Blood vessel segmentation reconstruction and bleeding analysis method and device

Abstract

The invention belongs to the technical field of medical image processing, and discloses a blood vessel segmentation reconstruction and bleeding analysis method and device. In the bleeding area detection model, detecting whether a bleeding area exists or not; and performing blood vessel segmentation and blood vessel three-dimensional reconstruction operation. According to the method, the time sequence image is adopted for comprehensive analysis, and the problems that blood vessel details are lost due to insufficient diffusion of a contrast agent and the image is blurred due to excessive diffusion are effectively solved. Based on an 8-neighborhood pixel tracking method, accurately identifying blood vessel cross points, and performing segmentation processing on the skeleton line by using the cross points. By constructing the topological structure of the blood vessel, the feature point matching efficiency is greatly improved, and the automatic three-dimensional reconstruction process is also realized. By introducing the architecture design of the hybrid convolutional neural network and the state space model, the local fine-grained feature extraction capability and the global long-range dependence modeling capability are effectively combined, so that the segmentation precision and robustness of the intracranial hemorrhage region are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a method and device for blood vessel segmentation, reconstruction and bleeding analysis. Background Art

[0002] With the rapid development of modern medical imaging technology, especially the widespread application of digital subtraction angiography (DSA), it has provided important technical support for clinical diagnosis and treatment. DSA angiography plays a key role in vascular disease detection, surgical planning and postoperative evaluation.

[0003] In clinical practice, various medical imaging technologies are widely used in the diagnosis and treatment of cerebrovascular diseases (such as aneurysms and cerebral hemorrhage), peripheral vascular diseases (such as lower limb artery occlusive disease) and cardiovascular diseases (such as coronary artery stenosis). Taking cerebrovascular disease as an example, doctors usually need to locate and analyze diseased blood vessels based on medical images, and rely on 3D reconstruction technology to generate vascular models to assist in formulating surgical strategies.

[0004] The technical solution described in the Chinese patent "CN117237536A A method for three-dimensional reconstruction of blood vessels based on image segmentation and centerline extraction" is as follows: obtain the medical image data to be reconstructed through medical imaging equipment as the basic material for three-dimensional reconstruction; then use the UNet neural network architecture to build an image segmentation model, and after multiple rounds of optimization training, achieve accurate segmentation of the medical image and extract the target area. The segmentation results are processed and converted by coordinates to ensure that they meet the coordinate system requirements of subsequent processing, and generate blood vessel surface slice data to accurately reflect the external morphology of the blood vessel. Finally, the centerline of the blood vessel is calculated by the Voronoi diagram algorithm, and the entry and exit points of the blood vessel are extracted to complete the extraction and three-dimensional reconstruction of key position information. This process realizes a complete closed loop from data acquisition to three-dimensional structure reconstruction, providing a high-precision modeling method for morphological analysis and functional research of blood vessels. However, it has the following problems: first, it is not flexible enough to meet the adjustment requirements in specific scenarios; second, the segmentation model and three-dimensional reconstruction process have high requirements for computing resources, including high-performance computers and large-capacity storage devices, which limits its applicability in resource-limited environments. In addition, the segmentation model has limited generalization capabilities on different data sets, which may lead to inaccurate segmentation or missed detections, requiring manual intervention for adjustment, reducing the degree of automation and actual efficiency.

[0005] The technical solution described in the Chinese patent "CN116894826A A method, device, electronic device and storage medium for vascular image recognition" is as follows: the image to be recognized is generated by preprocessing the original medical image such as normalization, ROI region segmentation, image cropping and data enhancement; then the vascular recognition model is used to extract vascular pixels through up and down sampling and attention mechanism to generate a vascular segmentation map; on this basis, the vascular segmentation map is introduced as prior knowledge, and the arteriovenous pixels are predicted by feature extraction and fusion technology to generate an arteriovenous segmentation map; finally, the construction of the three-dimensional structure of arteries and veins is completed through post-processing steps such as hole filling, maximum connected domain processing and surface rendering reconstruction, providing a high-precision model for medical research and diagnosis. However, it has the following problems: arteriovenous recognition is highly dependent on the results of vascular segmentation, and the early errors are easily magnified; the quality of preprocessing such as normalization and ROI segmentation directly affects the model effect; the multi-stage process is complex and difficult to implement; the computing resource requirements are high and the efficiency is low in resource-constrained environments; in addition, when dealing with complex vascular networks, the recognition speed and reconstruction integrity still need to be optimized.

[0006] The technical solution described in Chinese patent CN112614145B is as follows: obtain and preprocess intracranial hemorrhage CT images, use some preprocessed images as training samples; use training samples to train deep convolutional neural networks to obtain trained models; input the preprocessed images into the model for segmentation, generate hemorrhage area segmentation results, and display them through a GUI interface. This method has the following problems: First, this method mainly relies on convolution operations for local feature extraction. Although it performs well in capturing fine-grained features, it has limited global information modeling capabilities and is difficult to effectively handle long-range dependencies between hemorrhage areas, especially when the hemorrhage area distribution range is large or the morphology is complex, which may lead to a decrease in segmentation accuracy.

[0007] Medical image analysis technology is a key part of modern medical diagnosis and treatment, and plays a vital role in early detection, precise positioning and formulation of treatment plans for diseases. With the rapid growth of medical image data and the development of deep learning technology, how to efficiently and accurately analyze and process high-dimensional medical images has become a hot research topic in the field of intelligent medicine, and it is also the focus and difficulty.

[0008] As an important imaging technology, DSA technology performs subtraction processing on vascular images after intravenous injection of contrast agents, thereby providing high-resolution vascular images for clinical use. This technology is widely used in the diagnosis and treatment of cerebrovascular diseases (such as aneurysms, cerebral hemorrhage), peripheral vascular diseases (such as arteriosclerosis obliterans) and cardiovascular diseases (such as coronary artery stenosis). However, in traditional vascular feature point reconstruction algorithms, the acquisition and matching of feature points usually rely on manual selection, which is inefficient and greatly affected by human factors, resulting in reconstruction accuracy that cannot meet the needs of practical applications. Therefore, three-dimensional reconstruction technology based on vascular segmentation and automatic matching of feature points has become a hot topic in current vascular segmentation and reconstruction research, and is also the focus and difficulty of this field.

[0009] First, most of the existing blood vessel segmentation algorithms in China still rely on traditional geometric methods or some open source tools. Such methods have high requirements for data preprocessing and are difficult to adapt to the modeling requirements of high-resolution medical images. They also have a low degree of automation and rely on a lot of manual operations, resulting in low efficiency in practical applications. Although foreign studies have applied deep learning methods to the fields of blood vessel segmentation and three-dimensional reconstruction earlier, most algorithms have problems with low computational efficiency and high hardware resource requirements when processing high-resolution DSA images. At the same time, due to the strong dependence of deep learning models on labeled data, and the extremely high cost of acquiring and labeling medical imaging data, this has limited the versatility and promotion of foreign technologies to a certain extent. In addition, some foreign technologies are mostly developed based on specific scenarios or special equipment, and are highly dependent on image acquisition equipment. The migration ability and universality of the model are poor, making it difficult to meet the needs of diversified clinical applications.

[0010] Secondly, most of the existing medical image segmentation methods in China are based on traditional convolutional neural networks (CNNs). These methods are effective in extracting local features, but often perform poorly when dealing with large-scale, complex lesion areas, and it is difficult to accurately model global dependencies. In addition, some methods convert three-dimensional medical images into two-dimensional slices when processing them, which often leads to loss of spatial information and affects the consistency of segmentation results. In contrast, although foreign research has made some progress in introducing the Transformer architecture and multi-scale feature fusion, such methods are usually computationally complex and require large hardware resources, making it difficult to achieve high efficiency and lightweight in actual clinical applications. Summary of the invention

[0011] The first object of the present invention is to provide a comprehensive solution based on vascular segmentation and three-dimensional reconstruction. First, in the process of vascular segmentation, the present invention no longer relies on a single digital subtraction angiography image, but uses time series images for analysis. By comprehensively evaluating images at different times, the loss of details of the vascular end caused by insufficient diffusion of the contrast agent and the image blurring caused by excessive diffusion of the contrast agent are avoided. The adaptive threshold binarization method and the region growing algorithm are used to generate a high-precision vascular segmentation map. Secondly, in terms of three-dimensional reconstruction, the present invention traverses the vascular skeleton line, uses the 8-neighborhood pixel tracking method to accurately identify the intersection and segment the skeleton line accordingly, so that each segment of the blood vessel contains only two bifurcation points at the starting point and the end point, thereby constructing a vascular topology structure that is easy to store and process, and laying the foundation for automated feature point matching. The feature point matching adopts a hierarchical strategy, firstly aligning the vascular intersection through bifurcation point matching, and then combining the vascular topology structure to achieve segment-level matching. Finally, the present invention realizes automated three-dimensional reconstruction through the information of the vascular topology structure, gets rid of the dependence of the traditional method on the predefined vascular structure, and can flexibly adapt to the morphological changes of complex vascular networks. The present invention achieves full process optimization from blood vessel segmentation to three-dimensional reconstruction, which not only improves the accuracy and efficiency of segmentation and reconstruction, but also provides more reliable technical support for the diagnosis and treatment of clinical vascular lesions.

[0012] The second purpose of the present invention is to solve the problem that the traditional deep convolutional neural network has strong local feature extraction ability but insufficient global information modeling ability in the three-dimensional intracranial hemorrhage CT image segmentation task, as well as the defects of high computational complexity and low efficiency of existing methods when processing high-dimensional medical image data. By introducing the innovative hybrid convolutional neural network (CNN) and state space model (SSM) architecture design, it aims to take into account the modeling of local fine-grained features and global long-range dependencies, and comprehensively improve the accuracy and robustness of bleeding point detection. At the same time, the method optimizes the computational complexity, enabling it to efficiently process three-dimensional high-resolution medical imaging data, thereby reducing the demand for computing resources, and providing an efficient and practical new solution for three-dimensional medical image tasks, promoting its application in actual clinical scenarios.

[0013] The technical solution of the present invention is as follows: a blood vessel segmentation, reconstruction and bleeding analysis method, wherein an image to be detected is input into a bleeding area detection model to detect whether there is a bleeding area; if there is no bleeding area, the blood vessel segmentation and blood vessel three-dimensional reconstruction operations are directly performed; if there is a bleeding area, the blood vessel segmentation and blood vessel three-dimensional reconstruction operations are performed after hemostasis treatment;

[0014] The blood vessel segmentation and blood vessel three-dimensional reconstruction operations are specifically as follows: using time series image analysis to generate a blood vessel segmentation map; obtaining a blood vessel skeleton line based on the blood vessel segmentation map, traversing the blood vessel skeleton line, identifying intersections and segmenting the blood vessel skeleton line according to the intersections, each blood vessel segment only contains two bifurcation points, a starting point and an end point, to form a blood vessel topology structure; based on the blood vessel topology structure, obtaining the three-dimensional coordinates of the feature points and the blood vessel radius of the feature points, and reconstructing the blood vessel model.

[0015] The bleeding area detection model is designed based on the convolutional neural network (CNN) and the state space model (SSM) architecture;

[0016] The bleeding area detection model is mainly composed of an encoder and a decoder;

[0017] The encoder is a Mamba encoder; the features output by the Mamba encoder gradually reduce the resolution of the feature map through a 3D maximum pooling operation; in each layer of the Mamba encoder, the extracted local features are further processed by an embedded Mamba module to model global features and local features;

[0018] The decoder restores the resolution through deconvolution operation, and integrates the high-resolution features in the encoder while fusing local convolution features and global state space features to gradually reconstruct the segmented image; the prediction result is a voxel-level segmentation probability map with a shape of (B, N, D, H, W), where N is the number of categories, indicating the probability of each pixel belonging to each category.

[0019] The Mamba module is embedded in each layer of the encoder; each Mamba module includes two branches, namely an SSM branch and a CNN branch; the SSM branch is composed of a plurality of Mamba layers cascaded, and the CNN branch is composed of a plurality of convolutional layers;

[0020] The Mamba layer is used to capture global dependencies and model long-distance dependencies through the state space model SSM;

[0021] The outputs of the two branches are added together, and the local convolution features extracted by the CCN branch are combined with the global SSM features extracted by the SSM branch. The combined features are added to the features input by the Mamba module through a skip connection, retaining the original information of the input features and input to the decoder.

[0022] When traversing the blood vessel skeleton line, the 8-neighborhood pixel tracking method is used to extract the features of the bifurcation point, and each point on the blood vessel skeleton line is traversed for judgment; the pixel differences of adjacent points in the 8-neighborhood pixels are calculated in a clockwise manner, and when the sum of the absolute values ​​of these pixel differences is 6, the point is the bifurcation point of the blood vessel; the specific steps are as follows:

[0023] Determine the address offsets of the 8 points around the point to be judged: (-1,-1), (-1,0), (-1,1), (0,1), (1,1), (1,0), (1,-1), (0,-1);

[0024] Calculate the sum of the absolute values ​​of the pixel differences between any two adjacent points of these 8 points;

[0025] If the sum is 6, then this point is a bifurcation point, otherwise continue to judge;

[0026] After obtaining the bifurcation points of the vascular skeleton under different camera perspectives, the bifurcation points are matched according to the epipolar geometry principle; according to the epipolar constraint formula: Get the relationship between the two corresponding bifurcation points, the pixel coordinates p of the bifurcation point in the first image 1 Perform matrix multiplication with the basic matrix F, and then multiply it with the corresponding bifurcation point pixel coordinate p in the second image. 2 Perform multiplication operation, and the result of 0 represents successful matching; the basic matrix F is obtained by the essential matrix E and the camera intrinsic parameter K through the formula: F = K T EK -1 The essential matrix E is obtained by the eight-point method, and the essential matrix E is estimated by four pairs of known matching points, or by using the displacement matrix t and the rotation moment R through E = t Λ R obtains, t Λ is the antisymmetric matrix of t; the specific matching process is:

[0027] Cycle through the intersections in the two images, take the vascular skeleton line from one camera’s perspective as a reference, and use the intersections in the vascular skeleton line from the other perspective to perform epipolar constraint judgment with the intersections in the former. If satisfied, the match is successful, if not satisfied, continue matching. When the calculation result is not 0, select the pair with the smallest value as the matching point.

[0028] The extraction of vascular line segment points also uses the 8-neighborhood pixel tracking method, and the judgment is made based on whether the sum of the absolute values ​​of the pixel differences is equal to 4; matching: the vascular skeleton line is segmented according to the bifurcation points that have been matched; each line segment takes two different bifurcation points as the starting point and end point, and the line segments are matched according to the matching results of the bifurcation points to obtain the vascular line segment points.

[0029] The specific process of segmentation is as follows:

[0030] Define the search direction in 8 directions. The dx and dy arrays store the horizontal and vertical movement directions respectively. The directions are defined as 8 directions of up, down, left, right and diagonal.

[0031] Traverse and find the extension direction of the bifurcation point; point represents the current bifurcation point, traverse the 8 directions around the bifurcation point, and check whether the adjacent pixels are skeleton pixels, and the skeleton pixel back.at(x,y)==255. If so, it means that there is a line segment extending in this direction; for the diagonal direction, additionally check the two adjacent non-diagonal direction points to ensure that there is no repeated path; if the diagonal direction point is valid, that is, there is no repeated path, then add the diagonal direction point to the direction vector and mark the direction point as visited;

[0032] For each valid direction point d found, start to construct a line segment; the first point of the line segment is the current bifurcation point point, and the second point is d, which represents the extension direction found; during the extension process, move forward through the loop to find the next valid direction point. If the next point is found, add it to the line vector and update the auxiliary image used to track the processed area; when extending, first check whether the current point is a bifurcation point. If it is a bifurcation point: mark it as a bifurcation point, return, and end the extension of the line segment; otherwise, continue to extend to the next point until it cannot be extended;

[0033] When it is detected that the end point of a line segment is another bifurcation point, the end point is recorded as the starting point of a new line segment. If it extends to an end point without a bifurcation point, the end point is added to points and the drawing of the line segment ends; all line segments are segmented skeleton line segments;

[0034] Matching of the overall line segment;

[0035] The starting point of the segmented skeleton line segment is the bifurcation point, and the end point is the bifurcation point or the endpoint. After the epipolar constraint judgment is performed on the endpoint, the endpoint matching pair of the two images is obtained;

[0036] For the matched line segment pairs, the line segment with more points is used as the reference line segment, and the line segment with fewer points is matched on the reference line segment, and the matching points are selected using the epipolar constraint.

[0037] The process of obtaining the three-dimensional coordinates of the feature points and the blood vessel radius of the feature points is as follows:

[0038] Find the connection between pixel coordinates and real coordinates based on the camera model when acquiring the image;

[0039]

[0040] After obtaining the camera's intrinsic parameter K and the distance Z from the real point to the optical center, the three-dimensional coordinates of the pixel point in the real world are obtained; P = [X, Y, Z] is the world coordinate of the blood vessel point, u and v are the pixel coordinates of the blood vessel point, which are the pixel positions in the horizontal direction and the vertical direction, respectively, and f x 、fy is the focal length of the camera in pixels on the X and Y axes, c x 、c y is the coordinate of the optical center in the image coordinate system;

[0041] The vascular wall in the vascular segmentation image is obtained according to the edge detection algorithm, and the vascular radius corresponding to the matching point is obtained using the vascular wall and the matching point; for a certain matching point, the shortest distance to the vascular wall is searched on both sides of the left and right vascular walls, and this distance is used as the vascular radius of the bifurcation point or the vascular radius of the line segment point.

[0042] The reconstructed blood vessel model is constructed by sequentially modeling the line segments using topological sorting according to the matched bifurcation points or line segment points and the corresponding blood vessel radius;

[0043] The vascular topology structure starts from the identified bifurcation point and extends outward along the line segment point, stops at other bifurcation points or vascular endpoints and records the stop point as an extension direction of the bifurcation point, all directions are extended, the current bifurcation point completes the extension operation, and then extends from the next bifurcation point until all bifurcation points are extended;

[0044] The process of 3D reconstruction of blood vessels is as follows:

[0045] Initialization: calculate the in-degree of all nodes and whether there is a bifurcation point pointing to that point;

[0046] Select the starting point: add all nodes with in-degree 0 to the queue;

[0047] Processing nodes: taking a node from the queue each time, adding it to the topological sequence, and building the blood vessel model for that segment;

[0048] Update in-degree: When the in-degree of a node becomes 0, add it to the queue;

[0049] Repeat: until the queue is empty; the reconstruction of the vascular model is completed.

[0050] A blood vessel segmentation, reconstruction and bleeding analysis device comprises a bleeding analysis module, which is used to detect whether there is a bleeding area in an image to be detected; if there is a bleeding area, the bleeding area is located and marked, and after the hemostasis treatment is completed, the image of the patient rescanned is transmitted to the blood vessel segmentation and blood vessel reconstruction module; if no bleeding area is detected, the image is directly transmitted to the blood vessel segmentation module.

[0051] The blood vessel segmentation module is used to perform blood vessel segmentation operations on the input image and generate a blood vessel segmentation map; extract the blood vessel skeleton line, identify the bifurcation points and line segment points through the 8-neighborhood pixel tracking method, and establish the blood vessel topology structure.

[0052] The vascular reconstruction module is used to reconstruct the vascular three-dimensional model through a topological sorting method according to the vascular topological structure and the matched bifurcation points or line segment points, combined with their corresponding vascular radius, to generate a vascular three-dimensional structural model.

[0053] Beneficial effects of the present invention: The present invention provides a method for blood vessel segmentation, reconstruction and bleeding analysis, which achieves improvements in segmentation accuracy, modeling efficiency, dynamic update capability and system integration. First, in the process of blood vessel segmentation, the present invention uses time series images for comprehensive analysis, and realizes the segmentation of the vascular network through adaptive threshold binarization and region growing algorithm. Compared with the traditional single image segmentation method, the present invention effectively avoids the loss of vascular details due to insufficient diffusion of contrast agents and the image blurring problem caused by excessive diffusion, thereby generating a segmentation result with complete structure and strong coherence, laying a solid foundation for subsequent processing.

[0054] In terms of three-dimensional reconstruction, the present invention traverses the skeleton line, accurately identifies the intersection of blood vessels based on the 8-neighborhood pixel tracking method, and uses the intersection points to segment the skeleton line. By constructing the topological structure of the blood vessels, not only the efficiency of feature point matching is greatly improved, but also an automated three-dimensional reconstruction process is realized, getting rid of the dependence on predefined vascular structures and manual intervention in traditional methods, improving the efficiency and adaptability of modeling, and thus providing more reliable data support for clinical diagnosis, surgical planning and efficacy evaluation. The present invention also avoids the complexity of relying on multi-tool collaborative operations in traditional methods, and significantly improves the convenience and work efficiency of user operations. Users can efficiently complete the entire process from image input to model output, greatly reducing manual participation and technical difficulty, and improving ease of use in practical clinical applications.

[0055] The present invention introduces the architecture design of hybrid convolutional neural network (CNN) and state space model (SSM), effectively combining the local fine-grained feature extraction capability and the global long-range dependency modeling capability, so that the segmentation of intracranial hemorrhage area has achieved significant improvement in accuracy and robustness. Especially in hemorrhage areas with complex morphology and large distribution range, the present invention can provide more accurate detection results and significantly reduce the cases of false positive and false negative. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Provide an overall technical roadmap for vascular segmentation;

[0057] Figure 2 This is the normalized vascular image;

[0058] Figure 3 is the variance distribution diagram;

[0059] Figure 4 is the grayscale image transformed by variance;

[0060] Figure 5 It is a grayscale image after adaptive thresholding and binarization;

[0061] Figure 6 This is the effect diagram of blood vessel segmentation;

[0062] Figure 7 Provide a technical roadmap for vascular feature point matching and vascular 3D reconstruction;

[0063] Figure 8 Extract graph for bifurcation point;

[0064] Fig. 9 is the bifurcation point matching graph;

[0065] Fig.10 is a blood vessel segmentation diagram; (a)-(b) are schematic diagrams of different blood vessel segmentation;

[0066] Fig.11 This is a reconstruction rendering;

[0067] Fig.12 This is the overall flow chart of the bleeding area algorithm;

[0068] Fig.13 This is a structural diagram of the bleeding area detection model;

[0069] Fig.14 The following are the results of bleeding area detection; (a) and (c) are labels, and (b) and (d) are the results predicted by the model. DETAILED DESCRIPTION

[0070] 1. Bleeding area detection model Fig.13 As shown, the bleeding area detection model is used to detect whether there is a bleeding area; if there is no bleeding area, the blood vessel segmentation and blood vessel three-dimensional reconstruction operations are directly performed, and the reconstructed blood vessel model is used for subsequent processing; if there is a bleeding area, the blood vessel segmentation and blood vessel three-dimensional reconstruction operations are performed after the hemostasis treatment, and the reconstructed blood vessel model is used as a basis for subsequent judgment of the hemostasis result;

[0071] The overall process of bleeding area detection is as follows:

[0072] First, the input image goes through a preprocessing step, including normalization, resizing, and data augmentation. Then, the model training phase begins, where the model parameters are optimized through back-propagation using an annotated dataset. Once the model is trained, the performance is evaluated in the validation phase, usually using metrics such as Intersection over Union (IoU) and Dice coefficient. After that, it is applied to the test data for inference, and finally, the final segmentation result is generated through post-processing steps such as thresholding or edge smoothing.

[0073] The overall flow chart of the bleeding area detection algorithm is as follows: Fig.12 As shown:

[0074] 1) Dataset preparation:

[0075] The dataset used in the experiment is the INSTANCE challenge dataset, which includes 200 non-enhanced CT data of intracranial hemorrhage. These intracranial hemorrhage types include epidural, subdural, intraventricular, intraparenchymal, and subarachnoid.

[0076] We use 100 samples from the INSTANCE dataset, which are derived from the training set of the MICCAI2022Challenge, and another 11 CT images accurately annotated by experts. In the two datasets used, the original size of each CT image is 512×512 pixels in the cross section, and the slice thickness in the z-axis direction is 5 mm.

[0077] To make a convincing comparison, we extracted and randomized the slices in each dataset and performed 5-fold cross validation on this basis for both comparative and ablation experiments.

[0078] 2) Data preprocessing:

[0079] Image preprocessing is a key step in medical image analysis, which aims to reduce noise interference and improve the efficiency and accuracy of subsequent processing. By performing a series of preprocessing operations on the image, including binarization threshold processing, morphological operations, maximum connected domain processing, data normalization and image enhancement, the quality of the image can be effectively enhanced, providing more reliable input for subsequent analysis and segmentation tasks.

[0080] 3) Binarization threshold processing:

[0081] Brain images are usually composed of different types of tissues, such as gray matter, white matter, and cerebrospinal fluid. These tissues have different grayscale values ​​in the image. By observing the brain images, we can find that the grayscale value range of most brain tissues is usually in the lower range, such as between 0 and several hundred. Therefore, choosing an appropriate threshold can effectively distinguish brain tissue from the background. Finally, we set the lower limit to 100 and the upper limit to 5000 to achieve binarization.

[0082] 3) Morphological operations:

[0083] The open operation is often used in morphological processing to remove small interfering objects in a small area, that is, the process of corrosion followed by expansion. Therefore, this operation can remove the noise in the small edges and connected areas of the binary image, thereby cleaning up the image and improving the accuracy and stability of subsequent processing. In medical image processing, this preprocessing operation is particularly useful because medical images are often affected by noise, and the open operation can help reduce this noise and retain important structures in the image. Therefore, by performing an open operation on the binary image, the region of interest in the brain image can be accurately extracted and the error in subsequent processing can be reduced.

[0084] 4) Maximum Unicom domain processing:

[0085] In medical image processing, extracting the largest connected component is often used to extract the region of interest, such as the brain or other organ regions, for subsequent analysis and processing. By extracting the largest connected component from the input binary image and returning a binary image containing only this connected component, this can reduce the amount of computation and complexity of subsequent processing, ensure the accuracy of subsequent processing, and simplify the task of image analysis. Cropping and resizing images

[0086] The image and label are cropped by obtaining the bounding box coordinates of the largest connected component in the input binary image. Then, the image is resized to a specific size by resampling the image so that it can be input into the model for training and prediction.

[0087] 5) Data normalization:

[0088] Data normalization is a common step in image preprocessing. Its purpose is to reduce the differences in brightness, contrast, etc. between different images by standardizing the image pixel values. We count the pixel values ​​of all images in the dataset, calculate the global mean and variance, and standardize the pixel values ​​of each CT image to zero mean and unit variance. This operation helps to accelerate the convergence of the model and improve the stability of training.

[0089] 6) Data Augmentation:

[0090] In order to enhance the robustness and generalization ability of the model, we used a series of data augmentation methods. These enhancement methods include:

[0091] Gaussian blur: By applying a blur effect to the image, it simulates the blurred image boundary and improves the model's adaptability to blurred images.

[0092] Random pixel value multiplication: By adjusting the brightness and contrast of the image, it simulates the changes under different exposure conditions and improves the lighting adaptability of the model.

[0093] Random Angle Rotation: Randomly rotate the image to enhance the model's adaptability to rotation transformations.

[0094] Axial mirror flipping: By axially flipping the image, data with enhanced symmetry is generated, expanding the model's ability to recognize multi-directional features.

[0095] Gamma Correction: Adjusts image contrast to simulate different visual effects and exposure conditions.

[0096] Random Cropping and Scaling: Randomly cropping and scaling images to simulate local view changes and enhance the model's ability to adapt to scale changes.

[0097] These data enhancement methods help the model improve its adaptability to different scenes and image features by diversifying training data, thereby improving its generalization performance, reducing overfitting, and thus improving the actual application effect of the model.

[0098] Bleeding area detection model model related principles:

[0099] Mamaba:

[0100] State Space Sequence Models (SSMs) are a type of model that transforms a one-dimensional function or sequence into The mapped system can be expressed by the following linear ordinary differential equation (ODE):

[0101] x ′ (t)=Ax(t)+Bu(t),

[0102] y(t)=Cx(t)

[0103] Among them, the state matrix A∈R N×N ,B,C∈R N are the parameters of the model, x(t)∈R N Indicates an implicit potential state.

[0104] SSMs have several desirable properties, such as linear computational complexity per time step and parallel computing capabilities, which enable efficient training. However, original SSMs usually require more memory than equivalent CNNs and are prone to the gradient vanishing problem during training, limiting their widespread application in general sequence modeling.

[0105] The bleeding area detection model is a structured state space sequence model: the original SSMs are significantly improved by imposing a structured form on the state matrix and introducing an efficient algorithm. Specifically, the state matrix is ​​constructed and initialized using a high-order polynomial projection operator (HIPPO), so that a deep sequence model with rich capabilities and efficient long-range reasoning can be constructed. As a new network architecture, the bleeding area detection model surpasses Transformers by a significant margin on the challenging long-range reasoning benchmark (LongRangeArenaBenchmark).

[0106] Recently, Mamba has further advanced the capabilities of SSMs in modeling discrete data (e.g., text and genomic sequences) through two key improvements:

[0107] Input related selection mechanism:

[0108] Mamba introduces an input-based selection mechanism, which is different from traditional time- and input-invariant SSMs. This mechanism enables the model to efficiently filter information from the input by parameterizing the SSM parameters with the input data.

[0109] Hardware-aware algorithms:

[0110] Mamba developed a hardware-friendly algorithm that scales linearly with the sequence length, recursively computing the model through a scan mechanism, making it run faster than previous methods on modern hardware.

[0111] In addition, the Mamba architecture is more concise by combining SSM blocks with linear layers, and has demonstrated state-of-the-art performance in a variety of long sequence fields including language and genomics. Mamba also has significant computational efficiency during training and inference.

[0112] In the present invention, the bleeding area detection model is based on an innovative hybrid convolutional neural network (CNN) and state space model (SSM) architecture design. Thanks to this hybrid architecture, the model can not only accurately capture local fine-grained features in the image, but also effectively model long-range dependencies, thereby comprehensively improving detection accuracy and robustness. Compared with the current mainstream Transformer-based model architecture, our design shows significant advantages in many aspects: first, the model achieves linear expansion of computational complexity during feature processing, avoiding the common quadratic complexity problem in the Transformer architecture, thereby greatly reducing the demand for computing resources; second, the model can process high-dimensional medical image data more efficiently while maintaining lightweight. Therefore, this hybrid architecture not only solves the computational bottleneck of traditional methods, but also provides a new solution for efficient feature extraction in complex three-dimensional medical image tasks.

[0113] Compared with the traditional Unet3D model, the Mamba layer and Mamba module are introduced to combine the idea of ​​sequence modeling with 3D medical image processing. In addition, we introduced residual branches and multiple convolution operations. Through multi-branch fusion and parallel convolution processing and then merging, the feature expression can be enriched and the robustness of the network can be enhanced. These changes make the bleeding area detection model more accurate in detecting and segmenting complex targets such as small tumors, while retaining the lightweight advantage of the UNet3D model.

[0114] Evaluation indicators:

[0115] TP (TruePositive): The number of samples that are judged as positive samples by the model and are actually positive samples, that is, the part of the real positive examples that are correctly identified.

[0116] TN (TrueNegative): The number of samples that are judged as negative samples by the model and are actually negative samples, that is, the part of real negative examples that are correctly excluded.

[0117] FP (False Positive): The number of samples that are judged as positive samples by the model but are actually negative samples, that is, the number of negative samples that are mistakenly identified as positive samples.

[0118] FN (False Negative): The number of samples that are judged as negative samples by the model but are actually positive samples, that is, the part of positive samples that are mistakenly excluded as negative samples.

[0119] Dice: An evaluation metric commonly used in image segmentation tasks, especially in the field of medical image analysis. It is used to measure the similarity between the predicted result and the true label, with a value range of 0 to 1. The larger the value, the more similar the prediction is to the true label.

[0120] The specific definition is:

[0121]

[0122] Among them, A and B represent the predicted segmentation result and the true label area respectively. |A| and |B| represent the number of pixels in their respective areas, and |A∩B| represents the intersection area between the predicted result and the true label, that is, the number of pixels overlapping between the two.

[0123] Intersection over Union (IoU) is an evaluation metric widely used in image segmentation, object detection and other tasks to measure the overlap between the predicted result and the true label area. IoU evaluates the segmentation quality of the model by calculating the ratio of the intersection and union of the predicted area and the true label area.

[0124] IoU is defined as the ratio of the area of ​​the intersection of the predicted area and the true label area to the area of ​​their union:

[0125]

[0126] Where A represents the predicted area, B represents the true label area, |A∩B| represents the intersection area of ​​the predicted area and the true label area, and |A∪B| represents the union area of ​​the predicted area and the true label area.

[0127] Sensitivity, also known as Recall or TruePositiveRate (TPR), is an important evaluation indicator for measuring the model's ability to detect positive samples. It reflects the proportion of targets that are successfully predicted as positive samples among all targets that are actually positive samples. Sensitivity is one of the important performance metrics in tasks such as medical image analysis, classification, and target detection, especially in application scenarios where missed detections need to be minimized (such as disease diagnosis).

[0128] The calculation formula of Sensitivity is:

[0129]

[0130] PPV (Positive Predictive Value), also known as Precision, is an indicator that measures the proportion of actual positive samples among all the results predicted by the model as positive samples. PPV reflects the reliability of the model's prediction, especially its ability to predict the accuracy of positive samples, which is crucial for tasks that need to reduce false positives (FP).

[0131] The calculation formula for PPV is:

[0132]

[0133] The overall technical roadmap of vascular segmentation is as follows Figure 1 .

[0134] (1) Step 1: Linear stretching

[0135] Linear stretching maps the input value to the target range by scaling and translating. Assume that varMatrix is ​​the variance matrix and stretchedVarMatrix is ​​the stretched matrix:

[0136]

[0137] Here, minVal and maxVal are the minimum and maximum values ​​in the variance matrix, respectively. This formula will ensure

[0138] The values ​​in stretchedVarMatrix are in the range [0,255].

[0139] (2) Step 2: Normalization

[0140] Dividing each item by the maximum variance and then multiplying by 255 can map the values ​​of the variance matrix to the range [0,255]

[0141] Assume varMatrix is ​​the variance matrix and maxVal is the maximum value in the variance matrix:

[0142]

[0143] Here, the values ​​in normalizedMatrix will be mapped to the range [0,255]. Figure 2 .

[0144] (3) Step 3: Image enhancement based on sequence information

[0145] Angiography images are obtained by subtraction. As contrast agents flow in blood vessels, the images will change over time after the contrast agents are injected into the blood vessels. In the image sequence, the pixel values ​​on the blood vessels will fluctuate greatly, while the pixel values ​​of the background points will be relatively stable around a constant value with little change. This feature can be used to separate blood vessels from the background.

[0146] The change of gray value can be expressed by variance. The formula for calculating the gray variance of time series is:

[0147]

[0148] S (x,y)Represents the variance value at position (x, y), and T represents the time length of the sequence. (x,y,l) The grayscale value of the tth frame representing the variance at (x,y). is the average grayscale value over the entire sequence at (x,y).

[0149] After calculating the variance matrix of the entire image, points with large variance can be regarded as blood vessel points and points with small variance can be regarded as background points according to the characteristics of the image sequence. In addition, in order to facilitate the subsequent blood vessel segmentation work, the variance matrix needs to be mapped into a grayscale image with a grayscale value range of 0 to 255. Figure 3 , Figure 4 shown.

[0150] (4) Step 4: Adaptive Threshold Method

[0151] The adaptive thresholding method based on the local Gaussian weighted mean is a method that considers the local characteristics of the image and uses Gaussian weighted average to determine the binarization threshold. The following are the basic steps of the adaptive thresholding method based on the local Gaussian weighted mean:

[0152] ① Select neighborhood size and weight function: First, you need to select the size of a local neighborhood, usually a square or rectangular area. In addition, define a weight function, usually a Gaussian function. Gaussian weights make pixels farther from the center pixel contribute less to the threshold.

[0153] ② Calculate the local mean: For each pixel in the image, calculate the Gaussian weighted average of the grayscale values ​​in its neighborhood. The calculation formula for the weighted average is:

[0154]

[0155] Where: m(x,y) is the Gaussian weighted mean of the center pixel (x,y), w(i,j) is the Gaussian weight, and I(x+i,y+j) is the grayscale value of the pixel in the neighborhood.

[0156] ③ Calculate local variance: For each pixel, calculate the Gaussian weighted variance of the pixel grayscale values ​​in its neighborhood. The variance calculation formula is:

[0157]

[0158] ④ Determine the local threshold: For each pixel, the local mean m(x,y) is used as the local threshold of the pixel.

[0159] ⑤ Binarized image: Use these local thresholds to binarize the original image, that is, set the pixels less than or equal to the local threshold to one value (such as 0), and set the pixels greater than the local threshold to another value (such as 255). The effect is as follows: Figure 5 shown.

[0160] This method is more flexible to adapt to image changes, especially for images with uneven illumination, by considering local image characteristics and using Gaussian weighted averaging. However, attention should be paid to choosing the appropriate neighborhood size and weight function as well as the parameters of the weight function in order to achieve good results in different image scenarios.

[0161] (5) Step 5: Region Growing

[0162] ① Create a blank image (all black);

[0163] ②Store the seed point in a vector, which stores the seed point to be grown;

[0164] ③ Pop out the seed points one by one and judge the relationship between the seed points and the surrounding 8 neighborhoods (growth rule). The adjacent points are used as seed points for the next growth;

[0165] ④If there is no seed point in the vector, the growth stops. Figure 6 shown.

[0166] 3. Vascular feature point matching and vascular 3D reconstruction technology routes such as Figure 7 shown.

[0167] (1) Step 1: Feature extraction and matching

[0168] First, determine the bifurcation point

[0169] Extraction: Use the 8-neighborhood pixel tracking method to extract features of bifurcation points. Traverse each point on the vascular skeleton line and make a judgment. The judgment method for a certain point is: if the sum of the absolute values ​​of the differences between all two adjacent pixels in the 8 pixels around the point is 6 (the image has been binarized), it is determined that this is the bifurcation point of the blood vessel. The specific steps are as follows:

[0170] Determine the address offsets of the 8 points around the point to be judged: (-1,-1), (-1,0), (-1,1), (0,1), (1,1), (1,0), (1,-1), (0,-1)

[0171] Calculate the sum of the absolute values ​​of the differences between any two adjacent points of these 8 points; if the sum is 6, then the point is a bifurcation point, otherwise, continue to judge, the effect is as follows Figure 8 shown.

[0172] Secondly, matching: After obtaining the bifurcation points of the two images, the bifurcation points are matched using the principle of epipolar geometry. According to the epipolar constraint formula: p 1 T Fp 2= 0, we can get the relationship between the two corresponding bifurcation points, that is, the pixel coordinates p of the bifurcation point in the first image. 1 Perform matrix multiplication with the basic moment F, and then multiply it with the corresponding bifurcation point pixel coordinate p in the second image. 2 The multiplication operation is performed, and the result is 0. The basic matrix F can be obtained through the essential matrix E and the camera intrinsic parameter K through the formula: F = K T EK -1 The essential matrix E can be solved by the eight-point method (using four pairs of known matching points to estimate the essential matrix E), or by using the displacement matrix t and the rotation moment R through E = t Λ R, t Λ is the antisymmetric matrix of t. The specific matching process is:

[0173] ① Cycle through the intersections in the two images, take the vascular skeleton line from one camera's perspective as a reference, and use the intersections in the vascular skeleton line from the other perspective and the intersections in the former to perform epipolar constraint judgment;

[0174] ② If satisfied, the match is successful, if not satisfied, continue matching. (In fact, due to errors, even the matching points may not necessarily have a calculation result of 0, so the pair with the smallest value is selected as the matching point). Fig. 9 shown.

[0175] Finally, extraction: Similar to bifurcation points, it is also an 8-neighborhood pixel tracking method, and the judgment that the sum of the absolute values ​​of the difference is equal to 6 can be changed to equal to 4. Matching: It is necessary to use the matched bifurcation points to segment the vascular skeleton line, that is, each line segment takes two different bifurcation points as the starting point and end point, and then match the line segments according to the matching results of the bifurcation points.

[0176] The specific process of segmentation is as follows:

[0177] ①Define the search direction

[0178] The code first defines an 8-direction search mode, where the dx and dy arrays store the horizontal and vertical movement directions respectively. The directions are defined as 8 directions: up, down, left, right, and diagonal (-1, 0, +1 represent the change in direction).

[0179] ②Traverse and find the extension direction of the bifurcation point

[0180] point represents the current bifurcation point. The code traverses the eight directions around the bifurcation point and checks whether the adjacent pixels are skeleton pixels (back.at(x,y) == 255). If so, it means that there is a line segment extending in that direction. For the diagonal direction, the two adjacent non-diagonal direction points are additionally checked to ensure that there are no repeated paths (avoid repeated searches along the diagonal direction). If the direction point is valid (i.e., there are no repeated paths), the direction is added to the direction vector and the pixel value in that direction is set to 0, indicating that it has been visited.

[0181] ③Start extending the line segment

[0182] For each valid direction point d found, start constructing a line segment. The first point of the line segment is the current bifurcation point point, and the second point is d, which represents the extension direction found.

[0183] ④Extension along the line

[0184] During the extension process, the code moves forward through the loop to find the next point. If the next point is found, it is added to the line vector and the back image is updated (the value of the visited point is set to 0). When extending, first check whether the current point is a bifurcation point. If it is a bifurcation point: mark it as a bifurcation point, return, and end the extension of the line segment. Otherwise, continue to extend to the next point until it cannot be extended (that is, the pixel point of the next skeleton line is not found).

[0185] ⑤ Bifurcation point processing

[0186] When the code detects that the end point of a line segment is another bifurcation point, it records the end point as the head of a new line segment (stored in head for subsequent topological sorting). If it extends to an end point without a bifurcation point (that is, the path has no further direction to extend), the end point is added to points and the drawing of the line segment ends.

[0187] ⑥Save line segment

[0188] Each time a line segment is found, the code stores it in the lines vector. lines stores all the segmented skeleton line segments, each of which is a vector consisting of multiple PIIs (i.e. (x, y) coordinate pairs).

[0189] The example after segmentation is as follows Fig.10 shown.

[0190] Overall matching process:

[0191] ① After segmentation, several line segments are obtained. The starting point of the line segment is the bifurcation point, and the end point of the line segment is the bifurcation point or the endpoint. After the epipolar constraint judgment is performed on the endpoint, the endpoint matching pair of the two images is obtained. The starting point and end point of the line segment are both matched points, and the line segment matching is completed.

[0192] ③ For the matched line segment pairs, use the line segment points with more points as references, and perform point matching on the reference line segments with fewer line segment points, and use the epipolar constraint to select the matching points.

[0193] (2) Step 2: Obtain the three-dimensional coordinates of the feature points and the corresponding radius of the feature points

[0194] ① Obtain the three-dimensional coordinates of the feature points

[0195] Using the camera model, we can find the relationship between pixel coordinates and real coordinates.

[0196]

[0197] After obtaining the camera's intrinsic parameter K and the distance Z from the real point to the optical center, the three-dimensional coordinates of the pixel point in the real world can be obtained.

[0198] ② Obtain the radius corresponding to the feature point

[0199] The edge detection algorithm is used to obtain the vascular wall in the vascular segmentation image, and the vascular wall and matching points are used to obtain the vascular radius corresponding to the matching point. The specific steps are to search for the shortest distance to the vascular wall within a certain range on both sides of the left and right vascular walls for a certain matching point, and let this distance be the vascular radius of the feature point.

[0200] (3) Step 3: Reconstructing the vascular model

[0201] Using the previously matched feature points (stored in the form of line segments, that is, using a vector array to store multiple line segments, each line segment is a vector variable that stores the feature points on the line segment) and the corresponding radius, use topological sorting to construct the line segments in order.

[0202] First, construct the vascular topology structure. The vascular topology structure uses the identified bifurcation points, extending outward along the line segment points from the bifurcation points, stopping at other bifurcation points or vascular endpoints and recording the stopping point as an extension direction for the bifurcation point. After all directions are extended, the current bifurcation point completes the extension operation, and then extends from the next bifurcation point until all bifurcation points are extended. Because the blood vessels are connected, the vascular topology structure is stored at this time.

[0203] When modeling, using the stored blood vessel segments, it is easy to obtain a directed acyclic graph, which is a graph from a bifurcation point to another bifurcation point or a blood vessel endpoint. The line segments connecting the nodes are the line segment points on the blood vessel. The algorithm flow is as follows:

[0204] ① Initialization: Calculate the in-degree of all nodes (bifurcation points and endpoints) (whether there is a bifurcation point pointing to this point).

[0205] ②Select the starting point: add all nodes with in-degree 0 to the queue.

[0206] ③ Node processing: Take out a node from the queue each time, add it to the topological sequence, and build the blood vessel model for that section.

[0207] ④Update in-degree: If the in-degree of a node becomes 0, add it to the queue.

[0208] ⑤ Repeat: until the queue is empty. When the queue is empty, the vascular model is completed. The reconstruction effect is shown in the figure below. Fig.11 shown.

[0209] Result analysis:

[0210] Model DICE Sensitivity IOU Precision VNet3D 0.5282 0.5008 0.4271 0.7362 UNet3D 0.4732 0.5611 0.4732 0.7386 TransUnet3D 0.6163 0.5964 0.5097 0.7485 SegMamaba 0.6427 0.6142 0.5325 0.7667 SegResNet 0.5997 0.5735 0.4897 0.6965 UxNet 0.6322 0.5966 0.5201 0.7541 Ours 0.6956 0.6706 0.5780 0.7654

[0211] According to the evaluation indicators of the experimental results, the following analysis can be performed:

[0212] 1. DICE coefficient analysis: The segmentation method of the present invention achieved a DICE coefficient of 0.6956, which is significantly better than other models. Especially compared with VNet3D (0.5282) and UNet3D (0.4731), it showed a significant improvement. The DICE coefficient is an important indicator to measure the overlap between the segmentation result and the true label. The higher the DICE coefficient, the closer the segmentation result is to the true label. Therefore, the segmentation method of the present invention has obvious advantages in overall segmentation accuracy.

[0213] In comparison with TransUnet3D (0.6163) and SegMamaba (0.6426), the segmentation method of the present invention still has a relatively high advantage and shows good segmentation accuracy.

[0214] 2. Sensitivity analysis:

[0215] The sensitivity of the bleeding area detection model of the present invention is 0.6706, which is significantly higher than other models, such as VNet3D (0.5008) and UNet3D (0.5611). This shows that the bleeding area detection model of the present invention is more capable of identifying true positives (i.e. bleeding points), can more effectively avoid missed detection, and is suitable for medical tasks that are sensitive to missed detection.

[0216] Compared with SegMamaba (0.6142) and UxNet (0.5966), the bleeding area detection model of the present invention also shows stronger recall ability and can better identify actual bleeding points.

[0217] 3. Intersection over Union (IoU) analysis:

[0218] In terms of IoU index, the segmentation method of the present invention is 0.5780, which exceeds other models, especially TransUnet3D (0.5097) and SegMamaba (0.5325). IoU measures the degree of overlap between the predicted area and the true label area. The higher the value, the more accurate the target area can be predicted during segmentation. The segmentation method of the present invention can better handle the segmentation of the three-dimensional bleeding point area and improve the accuracy of the model.

[0219] 4. Precision analysis:

[0220] The accuracy of the bleeding area detection model of the present invention is 0.7654, which is close to SegMamaba (0.7669) and higher than other models (such as SegResNet 0.6965 and VNet3D 0.7362). This shows that the bleeding area detection model of the present invention can effectively reduce false positives, that is, when predicted as a positive sample, the actual positive ratio is higher, thereby improving the reliability of the model.

[0221] The vascular segmentation, reconstruction and bleeding analysis method proposed in this paper performs well in all indicators, especially in DICE coefficient and sensitivity, which is significantly better than other models, indicating that it has obvious advantages in accurately capturing intracranial hemorrhage areas and reducing missed detections. Compared with other mainstream models (such as VNet3D, UNet3D, TransUnet3D, SegMamaba and UxNet), the improvement in precision and recall makes it more practical and reliable in medical image segmentation tasks.

[0222] In summary, the vascular segmentation, reconstruction and bleeding analysis method proposed in the present invention demonstrates strong segmentation accuracy and robustness in the task of three-dimensional intracranial hemorrhage point detection, and is an effective and generalizable solution.

[0223] The model results are shown as Fig.14 (left original label, right predicted result).

[0224] Therefore, the present invention proposes a blood vessel segmentation method based on time series images, which realizes the comprehensive analysis of angiography images at multiple time points and overcomes the problem of loss of blood vessel details and image blurring caused by insufficient or excessive diffusion of contrast agents in single image segmentation. By combining the adaptive threshold binarization method and the region growing algorithm, the high accuracy and structural integrity of blood vessel segmentation are ensured.

[0225] The present invention proposes a 3D reconstruction method based on 8-neighborhood pixel tracking method and skeleton line segmentation, which realizes accurate identification and segmentation processing of the intersection points of the vascular skeleton lines, and overcomes the limitation of the traditional 3D reconstruction method that relies on manual pre-definition of the vascular topology structure. By constructing an automated vascular topology structure and combining the feature point matching strategy, the 3D reconstruction of the blood vessels is completed efficiently, significantly improving the reconstruction efficiency and adaptability of the model.

[0226] The present invention proposes a hybrid architecture combining convolutional neural network (CNN) and state space model (SSM), which is specially designed for the segmentation and detection tasks of three-dimensional intracranial hemorrhage CT images. The robustness of the model is improved through efficient data preprocessing and enhancement strategies (such as binarization, morphological operations, maximum connected domain extraction and diversity data enhancement), and the Mamba architecture is introduced to optimize the long-range dependency modeling capabilities, while significantly reducing the computational complexity. Experiments based on the INSTANCE dataset verified the high accuracy of the method in detecting bleeding point areas, and its superiority was comprehensively evaluated through indicators such as intersection over union (IoU) and Dice coefficient. The scheme has a high degree of modularity, efficient training, lightweight reasoning, and is suitable for actual clinical scenarios. It provides a novel and practical solution for the task of three-dimensional high-resolution medical image segmentation.

Claims

1. A blood vessel segmentation, reconstruction and bleeding analysis method, characterized in that: The image to be detected is input into the bleeding area detection model to detect whether there is a bleeding area; if there is no bleeding area, the blood vessel segmentation and blood vessel 3D reconstruction operations are directly performed; if there is a bleeding area, the blood vessel segmentation and blood vessel 3D reconstruction operations are performed after hemostasis treatment; The blood vessel segmentation and blood vessel three-dimensional reconstruction operations are specifically as follows: using time series image analysis to generate a blood vessel segmentation map; obtaining a blood vessel skeleton line based on the blood vessel segmentation map, traversing the blood vessel skeleton line, identifying intersections and segmenting the blood vessel skeleton line according to the intersections, each blood vessel segment only contains two bifurcation points, a starting point and an end point, to form a blood vessel topology structure; based on the blood vessel topology structure, obtaining the three-dimensional coordinates of the feature points and the blood vessel radius of the feature points, and reconstructing the blood vessel model.

2. The blood vessel segmentation, reconstruction and bleeding analysis method according to claim 1, characterized in that: The bleeding area detection model is designed based on the convolutional neural network (CNN) and the state space model (SSM) architecture; The bleeding area detection model is mainly composed of an encoder and a decoder; The encoder is a Mamba encoder; the features output by the Mamba encoder gradually reduce the resolution of the feature map through a 3D maximum pooling operation; in each layer of the Mamba encoder, the extracted local features are further processed by an embedded Mamba module to model global features and local features; The decoder restores the resolution through deconvolution operation, and integrates the high-resolution features in the encoder while fusing local convolution features and global state space features to gradually reconstruct the segmented image; the prediction result is a voxel-level segmentation probability map with a shape of (B, N, D, H, W), where N is the number of categories, indicating the probability of each pixel belonging to each category.

3. The blood vessel segmentation, reconstruction and bleeding analysis method according to claim 2, characterized in that: The Mamba module is embedded in each layer of the encoder; each Mamba module includes two branches, namely, an SSM branch and a CNN branch; The SSM branch is composed of a plurality of Mamba layers in cascade, and the CNN branch is composed of a plurality of convolutional layers; The Mamba layer is used to capture global dependencies and model long-distance dependencies through the state space model SSM; The outputs of the two branches are added together, and the local convolution features extracted by the CCN branch are combined with the global SSM features extracted by the SSM branch. The combined features are added to the features input by the Mamba module through a skip connection, retaining the original information of the input features and input to the decoder.

4. The blood vessel segmentation, reconstruction and bleeding analysis method according to claim 1, characterized in that: When traversing the blood vessel skeleton line, the 8-neighborhood pixel tracking method is used to extract the features of the bifurcation point, and each point on the blood vessel skeleton line is traversed for judgment; the pixel differences of adjacent points in the 8-neighborhood pixels are calculated in a clockwise manner, and when the sum of the absolute values ​​of these pixel differences is 6, the point is the bifurcation point of the blood vessel; the specific steps are as follows: Determine the address offsets of the 8 points around the point to be judged: (-1,-1), (-1,0), (-1,1), (0,1), (1,1), (1,0), (1,-1), (0,-1); Calculate the sum of the absolute values ​​of the pixel differences between any two adjacent points of these 8 points; If the sum is 6, then this point is a bifurcation point, otherwise continue to judge; After obtaining the bifurcation points of the vascular skeleton under different camera perspectives, the bifurcation points are matched according to the epipolar geometry principle; according to the epipolar constraint formula: p1 T Fp2 = 0, the relationship between the two corresponding bifurcation points is obtained. The pixel coordinates p1 of the bifurcation point in the first image are multiplied by the basic matrix F, and then multiplied by the corresponding pixel coordinates p2 of the bifurcation point in the second image. The result is 0, which means the match is successful. The basic matrix F is obtained by the essential matrix E and the camera intrinsic parameter K through the formula: F = K T EK -1 The essential matrix E is obtained by the eight-point method, and the essential matrix E is estimated by four pairs of known matching points, or by using the displacement matrix t and the rotation moment R through E = t Λ R obtains, t Λ is the antisymmetric matrix of t; the specific matching process is: Cycle through the intersections in the two images, take the vascular skeleton line from one camera’s perspective as a reference, and use the intersections in the vascular skeleton line from the other perspective to perform epipolar constraint judgment with the intersections in the former. If satisfied, the match is successful, if not satisfied, continue matching. When the calculation result is not 0, select the pair with the smallest value as the matching point. The extraction of blood vessel line segment points also uses the 8-neighborhood pixel tracking method, and the judgment is made based on whether the sum of the absolute values ​​of the pixel differences is equal to 4; Matching: According to the matched bifurcation points, the vascular skeleton line is segmented; each line segment takes two different bifurcation points as the starting point and the end point, and the line segments are matched according to the matching results of the bifurcation points to obtain the vascular line segment points.

5. The blood vessel segmentation, reconstruction and bleeding analysis method according to claim 4, characterized in that: The specific process of segmentation is as follows: Define the search direction in 8 directions. The dx and dy arrays store the horizontal and vertical movement directions respectively. The directions are defined as 8 directions of up, down, left, right and diagonal. Traverse and find the extension direction of the bifurcation point; point represents the current bifurcation point, traverse the 8 directions around the bifurcation point, and check whether the adjacent pixels are skeleton pixels, and the skeleton pixel back.at(x,y)==255. If so, it means that there is a line segment extending in this direction; for the diagonal direction, additionally check the two adjacent non-diagonal direction points to ensure that there is no repeated path; if the diagonal direction point is valid, that is, there is no repeated path, then add the diagonal direction point to the direction vector and mark the direction point as visited; For each valid direction point d found, start to construct a line segment; the first point of the line segment is the current bifurcation point point, and the second point is d, which represents the extension direction found; during the extension process, move forward through the loop to find the next valid direction point. If the next point is found, add it to the line vector and update the auxiliary image used to track the processed area; when extending, first check whether the current point is a bifurcation point. If it is a bifurcation point: mark it as a bifurcation point, return, and end the extension of the line segment; otherwise, continue to extend to the next point until it cannot be extended; When it is detected that the end point of a line segment is another bifurcation point, the end point is recorded as the starting point of a new line segment. If it extends to an end point without a bifurcation point, the end point is added to points and the drawing of the line segment ends; all line segments are segmented skeleton line segments; Matching of the overall line segment; The starting point of the segmented skeleton line segment is the bifurcation point, and the end point is the bifurcation point or the endpoint. After the epipolar constraint judgment is performed on the endpoint, the endpoint matching pair of the two images is obtained; For the matched line segment pairs, the line segment with more points is used as the reference line segment, and the line segment with fewer points is matched on the reference line segment, and the matching points are selected using the epipolar constraint.

6. The blood vessel segmentation, reconstruction and bleeding analysis method according to claim 5, characterized in that: The process of obtaining the three-dimensional coordinates of the feature points and the blood vessel radius of the feature points is as follows: Find the connection between pixel coordinates and real coordinates based on the camera model when acquiring the image; After obtaining the camera's intrinsic parameter K and the distance Z from the real point to the optical center, the three-dimensional coordinates of the pixel point in the real world are obtained; P = [X, Y, Z] is the world coordinate of the blood vessel point, u and v are the pixel coordinates of the blood vessel point, which are the pixel positions in the horizontal direction and the vertical direction, respectively, and f x 、f y is the focal length of the camera in pixels on the X and Y axes, c x 、c y is the coordinate of the optical center in the image coordinate system; The vascular wall in the vascular segmentation image is obtained according to the edge detection algorithm, and the vascular radius corresponding to the matching point is obtained using the vascular wall and the matching point; for a certain matching point, the shortest distance to the vascular wall is searched on both sides of the left and right vascular walls, and this distance is used as the vascular radius of the bifurcation point or the vascular radius of the line segment point.

7. The blood vessel segmentation, reconstruction and bleeding analysis method according to claim 6, characterized in that: The reconstructed blood vessel model is constructed by sequentially modeling the line segments using topological sorting according to the matched bifurcation points or line segment points and the corresponding blood vessel radius; The vascular topology structure starts from the identified bifurcation point and extends outward along the line segment point, stops at other bifurcation points or vascular endpoints and records the stop point as an extension direction of the bifurcation point, all directions are extended, the current bifurcation point completes the extension operation, and then extends from the next bifurcation point until all bifurcation points are extended; The process of 3D reconstruction of blood vessels is as follows: Initialization: calculate the in-degree of all nodes and whether there is a bifurcation point pointing to that point; Select the starting point: add all nodes with in-degree 0 to the queue; Processing nodes: taking a node from the queue each time, adding it to the topological sequence, and building the blood vessel model for that segment; Update in-degree: When the in-degree of a node becomes 0, add it to the queue; Repeat: until the queue is empty; the reconstruction of the vascular model is completed.

8. A blood vessel segmentation, reconstruction and bleeding analysis device, characterized in that: include: The bleeding analysis module is used to detect whether there is a bleeding area in the image to be detected; if there is a bleeding area, the bleeding area is located and marked, and after the hemostasis treatment is completed, the re-scanned image of the patient is transmitted to the vascular segmentation and vascular reconstruction module; if no bleeding area is detected, the image is directly transmitted to the vascular segmentation module; A blood vessel segmentation module is used to perform a blood vessel segmentation operation on an input image and generate a blood vessel segmentation map; Extract the vascular skeleton line, identify the bifurcation points and line segment points through the 8-neighborhood pixel tracking method, and establish the vascular topology structure; The vascular reconstruction module is used to reconstruct the vascular three-dimensional model through a topological sorting method according to the vascular topological structure and the matched bifurcation points or line segment points, combined with their corresponding vascular radius, to generate a vascular three-dimensional structural model.

Images