: Because the hematoma lesion region exhibits extremely low signal values ​​on magnetic resonance imaging (MRI), the MRI image is segmented using the Otsu threshold based on the signal value. The voxel with the smallest signal value among the two types of voxels is selected as the hematoma lesion voxel. Considering the potential signal inhomogeneity within the hematoma due to tissue texture or noise, directly using the region formed by the hematoma lesion voxels as the target region could lead to voids and other issues. To obtain a topologically closed geometric entity, a three-dimensional morphological closing operation is performed on the region formed by the hematoma lesion voxels to obtain the hematoma lesion region. The three-dimensional morphological closing operation can use a sphere with a radius of 3mm as the structuring element, filling internal gaps and smoothing boundaries through an expansion followed by erosion operation. The specific algorithm is a well-known technique and will not be elaborated here. The final hematoma lesion region marks the physical spatial extent of the hematoma interference source.

[0026] It should be noted that in other implementations of this invention, grayscale histograms of magnetic resonance amplitude images can also be obtained. Since hematoma and normal brain tissue exhibit a clear bimodal distribution in signal intensity, applying the Otsu thresholding algorithm to the grayscale histogram can yield a more effective global segmentation threshold. The specific algorithm is a well-known technique to those skilled in the art and will not be elaborated upon here.

[0027] Because paramagnetic substances (such as hemosiderin) accumulated within cerebral hemorrhage lesions produce a strong dipole effect, this large-scale background magnetic field is superimposed on local tissue signals, making it impossible to directly observe changes in the magnetic susceptibility of microvenous lesions. Therefore, the microvascular phase image extraction module 102 uses physical forward modeling, modeling, and numerical regression to remove the background magnetic field mainly caused by the macroscopic morphology of the hematoma.

[0028] Considering that the magnetic field interference generated by a hematoma is mainly determined by its macroscopic geometry, the influence of the non-uniformity of the microscopic magnetic susceptibility within the hematoma on the external magnetic field distribution rapidly decreases with distance in brain tissue regions within a certain range from the hematoma edge (far-field approximation). Therefore, treating the hematoma as a paramagnetic source with uniform magnetic susceptibility is sufficient to simulate the large-scale background field distribution characteristics it generates in surrounding tissues. The microvascular phase image extraction module 102 first performs theoretical calculations based on the location of the hematoma lesion area to obtain a unit hematoma magnetic field distribution map. This unit hematoma magnetic field distribution map quantifies the relative magnetic field intensity distribution generated by individual spatial pixels under the influence of only the macroscopic geometry of the hematoma. That is, the unit hematoma magnetic field distribution map has the same spatial information as the magnetic resonance amplitude image and the magnetic resonance phase image, and the coordinates are also in one-to-one correspondence.

[0029] Because the unit hematoma magnetic field distribution map is a unit magnetic field template driven by the geometry of the hematoma lesion area, it represents a theoretical magnetic field intensity distribution. However, the actual observed magnetic resonance phase image is a linear superposition of the large-scale hematoma background field and the local microvascular magnetic field. Compared to the theoretical magnetic field intensity distribution, the actual magnetic susceptibility distribution of the magnetic resonance phase image is an unknown scalar. Therefore, the microvascular phase image extraction module can use linear regression to model and calculate the regression, using the unit hematoma magnetic field distribution map as the independent variable and the magnetic resonance phase image as the dependent variable. The regression slope obtained from the linear regression can reflect the scaling factor of the actual magnetic susceptibility of the hematoma, and the regression result can characterize the system phase intercept. Therefore, the regression result can be regarded as the true physical magnetic susceptibility information of the hematoma. Furthermore, the residual of each voxel point is calculated, containing multiple dimensions and chaotic actual information. Subtracting the true physical magnetic susceptibility information of the hematoma yields the microvascular phase image. In microvascular phase imaging, the low-frequency background field dominated by hematoma geometry is removed, while the local negative phase signal caused by deep medullary veins and some high-frequency stripe artifacts remaining due to the strong gradient effect at the hematoma edge are preserved. Subsequent modules require further identification of the different distribution trends between deep medullary veins and artifacts.

[0030] Preferably, in this embodiment of the invention, considering that the unit hematoma magnetic field distribution map needs to be included as a unit magnetic field strength in the linear regression process in subsequent steps, the calculation method of the unit hematoma magnetic field distribution map includes: In the magnetic resonance amplitude image, the magnetic susceptibility of the voxel point in the hematoma lesion area is set to 1, and the magnetic susceptibility of other voxel points is set to 0. The magnetic field distribution map of the unit hematoma is calculated in the frequency domain based on the dipole field theory.

[0031] It should be noted that the method of calculating the magnetic field distribution map of a unit hematoma in the frequency domain based on dipole field theory is a well-known technique in the art, and the embodiments of the present invention are only briefly described here: First, define the expression for the unit dipole kernel function in the frequency domain as follows: ,in It is a frequency space coordinate vector. This represents the component along the direction of the main magnetic field; Coordinates in frequency space The unit dipole kernel function.

[0032] Then, convolution operations are performed using the Fast Fourier Transform (FFT) and its inverse (IFFT): in Let x be the magnetic field strength of voxel point x in the magnetic field distribution map of a unit hematoma; This serves as the identifier for the hematoma lesion region at voxel point x. If voxel point x is a voxel point representing the hematoma lesion region, then... It is 1 if it is not 1, otherwise it is 0.

[0033] In the specific implementation of this invention, considering that the T2 signal value inside the hematoma lesion area is extremely short, the magnetic resonance phase is usually random noise or severely entangled, belonging to outliers. If the information of the hematoma lesion area itself is used in the linear regression process, the obtained regression coefficient will deviate significantly from the true physical value. Since the background magnetic field generated by the hematoma diffuses outward to the whole brain, the aroma signal quality of the brain tissue around the hematoma is high and strictly follows the dipole field distribution diffused from the hematoma lesion area. Therefore, in the linear regression process, this invention only uses the area outside the hematoma lesion area, taking the area outside the hematoma lesion area as the brain parenchyma area. The image value of the brain parenchyma area in the magnetic resonance phase image is statistically analyzed and used as the vertical axis; the magnetic field strength of the brain parenchyma area in the unit hematoma magnetic field distribution map is statistically analyzed and used as the horizontal axis, and then linear regression is performed. The regression model is as follows: ,in voxel points in the brain parenchyma region Image values ​​in magnetic resonance phase images voxel points in the brain parenchyma region Magnetic field strength in the magnetic field distribution diagram of a unit hematoma The regression slope, The regression intercept is used. To reduce the interference of small veins (as outliers) on the background fitting, the system uses the minimum absolute deviation (L1 norm) regression algorithm to solve for the coefficients. The obtained regression slope and regression intercept are then applied globally to obtain the global linear fitting result. For any voxel point x, the magnetic field strength in the magnetic resonance phase image is... minus The residual for each voxel can then be obtained.

[0034] It should be noted that magnetic resonance phase images need to be unwound before linear regression. This process is a routine preprocessing operation when processing magnetic resonance phase images, and will not be elaborated here.

[0035] In the complex magnetic field environment of cerebral hemorrhage, signal intensity alone cannot distinguish between real deep veins and radial artifacts induced by the strong magnetization of the hematoma. The hemosiderin within the hematoma is strongly paramagnetic, generating significant local magnetic field shifts in the surrounding brain tissue. In space, lines or surfaces connecting points with equal magnetic field strength are called equipotential lines. On SWI images, equipotential lines around the hematoma typically exhibit a ring-like or bilobal distribution. Due to the processing of such strong local magnetic field variations by MRI (e.g., high-pass filtering effects or intravoxel dephase), noticeable signal loss bands or alternating stripe artifacts are generated in space. The extension and diffusion directions of these stripes are geometrically perpendicular to the equipotential lines. Mathematically, the direction perpendicular to the equipotential lines and pointing towards the most drastic numerical change is the gradient direction of the scalar field. Therefore, artifact stripes excited by abrupt changes in the hematoma's magnetization must exhibit a geometric shape extending outwards along the magnetic field gradient direction. In contrast, the anatomical course of real deep medullary veins is a naturally growing vascular tree, unconstrained by this temporary magnetic field gradient. The difference between the two in the direction of spatial extension is the underlying logic for the system to distinguish between true and false signals. Therefore, in this embodiment of the invention, the vein probability distribution map acquisition module 103 needs to extract the vascular axis vector of each voxel point and then compare it with the gradient direction of the magnetic field strength of each voxel point. When the vascular axis vector is nearly parallel to the gradient direction, it is proven to be an artifact.

[0036] Considering that deep medullary veins appear as small-diameter tubular low-signal structures in microvascular phase images and may exhibit discontinuous features under pathological conditions, gradient calculation based on single voxels is easily affected by noise and cannot capture multi-scale features. Different voxel points on a vein may correspond to different sizes. Therefore, the vein probability distribution map acquisition module 103 extracts the structural tensor of each voxel point on the gastric vascular phase image based on high-speed space at multiple spatial scales and combined with the Hessian matrix algorithm.

[0037] The Hessian matrix primarily operates on data using the second derivative. Ordinary gradient operators, i.e., the first derivative, can only detect edge intensity and cannot distinguish whether an edge belongs to a tubular vessel, patchy artifact, or spherical noise. In SWI images, hematoma artifacts and veins are very dark with similar signal intensities, making them indistinguishable based solely on intensity or edge characteristics. The Hessian matrix, however, describes the curvature and shape of image brightness. Eigenvalue decomposition can decompose the local structure into three orthogonal directions, yielding three eigenvalues ​​and three eigenvectors. If the smallest eigenvalue approaches 0, it indicates minimal brightness variation along the vessel axis. If the other two eigenvalues ​​are much greater than 0, it indicates significant brightness curvature along the vessel radial direction. Based on the magnitude of the eigenvalues, the accurate tubular probability of each voxel at the most suitable scale is determined in multi-scale space. Furthermore, the axial vector of the vessel is directly selected from the eigenvectors.

[0038] In one specific implementation of this invention, four scales are selected in the multi-scale Gaussian space: 0.5, 1.0, 1.5, and 2.0, in millimeters. Furthermore, to eliminate the influence of image boundary phase folds or extreme noise on subsequent calculations, the pixel values ​​on the microvascular phase image need to be normalized. The absolute value distribution of the whole-brain voxel signal intensity is statistically analyzed, and the 99th percentile value is selected as the truncation threshold. The pixel values ​​of all voxel points are mapped to values ​​between 0 and 1 using a maximization method. If the pixel value is greater than the truncation threshold, the normalization result is directly set to 1.

[0039] Preferably, in this embodiment of the invention, the method for extracting the structure tensor includes: Because the standard deviation in a Gaussian kernel function corresponds to the scale of the function and controls the coarseness of the observed features in image processing, for each scale, the corresponding scale is used as the standard deviation in the Gaussian kernel function. The second derivative of the Gaussian kernel function is then convolved with the normalized microvascular phase image to obtain the Hessian matrix, which serves as the tubular structure tensor at the corresponding scale. It should be noted that the specific Hessian matrix algorithm and the structure of the Gaussian kernel function are techniques well-known to those skilled in the art and will not be elaborated upon or limited here.

[0040] Preferably, in this embodiment of the invention, the method for obtaining the tubular probability includes: For each voxel, a tubular structure tensor can be obtained at each scale.

[0041] For each tubular structure tensor, three eigenvalues ​​are obtained: a first eigenvalue, a second eigenvalue, and a third eigenvalue. The first eigenvalue has the smallest absolute value, and the third eigenvalue has the largest absolute value. The absolute value of the second eigenvalue is used as the numerator, and the absolute value of the third eigenvalue as the denominator to obtain the tubular roundness. The absolute value of the first eigenvalue is used as the numerator, and the absolute value of the second eigenvalue as the denominator to obtain the irregular structure probability. The irregular structure probability is negatively correlated and then multiplied by the tubular roundness to obtain the initial tubular probability. The closer the absolute values ​​of the second and third eigenvalues ​​are, the more uniform the tubular cross-section, i.e., the greater the tubular roundness, and the more likely the corresponding voxel point is a tubular vein voxel point. The greater the difference between the absolute values ​​of the first and second eigenvalues, the smaller the absolute value of the first eigenvalue, and the lower the irregular structure probability. This indicates that the signal value change is more consistent with the axial variation, meaning the voxel point is more likely to be a tubular vein, and the corresponding tubular probability is greater.

[0042] As a concrete example, the initial tubular probability is expressed by the formula: ;in Let x be the initial tubular probability of voxel point x. The first eigenvalue, The second eigenvalue, The third eigenvalue, It should be noted that the above calculation is only performed when the second and third eigenvalues ​​have the same sign and are not 0; otherwise, the tubular probability is directly set to 0.

[0043] Each voxel has a tubular probability at each scale; therefore, the maximum initial tubular probability across all scales is taken as the tubular probability. The scale corresponding to this tubular probability is also the optimal scale for that voxel.

[0044] Furthermore, in the tubular structure tensor at the scale corresponding to the tubular probability, the direction of the eigenvector corresponding to the first eigenvalue is selected as the vascular axial vector. This is because tubular blood vessels exhibit strong anisotropy in 3D images. Moving along the centerline of the blood vessel results in weak changes in grayscale values, while abrupt changes occur in grayscale values ​​perpendicular to the vessel's cross-section. The three eigenvalues ​​of the Hessian matrix represent the intensity of the second derivative of the image in three mutually orthogonal directions. The direction corresponding to the smallest eigenvalue is the direction with the most gradual signal change. In an ideal tubular structure, the second derivative along the axial direction approaches 0; therefore, the direction of the eigenvector corresponding to the first eigenvalue should be used as the vascular axial vector.

[0045] The vein probability distribution map acquisition module 103 further utilizes the directional difference between the gradient vector of each voxel position on the unit hematoma magnetic field distribution map and the axial vector of the blood vessel to obtain the orientation difference weight of each voxel position. If the axial vector of the blood vessel at a voxel point is highly parallel to the direction of the magnetic field gradient, the structure of that voxel point is likely an artifact induced by the magnetic field, and the orientation difference weight is small; if the two tend to be orthogonal or have no obvious correlation, the structure is more likely to be an anatomically independent real vein, and the orientation difference weight is large.

[0046] It should be noted that, in this embodiment of the invention, considering the possible anisotropy of voxels in physical space (such as differences between layer thickness and intra-layer resolution), the system uses the central difference method combined with the physical dimensions of the voxels to calculate the magnetic field gradient vector of voxel point x, with the coordinate point as... Taking the x-component of a voxel point along the x-axis as an example: in, is the gradient component of the voxel point x along the x-axis; On the magnetic field distribution map of a unit hematoma, the coordinates are... Magnetic field strength; On the magnetic field distribution map of a unit hematoma, the coordinates are... Magnetic field strength; Let x be the physical dimension of the voxel point x along the x-axis in the real coordinate system.

[0047] Similarly, the gradient components along the y-axis and z-axis can be obtained, and then combined to obtain the gradient vector.

[0048] Preferably, in this embodiment of the invention, the method for obtaining the difference weights includes: For each voxel point, the cosine similarity between the gradient vector and the blood vessel axis vector is obtained. That is, the larger the cosine similarity, the more similar the two directions are. It should be noted that, considering the cosine similarity value range is between -1 and 1, the results obtained in the specific implementation of this embodiment are normalized using range standardization for ease of subsequent calculations.

[0049] The magnetic field intensity in the unit hematoma magnetic field distribution map is normalized and used as the weight of the cosine similarity to obtain a weighted cosine similarity. The negative correlation mapping of the weighted cosine similarity yields the orientation difference weight. The weight design ensures that the weight value significantly decays to 0 only in regions with strong magnetic field interference and high coupling between the structural direction and the gradient direction, thereby maximizing the preservation of true venous signals located within the penumbra surrounding the hematoma but with independent orientation. The magnetic field intensity normalization operation can be the same as the normalization method for microvascular phase imaging, and will not be elaborated or limited here.

[0050] As a concrete example, the weighting of differences can be expressed by the formula: in, The difference weights for the orientation of voxel point x; To suppress the intensity coefficient and adjust the value range of the result, the value in this embodiment of the invention is 5. In other embodiments, the value can be selected between 5 and 10, which will not be elaborated or limited here. The cosine similarity of voxel point x is calculated. The nonlinear order is used to amplify the information of cosine similarity. In this embodiment of the invention, the value is 2, but in other embodiments, it can be selected between 2 and 4. The normalized magnetic field strength of voxel point x in the unit hematoma magnetic field distribution map is given. Because the normalization operation is implemented by maximizing the normalization of the truncation threshold, negative values ​​may exist, so an absolute value sign needs to be added. The positive integer 1 in the denominator is to prevent the denominator from being 0, that is, the formula uses the reciprocal form to implement the negative correlation mapping of the weighted result.

[0051] At this point, the vein probability distribution acquisition module 103 has obtained the tubular probability and direction difference weight for each voxel point. A higher tubular probability and a higher direction difference weight indicate that the voxel point is more likely to be a true tubular vein. Since the risk assessment of vasospasm after cerebral hemorrhage should not only focus on the intensity of venous visualization, but also on whether the venous network possesses effective topological connectivity to the deep drainage system (i.e., the ventricular region), broken, isolated low-signal points usually correspond to noise or local artifacts, while continuous low-impedance paths correspond to functional hypoxic veins. Therefore, this embodiment of the invention further employs a path simulation algorithm. First, based on the direction difference weight and the tubular probability, a conduction velocity field is constructed for all voxel locations; that is, the conduction velocity is higher in areas belonging to true vein vessels and lower in locations belonging to noise artifacts. Then, using a preset vein convergence area as the source point, conduction simulation is performed to obtain the vein probability distribution map for each voxel point location. The vein probability distribution map has the same dimensions as the images mentioned above, with each voxel point corresponding to a specific one. The pixel value represents the vein probability corresponding to the voxel point. The higher the vein probability, the more likely the voxel point at that location belongs to a continuous, high-confidence vein path that is not contaminated by hematoma magnetic field artifacts and is connected to the deep drainage system. This path can specifically reflect the hypoxic load in the relevant area.

[0052] Preferably, in this embodiment of the invention, the method for obtaining the conduction velocity field includes: For each voxel, the corresponding tubular probability is multiplied by the direction difference weight, and then added to a preset minimum conduction velocity to obtain the final conduction velocity of each voxel, thus forming the conduction velocity field. The minimum conduction velocity is set to 0.01. It should be noted that the conduction velocity proposed in this embodiment is only for conduction simulation and does not represent the actual conduction velocity of blood. This embodiment aims to simulate effective conduction paths through different conduction velocities.

[0053] Preferably, in this embodiment of the invention, the method for obtaining the vein probability distribution map includes: Using a pre-defined venous convergence area as the source point, an equation is constructed based on the conduction velocity field. The equation is then solved using the fast travel method to calculate the minimum geodesic distance from the source point to each voxel. This minimum geodesic distance constitutes a venous return impedance map. The value at each position in the venous return impedance map represents the minimum geodesic distance; a smaller value indicates a smoother path during conduction simulation, and thus a higher probability of venous return. The fast travel method for solving the equation is a well-known technique. It can be considered as a signal source existing within the ventricles, spreading throughout the brain. The signal propagation speed is constrained by the conduction velocity field. The equation is in the form of: in Let x be the conduction velocity at voxel point x. Let x be the spatial gradient modulus of the venous return impedance. The essence of the equation is to find the shortest time for the wavefront to arrive at each voxel. In this embodiment of the invention, the initial time of all voxels in the preset venous convergence area is set to 0, and the time of other voxels is set to infinity. The equation is solved by wavefront evolution. The solution is the cumulative time of each voxel x. If a voxel x is on a continuous venous path with high conduction velocity, the cumulative time obtained will be low because the path is unobstructed, that is, the venous return impedance will be low. Conversely, if the voxel x is cut off from the ventricle by a magnetic field artifact, or if there is solid brain tissue in between instead of a venous structure, the impedance obtained will be very high or even approach infinity because conduction is impossible.

[0054] The vein probability distribution map is obtained by negatively correlated mapping and normalizing each pixel value in the venous return impedance map. The normalization method in the negative correlation mapping and normalization method can be the same as that used for microvascular phase imaging. After normalization, the negative correlation is achieved by subtracting the normalization result from the positive integer 1. The final numerical value range is between 0 and 1.

[0055] In this embodiment of the invention, the method for setting the preset vein convergence area is as follows: The lateral ventricle and third ventricle regions are extracted using an anatomical atlas, and the set of voxel points belonging to the extracted regions but not to the hematoma lesion region is used as the preset vein convergence area. If voxel points meeting the above conditions cannot be found due to global cerebral hemorrhage, the top 5% of voxels with the smallest absolute value of magnetic field strength in the unit hematoma magnetic field distribution map in the central brain region are automatically selected as the vein convergence area.

[0056] When a specific arterial branch spasms, causing a decrease in perfusion pressure, the brain tissue in its supply area compensates by increasing the fractional oxygen uptake (OEF) to maintain metabolic needs, leading to an increase in the concentration of deoxyhemoglobin in the draining veins. This phenomenon is manifested on imaging as significant visualization of deep veins. A vein probability distribution map can characterize the visualized structure of real veins with anatomical connectivity; therefore, the risk prediction module 104 further determines the venous visualization volume based on the vein probability distribution map and provides feedback on the risk prediction result based on the size of the venous visualization volume. Vasospasm leads to insufficient blood supply, causing cerebral ischemia and hypoxia. To maintain survival, brain cells will take up oxygen from the hematoma, resulting in a significant increase in the concentration of deoxyhemoglobin in the venous blood. Deoxyhemoglobin is a paramagnetic substance and appears as a significant low signal on SWI images. Therefore, the larger the final venous visualization volume, the higher the degree of deoxygenation in the venous blood in more areas, indicating more severe hypoxia compensation in the supply area and a higher risk.

[0057] Preferably, in this embodiment of the invention, voxel points with a vein probability greater than a preset probability threshold in the vein probability distribution map are used as vein imaging voxel points, and the volume of the three-dimensional region formed by the vein imaging voxel points is used as the vein imaging volume.

[0058] The probability threshold can be set by statistically analyzing historical medical data under known conditions of significant hypoxia, determining the average intravenous probability in the historical data using the same algorithm, and then using this probability threshold as the threshold. Specific settings can be made according to the specific implementation scenario, and will not be elaborated or limited here.

[0059] It should be noted that in clinical practice, the standard brain space is divided into multiple non-overlapping independent anatomical sub-regions, such as the bilateral anterior cerebral artery region, the middle cerebral artery region, and the posterior cerebral artery region. This embodiment of the invention utilizes a three-dimensional rigid body registration algorithm based on maximizing mutual information to calculate the geometric transformation matrix that transforms the standard template space to the current patient coordinate space. This matrix is ​​used to resample the anatomical sub-regions, thereby determining each voxel point belonging to each anatomical sub-region in the venous probability distribution map. Then, in each anatomical sub-region, the proportion of venous visualization volume in the total volume of the sub-region is statistically analyzed. Based on this proportion, a risk comparison can be performed. If the volume proportion is less than a preset safety threshold, it is considered low-risk, indicating sufficient arterial blood supply and normal venous metabolism. If the volume proportion is greater than the preset safety threshold but less than a preset danger threshold, it is considered medium-risk, indicating that the artery may have mild spasm or reduced perfusion. If the volume proportion is greater than the preset danger threshold, it is considered high-risk, representing significant arterial spasm leading to widespread metabolic hypoxia compensation in the blood supply area.

[0060] The safety threshold and danger threshold can also be obtained by statistically analyzing historical clinical information, and will not be elaborated or limited here.

[0061] In summary, this invention utilizes magnetic resonance amplitude imaging to extract hematoma lesion regions and performs theoretical calculations. The ideal distribution map is used as the independent variable in a linear regression with magnetic resonance phase imaging to extract microvascular phase images that retain local venous features. The Hessian matrix algorithm is used to extract the structural tensor from the microvascular phase images, and combined with the gradient vector of the unit hematoma magnetic field distribution map, the directional difference between the vessel axis vector and the magnetic field gradient vector is calculated. This difference is used to construct a direction difference weight. By constructing a conduction velocity field and performing conduction simulation, local features are transformed into a venous probability distribution map reflecting connectivity, ensuring that only structures with continuous paths leading to the preset venous convergence area are statistically analyzed. Finally, based on the risk prediction results fed back by venous imaging volume, the hypoxic load on brain tissue caused by vasospasm can be specifically quantified, avoiding artifact interference with clinical assessment and improving the accuracy of cerebral vasospasm risk prediction.

[0062] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0063] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A system for predicting the risk of cerebral vasospasm after intracerebral hemorrhage based on magnetic resonance imaging (SWI), characterized in that, The system includes: The magnetic resonance imaging acquisition module is used to acquire magnetic resonance amplitude images and magnetic resonance phase images of the patient's brain; the magnetic resonance amplitude image includes the hematoma lesion area; The microvascular phase image extraction module is used to perform theoretical calculations based on the location of the hematoma lesion area to obtain a unit hematoma magnetic field distribution map; the unit hematoma magnetic field distribution map is used as the independent variable and the magnetic resonance phase image is used as the dependent variable to perform linear regression, and the residual of each voxel point is used as the microvascular phase image. The vein probability distribution map acquisition module is used to extract the structure tensor from the microvascular phase image in a multi-scale Gaussian space using the Hessian matrix algorithm. It determines the tubular probability of each voxel point based on the eigenvalue distribution of the structure tensor at each scale and filters out the vessel axis vector from the feature vector. It obtains the orientation difference weight of each voxel position by utilizing the directional difference between the gradient vector of each voxel position on the unit hematoma magnetic field distribution map and the vessel axis vector. It constructs the conduction velocity field for all voxel positions based on the orientation difference weight and the tubular probability. Based on the conduction velocity field and a preset vein convergence region, it performs conduction simulation to obtain the vein probability distribution map. The risk prediction module determines the venous contrast volume based on the venous probability distribution map and feeds back the risk prediction result based on the size of the venous contrast volume. The method for obtaining the tubular probability includes: for each tubular structure tensor, obtaining three eigenvalues, namely a first eigenvalue, a second eigenvalue, and a third eigenvalue, where the absolute value of the first eigenvalue is the smallest and the absolute value of the third eigenvalue is the largest; using the absolute value of the second eigenvalue as the numerator and the absolute value of the third eigenvalue as the denominator, obtaining the tubular roundness; using the absolute value of the first eigenvalue as the numerator and the absolute value of the second eigenvalue as the denominator, obtaining the irregular structure probability; mapping the irregular structure probability negatively and multiplying it by the tubular roundness to obtain the initial tubular probability; using the maximum initial tubular probability at all scales as the tubular probability. The method for obtaining the orientation difference weight includes: for each voxel point, obtaining the cosine similarity between the gradient vector and the blood vessel axis vector, normalizing the magnetic field strength in the unit hematoma magnetic field distribution map and using it as the weight of the cosine similarity to obtain the weighted cosine similarity, and negatively mapping the weighted cosine similarity to obtain the orientation difference weight. The method for obtaining the conduction velocity field includes: for each voxel point, multiplying the corresponding tubular probability and the orientation difference weight, and then adding it to the preset minimum conduction velocity to obtain the final conduction velocity of each voxel point, thus forming the conduction velocity field; The method for obtaining the vein probability distribution map includes: taking a preset vein convergence area as the source point, constructing an equation based on the conduction velocity field, solving the equation using the fast travel method, calculating the minimum geodesic distance from the source point to each voxel point, and using the minimum geodesic distance of each voxel point to form a vein return impedance map; performing negative correlation mapping on each pixel value in the vein return impedance map and normalizing it to obtain the vein probability distribution map. In the linear regression process, only the brain parenchyma region outside the hematoma lesion area is used, and the minimum absolute deviation regression algorithm is used to obtain the regression slope and regression intercept. The obtained regression slope and regression intercept are applied to the global space to obtain the residual.

2. The system for predicting the risk of cerebral vasospasm after intracerebral hemorrhage based on magnetic resonance imaging (SWI) according to claim 1, characterized in that, The method for extracting the hematoma lesion area includes: The magnetic resonance amplitude image is segmented using the Otsu threshold based on the signal value. The voxel with the smallest signal value is selected as the voxel of the hematoma lesion. The region formed by the voxel of the hematoma lesion is subjected to a three-dimensional morphological closing operation to obtain the hematoma lesion region.

3. The system for predicting the risk of cerebral vasospasm after intracerebral hemorrhage based on magnetic resonance imaging (SWI) according to claim 1, characterized in that, The calculation method for the unit hematoma magnetic field distribution map includes: In the magnetic resonance amplitude image, the magnetic susceptibility of the voxel point in the hematoma lesion area is set to 1, and the magnetic susceptibility of other voxel points is set to 0. The magnetic field distribution map of the unit hematoma is calculated in the frequency domain based on the dipole field theory.

4. The system for predicting the risk of cerebral vasospasm after intracerebral hemorrhage based on magnetic resonance imaging (SWI) according to claim 1, characterized in that, The method for extracting the structure tensor includes: For each scale, the corresponding scale is used as the standard deviation in the Gaussian kernel function. The second derivative of the Gaussian kernel function is convolved with the normalized microvascular phase image to obtain the Hessian matrix as the tubular structure tensor at the corresponding scale.

5. A system for predicting the risk of cerebral vasospasm after intracerebral hemorrhage based on magnetic resonance imaging (SWI) according to claim 1, characterized in that, The method for filtering the vascular axial vector includes: In the tubular structure tensor at the scale corresponding to the tubular probability, the direction of the eigenvector corresponding to the first eigenvalue is selected as the vascular axis vector.

6. A system for predicting the risk of cerebral vasospasm after intracerebral hemorrhage based on magnetic resonance imaging (SWI) according to claim 1, characterized in that, The method for obtaining the venous contrast volume includes: In the vein probability distribution map, voxel points with vein probabilities greater than a preset probability threshold are used as vein imaging voxel points, and the volume of the three-dimensional region formed by the vein imaging voxel points is used as the vein imaging volume.

Images