Deep learning based segmentation method for brain hemorrhage hematoma region swi image

Abstract

The present application relates to the technical field of image segmentation, and particularly relates to a cerebral hemorrhage hematoma region SWI image segmentation method based on deep learning, which extracts a complex phase gradient in a signal phase diagram and a Laplace source eigenvalue reflecting a magnetization rate source mutation characteristic, acquires first and second order spatial derivative information of a magnetic field distribution, and further couples a direction consistency factor of an anatomical edge and a magnetic field edge to establish a boundary resistance strength value with physical constraints in a voxel grid space; based on the boundary resistance strength value and a hematoma center voxel, a ghost separation feature image is constructed, a cumulative path cost difference from outward diffusion of the hematoma center is used to map a blooming ghost region adhered to the hematoma entity in space to a feature distribution with significant numerical difference, thereby providing a physical discrimination basis for distinguishing the hematoma target from a passive magnetic field ghost, and making the cerebral hemorrhage hematoma region SWI digital image segmentation based on a deep convolutional neural network more accurate.

Description

Technical Field

[0001] This invention relates to the field of image segmentation technology, specifically to a deep learning-based SWI image segmentation method for brain hemorrhage hematoma regions. Background Technology

[0002] Susceptibility-weighted imaging (SWI) digital image processing technology plays a crucial role in medical image analysis. Accurate extraction of the hematoma target region using computer algorithms is fundamental for subsequent data quantification and image feature analysis. However, magnetic field disturbances induced by strongly paramagnetic materials often produce a significant "flowering" effect in the image, causing the low-signal area to extend beyond the actual anatomical structure. Existing techniques typically utilize deep convolutional neural networks to process SWI digital image sequences, segmenting the hematoma region by extracting grayscale distribution and visual texture features from the signal intensity map.

[0003] However, when directly processing SWI image data, deep convolutional neural networks typically lack the ability to perceive the distribution characteristics of physical sources in the local phase field. Since both real hematoma entities and flowering artifacts exhibit low signal intensity and spatial adhesion, relying solely on visual features cannot perceive the non-harmonic Laplace second derivative response of the magnetic susceptibility source, nor can it identify the directional consistency of the magnetic field physical boundary relative to the anatomical edge. This prevents deep convolutional neural networks from establishing a physical topological impedance in the numerical dimension that can distinguish between passive artifacts and active entities, making it difficult to effectively strip and suppress flowering artifact regions. Consequently, existing techniques using deep convolutional neural networks for SWI digital image segmentation of hematoma regions in brain hemorrhage typically have low accuracy. Summary of the Invention

[0004] To address the low accuracy of existing technologies using deep convolutional neural networks for SWI image segmentation of hematoma regions in brain hemorrhage, this invention aims to provide a deep learning-based SWI image segmentation method for brain hemorrhage regions. The specific technical solution adopted is as follows: The first aspect of this invention provides a deep learning-based SWI image segmentation method for hematoma regions in brain hemorrhage, comprising: In a medical database, magnetic susceptibility-weighted imaging scan data of the cranial region are resampled to a voxel grid space to determine the signal intensity map and signal phase map; based on the signal intensity distribution of all voxels in the signal intensity map, the central voxel of the hematoma is determined. Based on the spatial variation of the signal phase of all voxels in the signal phase diagram, the complex phase gradient of each voxel is determined; based on the overall deviation of the complex phase gradient between each voxel and other voxels in its spatial neighborhood, the Laplace source eigenvalue of each voxel is determined. Based on the local variations in complex phase gradient and signal intensity between adjacent voxels, the orientation consistency factor of each voxel is determined; the orientation consistency factor, the complex phase gradient, and the Laplace source eigenvalue are coupled and calculated to determine the boundary drag strength value of each voxel. Based on the distribution of boundary resistance strength values ​​of each voxel and their relative positions with each hematoma center voxel, an artifact separation feature image is constructed; the signal intensity map and the artifact separation feature image are input into a deep learning model to output the hematoma region to be segmented.

[0005] Furthermore, the process of obtaining the hematoma center voxel includes: In the signal intensity diagram, the signal intensities of all voxels are arranged in descending order to determine the signal intensity sequence; the voxels corresponding to the first preset number of signal intensities in the signal intensity sequence are taken as the hematoma center voxels.

[0006] Furthermore, the process of obtaining the complex phase gradient includes: Input the signal phase of each voxel into Euler's formula and output the corresponding complex unit vector; In the three-dimensional space corresponding to the complex unit vectors of all voxels, the complex gradient vector of each voxel is calculated by the central difference method; the complex phase gradient of each voxel is determined according to the magnitude of the complex gradient vector.

[0007] Furthermore, the process of obtaining the Laplace source eigenvalues ​​includes: The complex unit vectors of all voxels within a preset neighborhood of each voxel are summed to determine the neighborhood summation vector of each voxel; the number of other voxels within the preset neighborhood of each voxel is used as the corresponding neighborhood quantity weight; and the corresponding weighted reference vector is determined based on the product between the complex unit vector of each voxel and the neighborhood quantity weight. The Laplacian source eigenvalue of each voxel is determined based on the magnitude of the difference vector between the weighted reference vector and the neighborhood accumulation vector.

[0008] Furthermore, the process of obtaining the directional consistency factor includes: In the signal intensity map, the intensity gradient vector of each voxel is calculated using the central difference method; the reference gradient vector of each voxel is calculated using the central difference method based on the complex phase gradients of all voxels; the cosine similarity between the intensity gradient vector and the reference gradient vector is positively correlated to determine the corresponding directional consistency factor.

[0009] Furthermore, the process of obtaining the boundary resistance strength value includes: The first standard deviation is determined based on the standard deviation of the complex phase gradients of all voxels; the second standard deviation is determined based on the standard deviation of the Laplace source eigenvalues ​​of all voxels. The standard phase gradient is determined based on the ratio between the complex phase gradient of each voxel and the first standard deviation; the standard eigenvalue is determined by positively correlating the ratio between the Laplacian source eigenvalue of each voxel and the second standard deviation; and the fusion boundary eigenvalue is determined based on the sum of the standard phase gradient and the standard eigenvalue. The boundary resistance strength value of each voxel is determined by performing a positive correlation index mapping on the product between the fusion boundary feature value and the direction consistency factor.

[0010] Furthermore, the process of acquiring the artifact separation feature image includes: Using each hematoma center voxel as the source point for wavefront propagation, and combining the boundary resistance strength value of each voxel, the minimum cumulative resistance of each voxel is calculated using a fast travel algorithm. The minimum cumulative resistance is normalized to determine the artifact separation feature value of each voxel; an artifact separation feature image is constructed based on the artifact separation feature values ​​of all voxels.

[0011] Furthermore, the process of inputting the signal intensity map and the artifact separation feature image into the deep learning model and outputting the hematoma region to be segmented includes: The signal intensity map and the artifact separation feature image are input into the trained deep convolutional neural network model, and the hematoma probability of each voxel is output. Voxels with hematoma probabilities greater than a preset probability threshold are designated as hematoma voxels. In all three-dimensional connected regions formed by all hematoma voxels, three-dimensional connected regions with a number of voxels greater than a preset number threshold are marked as hematoma regions to be segmented.

[0012] Furthermore, the preset probability threshold is set to 0.5.

[0013] Furthermore, the preset quantity threshold is set to 15.

[0014] Secondly, the present invention provides a deep learning-based SWI image segmentation system for hematoma regions in brain hemorrhage, the system comprising: The data acquisition and preprocessing module is used to resample the magnetic susceptibility-weighted imaging scan data of the cranial region into a voxel grid space from a medical database to determine the signal intensity map and the signal phase map; and to determine the central voxel of the hematoma based on the signal intensity distribution of all voxels in the signal intensity map. The first determining module is used to determine the complex phase gradient of each voxel based on the spatial change of the signal phase of all voxels in the signal phase diagram; and to determine the Laplace source eigenvalue of each voxel based on the overall deviation of the complex phase gradient between each voxel and other voxels in the spatial neighborhood. The second determining module is used to determine the orientation consistency factor of each voxel based on the local changes in complex phase gradient and signal intensity between adjacent voxels; and to perform coupled calculations on the orientation consistency factor, the complex phase gradient and the Laplace source eigenvalue to determine the boundary drag strength value of each voxel. The image segmentation module is used to construct an artifact separation feature image based on the distribution of boundary resistance strength values ​​of each voxel and their relative positions with each hematoma center voxel; the signal intensity map and the artifact separation feature image are input into a deep learning model to output the hematoma region to be segmented.

[0015] Thirdly, the present invention provides a computer device including a memory and a processor. The memory is used to store computer program code, and the processor is used to call and run the computer program code from the memory to perform the method as described in the first aspect or any embodiment of the first aspect of the present invention.

[0016] Fourthly, the present invention provides a computer program product comprising computer program code, which, when executed, performs the method as described in the first aspect or any embodiment of the first aspect of the present invention.

[0017] Fifthly, the present invention provides a computer-readable storage medium storing computer program code that, when executed, performs the method as described in the first aspect or any embodiment of the first aspect of the present invention.

[0018] The present invention has the following beneficial effects: This invention extracts the complex phase gradient from the signal phase image and the Laplace source feature value reflecting the abrupt change characteristics of the magnetization source to obtain the first and second spatial derivative information of the magnetic field distribution. Furthermore, it couples the directional consistency factor between the anatomical edge and the magnetic field edge, establishing a physically constrained boundary resistance strength value within the voxel grid space. Based on the boundary resistance strength value and the voxel at the hematoma center, an artifact separation feature image is constructed. Utilizing the cumulative path cost difference spreading outward from the hematoma center, the flowering artifact region spatially adhered to the hematoma entity is mapped to a feature distribution with significant numerical differences. This provides a physical basis for deep learning models to distinguish between the actual hematoma target and passive magnetic field artifacts, overcoming the shortcomings of existing technologies that rely solely on visual features and are prone to artifact missegmentation. This results in higher accuracy for SWI digital image segmentation of the brain hematoma region based on deep convolutional neural networks. Attached Figure Description

[0019] 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.

[0020] Figure 1 The flowchart illustrates a deep learning-based SWI image segmentation method for hematoma regions in brain hemorrhage, as provided in one embodiment of the present invention. Figure 2 This is a structural diagram of a deep learning-based SWI image segmentation system for brain hemorrhage hematoma regions provided in one embodiment of the present invention. Figure 3 This is a schematic diagram of a computer device structure provided in one embodiment of the present invention. Detailed Implementation

[0021] 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 deep learning-based SWI image segmentation method for hematoma regions in cerebral hemorrhage proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment, and specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature.

[0022] 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.

[0023] The following description, in conjunction with the accompanying drawings, details a specific scheme for a deep learning-based SWI image segmentation method for hematoma regions in brain hemorrhage provided by this invention.

[0024] This invention provides a deep learning-based SWI image segmentation method for hematoma regions in brain hemorrhage. Please refer to [link to relevant documentation]. Figure 1 The diagram illustrates a flowchart of a deep learning-based SWI image segmentation method for hematoma regions in brain hemorrhage, according to an embodiment of the present invention. The method includes: Step S101: In the medical database, resample the magnetic susceptibility-weighted imaging scan data of the cranial region to a voxel grid space to determine the signal intensity map and the signal phase map; based on the signal intensity distribution of all voxels in the signal intensity map, determine the central voxel of the hematoma.

[0025] Susceptibility-weighted imaging (SWI) scan sequences of the cranial region were obtained from a medical database, and the Digital Imaging and Communication (DICOM) standard file containing patient scan parameters and anatomical spatial coordinate information was read. From these, the original amplitude image volume containing structural texture information and the original phase image volume recording magnetic field distribution information were extracted. In the original amplitude image volume, each voxel corresponds to an initial signal intensity, and in the original phase image volume, each voxel corresponds to an initial phase grayscale value (this grayscale value is dimensionless machine code generated by linearly compressing the actual physical phase angle for easy storage). Since the slice thickness of clinical MRI scans is typically greater than the in-plane imaging resolution, the voxels of the original data are often anisotropic. To eliminate the scale bias caused by spatial dimensionality asymmetry in subsequent 3D gradient calculations, spatial interpolation algorithms such as cubic spline interpolation were used to uniformly resample the original amplitude image volume and the original phase image volume to a preset 3D voxel grid space, ensuring that the resampled data has identical physical resolution in all three spatial dimensions.

[0026] In one specific implementation of this invention, the preset three-dimensional voxel grid space is a three-dimensional Cartesian coordinate system grid with isotropic physical resolution. Specifically: the physical field of view, spatial origin coordinates, and affine direction matrix of the original magnetic susceptibility weighted imaging scan data are extracted; the set isotropic physical resolution (set to 1mm×1mm×1mm in this embodiment, which can be adjusted according to the resolution requirements of the specific implementation environment) is used as the resampling step size, and the length, width, and height dimensions of the grid are recalculated and divided; the original scan data is mapped to the coordinate system grid using a spatial interpolation algorithm, so that the resampled voxels have completely equal physical absolute distances in the three spatial dimensions of the sagittal, coronal, and transverse planes, thereby eliminating the spatial scale distortion caused by the large layer thickness in the original imaging and ensuring that the calculation weights of the subsequent three-dimensional spatial differential operators remain consistent in all directions.

[0027] For the resampled amplitude image volume, to unify the differences in absolute signal values ​​from different scanning devices and highlight the characteristics of hemorrhage foci, negative correlation normalization was performed on the initial signal intensity of all voxels to determine the signal intensity corresponding to each voxel. A signal intensity map for analysis was then constructed based on the signal intensities of all voxels. This processing maps hematoma regions that originally exhibited extremely low initial signal intensities in the amplitude image to a high signal range with values ​​close to the upper limit, while suppressing the background of normal brain tissue with high initial signal intensities to a low value range. This results in higher accuracy for subsequent analysis based on the preprocessed signal intensity map.

[0028] In one specific implementation of this invention, the process of obtaining signal strength is expressed by the formula: ;in, voxels The signal strength; This represents the maximum initial signal intensity of all voxels in the resampled amplitude image volume. When the maximum initial signal intensity of all voxels in the resampled amplitude image volume is equal to 0, it indicates that there is serious data loss, DICOM file corruption, or scanning equipment failure in the current scanning sequence, resulting in the image not containing any valid anatomical tissue signals. In this case, the image needs to be acquired again.

[0029] For the resampled phase image volume, since the initial phase gray values ​​do not have true physical angular meaning and cannot be directly substituted into the complex domain for calculation, it is first necessary to read the scaling and intercept parameters recorded in the scan data (DICOM) header file, and then perform a linear transformation to restore the physical dimensions of the initial phase gray values ​​of each voxel, converting them to a value within the complex domain. The actual physical radian value of the interval.

[0030] In one specific implementation of this invention, the process of obtaining the actual physical radian value is expressed by the following formula: ;in, voxels The actual physical radian value; voxels The initial phase grayscale value; The scaling parameters are recorded in the scan header file; This refers to the intercept parameter recorded in the scan header file.

[0031] Furthermore, due to the inhomogeneity of the main magnetic field, a macroscopic low-frequency background field is superimposed in space. The low-frequency phase component corresponding to this background field often masks the minute high-frequency phase fluctuations caused by local hemorrhages (paramagnetic materials). To accurately remove this interference, a filter size that can adapt to the patient's anatomical structure needs to be determined. Specifically, the voxel dimensions of the resampled phase image roll are extracted in three spatial dimensions. The smallest dimension is selected, and the product of this smallest dimension and a preset scaling factor is determined as the adaptive kernel feature size of the three-dimensional spatial filter. For each voxel, with the voxel as the spatial center, a three-dimensional Gaussian low-pass filter constructed using the obtained adaptive kernel feature size is used to perform weighted smoothing calculations on the real physical radian values ​​in its spatial neighborhood to determine the low-frequency background phase value corresponding to the voxel. Finally, the real physical radian value of the voxel is subtracted from its corresponding low-frequency background phase value to filter out the macroscopic low-frequency trend, determine the signal phase of each voxel after background removal, and construct a signal phase map for subsequent physical feature extraction based on the final signal phases of all voxels. In a specific implementation of this invention, the preset scaling factor is set to The scaling factor can be adjusted according to the expected hematoma size requirements in the specific implementation environment. That is, the larger the expected hematoma size, the larger the value of the preset scaling factor should be to expand the stripping scale of low-frequency background. This will not be elaborated further here.

[0032] After acquiring the signal intensity map and signal phase map in a unified physical space, in order to address the technical problem that existing deep convolutional neural networks cannot effectively remove spatial adhesion artifacts based solely on visual features, this invention cannot simply stop at the isolated identification of local pixels. Instead, it needs to start from the real physical source of the lesion to construct a topological potential that can distinguish between active entities and passive artifacts. Therefore, this embodiment of the invention further accurately determines the hematoma center voxel with a high signal intensity, representing the core of the physical source, based on the signal intensity distribution of all voxels in the signal intensity map. The execution of this step provides an absolutely reliable spatial propagation origin for subsequent construction of boundary resistance based on magnetic field spatial changes and calculation of topological cumulative cost. This ensures that the construction of the artifact separation feature image can strictly revolve around the real hematoma physical entity, thereby laying a key topological spatial benchmark for effectively suppressing false positive predictions and improving the accuracy of cerebral hemorrhage segmentation.

[0033] Step S102: Determine the complex phase gradient of each voxel based on the spatial variation of the signal phase of all voxels in the signal phase diagram; determine the Laplace source eigenvalue of each voxel based on the overall deviation of the complex phase gradient between each voxel and other voxels in the spatial neighborhood.

[0034] While the signal intensity map provides the anatomical morphology of the lesion, it does not reveal the true distribution of magnetic susceptibility sources, preventing deep convolutional neural networks from distinguishing the physical nature behind visually low-signal regions. To establish a physical criterion that can distinguish between active hematomas and passive artifacts, this invention starts with the spatial variation characteristics of the magnetic field. Next, based on the spatial variation of the signal phase of all voxels in the signal phase map, the complex phase gradient of each voxel is determined. This step aims to mine the first-order spatial derivative characteristics of the magnetic resonance phase field, effectively avoiding mathematical singularities caused by phase entanglement by utilizing the continuity of calculations in the complex domain. This provides a precise physical field basis for further combining second-order Laplace source features to construct complete boundary drag strength values, providing deep learning models with physical constraint information that can perceive abrupt changes in magnetic field distribution, thereby effectively improving their ability to identify and segment flowering artifacts.

[0035] After extracting the first-order spatial distribution features of the magnetic field through complex phase gradients, the anisotropy of the dipole magnetic field in magnetic susceptibility-weighted imaging makes it easy to generate numerical zeros at specific angles by simply relying on the first-order gradient. This leads to a break in the extraction of physical boundaries, making it impossible for deep convolutional neural networks to achieve closed-loop identification of artifact boundaries. To overcome this deficiency, this invention utilizes the physical properties of material entities as non-harmonic magnetic fields. Next, based on the overall deviation of the complex phase gradient between each voxel and other voxels in its spatial neighborhood, the Laplace source feature value of each voxel is determined. This feature value, through the calculation of the second-order differential response in the complex domain, highlights the physical source with a true abrupt change in magnetic susceptibility (hematoma entity) while suppressing the passive radiative magnetic field (flowering artifact). This not only fills the gap in the physical boundary of the first-order gradient but also essentially provides the deep learning model with a second-order non-harmonic constraint feature that can separate passive artifacts from active entities, thereby effectively improving the accuracy of hematoma segmentation in cerebral hemorrhage based on deep convolutional neural networks.

[0036] Step S103: Determine the orientation consistency factor for each voxel based on the local variations in complex phase gradient and signal intensity between adjacent voxels; perform coupled calculations on the orientation consistency factor, complex phase gradient, and Laplace source eigenvalues ​​to determine the boundary resistance strength value for each voxel.

[0037] Considering that relying solely on magnetic field signals is susceptible to interference from non-target voxels in complex environmental backgrounds, leading to missegmentation in deep learning models, this invention aims to improve the robustness of physical boundaries in complex environments. This embodiment uses the overlap between intensity change edges and magnetic field change edges as a key evaluation metric. Furthermore, based on the local variations in complex phase gradients and signal intensities between adjacent voxels, a directional consistency factor is determined for each voxel. This factor, by calculating the consistency between the magnetic field gradient direction and the intensity gradient direction, filters and strengthens target boundary voxels exhibiting significant abrupt changes at both the anatomical and magnetic field edges. Simultaneously, it suppresses background noise interference that only responds at a single edge, providing precise constraints on the physical spatial orientation dimension for subsequently constructing boundary resistance strength values ​​with strong noise resistance and high boundary fit.

[0038] After calculating the directional consistency factor, although spatial orientation information reflecting the coordinated changes of the anatomical and magnetic field edges has been obtained, the single directional feature and the discrete complex phase gradient and Laplace source eigenvalues ​​still cannot directly form a continuous physical barrier that can block the topological propagation of flowering artifacts. This results in the deep learning model still lacking an overall perception of the physical boundary between active entities and passive artifacts. To overcome this deficiency, this invention then couples the directional consistency factor, complex phase gradient, and Laplace source eigenvalues ​​to determine the boundary resistance strength value of each voxel. This step integrates the first-order rate of change of the magnetic field, the second-order anharmonic abrupt response, and the anatomical-magnetic field directional coordination to construct a continuous damped distribution field in the voxel grid space that exhibits high impedance at the real physical boundary of the hematoma and low impedance in the passive artifact region. This provides a solid medium foundation for the subsequent establishment of artifact isolation features based on geodesic path integrals, thereby providing accurate physical discrimination criteria for deep networks and significantly improving the accuracy of hematoma region segmentation in cerebral hemorrhage.

[0039] Step S104: Construct artifact separation feature images based on the distribution of boundary resistance strength values ​​of each voxel and their relative positions with each hematoma center voxel; input the signal intensity map and artifact separation feature images into the deep learning model and output the hematoma region to be segmented.

[0040] After calculating the boundary resistance strength values ​​of each voxel, although a high-impedance barrier reflecting the true physical boundary of the hematoma has been established in voxel space, this is only a static local medium property distribution. Deep networks still find it difficult to directly break the spatial adhesion between the hematoma entity and the flowering artifact in terms of visual features. In order to transform this local physical impedance into a feature distribution that the network can perceive globally, this invention next constructs an artifact separation feature image based on the distribution of the boundary resistance strength values ​​of each voxel and its relative position with each central voxel of the hematoma. This step uses the real physical source core (the central voxel of the hematoma) as a spatial reference and utilizes the natural blocking effect of the high-resistance physical boundary on the propagation of spatial position to forcibly map the flowering artifact region, which is originally closely adhered to the hematoma in the anatomical position but belongs to the passive field in terms of physical properties, into a topological isolation feature with a significant drop in the numerical dimension. This processing transforms the high similarity between artifacts and hematoma entities in conventional visual signals into topological differences in numerical dimensions, providing an intuitive and physically constrained basis for isolation features in subsequent deep learning models. It effectively compensates for the technical defects of artifact missegmentation caused by relying solely on visual texture extraction, and further ensures the accuracy of SWI image segmentation of the brain hemorrhage hematoma region.

[0041] After successfully constructing the artifact separation feature image, although the active hematoma entity and the passive flowering artifact have been effectively separated in the numerical topological dimension, this only introduces constraint information based on the physical distribution of the magnetic susceptibility source. A single physical feature is insufficient to cover the complete anatomical morphological variations of the hematoma. This invention then inputs the signal intensity map and the artifact separation feature image into a deep learning model, outputting the hematoma region to be segmented. This step performs multimodal fusion of signal intensity information reflecting the lesion's anatomical texture distribution and artifact separation features reflecting the physical barrier properties of the magnetic field. This allows the deep convolutional neural network to perceive the general outline of the hematoma using visual features while being guided and constrained by physical topological impedance features, thereby accurately suppressing flowering artifact regions with high isolation feature values ​​during the inference stage. This process achieves complementary advantages between structural texture features and magnetic field physical properties, significantly improving the accuracy and robustness of SWI image segmentation of the hematoma region in cerebral hemorrhage while preserving the model's ability to generalize and adapt to hematoma morphology.

[0042] Preferably, in some possible implementations of the embodiments of the present invention, the process of obtaining the hematoma center voxel includes: In the signal intensity map, the signal intensities of all voxels are arranged in descending order to determine the signal intensity sequence. The voxels corresponding to the first predetermined number of signal intensities in the signal intensity sequence are taken as the hematoma center voxels. Because the hematoma center voxels serve as the starting anchor points for subsequent construction of boundary drag distribution and geodesic wavefront propagation, and the hematoma core usually corresponds to the region with the most severe signal attenuation in magnetic susceptibility-weighted imaging, i.e., the region with the highest signal intensity after negative correlation normalization, selecting the voxels with the highest signal intensity can directly and stably anchor the physical source center of the lesion, thus laying a reliable spatial reference for subsequent calculations of the topological potential energy that reflects the true physical properties.

[0043] In one specific implementation of this invention, the range of values ​​for the preset quantity is set as follows: In this embodiment of the invention, the number is preferably set to 20, but it can be adjusted according to the specific implementation environment. The larger the preset number, the larger the core lesion range of the hematoma center voxel cluster can be covered, which is beneficial for subsequent algorithms to obtain a more stable initial wavefront when facing large hematomas. Conversely, it is beneficial to accurately locate the core of small hemorrhage foci and avoid excessive interference from non-hemorrhagic high signal noise points to the propagation starting point.

[0044] Preferably, in some possible implementations of the embodiments of the present invention, the process of obtaining the complex phase gradient includes: The signal phase of each voxel is input into Euler's formula, and the corresponding complex unit vector is output; the calculation process of the complex unit vector is expressed by the formula: ;in, voxels The complex unit vector; voxels The actual physical radian value; The imaginary unit; It is a cosine function; Let be a sine function. In the three-dimensional space corresponding to the complex unit vectors of all voxels, the complex gradient vector of each voxel is calculated using the central difference method. The complex phase gradient of each voxel is determined based on the magnitude of the complex gradient vector. Specifically: for any voxel in the voxel grid space, the complex unit vectors of its directly adjacent voxels in the positive and negative directions of the three orthogonal spatial coordinate axes are obtained. Using the conventional three-dimensional central difference algorithm, the complex spatial partial derivatives of the voxel in the three coordinate axis directions are calculated respectively, and the complex spatial partial derivatives in the three coordinate axis directions are used to construct the three-dimensional complex gradient vector corresponding to the voxel. Subsequently, the Euclidean magnitude of the complex gradient vector is calculated (i.e., taking the modulus of the three complex spatial partial derivatives, summing the squares, and then performing the square root operation), and the final real scalar value is determined as the complex phase gradient of the voxel. It should be noted that the central difference method is a technique well known to those skilled in the art, and will not be further limited or elaborated upon below.

[0045] Because the original signal phase is usually limited to Within this range, when the local magnetic susceptibility of the hemorrhage site changes drastically, phase winding is highly likely to occur, meaning the value changes from... Mutation to If spatial differentiation is performed directly in the real number domain, such discontinuous numerical jumps will be misjudged as high-intensity false edges (mathematical singularities). Therefore, by mapping the scalar phase to a complex unit vector and utilizing the absolute continuity of the complex vector rotating on the unit circle, the discontinuous jumps caused by the winding are fundamentally eliminated. Furthermore, by obtaining the magnitude of the complex gradient through three-dimensional central difference, the intensity of the first-order spatial abrupt change in the local magnetic susceptibility field can be reflected in a true, continuous, and artifact-free manner, thus eliminating mathematical interference for the subsequent construction of accurate physical boundary impedance characteristics.

[0046] Preferably, in some possible implementations of the embodiments of the present invention, the process of obtaining the Laplace source feature values ​​includes: The complex unit vectors of all voxels within a preset neighborhood of each voxel are summed to determine the neighborhood summation vector of each voxel. The number of other voxels within the preset neighborhood of each voxel is used as the corresponding neighborhood quantity weight. The corresponding weighted reference vector is determined based on the product between the complex unit vector of each voxel and the neighborhood quantity weight. The Laplacian source eigenvalue of each voxel is determined based on the magnitude of the difference vector between the weighted reference vector and the neighborhood summation vector.

[0047] In one specific implementation of this invention, the preset neighborhood range is set as a three-dimensional spatial cross-shaped six-neighbor area (up, down, left, right, front, and back) centered on the current voxel, that is, six directly adjacent voxels. It can be adjusted according to the requirements of the magnetic susceptibility source calculation accuracy in the specific implementation environment. The larger the preset neighborhood range, the more voxels it contains, the smoother the calculated Laplace response, and the stronger the robustness to noise, but it will reduce the sensitivity to fine boundaries; conversely, it will be more sensitive to the boundary identification of small lesions.

[0048] Because the Laplace operator in the complex domain essentially reflects the degree of deviation between the local vector field and the mean of its neighborhood, and this degree of deviation precisely corresponds to the essential difference between the magnetic susceptibility source (nonharmonic field) and the passive background (harmonic field) in the second derivative, by calculating the magnitude of the difference between the current voxel vector and the neighborhood accumulated vector, the hematoma entity with real physical source properties can be directly highlighted, thereby effectively suppressing the flowering artifact region with only first-order gradient response but no passive properties.

[0049] Preferably, in some possible implementations of the embodiments of the present invention, the process of obtaining the direction consistency factor includes: In the signal intensity map, the intensity gradient vector of each voxel is calculated using the central difference method; the reference gradient vector of each voxel is calculated using the central difference method based on the complex phase gradient of all voxels. The calculation principle of the intensity gradient vector and the reference gradient vector here is the same as the calculation principle of the three-dimensional complex gradient vector mentioned above. That is, for the signal intensity map or the complex phase gradient map, the conventional three-dimensional central difference algorithm is used to calculate the difference between the values ​​of adjacent voxels in the positive and negative directions of the three orthogonal spatial coordinate axes of each voxel and halve them to construct the corresponding three-dimensional spatial vector. This is a central difference method technique well known to those skilled in the art, and will not be further limited or elaborated here.

[0050] Furthermore, the cosine similarity between the intensity gradient vector and the reference gradient vector is positively correlated to determine the corresponding orientation consistency factor. In a specific implementation of this invention, the process of obtaining the orientation consistency factor is expressed by the formula: ;in, voxels directional consistency factor; voxels The cosine similarity between the intensity gradient vector and the reference gradient vector.

[0051] In magnetic resonance imaging (MRI), the true physical boundary of a hematoma should be the region where the abrupt change in anatomical morphology (signal intensity gradient) and the abrupt change in local magnetic susceptibility (complex phase gradient and its spatial rate of change) highly coincide in direction. This can be achieved by calculating the cosine similarity of these two gradient vectors and mapping it to... The positive correlation interval can significantly enhance the target boundary voxel whose anatomical edge and magnetic field edge are aligned, while greatly suppressing the noise response that appears only in a single modal image. This provides a precise physical orientation dimension constraint for the subsequent construction of the boundary impedance distribution field, and greatly improves the anti-interference ability of the resistance field under complex physiological backgrounds.

[0052] Preferably, in some possible implementations of the embodiments of the present invention, the process of obtaining the boundary resistance strength value includes: The first standard deviation is determined based on the standard deviation of the complex phase gradients of all voxels; the second standard deviation is determined based on the standard deviation of the Laplace source eigenvalues ​​of all voxels; the standard phase gradient is determined based on the ratio between the complex phase gradient and the first standard deviation of each voxel; the standard eigenvalue is determined by positively correlated mapping of the ratio between the Laplace source eigenvalue and the second standard deviation of each voxel; the fusion boundary eigenvalue is determined based on the sum of the standard phase gradient and the standard eigenvalue; and the boundary drag strength value of each voxel is determined by positively correlated exponential mapping of the product between the fusion boundary eigenvalue and the directional consistency factor.

[0053] In one specific implementation of this invention, the process of obtaining the boundary resistance strength is expressed by the formula: ;in, voxels Boundary resistance strength; voxels directional consistency factor; voxels The complex phase gradient; The standard deviation of the complex phase gradient of all voxels is the first standard deviation. voxels The standard phase gradient; voxels Laplace source eigenvalues; The standard deviation of the Laplace source eigenvalues ​​of all voxels is the second standard deviation. voxels Standard eigenvalues; voxels fusion boundary eigenvalues; It is an exponential function with the natural constant as its base; The sensitivity coefficient is preferably set to 2 in this embodiment of the invention, but can be adjusted according to the specific implementation environment. The sensitivity coefficient is used to control the degree of nonlinear amplification of the drag field to the boundary response. The larger the value, the more sensitive it is to the impedance response of weak boundaries, and the stronger it can isolate artifacts, but it may also introduce more background noise; conversely, the drag field is smoother, the noise resistance is enhanced, but the ability to isolate subtle artifacts is weakened.

[0054] It should be noted that in conventional MRI scans, due to the inherent differences in the magnetic susceptibility of human tissues and the objective existence of equipment thermal noise, the complex phase gradient and Laplace source eigenvalues ​​of the entire image will inevitably fluctuate. Therefore, the first and second standard deviations are usually strictly greater than zero in a physical sense. A first or second standard deviation of 0 usually corresponds to extreme anomalies, such as completely empty scan data or equipment malfunction. In such cases, the analyzed image is considered invalid. It is necessary to re-acquire the signal intensity map and signal phase map, which will not be further limited or elaborated here.

[0055] Because the first-order complex phase gradient and the second-order Laplace source eigenvalues ​​have different physical dimensions and numerical distribution ranges, direct addition would cause one of the features to be submerged. Furthermore, the directional consistency factor, as an independent modulation term, needs to be coupled with the aforementioned features to impose spatial constraints. Therefore, the first-order and second-order features are first standardized using their respective standard deviations to eliminate dimensional differences and balance their contribution weights, generating fused boundary eigenvalues. Then, the directional consistency factor is used to modulate this fused feature, strengthening the response in the anatomically and magnetically consistent region. Finally, the modulated value is mapped to boundary drag strength using an exponential function. This process unifies and transforms multi-dimensional, heterogeneous physical features into a continuous, interpretable scalar drag field, where exponentially high drag values ​​will be exhibited at the actual hematoma physical boundary, providing a clear and quantitative definition of the medium properties for subsequent topology propagation algorithms. In addition, the introduction of a positively correlated exponential mapping is to nonlinearly amplify the boundary response at the numerical level. In linear models, the resistance difference between the hematoma boundary and the internal tissue may not be sufficient to produce a decisive truncation effect in subsequent wavefront propagation algorithms. However, through exponential mapping, even a slight high response (i.e., a potential hematoma boundary) in the fused boundary eigenvalues ​​is nonlinearly amplified into extremely high resistance, thus constructing an insurmountable resistance wall in space. This effectively prevents the wavefront from penetrating the hematoma boundary and spreading to the passive artifact region, thereby providing a more extreme and robust physical medium basis for subsequent geodesic path integration. Further elaboration is omitted here.

[0056] Preferably, in some possible implementations of the embodiments of the present invention, the process of obtaining the artifact separation feature image includes: Using each hematoma center voxel as the source point of wavefront propagation, the minimum cumulative resistance of each voxel is calculated using a fast travel algorithm, combined with the boundary resistance strength value of each voxel. The minimum cumulative resistance is then normalized to determine the artifact separation feature value of each voxel. An artifact separation feature image is constructed based on the artifact separation feature values ​​of all voxels. In a specific implementation of this invention, the method for normalizing the minimum cumulative resistance uses minimum-maximum normalization, and the normalization scope covers all voxels. It should be noted that the fast travel algorithm is a well-known technique. Here, the hematoma center voxel is considered the source point of wavefront propagation in the fast travel algorithm, and the boundary resistance strength value is the cost of the spatial medium that the wavefront propagation must penetrate in the fast travel algorithm. That is, the difference in minimum cumulative resistance between any two adjacent voxels is controlled by the boundary resistance strength value of that region, and this will not be further limited or elaborated upon here.

[0057] Furthermore, it should be noted that, unless otherwise specified, the normalization function or normalization method mentioned in this invention employs minimum-maximum normalization to normalize the normalization result to... The range. The maximum and minimum values ​​used in minimum-maximum normalization can be obtained based on the actual situation. For example, when multiple values ​​can be obtained during implementation and it is necessary to compare the size relationship between different values, multiple values ​​can be counted to obtain the maximum and minimum values. However, when only a single value can be obtained during implementation or it is not necessary to compare the size relationship between different values, the maximum and minimum values ​​can be obtained by counting based on a large amount of historical experimental data or prior data obtained earlier.

[0058] Because the fast-progression algorithm transforms the static local boundary resistance intensity into a global topological potential distribution through the outward diffusion process from the hematoma center, for flowering artifacts that are closely connected to the hematoma entity in anatomical space, wavefront propagation inevitably accumulates a high resistance cost on the geodesic integration path due to the need to cross the high-impedance physical boundary of the hematoma. This leads to a significant increase in the minimum cumulative resistance in the artifact region. By performing min-max normalization, this difference in cumulative cost can be mapped to a standardized numerical range. This allows active hematomas and passive artifacts, which are difficult to distinguish in conventional signal intensity maps, to exhibit significant distributional differences in the numerical dimension of the constructed artifact separation feature image. This provides a direct and reliable physical constraint criterion for deep networks.

[0059] Preferably, in some possible implementations of the embodiments of the present invention, the process of inputting the signal intensity map and artifact separation feature image into a deep learning model and outputting the hematoma region to be segmented includes: The signal intensity map and artifact separation feature image are input into a trained deep convolutional neural network model, which outputs the hematoma probability for each voxel. In a specific implementation of this invention, the deep convolutional neural network model adopts an encoder-decoder network architecture based on a three-dimensional U-Net structure. The model training process specifically includes: concatenating the signal intensity map of the sample data with the corresponding artifact separation feature image in the channel dimension, and feeding it as a dual-channel input tensor into the constructed initial deep convolutional neural network; calculating the basic segmentation similarity loss (such as Dice loss) using a manually labeled real hematoma mask and the predicted probability map output by the network's forward propagation; simultaneously, extracting the input artifact separation feature image as a spatial penalty weight, multiplying it by the predicted probability map voxel by voxel to calculate the artifact suppression loss, so that if the network outputs a high probability in a region with high artifact feature values, a huge penalty is incurred; the segmentation similarity loss and artifact suppression loss are weighted and summed to obtain a joint loss function, and the network parameters are iteratively updated using a gradient descent optimization algorithm (such as the Adam optimizer) until the joint loss function converges, resulting in a trained deep convolutional neural network model. It should be noted that the encoder-decoder architecture based on the 3D U-Net structure, the calculation principle of Dice loss, and the parameter iteration and convergence determination process based on gradient descent are all conventional technical means and technical terms well known to those skilled in the art, and will not be elaborated further here.

[0060] In one specific implementation of this invention, the formula corresponding to the joint loss function includes: ;in, This represents the joint loss function value for a single iteration of the network. The segmentation similarity loss (i.e., Dice loss) between the predicted probability map of the sample data and the manually labeled real hematoma mask is used to constrain the network to learn the general structural contour of the hematoma. Predict the probability map of the sample data output by the network forward propagation; This is a manually labeled real hematoma mask corresponding to the sample data; The number of voxels in the entire voxel grid space of the sample data; For the sample data The probability value of an individual being predicted as a hematoma; For the sample data The artifact separation feature value corresponding to the individual element in the input artifact separation feature image (this value is close to 1 in the flowering artifact region and close to 0 in the hematoma core region). The preset penalty weight coefficient is used to balance the proportion between the basic segmentation task and the artifact suppression task (in this embodiment, the value is 0.5, which can be fine-tuned according to the convergence state during specific model training).

[0061] Furthermore, voxels corresponding to hematoma probabilities greater than a preset probability threshold are designated as hematoma voxels. In all three-dimensional connected regions formed by all hematoma voxels, three-dimensional connected regions with a voxel count greater than a preset quantity threshold are marked as hematoma regions to be segmented. In a specific implementation of this invention, the preset probability threshold is set to 0.5, and the preset quantity threshold is set to 15, which can be adjusted according to the specific implementation environment. Since the hematoma probability ranges from 0 to 1, the preset probability threshold can be adjusted from 0 to 1. A larger preset probability threshold indicates a higher confidence requirement for the model's prediction of a hematoma entity, resulting in a more conservative segmentation result. This allows for more rigorous removal of suspected artifact regions with blurred edges, but may lead to missed segmentation of the true edge of the lesion. Conversely, a smaller threshold results in a more aggressive segmentation result, preserving more edge details and weak signal areas, but may introduce background noise. Correspondingly, the larger the preset number threshold, the stricter the minimum volume restriction on the spatial connected domain, which can more effectively filter out isolated noise voxels caused by free noise or sporadic false positive predictions, but may cause real multiple micro-hemorrhagic lesions (satellite foci) to be mistakenly deleted as noise; conversely, the ability to preserve micro-lesions is stronger, but the overall noise resistance of the final segmentation result will be reduced accordingly.

[0062] The deep convolutional neural network simultaneously receives anatomical texture information (signal intensity map) and physical constraint information (artifact separation feature image) through dual-channel input. During the inference stage, it can capture the general outline of the hematoma using the visual features of the first channel, while being forced to be guided by the physical topological high impedance features of the second channel. This allows it to accurately suppress the spatially connected flowering artifact regions in the feature mapping within the model. Finally, by combining probability threshold truncation and morphological cleaning operations based on the volume of three-dimensional connected regions, it can remove discretely distributed false positive prediction noise while fully preserving the core hematoma and satellite lesions with actual clinical significance, ensuring high accuracy and robustness of SWI image segmentation of the hematoma region in cerebral hemorrhage.

[0063] In summary, a deep learning-based SWI image segmentation method for hematoma regions in cerebral hemorrhage extracts the complex phase gradient and Laplacian source feature values ​​reflecting the abrupt change characteristics of the magnetic susceptibility source from the signal phase map to obtain the first and second spatial derivative information of the magnetic field distribution. Furthermore, it couples the directional consistency factors of the anatomical edge and the magnetic field edge, establishing physically constrained boundary resistance strength values ​​within the voxel grid space. Based on these boundary resistance strength values ​​and the voxel at the hematoma center, an artifact separation feature image is constructed. Utilizing the cumulative path cost difference spreading outward from the hematoma center, the flowering artifact region spatially adhered to the hematoma entity is mapped to a feature distribution with significant numerical differences. This provides a physical basis for the deep learning model to distinguish between the actual hematoma target and passive magnetic field artifacts, overcoming the shortcomings of existing technologies that rely solely on visual features and are prone to artifact missegmentation. This results in higher accuracy for SWI digital image segmentation of hematoma regions in cerebral hemorrhage based on deep convolutional neural networks.

[0064] This invention also provides a deep learning-based SWI image segmentation system for hematoma regions in brain hemorrhage. Please refer to [link to relevant documentation]. Figure 2 The diagram shows a structural diagram of a deep learning-based SWI image segmentation system for brain hemorrhage hematoma regions provided by an embodiment of the present invention. The system includes: a data acquisition and preprocessing module 201, a first determination module 202, a second determination module 203, and an image segmentation module 204.

[0065] The data acquisition and preprocessing module 201 is used to resample the magnetic susceptibility-weighted imaging scan data of the cranial region into a voxel grid space from a medical database to determine the signal intensity map and the signal phase map; and to determine the central voxel of the hematoma based on the signal intensity distribution of all voxels in the signal intensity map. The first determining module 202 is used to determine the complex phase gradient of each voxel based on the spatial change of the signal phase of all voxels in the signal phase diagram; and to determine the Laplace source eigenvalue of each voxel based on the overall deviation of the complex phase gradient between each voxel and other voxels in the spatial neighborhood. The second determining module 203 is used to determine the orientation consistency factor of each voxel based on the local changes in complex phase gradient and signal intensity between adjacent voxels; and to perform coupled calculations on the orientation consistency factor, complex phase gradient and Laplace source eigenvalues ​​to determine the boundary drag strength value of each voxel. The image segmentation module 204 is used to construct an artifact separation feature image based on the distribution of boundary resistance strength values ​​of each voxel and their relative positions with each hematoma center voxel; the signal intensity map and the artifact separation feature image are input into the deep learning model, and the hematoma region to be segmented is output.

[0066] It should be noted that the system provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above. In addition, the deep learning-based SWI image segmentation system for brain hemorrhage hematoma regions and the deep learning-based SWI image segmentation method for brain hemorrhage hematoma regions provided in the above embodiments belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0067] This invention also provides a computer device; please refer to [link / reference]. Figure 3 The illustration shows a schematic diagram of a computer device structure according to an embodiment of the present invention. The computer device includes a memory 301, a processor 302, and a computer program 303 stored in the memory 301 and running on the processor 302. When the processor 302 executes the computer program 303, the computer device can execute any of the deep learning-based SWI image segmentation methods for brain hemorrhage hematoma regions described above.

[0068] This invention also provides a computer program product that, when run on a computer device, enables the computer device to execute any of the deep learning-based SWI image segmentation methods for brain hemorrhage hematoma regions described above.

[0069] This invention also provides a computer-readable storage medium storing computer program code. When the computer program code is run on a computer device, the computer device can execute any of the deep learning-based SWI image segmentation methods for brain hemorrhage hematoma regions described above.

[0070] In the embodiments provided by the present invention, it should be understood that the computer device, computer program product and computer-readable storage medium provided are all used to perform the corresponding methods provided above, and therefore the beneficial effects they can achieve can be referred to the beneficial effects of the methods provided above, which will not be repeated here.

[0071] 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.

[0072] 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 deep learning-based SWI image segmentation method for hematoma regions in brain hemorrhage, characterized in that, The method includes: In a medical database, magnetic susceptibility-weighted imaging scan data of the cranial region are resampled to a voxel grid space to determine the signal intensity map and signal phase map; based on the signal intensity distribution of all voxels in the signal intensity map, the central voxel of the hematoma is determined. Based on the spatial variation of the signal phase of all voxels in the signal phase diagram, the complex phase gradient of each voxel is determined; based on the overall deviation of the complex phase gradient between each voxel and other voxels in its spatial neighborhood, the Laplace source eigenvalue of each voxel is determined. Based on the local variations in complex phase gradient and signal intensity between adjacent voxels, the orientation consistency factor of each voxel is determined; the orientation consistency factor, the complex phase gradient, and the Laplace source eigenvalue are coupled and calculated to determine the boundary drag strength value of each voxel. Based on the distribution of boundary resistance strength values ​​of each voxel and their relative positions with each hematoma center voxel, an artifact separation feature image is constructed; the signal intensity map and the artifact separation feature image are input into a deep learning model to output the hematoma region to be segmented; The process of obtaining the boundary resistance strength value includes: The first standard deviation is determined based on the standard deviation of the complex phase gradients of all voxels; the second standard deviation is determined based on the standard deviation of the Laplace source eigenvalues ​​of all voxels. The standard phase gradient is determined based on the ratio between the complex phase gradient of each voxel and the first standard deviation; the standard eigenvalue is determined by positively correlating the ratio between the Laplacian source eigenvalue of each voxel and the second standard deviation; and the fusion boundary eigenvalue is determined based on the sum of the standard phase gradient and the standard eigenvalue. The boundary resistance strength value of each voxel is determined by performing a positive correlation index mapping on the product between the fusion boundary feature value and the direction consistency factor.

2. The SWI image segmentation method for brain hemorrhage hematoma region based on deep learning according to claim 1, characterized in that, The process of obtaining the hematoma center voxel includes: In the signal intensity diagram, the signal intensities of all voxels are arranged in descending order to determine the signal intensity sequence; the voxels corresponding to the first preset number of signal intensities in the signal intensity sequence are taken as the hematoma center voxels.

3. The SWI image segmentation method for brain hemorrhage hematoma region based on deep learning according to claim 1, characterized in that, The process of obtaining the complex phase gradient includes: Input the signal phase of each voxel into Euler's formula and output the corresponding complex unit vector; In the three-dimensional space corresponding to the complex unit vectors of all voxels, the complex gradient vector of each voxel is calculated by the central difference method; the complex phase gradient of each voxel is determined according to the magnitude of the complex gradient vector.

4. The SWI image segmentation method for brain hemorrhage hematoma region based on deep learning according to claim 1, characterized in that, The process of obtaining the Laplace source eigenvalues ​​includes: The complex unit vectors of all voxels within a preset neighborhood of each voxel are summed to determine the neighborhood summation vector of each voxel; the number of other voxels within the preset neighborhood of each voxel is used as the corresponding neighborhood quantity weight; and the corresponding weighted reference vector is determined based on the product between the complex unit vector of each voxel and the neighborhood quantity weight. The Laplacian source eigenvalue of each voxel is determined based on the magnitude of the difference vector between the weighted reference vector and the neighborhood accumulation vector.

5. The SWI image segmentation method for brain hemorrhage hematoma region based on deep learning according to claim 1, characterized in that, The process of obtaining the directional consistency factor includes: In the signal intensity map, the intensity gradient vector of each voxel is calculated using the central difference method; the reference gradient vector of each voxel is calculated using the central difference method based on the complex phase gradients of all voxels; the cosine similarity between the intensity gradient vector and the reference gradient vector is positively correlated to determine the corresponding directional consistency factor.

6. The SWI image segmentation method for brain hemorrhage hematoma region based on deep learning according to claim 1, characterized in that, The process of obtaining the artifact separation feature image includes: Using each hematoma center voxel as the source point for wavefront propagation, and combining the boundary resistance strength value of each voxel, the minimum cumulative resistance of each voxel is calculated using a fast travel algorithm. The minimum cumulative resistance is normalized to determine the artifact separation feature value of each voxel; an artifact separation feature image is constructed based on the artifact separation feature values ​​of all voxels.

7. The SWI image segmentation method for brain hemorrhage hematoma region based on deep learning according to claim 1, characterized in that, The process of inputting the signal intensity map and the artifact separation feature image into the deep learning model and outputting the hematoma region to be segmented includes: The signal intensity map and the artifact separation feature image are input into the trained deep convolutional neural network model, and the hematoma probability of each voxel is output. Voxels with hematoma probabilities greater than a preset probability threshold are designated as hematoma voxels. In all three-dimensional connected regions formed by all hematoma voxels, three-dimensional connected regions with a number of voxels greater than a preset number threshold are marked as hematoma regions to be segmented.

8. The SWI image segmentation method for brain hemorrhage hematoma region based on deep learning according to claim 7, characterized in that, The preset probability threshold is set to 0.

5.

9. A deep learning-based SWI image segmentation method for hematoma regions in brain hemorrhage according to claim 7, characterized in that, The preset quantity threshold is set to 15.

Images