Cerebral hemorrhage space positioning and prognosis prediction network system based on three-dimensional point cloud

Abstract

The invention relates to the technical field of internet big data, in particular to a cerebral hemorrhage space positioning and prognosis prediction network system based on three-dimensional point cloud. The cerebral hemorrhage space positioning system comprises an image acquisition module used for acquiring a brain CT image to be processed; the offset correction module is used for performing offset correction on the brain CT image to obtain a corrected brain CT image; the hematoma region segmentation model is used for segmenting the corrected brain CT image to obtain a hematoma region; the point cloud calculation module is used for calculating three-dimensional space coordinates of hematoma voxels in a hematoma area in a brain coordinate system based on the corrected brain CT image and taking the three-dimensional space coordinates as point cloud data of the hematoma area; and the point cloud data output module is used for performing normalization processing on the point cloud data of the hematoma area to obtain hematoma three-dimensional point cloud data representing the hematoma three-dimensional space position. According to the method, an accurate and comprehensive data basis is provided for cerebral hemorrhage prognosis prediction, and the accuracy and comprehensiveness of subsequent cerebral hemorrhage prognosis prediction are improved.

Description

Technical Field

[0001] The present invention relates to the field of Internet big data technology, and in particular to a 3D point cloud-based spatial positioning and prognosis prediction network system for cerebral hemorrhage. Background Art

[0002] In the medical field, cerebral hemorrhage, a serious cerebrovascular disease, has a crucial impact on the patient's health and quality of life. Accurately predicting the prognosis of patients with cerebral hemorrhage can provide a key basis for clinical treatment decisions, helping doctors develop personalized treatment plans, rationally allocate medical resources, and provide timely health guidance to patients and their families.

[0003] Currently, clinicians assess the prognosis of patients with cerebral hemorrhage based primarily on basic clinical information such as the patient's age, Glasgow Coma Scale (GCS) upon admission, the volume of cerebral hemorrhage, and the presence of intraventricular hemorrhage. However, these indicators have certain limitations. For example, even if different patients have the same clinical indicator values, their prognosis may still vary significantly, indicating that relying solely on these traditional indicators is difficult to fully and accurately reflect the patient's true prognosis.

[0004] With the advancement of medical imaging technology, computer-aided diagnosis (CAD) has been gradually applied to the prognosis prediction of intracerebral hemorrhage. Early methods primarily focused on simple analysis of brain CT images, such as predicting prognosis by measuring two-dimensional morphological parameters such as the hematoma's area and perimeter. However, these existing methods ignored the hematoma's three-dimensional location and complex geometric features, failing to fully exploit the rich information contained in the imaging data, making it difficult to further improve prediction accuracy.

[0005] In recent years, the application of deep learning technology in medical image analysis has made significant progress. Some studies have attempted to use convolutional neural networks (CNNs) to extract features and predict prognosis from CT images of intracerebral hemorrhage. Although CNNs can automatically learn image features to a certain extent, because CT images do not provide the three-dimensional spatial location of hematomas, the model is still unable to effectively utilize this information. Consequently, it is difficult to directly capture the three-dimensional spatial structure and local geometric variations of the hematoma. Prediction performance still needs to be improved when dealing with complex hematomas and data with large individual differences.

[0006] In summary, the existing prognosis prediction methods that directly use CT images as input have shortcomings in accuracy and comprehensiveness. There is an urgent need for a method that can effectively extract hematoma three-dimensional point cloud data that represents the three-dimensional spatial position of the hematoma from CT images, so as to provide an accurate and comprehensive data basis for the prognosis prediction of cerebral hemorrhage, thereby helping to improve the accuracy and comprehensiveness of subsequent cerebral hemorrhage prognosis prediction. Summary of the Invention

[0007] In view of the deficiencies of the above-mentioned existing technologies, the technical problem to be solved by the present invention is: how to provide a cerebral hemorrhage spatial positioning and prognosis prediction network system based on three-dimensional point cloud, which can effectively extract hematoma three-dimensional point cloud data representing the three-dimensional spatial position of the hematoma from the brain CT image, and provide an accurate and comprehensive data basis for the prognosis prediction of cerebral hemorrhage, thereby helping to improve the accuracy and comprehensiveness of subsequent cerebral hemorrhage prognosis prediction.

[0008] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0009] The spatial positioning system for cerebral hemorrhage based on three-dimensional point cloud includes:

[0010] An image acquisition module, used for acquiring brain CT images to be processed;

[0011] An offset correction module is used to perform offset correction on the brain CT image to obtain a corrected brain CT image;

[0012] A hematoma region segmentation model is used to segment the hematoma region of the rectified brain CT image to obtain the hematoma region;

[0013] A point cloud computing module is used to calculate the three-dimensional spatial coordinates of hematoma voxels in the hematoma area in the brain coordinate system based on the corrected brain CT image as point cloud data of the hematoma area;

[0014] The point cloud data output module is used to normalize the point cloud data of the hematoma area to obtain three-dimensional point cloud data of the hematoma that represents the three-dimensional spatial position of the hematoma.

[0015] Preferably, in the point cloud data output module, after obtaining the hematoma three-dimensional point cloud data, the normal vector information of the three-dimensional coordinates of each point in the hematoma three-dimensional point cloud data is further calculated, and the normal vector information is associated with the three-dimensional coordinates of the corresponding point;

[0016] For each point P in the hematoma 3D point cloud data i ’ , select several points P in its neighborhood j ’ To construct a local plane, point P i ’ The neighborhood point cloud is {P j ’}, j∈{1,2,…,K} is the index of the neighborhood point; the normal vector n of the corresponding local plane is obtained by performing least squares plane fitting on the neighborhood points i ;

[0017] The formula is:

[0018]

[0019] Where: n i represents the normal vector to be found; P′ i is the three-dimensional coordinate of point i, P j ′ is the three-dimensional coordinate of point j in the neighborhood of point i.

[0020] Preferably, in the offset correction module, the offset correction performed on the brain CT image includes any one or more of binarization, erosion, generating an external rectangular frame, and translation and rotation correction;

[0021] 1) Binarization

[0022] The binarization process divides the pixel values in the brain CT image into two categories by setting a threshold T: the part greater than the threshold represents the target area, and the other part is the background;

[0023] The formula for binarization is expressed as:

[0024]

[0025] Where: I(x,y) is the grayscale value of the original brain CT image; I b (x, y) is the binarized brain CT image; T is the set threshold;

[0026] 2) Corrosion operation

[0027] The erosion operation processes the brain CT image by using a structure element to reduce the boundaries of the white area;

[0028] The formula for the corrosion operation is expressed as:

[0029]

[0030] Where: S is the structural element of the corrosion operation, I e (x,y) is the brain CT image after corrosion;

[0031] 3) Generate an external rectangular frame

[0032] The bounding rectangle is the smallest rectangle that can completely contain the target area. The coordinates of the bounding rectangle are (x min ,y min ) and (x max ,y max ), which represent the coordinates of the upper left corner and lower right corner of the rectangular box respectively;

[0033] 4) Translation and rotation correction

[0034] Calculate the center point C of the circumscribed rectangle, whose coordinates are:

[0035]

[0036] By using the rotation matrix and displacement transformation, the bounding rectangle is rotated to align with the positive coordinate axis;

[0037] The rotation matrix R(θ) is defined as:

[0038] P new =R(θ)(PC)+C+t;

[0039] Where: P is the coordinate of any point in the circumscribed rectangular frame, t=(t x ,t y ) is the translation vector, P new are the coordinates of the point after rotation and translation.

[0040] Preferably, in the point cloud data output module, the point cloud data of the hematoma area is P = {P1, P2, ..., P N}, the point cloud centroid of the hematoma area is

[0041] The formula for normalization is expressed as:

[0042] P′ i =P i -Cfori∈{1,2,…,N};

[0043] Where: P i ’ is the three-dimensional coordinate of the normalized hematoma point cloud.

[0044] Preferably, in the point cloud data output module, after obtaining the hematoma three-dimensional point cloud data, the hematoma three-dimensional point cloud data is further subjected to spatial transformation processing by a local neighborhood spatial transformation network to obtain the hematoma three-dimensional point cloud data after spatial transformation;

[0045] The local neighborhood spatial transformation network includes a local spatial transformation module for performing local enhancement operations on hematoma three-dimensional point cloud data, and a local neighborhood STN module for dynamically modeling the neighborhood features of each point in the hematoma three-dimensional point cloud data.

[0046] Preferably, the local space transformation module is used to perform local enhancement operations on the hematoma three-dimensional point cloud data, including introducing random noise, local translation, local rotation and random perturbation, to obtain the hematoma three-dimensional point cloud data after local space transformation;

[0047] 1) Introducing random noise

[0048] Random noise with controllable amplitude is added to the spatial coordinates of the hematoma 3D point cloud data. The noise amplitude is dynamically adjusted according to the size of the hematoma area to simulate the slight deviation in the brain CT imaging process.

[0049] 2) Local translation

[0050] Performing an adaptive random translation operation on the local area of the hematoma 3D point cloud data. The translation range is dynamically adjusted according to the density and geometric complexity of the local area.

[0051] 3) Local rotation

[0052] Perform angle-restricted random rotation around the centroid of the local area of the hematoma 3D point cloud data, and the rotation angle range is dynamically adjusted according to the geometric characteristics of the local area;

[0053] 4) Random perturbations

[0054] The point cloud of the local area of the hematoma three-dimensional point cloud data is subjected to amplitude-controllable random perturbation, and the perturbation amplitude is dynamically adjusted according to the geometric complexity of the local area.

[0055] Preferably, the processing steps of the local neighborhood STN module include:

[0056] S01: Three-dimensional point cloud data of hematoma after local space transformation Through three consecutive one-dimensional convolutional layers, the dimension of the output feature is gradually increased from C to 64, 128 and 1024, and feature F3 is obtained;

[0057] The formula is:

[0058] F1 = ReLU(BN(Conv1d(X)));

[0059] F2 = ReLU(BN(Conv1d(F1)));

[0060] F3 = ReLU(BN(Conv1d(F2)));

[0061] Where: BN represents batch normalization operation; ReLU represents activation function; Cond1d represents one-dimensional convolution;

[0062] S02: Use global maximum pooling to aggregate the feature F3 in point dimensions to obtain the global feature vector

[0063] The formula is:

[0064] z = MaxPool(F3, dim = 2);

[0065] S03: Input the global feature z into three fully connected layers, gradually compressing its feature dimensions to 512, 256 and 9, and obtaining the transformation matrix in tiled form

[0066] The formula is:

[0067] g1 = ReLU(BN(W1z+b1));

[0068] g2 = ReLU(BN(W2g1+h2));

[0069] T=W3g2+b3;

[0070] Where: W1, W2, W3 and b1, b2, b3 are the weights and bias parameters of the fully connected layer;

[0071] S04: Reconstruct the transformation matrix T into a 3×3 affine transformation matrix A and add the identity matrix I;

[0072] The formula is:

[0073] A=Teshape(T)+I;

[0074] I = diag(1,1,1);

[0075] S05: Set the diagonal elements in the affine transformation matrix A to be fixed to 1; set range constraints for the non-diagonal elements in the affine transformation matrix A, which is expressed as follows:

[0076] A ij =Clamp(A ij ,-θ max ,θ max ),i≠j;

[0077] Where: θ max =0.3 is the maximum rotation and translation amplitude;

[0078] S06: Performing local spatial transformation and alignment on the hematoma three-dimensional point cloud data after local spatial transformation using the affine transformation matrix A to obtain the hematoma three-dimensional point cloud data after spatial transformation.

[0079] The present invention also discloses a cerebral hemorrhage prognosis prediction network system based on three-dimensional point cloud, comprising:

[0080] The input layer is used to take the hematoma three-dimensional point cloud data obtained by the point cloud data output module as input;

[0081] The first processing module is used to input the hematoma three-dimensional point cloud data into the MLP and multi-scale point cloud attention module in sequence to model and fuse global features and local features to obtain initial point cloud features;

[0082] The feature transformation module is used to perform feature transformation on the initial point cloud features to obtain the transformed point cloud feature map;

[0083] The second processing module is used to input the transformed point cloud feature map into the MLP and multi-scale point cloud attention module in sequence to model and fuse global features and local features to obtain secondary point cloud features;

[0084] A global multi-graph self-attention module is used to build graph structures by learning node relationships from various channel dimensions and capture complex spatial and semantic information in secondary point cloud features to obtain global features;

[0085] The classification module is used to classify global features through MLP and Softmax to obtain corresponding prognosis prediction results.

[0086] Preferably, the multi-scale point cloud attention module in the first processing module and the second processing module includes two parts: channel attention and multi-scale spatial attention, which are used to capture key feature information in the hematoma three-dimensional point cloud data at different levels;

[0087] 1) The processing steps of channel attention include:

[0088] S11: Input features Perform global average pooling and global maximum pooling respectively to obtain the pooling feature X avg and X max ;

[0089] The formula is:

[0090] X avg =AdaptiveAvgPool(X);

[0091] X max =AdaptiveMaxPool(X);

[0092] Where: B, C and N represent the batch size, the number of feature channels and the number of points respectively; AvgPool represents the global average pooling operation; MaxPool represents the global maximum pooling operation; Adaptive represents adaptive;

[0093] S12: Pooling feature X through two layers of shared fully connected network avg and X max Processing is performed to obtain feature X′ avg and X′ max ;

[0094] The formula is:

[0095] X′ avg =σ(W2·ReLU(W1·X avg ));

[0096] X′ max =σ(W2·ReLU(W1·Xmax ));

[0097] Where: W1 and W2 are the weights of the two fully connected layers respectively; σ represents the Sigmoid activation function;

[0098] S13: For feature X′ avg and X′ max Perform weighted fusion to obtain the channel attention weight A c ;

[0099] The formula is:

[0100] A c =X′ avg +X′ max ;

[0101] S14: Input feature X and channel attention weight A c Perform element-by-element multiplication to obtain the channel attention feature X c ;

[0102] The formula is:

[0103] X c =X⊙A c ;

[0104] Where: ⊙ represents the element-by-element multiplication operation;

[0105] 2) The processing steps of multi-scale spatial attention include:

[0106] S21: Use depth convolution kernels of multiple scales of size 5×1, 7×1, 11×1 and 21×1 to focus on channel features X c Perform multi-scale feature extraction to obtain the features S of each scale i ;

[0107] The formula is:

[0108]

[0109] Where: Indicates that the kernel size is k i Depthwise separable convolution operation;

[0110] S22: The features S of each scale i After superposition, the spatial attention weight A is generated by channel mixing convolution s ;

[0111] The formula is:

[0112]

[0113] S23: Channel attention feature X c With the spatial attention weight A s After element-wise multiplication, it is then multiplied with the spatial attention weight A s Perform channel-level feature fusion to obtain the channel multi-scale spatial attention feature Y as the output of the multi-scale point cloud attention module;

[0114] The formula is:

[0115] X s =X c ⊙A s ;

[0116] Y = Conv(X s ⊙A s );

[0117] Where: Conv is the convolution operation for channel mixing.

[0118] Preferably, the processing steps of the global multi-image self-attention module include:

[0119] S31: For the input secondary point cloud features Generate query matrix Q, key matrix K and value matrix V through one-dimensional convolution;

[0120] The formula is:

[0121] Q = Conv1D(X);

[0122] K = Conv1D(X);

[0123] V = Conv1D(X);

[0124] Where:

[0125] S32: Decompose the query matrix Q into a multi-head form Q through reshaping and grouping operations h ;

[0126] The formula is:

[0127]

[0128] Where: H is the number of long positions, C k The query channel dimension for each head;

[0129] S33: Through the Softmax normalization operation, the key matrix K is converted into the probability distribution of the key and the feature interaction is performed with the value matrix V, combined with Q h Generate initial global interaction result Y c ;

[0130] The formula is:

[0131] K V =K T V;

[0132] Y c =Q h ·K V ;

[0133] Where: It is the global preliminary interaction result;

[0134] S34: Define the graph structure as G = (V, ε), treat each feature point as a node of the graph, and the node features come from the value matrix V; multidimensional adjacency matrix are initialized as learnable parameters to represent various relationships between nodes; the edge ε is modeled by the learnable adjacency matrix A;

[0135] S35: Calculate the weighted connections between feature points through matrix multiplication to obtain the relationship matrix G of the graph;

[0136] The formula is:

[0137]

[0138] Where: each channel dimension of the adjacency matrix A corresponds to a global relationship type; the edge weight between the i-th node and the j-th node in the relationship matrix G is obtained by mapping the features of node i and node j through the adjacency matrix A;

[0139] S36: query matrix Q h Interact with the relationship matrix G to generate multi-graph reasoning results Y p ;

[0140] The formula is:

[0141] Y p =Q h G;

[0142] Where: The representation captures the global semantic relationships between nodes generated by graph reasoning;

[0143] S37: The global interaction result Y c and multi-graph reasoning result Y p Add together to get the output feature Y;

[0144] The formula is:

[0145] Y=Y c +Y p ;

[0146] S38: Add the output feature Y to the input secondary point cloud feature X through residual connection and batch normalization operations to obtain the global feature O;

[0147] The public notice is:

[0148] O=ReLU(W skip (Y)+X);

[0149] Where: W skip is the linear transformation matrix used for skip connections.

[0150] Compared with the existing technology, the cerebral hemorrhage spatial positioning and prognosis prediction network system based on three-dimensional point cloud in the present invention has the following beneficial effects:

[0151] The present invention obtains a hematoma region by segmenting the brain CT image and calculating the three-dimensional spatial coordinates of the hematoma voxels in the hematoma region in the brain coordinate system, thereby modeling the hematoma of cerebral hemorrhage as a three-dimensional point cloud data structure. This can effectively extract the hematoma three-dimensional point cloud data representing the three-dimensional spatial position of the hematoma from the brain CT image. The three-dimensional point cloud data can directly represent the three-dimensional geometric shape and spatial position information of the hematoma, avoiding the information loss when using two-dimensional CT images. In the prognosis prediction of cerebral hemorrhage, the three-dimensional spatial position and morphological characteristics of the hematoma are crucial to the prognosis judgment. Therefore, by modeling the hematoma as a three-dimensional point cloud, this rich geometric information can be directly utilized, thereby providing an accurate and comprehensive data foundation for the prognosis prediction of cerebral hemorrhage. At the same time, during the processing of the hematoma three-dimensional point cloud data, offset correction and normalization operations are effectively eliminated to effectively eliminate the spatial inconsistency caused by imaging and individual differences, thereby improving the accuracy of the hematoma three-dimensional point cloud data. Ultimately, by using the hematoma three-dimensional point cloud data as the input of the subsequent prognosis prediction model, the model is able to utilize the hematoma's position information in three-dimensional space to capture the hematoma's three-dimensional spatial structure and local geometric change characteristics. The hematoma's spatial volume characteristics are associated with poor prognostic heterogeneity, thereby helping to improve the accuracy and comprehensiveness of cerebral hemorrhage prognosis prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0152] In order to make the purpose, technical solutions and advantages of the invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0153] Figure 1 This is the logical block diagram of the cerebral hemorrhage spatial positioning system based on three-dimensional point cloud.

[0154] Figure 2 This is the network structure diagram of the cerebral hemorrhage prognosis prediction network system based on three-dimensional point cloud.

[0155] Figure 3 This is the network structure diagram of the multi-scale point cloud attention module. DETAILED DESCRIPTION

[0156] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but only represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0157] It should be noted that similar reference numerals and letters denote similar items in the following figures. Therefore, once an item is defined in one figure, it does not require further definition or explanation in subsequent figures. In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer" indicate positions or relationships based on the positions or relationships shown in the figures, or the positions or relationships in which the inventive product is typically placed when in use. These terms are intended solely to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or components referred to must have a specific orientation, be constructed, or operate in a specific orientation, and are therefore not to be construed as limiting the present invention. Furthermore, the terms "first," "second," and "third," etc., are used solely to distinguish descriptions and are not to be construed as indicating or implying relative importance. Furthermore, terms such as "horizontal" and "vertical" do not imply that a component is absolutely horizontal or overhanging, but rather may be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but rather may be slightly tilted. In the description of the present invention, it should also be noted that, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; they may refer to mechanical connections or electrical connections; they may refer to direct connections or indirect connections through an intermediate medium; and they may refer to internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0158] The following is a further detailed description through specific implementation methods:

[0159] Example 1:

[0160] This embodiment discloses a spatial positioning system for cerebral hemorrhage based on three-dimensional point cloud.

[0161] like Figure 1 As shown, the spatial positioning system for cerebral hemorrhage based on three-dimensional point cloud includes:

[0162] An image acquisition module, used for acquiring brain CT images to be processed;

[0163] An offset correction module is used to perform offset correction on the brain CT image to obtain a corrected brain CT image;

[0164] A hematoma region segmentation model is used to segment the hematoma region of the rectified brain CT image to obtain the hematoma region;

[0165] like Figure 2 As shown in (A), the trained existing 3D UNet segmentation network is used as the hematoma region segmentation model to perform hematoma region segmentation.

[0166] A point cloud computing module is used to calculate the three-dimensional spatial coordinates of hematoma voxels in the hematoma area in the brain coordinate system based on the corrected brain CT image as point cloud data of the hematoma area;

[0167] In this embodiment, existing means are used to calculate the three-dimensional spatial coordinates of the hematoma voxels in the hematoma area in the brain coordinate system. The general process includes: generating a three-dimensional binary mask of the hematoma area, extracting the voxel indexes of all voxels marked as hematomas in the mask, combining the spatial positioning parameters in the DICOM file (including origin coordinates, voxel spacing, and direction matrix), and converting the voxel indexes into three-dimensional physical coordinates in the brain coordinate system through a coordinate mapping formula, and finally outputting a set of millimeter-level spatial coordinates corresponding to each hematoma voxel.

[0168] The point cloud data output module is used to normalize the point cloud data of the hematoma area to obtain three-dimensional point cloud data of the hematoma that represents the three-dimensional spatial position of the hematoma.

[0169] The present invention obtains a hematoma region by segmenting the brain CT image and calculating the three-dimensional spatial coordinates of the hematoma voxels in the hematoma region in the brain coordinate system, thereby modeling the hematoma of cerebral hemorrhage as a three-dimensional point cloud data structure. This can effectively extract the hematoma three-dimensional point cloud data representing the three-dimensional spatial position of the hematoma from the brain CT image. The three-dimensional point cloud data can directly represent the three-dimensional geometric shape and spatial position information of the hematoma, avoiding the information loss when using two-dimensional CT images. In the prognosis prediction of cerebral hemorrhage, the three-dimensional spatial position and morphological characteristics of the hematoma are crucial to the prognosis judgment. Therefore, by modeling the hematoma as a three-dimensional point cloud, this rich geometric information can be directly utilized, thereby providing an accurate and comprehensive data foundation for the prognosis prediction of cerebral hemorrhage. At the same time, during the processing of the hematoma three-dimensional point cloud data, offset correction and normalization operations are effectively eliminated to effectively eliminate the spatial inconsistency caused by imaging and individual differences, thereby improving the accuracy of the hematoma three-dimensional point cloud data. Ultimately, by using the hematoma three-dimensional point cloud data as the input of the subsequent prognosis prediction model, the model is able to utilize the hematoma's position information in three-dimensional space to capture the hematoma's three-dimensional spatial structure and local geometric change characteristics. The hematoma's spatial volume characteristics are associated with poor prognostic heterogeneity, thereby helping to improve the accuracy and comprehensiveness of cerebral hemorrhage prognosis prediction.

[0170] In order to better introduce the technical solution of the present invention, this embodiment is described through the following parts.

[0171] 1. Offset Correction Module

[0172] In the offset correction module, the offset correction performed on the brain CT image includes any one or more of binarization, corrosion, external rectangular frame generation, and translation and rotation correction; the specific process is: first, the image is binarized; then the corrosion operation is applied to remove noise; then the external rectangular frame of the hematoma area is generated; finally, the rectangular frame is aligned to the standard position through rotation and translation adjustment.

[0173] 1) Binarization

[0174] The brain CT image is binarized to separate the target area (i.e., brain tissue and hematoma area) from the background. The binarization process divides the pixel values in the brain CT image into two categories by setting a threshold T: the part greater than the threshold represents the target area, and the rest is the background.

[0175] The formula for binarization is expressed as:

[0176]

[0177] Where: I(x,y) is the grayscale value of the original brain CT image; I b (x, y) is the binarized brain CT image; T is the set threshold.

[0178] 2) Corrosion operation

[0179] After binarization, there may be noise or small stray points in the image. To remove this noise and enhance the target area, the present invention uses an erosion operation. The erosion operation processes the brain CT image by using a small structural element (such as a circular or square kernel) to reduce the boundaries of the white area and remove small interference points.

[0180] The formula for the corrosion operation is expressed as:

[0181]

[0182] Where: S is the structural element of the corrosion operation, I e (x,y) is the brain CT image after corrosion.

[0183] 3) Generate an external rectangular frame

[0184] After the corrosion process, the bounding rectangle of the target area needs to be determined by the contour extraction method; the bounding rectangle is the smallest rectangle that can completely contain the target area, and the coordinates of the bounding rectangle are (x min ,y min ) and (x max ,y max ), which represent the coordinates of the upper left corner and lower right corner of the rectangle, respectively. This step provides a reference frame for subsequent translation and rotation correction.

[0185] 4) Translation and rotation correction

[0186] Once the position and orientation of the bounding rectangle are determined, the rectangle needs to be rotated and translated to align its orientation with the standard coordinate axes. Specifically, first calculate the center point C of the bounding rectangle, whose coordinates are:

[0187]

[0188] By using the rotation matrix and displacement transformation, the bounding rectangle is rotated to align with the positive coordinate axis;

[0189] The rotation matrix R(θ) is defined as:

[0190] P new =R(θ)(PC)+C+t;

[0191] Where: P is the coordinate of any point in the circumscribed rectangular frame, t=(t x ,t y ) is the translation vector, P new are the coordinates of the point after rotation and translation.

[0192] 5) CT image rotation and translation applications

[0193] Apply rotation and translation transformations to the entire brain CT image to ensure that all pixels in the image are adjusted according to the same transformation rules. The image transformation formula is:

[0194] I corrected (x′,y′)=I(x,y);

[0195] Where: I corrected (x′, y′) represents the rectified brain CT image, and I(x, y) represents the original brain CT image. After image offset correction, the next step is to normalize the hematoma region point cloud data. At this point, the hematoma region has been accurately extracted and aligned from the CT image, providing high-quality basic data for subsequent spatial position extraction, voxelization, and spatial normalization.

[0196] 2. Normalization

[0197] After the CT image correction is completed, the point cloud data of the hematoma area can be normalized.

[0198] Let P i =(x i ,y i ,z i ) is the spatial coordinate of the i-th hematoma voxel in the brain coordinate system, where i∈{1,2,…,N}, and N represents the total number of voxels in the hematoma area. On this basis, in order to eliminate the spatial inconsistency caused by individual differences, equipment errors or different imaging methods, the point cloud data of the hematoma area will be standardized through normalization. The purpose of normalization is to map the point cloud of the hematoma area to a unified spatial scale and position coordinate system to ensure that the relative spatial position of the hematoma remains consistent between different data sets. The point cloud data of the hematoma area is defined as P={P1,P2,…,P N}, the point cloud centroid of the hematoma area is

[0199] The formula for normalization is expressed as:

[0200] P′ i =P i -Cfori∈{1,2,…,N};

[0201] Where: P i ’ is the three-dimensional coordinate of the normalized hematoma point cloud, and the center of mass C moves the center of gravity of the entire hematoma area to the origin of the coordinate system.

[0202] 3. Normal vector calculation

[0203] In order to further enrich the representation information of point cloud data and enhance the geometric features of the hematoma area in three-dimensional space, after obtaining the three-dimensional point cloud data of the hematoma, the present invention further calculates the normal vector information of the three-dimensional coordinates of each point in the three-dimensional point cloud data of the hematoma, and associates the normal vector information with the three-dimensional coordinates of the corresponding point.

[0204] The calculation of the normal vector can effectively reflect the local geometric shape and surface features of the point cloud, and thus play an important role in the spatial positioning of the hematoma area. By analyzing the normal vector of each point cloud, it can help identify and distinguish the different parts of the hematoma area, especially for hematoma areas with complex morphology. The normal vector provides a quantitative description of the local surface shape, which helps to distinguish the boundary, central area and local morphology of the hematoma. For example, the boundary of the hematoma usually appears as a sharper surface angle, while the central area may present a smoother surface. By analyzing the direction and size of the normal vector, the different morphological characteristics of these areas can be effectively distinguished. Especially when dealing with hematomas with complex or irregular shapes, the normal vector can provide more detailed local geometric information to support subsequent prognosis prediction tasks.

[0205] Specifically, the calculation of the normal vector is usually achieved by fitting the local plane using the least squares method. i ’ , select several points P in its neighborhood j ’ To construct a local plane, point P i ’ The neighborhood point cloud is {P j ’}, j∈{1,2,…,K} is the index of the neighborhood point; the normal vector n of the corresponding local plane is obtained by performing least squares plane fitting on the neighborhood points i ;

[0206] The formula is:

[0207]

[0208] Where: n i represents the normal vector to be found; P′ i is the three-dimensional coordinate of point i, P j ′ is the three-dimensional coordinate of point j in the neighborhood of point i; K=50.

[0209] The goal of the least squares fitting method is to make the normal vector n i The vertical distance between the point cloud and the plane is minimized. Subsequently, the normalized hematoma coordinates corresponding to each case are stored as a text file (txt) in a specified format, where each line represents the normalized three-dimensional coordinates of a point and its corresponding normal vector information, which can be used for subsequent prognosis prediction.

[0210] 4. Local Neighborhood Spatial Transformation Network

[0211] In the point cloud data output module, after obtaining the hematoma three-dimensional point cloud data, the hematoma three-dimensional point cloud data is further spatially transformed through the Local Neighborhood Spatial Transformation Network (LN-STN) to obtain the spatially transformed hematoma three-dimensional point cloud data; the local neighborhood spatial transformation network is designed to retain the position information and volume characteristics of the hematoma point cloud while enhancing the model's robustness to local geometric changes.

[0212] like Figure 2 As shown in (B) and (C), the local neighborhood spatial transformation network includes a local spatial transformation module for performing local enhancement operations on the hematoma 3D point cloud data, and a local neighborhood STN module for dynamically modeling the neighborhood features of each point in the hematoma 3D point cloud data.

[0213] 1. Local Space Transformation Module

[0214] The local spatial transformation module is used to perform local enhancement operations on the hematoma three-dimensional point cloud data, including the introduction of random noise, local translation, local rotation and random perturbation, to obtain the hematoma three-dimensional point cloud data after local spatial transformation; the purpose is to simulate the possible subtle changes in the hematoma morphology without changing the overall structure of the point cloud.

[0215] 1) Introducing random noise

[0216] Random noise with controllable amplitude is added to the spatial coordinates of the hematoma three-dimensional point cloud data. The noise amplitude is dynamically adjusted according to the size of the hematoma area to simulate the slight deviation in the brain CT imaging process while avoiding damage to key geometric features.

[0217] 2) Local translation

[0218] A random translation operation with an adaptive range is performed on the local area in the three-dimensional point cloud data of the hematoma. The translation range is dynamically adjusted according to the density and geometric complexity of the local area to enhance the adaptability of the model to hematomas of different shapes.

[0219] 3) Local rotation

[0220] Angle-restricted random rotation is performed around the centroid of the local area of the hematoma three-dimensional point cloud data. The rotation angle range is dynamically adjusted according to the geometric characteristics of the local area to enhance the network's adaptability to hematoma point clouds at different angles while avoiding geometric distortion caused by excessive rotation.

[0221] 4) Random perturbations

[0222] The point cloud of the local area of the hematoma 3D point cloud data is subjected to amplitude-controllable random perturbation. The perturbation amplitude is dynamically adjusted according to the geometric complexity of the local area to simulate possible morphological irregularities while retaining key geometric information.

[0223] The present invention performs local spatial transformation and data enhancement on hematoma point cloud data by introducing random noise, local translation, rotation and perturbation operations. These operations not only increase the diversity of the data, but also simulate the natural variation of hematomas in different patients, thereby improving the model's adaptability to hematomas of different shapes and locations and effectively preventing overfitting. At the same time, it can enhance the model's robustness to local geometric changes while retaining the key geometric features of the hematoma point cloud, thereby improving the accuracy of prognosis prediction.

[0224] 2. Local neighborhood STN module

[0225] After performing local spatial transformation, the local neighborhood STN dynamically models the neighborhood features of each point, enabling it to flexibly adapt to the changes in the different morphologies and spatial distributions of the hematoma point cloud while strictly preserving the overall spatial position and size information of the hematoma. Specifically, the processing steps of the local neighborhood STN module include:

[0226] S01: Three-dimensional point cloud data of hematoma after local space transformation (Where B is the batch size, C is the feature dimension of the point cloud, and N is the number of points) Through three consecutive one-dimensional convolutional layers, the dimension of the output feature is gradually increased from C to 64, 128, and 1024, and feature F3 is obtained;

[0227] The formula is:

[0228] F1 = ReLU(BN(Conv1d(X)));

[0229] F2 = ReLU(BN(Conv1d(F1)));

[0230] F3 = ReLU(BN(Conv1d(F2)));

[0231] Where: BN represents batch normalization operation; ReLU represents activation function; Conv1d represents one-dimensional convolution;

[0232] S02: Use global maximum pooling to aggregate the feature F3 in point dimensions to obtain the global feature vector

[0233] The formula is:

[0234] z = MaxPool(F3, dim = 2);

[0235] S03: Input the global feature z into three fully connected layers, gradually compressing its feature dimensions to 512, 256 and 9, and obtaining the transformation matrix in tiled form

[0236] The formula is:

[0237] g1 = ReLU(BN(W1z+b1));

[0238] g2 = ReLU(BN(W2g1+b2));

[0239] T=W3g2+b3;

[0240] Where: W1, W2, W3 and b1, b2, b3 are the weights and bias parameters of the fully connected layer;

[0241] S04: Reconstruct the transformation matrix T into a 3×3 affine transformation matrix A and add the identity matrix I;

[0242] The formula is:

[0243] A=Reshape(T)+I;

[0244] I = diag(1,1,1);

[0245] S05: To further limit the impact of the transformation on the volume size characteristics of the hematoma point cloud, the diagonal elements in the affine transformation matrix A are fixed to 1 to avoid scaling operations; range constraints are set for the non-diagonal elements (rotation and translation components) in the affine transformation matrix A. The formula is expressed as follows:

[0246] A ij =Clamp(A ij ,-θ max ,θ max ),i≠j;

[0247] Where: θ max =0.3 is the maximum rotation and translation amplitude;

[0248] S06: Performing local spatial transformation and alignment on the hematoma three-dimensional point cloud data after local spatial transformation using the affine transformation matrix A to obtain the hematoma three-dimensional point cloud data after spatial transformation.

[0249] In the present invention, LN-STN increases data diversity while retaining key information through local spatial transformation and neighborhood feature modeling, which helps to improve the generalization ability and robustness of the model and prevent overfitting.

[0250] Example 2:

[0251] This embodiment discloses a cerebral hemorrhage prognosis prediction network system based on three-dimensional point cloud.

[0252] like Figure 2 (B) shows a 3D point cloud-based network system for predicting the prognosis of cerebral hemorrhage, including:

[0253] An input layer, configured to take the (spatially transformed) hematoma three-dimensional point cloud data obtained by the point cloud data output module in the first embodiment as input;

[0254] The first processing module is used to input the hematoma three-dimensional point cloud data into the MLP and multi-scale point cloud attention module in sequence to model and fuse global features and local features to obtain initial point cloud features;

[0255] The feature transformation module is used to transform the initial point cloud features (through T-Net) to obtain the transformed point cloud feature map;

[0256] In this embodiment, the process of T-Net for point cloud feature transformation is as follows: first, a small sub-network is designed to encode and learn the input point cloud features through a fully connected layer or a convolutional layer; then the sub-network outputs a transformation matrix, which is obtained through orthogonality constraint training to ensure transformation stability; finally, the transformation matrix is multiplied by the original features to complete the feature space alignment operation and enhance the model's robustness to rigid transformations.

[0257] The second processing module is used to input the transformed point cloud feature map into the MLP and multi-scale point cloud attention module in sequence to model and fuse global features and local features to obtain secondary point cloud features;

[0258] A global multi-graph self-attention module is used to build graph structures by learning node relationships from various channel dimensions and capture complex spatial and semantic information in secondary point cloud features to obtain global features;

[0259] The classification module is used to classify global features through MLP and Softmax to obtain corresponding prognosis prediction results.

[0260] In this embodiment, the prognosis prediction result includes prediction labels such as good prognosis and poor prognosis.

[0261] By using hematoma three-dimensional point cloud data as input for subsequent prognostic prediction models, the present invention facilitates the model's utilization of the hematoma's positional information in three-dimensional space, capturing the hematoma's three-dimensional spatial structure and local geometric variation characteristics. The hematoma's spatial volumetric characteristics are associated with poor prognostic heterogeneity, thereby helping to improve the accuracy and comprehensiveness of intracerebral hemorrhage prognosis prediction. Furthermore, the introduction of a multi-scale point cloud attention mechanism and a multi-graph-guided global attention mechanism enables weighted processing and fusion of hematoma point cloud features from multiple dimensions. This efficient feature extraction and fusion mechanism not only enhances the expressive power of features, but also further improves the accuracy of prognostic prediction.

[0262] In order to better introduce the technical solution of the present invention, this embodiment is described through the following parts.

[0263] 1. Multi-scale Point Cloud Attention Module

[0264] The multi-scale point cloud attention module takes channel attention and multi-scale spatial attention as its core, and realizes the efficient modeling and fusion of global and local features. Figure 3 As shown in (A), the multi-scale attention module is specifically designed for processing point cloud features. It consists of two components: channel attention and multi-scale spatial attention. These components are used to capture key feature information from hematoma 3D point cloud data at different levels. Channel attention primarily operates on the high-dimensional feature vector of each point, aiming to deeply explore the relative importance of feature channels. It also suppresses the influence of redundant information by enhancing the representation of key information, thereby improving feature extraction accuracy and model expressiveness.

[0265] Specifically, the channel attention module extracts global feature information from different perspectives through global average pooling and maximum pooling. These two pooling operations can extract deep features of point cloud data at both local and global scales. These extracted features are then processed through a two-layer fully connected network with shared weights to better model the relative importance of channels and achieve precise weighting of information. The core advantage of this design is that it can effectively highlight key information while suppressing noise and redundant data, improving the accuracy of point cloud feature representation.

[0266] like Figure 3 As shown in (B), the processing steps of channel attention include:

[0267] S11: Input features (or hematoma 3D point cloud data) Perform global average pooling and global maximum pooling respectively to obtain the pooling feature X avg and X max ;

[0268] The formula is:

[0269] Xavg =AdaptiveAvgPool(X);

[0270] X max =AdaptiveMaxPool(X);

[0271] Where: B, C and N represent the batch size, the number of feature channels and the number of points respectively; AvgPool represents the global average pooling operation; MaxPool represents the global maximum pooling operation; Adaptive represents adaptive;

[0272] S12: Pooling feature X through two layers of shared fully connected network avg and X max Processing is performed to obtain feature X′ avg and X′ max ;

[0273] The formula is:

[0274] X′ avg =σ(W2·ReLU(W1·X avg ));

[0275] X′ max =σ(W2·ReLU(W1·X max ));

[0276] Where: W1 and W2 are the weights of the two fully connected layers respectively; σ represents the Sigmoid activation function;

[0277] S13: For feature X′ avg and X′ max Perform weighted fusion to obtain the channel attention weight A c ;

[0278] The formula is:

[0279] A c =X′ avg +X′ max ;

[0280] S14: Input feature X and channel attention weight A c Perform element-by-element multiplication to obtain the channel attention feature X c ;

[0281] The formula is:

[0282] X c =X⊙A c ;

[0283] Where: ⊙ represents the element-by-element multiplication operation.

[0284] Multi-scale spatial attention employs a multi-scale convolution kernel design, extracting multi-scale spatial information through deep convolutions with varying receptive fields. This process utilizes convolution kernels of varying scales to extract both local and global spatial information. This allows the model to not only identify tiny local geometric features when processing point cloud data, but also effectively capture details of the global spatial structure, significantly enhancing its ability to understand complex spatial relationships. This multi-scale design allows the model to flexibly adapt to input data of varying sizes, thereby improving the accuracy and robustness of feature representation.

[0285] like Figure 3 As shown in (C), the processing steps of multi-scale spatial attention include:

[0286] S21: Use depth convolution kernels of multiple scales of size 5×1, 7×1, 11×1 and 21×1 to focus on channel features X c Perform multi-scale feature extraction to obtain the features S of each scale i ;

[0287] The formula is:

[0288]

[0289] Where: Indicates that the kernel size is k i Depthwise separable convolution operation;

[0290] S22: The features S of each scale i After superposition, the spatial attention weight A is generated by channel mixing convolution s ;

[0291] The formula is:

[0292]

[0293] S23: Channel attention feature X c With the spatial attention weight A s After element-wise multiplication, it is then multiplied with the spatial attention weight A s Perform channel-level feature fusion to obtain the channel multi-scale spatial attention feature Y as the output of the multi-scale point cloud attention module;

[0294] The formula is:

[0295] X s =X c ⊙A s ;

[0296] Y = Conv(X s ⊙A s );

[0297] Where: Conv is the convolution operation for channel mixing.

[0298] In the present invention, the multi-scale point cloud attention mechanism performs weighted processing on features from two dimensions: channel and space, and can focus on the key feature areas of the hematoma, which is in line with the principle of focusing on important information in human vision and cognitive mechanisms, thereby improving the effectiveness of feature extraction.

[0299] 2. Global Multi-Image Self-Attention Module

[0300] After two multi-scale point cloud attention operations, the input data is fed into the multi-graph guided global attention module. This module employs a self-attention mechanism based on multi-graph reasoning. By learning graph node relationships from various channel dimensions, it aims to further construct a precise graph structure to capture the complex spatial and semantic information in the point cloud. Self-attention blocks are applied to the global feature space for efficient feature interaction and information fusion, significantly enhancing the expressive power of point cloud features. This mechanism enables the module to dynamically model long-range dependencies between different features and effectively improve the diversity and robustness of feature representation.

[0301] In the design of this module, a learnable adjacency matrix is introduced to better adapt to the complexity and irregularity of point cloud data. This matrix is used to initialize multiple graphs, thereby enhancing the diversity of node features. Learning the adjacency matrix not only provides a flexible mechanism for adjusting the graph's topology but also enables the graph to flexibly switch between multiple relationships, thereby obtaining a more fine-grained global feature representation in the hematoma point cloud feature extraction task. Furthermore, to further optimize the fusion of the self-attention mechanism with global information, the module makes special adjustments to the value. Specifically, in the design of the self-attention layer, the original position encoding component of the last attention layer is replaced, and the enhancement of neighboring information is intentionally ignored. This adjustment helps focus on the aggregation of global information and the interaction of relationships, rather than local adjacency information. This enables the model to more effectively capture long-range dependencies when processing hematoma point clouds with complex spatial relationships, improving the model's global understanding and accuracy.

[0302] like Figure 2 As shown in (D), the processing steps of the global multi-image self-attention module include:

[0303] S31: For the input secondary point cloud features (Where B is the batch size, C is the feature dimension of the point cloud, and N is the number of points), generate the query matrix Q, key matrix K, and value matrix V through one-dimensional convolution;

[0304] The formula is:

[0305] Q = Conv1D(X);

[0306] K = Conv1D(X);

[0307] V = Conv1D(X);

[0308] Where:

[0309] S32: Decompose the query matrix Q into a multi-head form Q through reshaping and grouping operations h ;

[0310] The formula is:

[0311]

[0312] Where: H is the number of long positions, C k The query channel dimension for each head;

[0313] S33: Through the Softmax normalization operation, the key matrix K is converted into the probability distribution of the key and the feature interaction is performed with the value matrix V, combined with Q h Generate initial global interaction result Y c ;

[0314] The formula is:

[0315] K v =K T V;

[0316] Y c =Q h ·K V ;

[0317] Where: It is the global preliminary interaction result, which captures the global relationship between each point;

[0318] S34: Define the graph structure as G = (V, ε), treat each feature point as a node of the graph, and the node features come from the value matrix V; multidimensional adjacency matrix are initialized as learnable parameters to represent various relationships between nodes; the edge ε is modeled by the learnable adjacency matrix A, which captures various semantic relationships (such as proximity and curvature correlation);

[0319] S35: Calculate the weighted connections between feature points through matrix multiplication to obtain the relationship matrix G of the graph;

[0320] The formula is:

[0321]

[0322] Where: each channel dimension of the adjacency matrix A corresponds to a global relationship type; the edge weight between the i-th node and the j-th node in the relationship matrix G is obtained by mapping the features of node i and node j through the adjacency matrix A;

[0323] S36: query matrix Q h Interact with the relationship matrix G to generate multi-graph reasoning results Y p ;

[0324] The formula is:

[0325] Y p =Q h G;

[0326] Where: The representation captures the global semantic relationships between nodes generated by graph reasoning;

[0327] S37: The global interaction result Y c and multi-graph reasoning result Y p Add together to get the output feature Y;

[0328] The formula is:

[0329] Y=Y c +Y p ;

[0330] S38: The output feature Y is added to the input secondary point cloud feature X through residual connection and batch normalization operations to enhance gradient fluidity, and then processed by activation function to obtain the global feature O;

[0331] The public notice is:

[0332] O=ReLU(W skip (Y)+X);

[0333] Where: W skip is the linear transformation matrix used for skip connections.

[0334] In this embodiment, the global features can also be saved as intermediate data representing the three-dimensional spatial position of the hematoma, that is, the saved intermediate data (data basis) is used as the input of the prognosis prediction model (at this time the prognosis prediction model is a classification model), and the prognosis prediction results are directly predicted by the prognosis prediction model (classification model).

[0335] In the present invention, the multi-graph guided global attention mechanism is based on graph theory and self-attention mechanism. By constructing graph structure and learning node relationships, it explicitly captures the global spatial positioning information of the hematoma in the skull, making up for the overall information that may be ignored by local feature analysis, further improving the model's understanding and prediction ability of complex hematoma data, and overall improving the accuracy and reliability of cerebral hemorrhage prognosis prediction.

[0336] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the technical solutions. Those skilled in the art should understand that modifications or equivalent replacements of the technical solutions of the present invention that do not depart from the purpose and scope of the technical solutions of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A cerebral hemorrhage spatial positioning system based on three-dimensional point cloud, characterized by: include: An image acquisition module, used for acquiring brain CT images to be processed; An offset correction module is used to perform offset correction on the brain CT image to obtain a corrected brain CT image; A hematoma region segmentation model is used to segment the hematoma region of the rectified brain CT image to obtain the hematoma region; A point cloud computing module is used to calculate the three-dimensional spatial coordinates of hematoma voxels in the hematoma area in the brain coordinate system based on the corrected brain CT image as point cloud data of the hematoma area; The point cloud data output module is used to normalize the point cloud data of the hematoma area to obtain three-dimensional point cloud data of the hematoma that represents the three-dimensional spatial position of the hematoma.

2. The cerebral hemorrhage spatial positioning system based on three-dimensional point cloud according to claim 1, characterized in that: In the point cloud data output module, after obtaining the three-dimensional point cloud data of the hematoma, the normal vector information of the three-dimensional coordinates of each point in the three-dimensional point cloud data of the hematoma is further calculated, and the normal vector information is associated with the three-dimensional coordinates of the corresponding point; For each point P in the hematoma 3D point cloud data i ’ , select several points P in its neighborhood j ’ To construct a local plane, point P i ’ The neighborhood point cloud is {P j ’ }, j∈{1,2,…,K} is the index of the neighborhood point; the normal vector n of the corresponding local plane is obtained by performing least squares plane fitting on the neighborhood points i ; The formula is: Where: n i represents the normal vector to be found; P ′ ′ is the three-dimensional coordinate of point i, P j ′ is the three-dimensional coordinate of point j in the neighborhood of point i.

3. The cerebral hemorrhage spatial positioning system based on three-dimensional point cloud according to claim 1, characterized in that: In the offset correction module, the offset correction performed on the brain CT image includes any one or more of binarization, erosion, external rectangular frame generation, and translation and rotation correction; 1) Binarization By setting a threshold T, the pixel values in the brain CT image are divided into two categories: the part greater than the threshold represents the target area, and the other part is the background; The formula for binarization is expressed as: Where: I(x,y) is the grayscale value of the original brain CT image; I b (x, y) is the binarized brain CT image; T is the set threshold; 2) Corrosion operation The erosion operation processes the brain CT image by using a structuring element; The formula for the corrosion operation is expressed as: Where: S is the structural element of the corrosion operation, I e (x,y) is the brain CT image after corrosion; 3) Generate an external rectangular frame The bounding rectangle is the smallest rectangle that can completely contain the target area. The coordinates of the bounding rectangle are (x min ,y min ) and (x max ,y max ), which represent the coordinates of the upper left corner and lower right corner of the rectangular box respectively; 4) Translation and rotation correction Calculate the center point C of the circumscribed rectangle, whose coordinates are: By using the rotation matrix and displacement transformation, the bounding rectangle is rotated to align with the positive coordinate axis; The rotation matrix R(θ) is defined as: P new =R(θ)(P-C)+C+t; Where: P is the coordinate of any point in the circumscribed rectangular frame, t=(t x ,t y ) is the translation vector, P new are the coordinates of the point after rotation and translation.

4. The cerebral hemorrhage spatial positioning system based on three-dimensional point cloud according to claim 1, characterized in that: In the point cloud data output module, the point cloud data of the hematoma area is P = {P1, P2, ..., P N }, the point cloud centroid of the hematoma area is The formula for normalization is expressed as: P′ i =P i -C for i∈{1,2,…,N}; Where: P′ i is the three-dimensional coordinate of the normalized hematoma point cloud.

5. The cerebral hemorrhage spatial positioning system based on three-dimensional point cloud according to claim 1, characterized in that: In the point cloud data output module, after obtaining the hematoma three-dimensional point cloud data, the hematoma three-dimensional point cloud data is further spatially transformed through the local neighborhood spatial transformation network to obtain the spatially transformed hematoma three-dimensional point cloud data; The local neighborhood spatial transformation network includes a local spatial transformation module for performing local enhancement operations on hematoma three-dimensional point cloud data, and a local neighborhood STN module for dynamically modeling the neighborhood features of each point in the hematoma three-dimensional point cloud data.

6. The cerebral hemorrhage spatial positioning system based on three-dimensional point cloud according to claim 5, characterized in that: A local space transformation module is used to perform local enhancement operations on the hematoma three-dimensional point cloud data, including introducing random noise, local translation, local rotation and random perturbation, to obtain the hematoma three-dimensional point cloud data after local space transformation; 1) Introducing random noise Random noise with controllable amplitude is added to the spatial coordinates of the hematoma 3D point cloud data. The noise amplitude is dynamically adjusted according to the size of the hematoma area to simulate the slight deviation in the brain CT imaging process. 2) Local translation Performing an adaptive random translation operation on the local area of the hematoma 3D point cloud data. The translation range is dynamically adjusted according to the density and geometric complexity of the local area. 3) Local rotation Perform angle-restricted random rotation around the centroid of the local area of the hematoma 3D point cloud data, and the rotation angle range is dynamically adjusted according to the geometric characteristics of the local area; 4) Random perturbations The point cloud of the local area of the hematoma three-dimensional point cloud data is subjected to amplitude-controllable random perturbation, and the perturbation amplitude is dynamically adjusted according to the geometric complexity of the local area.

7. The cerebral hemorrhage spatial positioning system based on three-dimensional point cloud according to claim 5, characterized in that: The processing steps of the local neighborhood STN module include: S01: Three-dimensional point cloud data of hematoma after local space transformation Through three consecutive one-dimensional convolutional layers, the dimension of the output feature is gradually increased from C to 64, 128 and 1024, and feature F3 is obtained; The formula is: F1 = ReLU(BN(Conv1d(X))); F2 = ReLU(BN(Conv1d(F1))); F3 = ReLU(BN(Conv1d(F2))); Where: BN represents batch normalization operation; ReLU represents activation function; Conv1d represents one-dimensional convolution; S02: Use global maximum pooling to aggregate the feature F3 in point dimensions to obtain the global feature vector The formula is: z = MaxPool(F3, dim = 2); S03: Input the global feature z into three fully connected layers, gradually compressing its feature dimensions to 512, 256 and 9, and obtaining the transformation matrix in tiled form The formula is: g1 = ReLU(BN(W1z+b1)); g2 = ReLU(BN(W2g1+b2)); T=W3g2+b3; Where: W1, W2, W3 and b1, b2, b3 are the weights and bias parameters of the fully connected layer; S04: Reconstruct the transformation matrix T into a 3×3 affine transformation matrix A and add the identity matrix I; The formula is: A = Reshape(T) + i; I = diag(1,1,1); S05: Set the diagonal elements in the affine transformation matrix A to be fixed to 1; set range constraints for the non-diagonal elements in the affine transformation matrix A, which is expressed as follows: A ij (Clamp(A). ij ,-θ max ,θ max ),i≠j. Where: θ max =0.3 is the maximum rotation and translation amplitude; S06: Performing local spatial transformation and alignment on the hematoma three-dimensional point cloud data after local spatial transformation using the affine transformation matrix A to obtain the hematoma three-dimensional point cloud data after spatial transformation.

8. A 3D point cloud-based network system for predicting the prognosis of cerebral hemorrhage, characterized by: include: An input layer, configured to take the three-dimensional point cloud data of the hematoma obtained by the point cloud data output module in claim 1 as input; The first processing module is used to input the hematoma three-dimensional point cloud data into the MLP and multi-scale point cloud attention module in sequence to model and fuse global features and local features to obtain initial point cloud features; The feature transformation module is used to perform feature transformation on the initial point cloud features to obtain the transformed point cloud feature map; The second processing module is used to input the transformed point cloud feature map into the MLP and multi-scale point cloud attention module in sequence to model and fuse global features and local features to obtain secondary point cloud features; A global multi-graph self-attention module is used to build graph structures by learning node relationships from various channel dimensions and capture complex spatial and semantic information in secondary point cloud features to obtain global features; The classification module is used to classify global features through MLP and Softmax to obtain corresponding prognosis prediction results.

9. The cerebral hemorrhage prognosis prediction network system based on three-dimensional point cloud according to claim 8, characterized in that: The multi-scale point cloud attention module in the first and second processing modules includes two parts: channel attention and multi-scale spatial attention, which are used to capture key feature information in the hematoma 3D point cloud data at different levels; 1) The processing steps of channel attention include: S11: Input features Perform global average pooling and global maximum pooling respectively to obtain the pooling feature X avg and X max ; The formula is: X avg =AdaptiveMaxPool(X); X max =AdaptiveMaxPool(X); Where: B, C and N represent the batch size, the number of feature channels and the number of points respectively; AvgPool represents the global average pooling operation; MaxPool represents the global maximum pooling operation; Adaptive represents adaptive; S12: Pooling feature X through two layers of shared fully connected network avg and X maX Processing is performed to obtain feature X ′ avg and X′ max ; The formula is: X′ avg =σ(W2·ReLU(W1·X avg )); X′ max =σ(W2·ReLU(W1·X max )); Where: W1 and W2 are the weights of the two fully connected layers respectively; σ represents the Sigmoid activation function; S13: For feature X′ avg and X′ max Perform weighted fusion to obtain the channel attention weight A c ; The formula is: A c =X′ avg +X′ max ; S14: Input feature X and channel attention weight A c Perform element-by-element multiplication to obtain the channel attention feature X c ; The formula is: X c =X⊙A c ; Where: ⊙ represents the element-by-element multiplication operation; 2) The processing steps of multi-scale spatial attention include: S21: Use depth convolution kernels of multiple scales of size 5×1, 7×1, 11×1 and 21×1 to focus on channel features X c Perform multi-scale feature extraction to obtain the features S of each scale i ; The formula is: Where: Indicates that the kernel size is k i Depthwise separable convolution operation; S22: The features S of each scale i After superposition, the spatial attention weight A is generated by channel mixing convolution s ; The formula is: S23: Channel attention feature X c With the spatial attention weight A s After element-wise multiplication, it is then multiplied with the spatial attention weight A s Perform channel-level feature fusion to obtain the channel multi-scale spatial attention feature Y as the output of the multi-scale point cloud attention module; The formula is: X s =X c ⊙A s ; Y=Conv(X s ⊙A s ); Where: Conv is the convolution operation for channel mixing.

10. The cerebral hemorrhage prognosis prediction network system based on three-dimensional point cloud according to claim 8, characterized in that: The processing steps of the global multi-image self-attention module include: S31: For the input secondary point cloud features Generate query matrix Q, key matrix K and value matrix V through one-dimensional convolution; The formula is: Q = Conv1D(X); K = Conv1D(X); V = Conv1D(X); Where: S32: Decompose the query matrix Q into a multi-head form Q through reshaping and grouping operations h ; The formula is: Where: H is the number of long positions, C k The query channel dimension for each head; S33: Through the Softmax normalization operation, the key matrix K is converted into the probability distribution of the key and the feature interaction is performed with the value matrix V, combined with Q h Generate initial global interaction result Y c ; The formula is: K V =K T ·V; Y c =Q h ·K V ; Where: It is the global preliminary interaction result; S34: Define the graph structure as G = (V, ε), treat each feature point as a node of the graph, and the node features come from the value matrix V; multidimensional adjacency matrix are initialized as learnable parameters to represent various relationships between nodes; the edge ε is modeled by the learnable adjacency matrix A; S35: Calculate the weighted connections between feature points through matrix multiplication to obtain the relationship matrix G of the graph; The formula is: Where: each channel dimension of the adjacency matrix A corresponds to a global relationship type; the edge weight between the i-th node and the j-th node in the relationship matrix G is obtained by mapping the features of node i and node j through the adjacency matrix A; S36: query matrix Q h Interact with the relationship matrix G to generate multi-graph reasoning results Y p ; The formula is: Y p =Q h ·G; Where: The representation captures the global semantic relationships between nodes generated by graph reasoning; S37: The global interaction result Y c and multi-graph reasoning result Y p Add together to get the output feature Y; The formula is: Y=Y c +Y p ; S38: Add the output feature Y to the input secondary point cloud feature X through residual connection and batch normalization operations to obtain the global feature O; The public notice is: O=ReLU(W skip (Y)+X)? Where: W skip is the linear transformation matrix used for skip connections.

Images