Cerebrovascular image recognition and analysis system based on deep learning

Abstract

The invention relates to the technical field of image processing, in particular to a cerebrovascular image recognition and analysis system based on deep learning. The system comprises a cerebrovascular image registration module, a CT pixel communication module, a lesion area identification and segmentation module and a lesion risk assessment module, and can obtain a cerebrovascular CT angiography image, perform image enhancement and multi-modal image registration, and generate a cerebrovascular CT multi-modal registration image sequence; acquiring cerebrovascular CT pixel point clouds in the corresponding modals and performing blood vessel cluster connectivity analysis to generate corresponding cerebrovascular CT pixel cluster connectivity graphs in the respective modals; and based on a deep learning multi-branch network architecture, introducing an attention mechanism to carry out lesion region identification and segmentation, obtaining a corresponding cerebrovascular stenosis degree and a lesion thrombus area, and carrying out lesion risk number evaluation on the corresponding cerebrovascular CT image lesion region to obtain a cerebrovascular CT lesion risk number. According to the invention, accurate recognition of the cerebral vascular image lesion condition can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a cerebrovascular image recognition and analysis system based on deep learning. Background Art

[0002] In recent years, with the rapid development of artificial intelligence and deep learning technologies, medical image analysis based on deep learning has gradually become an important auxiliary tool for the recognition of cerebrovascular image lesions. Deep learning, especially convolutional neural networks (CNNs), has shown excellent performance in the application of medical images, and can automatically extract features in images and perform efficient recognition. However, traditional cerebrovascular image analysis methods, such as the manual analysis of computed tomography (CT) and magnetic resonance imaging (MRI) images, often require the experience of professional doctors. There are high noises and artifacts in the cerebrovascular images, resulting in unstable image quality. This makes the accuracy of deep learning models drop significantly when processing low-quality or complex images. In addition, due to the complexity of the cerebrovascular structure and individual differences, existing deep learning models are difficult to uniformly adapt to the image features of different patients, and the generalization ability of the models is limited. Summary of the Invention

[0003] Based on this, it is necessary for the present invention to provide a cerebrovascular image recognition and analysis system based on deep learning to solve at least one of the above technical problems.

[0004] To achieve the above object, a cerebrovascular image recognition and analysis system based on deep learning includes the following modules: A cerebrovascular image registration module, configured to obtain a cerebrovascular CT angiography image, and perform image enhancement and multi-modal image registration on the cerebrovascular CT angiography image to generate a cerebrovascular CT multi-modal registration image sequence; A CT pixel connectivity module, configured to obtain a cerebrovascular CT pixel point cloud in the corresponding modality through the cerebrovascular CT multi-modal registration image sequence, and perform vascular clustering connectivity analysis on the cerebrovascular CT image in the corresponding modality in the cerebrovascular CT multi-modal registration image sequence based on the cerebrovascular CT pixel point cloud in the corresponding modality to generate a corresponding cerebrovascular CT pixel cluster connectivity map in each modality; A lesion area recognition and segmentation module, configured to perform lesion area recognition and segmentation on the corresponding cerebrovascular CT pixel cluster connectivity map in each modality based on a deep learning multi-branch network architecture and by introducing an attention mechanism to generate a cerebrovascular CT image lesion area; A lesion risk assessment module, configured to obtain the corresponding cerebrovascular stenosis degree and lesion thrombus area through the cerebrovascular CT image lesion area, and perform lesion risk assessment on the corresponding cerebrovascular CT image lesion area based on the cerebrovascular stenosis degree and lesion thrombus area to obtain the size of the cerebrovascular CT lesion risk.

[0005] Furthermore, the cerebrovascular image registration module includes the following functions: Obtain cerebral vascular CT angiography images; Perform grayscale processing on the cerebral vascular CT angiography images to obtain cerebral vascular CT grayscale images; Quantify the pixel blurriness of the cerebral vascular CT grayscale images to obtain the pixel blurriness of the cerebral vascular CT images; based on the pixel blurriness of the cerebral vascular CT images, perform pixel blurring denoising on the cerebral vascular CT grayscale images to generate cerebral vascular CT blurred denoised images; Perform image enhancement processing on the cerebral vascular CT blurred denoised images to adopt random rotation, flipping, scaling, and translation to perform modality augmentation on the cerebral vascular CT images at different time points, and generate a cerebral vascular CT multi-modal image sequence; Perform multi-modal image registration on the cerebral vascular CT images corresponding to different time points and different modalities within the cerebral vascular CT multi-modal image sequence to generate a cerebral vascular CT multi-modal registered image sequence.

[0006] Furthermore, the multi-modal image registration of the cerebral vascular CT images corresponding to different time points and different modalities within the cerebral vascular CT multi-modal image sequence includes: Perform temporal synchronization annotation on the cerebral vascular CT images corresponding to different time points within the cerebral vascular CT multi-modal image sequence to obtain cerebral vascular CT multi-modal temporal annotation images; Obtain the contrast and resolution of the cerebral vascular CT images under the corresponding modality through the cerebral vascular CT multi-modal temporal annotation images, and perform modality-level sorting analysis on the cerebral vascular CT images corresponding to different time points of the cerebral vascular CT multi-modal temporal annotation images based on the contrast and resolution of the cerebral vascular CT images under the corresponding modality, so as to obtain the modality characteristic complementary levels between the corresponding modalities according to the sorted contrast and resolution, and obtain the corresponding modality-level relationships between the cerebral vascular CT images at different time points; Based on the corresponding modality-level relationships between the cerebral vascular CT images at different time points, perform image spatial alignment on the cerebral vascular CT images corresponding to different time points of the cerebral vascular CT multi-modal temporal annotation images to obtain the cerebral vascular CT image spatial transformation matrix between different time points and modalities; Based on the cerebral vascular CT image spatial transformation matrix between different time points and modalities, perform global optimization registration transformation on the local regions of the cerebral vascular CT images corresponding to different time points within the cerebral vascular CT multi-modal image sequence, so as to align and correct the errors corresponding to the spatial transformation matrix at the cerebrovascular junction of the cerebral vascular CT images at different time points, and generate a globally optimized cerebral vascular CT registration transformation function; Perform full-modal synchronous alignment on the cerebrovascular CT multimodal image sequence using the cerebrovascular CT registration transformation function optimized based on global optimization to generate a cerebrovascular CT multimodal registration image sequence.

[0007] Furthermore, the CT pixel connectivity module includes the following functions: Perform pixel-level segmentation on each modal image in the cerebrovascular CT multimodal registration image sequence to obtain the pixel spatial positions and structure points of the cerebrovascular structure in each modality through automatic threshold segmentation or region growing method and form the corresponding pixel matrix, obtaining the cerebrovascular CT regional pixel matrix; Perform voxel density analysis on the corresponding cerebrovascular region voxels in each modal image based on the cerebrovascular CT regional pixel matrix to generate a cerebrovascular modal feature volume cloud map; Perform multimodal contrast enhancement on each voxel structure in the cerebrovascular modal feature volume cloud map to obtain a cerebrovascular multimodal structure point cloud; Perform spatial point cloud fitting optimization on the cerebrovascular multimodal structure point cloud to generate a cerebrovascular CT pixel point cloud corresponding to the corresponding modality; Perform vascular cluster connectivity analysis on the cerebrovascular CT image corresponding to the corresponding modality in the cerebrovascular CT multimodal registration image sequence based on the cerebrovascular CT pixel point cloud corresponding to the corresponding modality to generate a cerebrovascular CT pixel cluster connectivity map corresponding to each modality.

[0008] Furthermore, the performing spatial point cloud fitting optimization on the cerebrovascular multimodal structure point cloud includes: Perform pixel density statistical analysis on the cerebrovascular CT structure point cloud corresponding to the corresponding modality in the cerebrovascular multimodal structure point cloud to obtain the pixel density values corresponding to each cerebrovascular CT point cloud in the corresponding modality; Perform pixel outlier removal on the cerebrovascular CT structure point cloud corresponding to the corresponding modality in the cerebrovascular multimodal structure point cloud based on the pixel density values corresponding to each cerebrovascular CT point cloud in the corresponding modality to obtain a cerebrovascular CT pixel outlier removal point cloud corresponding to the corresponding modality; Perform local smoothing fitting processing on the cerebrovascular CT pixel outlier removal point cloud corresponding to the corresponding modality to generate a cerebrovascular CT pixel point cloud corresponding to the corresponding modality.

[0009] Furthermore, the performing vascular cluster connectivity analysis on the cerebrovascular CT image corresponding to the corresponding modality in the cerebrovascular CT multimodal registration image sequence based on the cerebrovascular CT pixel point cloud corresponding to the corresponding modality includes: Perform spatial neighborhood division on the cerebrovascular CT pixel point cloud corresponding to the corresponding modality to divide each local spatial small branch of the cerebrovascular using the local density distribution based on the point cloud, obtaining a cerebrovascular CT point cloud local spatial structure diagram; Perform vascular topology node connection analysis based on the local spatial structure diagram of cerebral vascular CT point cloud to identify the corresponding branch points, intersection points and end points of cerebral blood vessels, and topologically connect them to generate the corresponding vascular node connection structure, so as to generate the cerebral vascular CT point cloud node connection diagram; Perform progressive connectivity analysis on the cerebral vascular CT point cloud node connection diagram to calculate the connectivity path between cerebral vascular nodes by using the shortest path recursion, and gradually merge local vascular nodes into corresponding vascular clusters to generate the cerebral vascular cluster connectivity sequence; Based on the cerebral vascular cluster connectivity sequence, optimize the details of the cerebral vascular CT pixel point cloud in the corresponding modality, eliminate the tiny connection parts between cerebral vascular clusters according to the cerebral vascular cluster connectivity structure corresponding to different modalities, and ensure the high connectivity of different modalities in the cerebral vascular network connection, so as to generate the corresponding cerebral vascular CT pixel cluster connectivity diagram in each modality.

[0010] Furthermore, the lesion area recognition and segmentation module includes the following functions: Perform cerebral vascular region feature analysis on the cerebral vascular CT pixel cluster connectivity diagram corresponding to each modality to obtain the corresponding cerebral vascular region pixel difference and morphological features in each modality; Input the corresponding cerebral vascular region pixel difference and morphological features in each modality into the corresponding modality network branch in the deep learning multi-branch network architecture for modality feature fusion, so as to enhance and fuse the complementary information between different modality cerebral vascular images to generate the corresponding structural feature map of cerebral blood vessels, and highlight and label the lesion area contained in the structural feature map of cerebral blood vessels corresponding to the attention mechanism to generate the cerebral vascular lesion area recognition model, and output the corresponding cerebral vascular lesion sensitive area map; Perform region segmentation and contour extraction on the corresponding lesion area in the cerebral vascular lesion sensitive area map, extract the lesion area boundary by combining edge detection and region growth, and form a clear lesion area contour in the cerebral vascular lesion sensitive area map to generate the cerebral vascular lesion area precise segmentation map; Perform morphological optimization processing on the cerebral vascular lesion area precise segmentation map to repair the tiny breaks in the lesion area through erosion and dilation operations, smooth the segmentation boundary, and strengthen the detail expression of the lesion area to generate the cerebral vascular CT image lesion area.

[0011] Furthermore, the lesion risk number assessment module includes the following functions: Obtain the corresponding cerebral vascular stenosis degree through the cerebral vascular CT image lesion area; Locate the thrombus area in the cerebral vascular CT image lesion area to obtain the cerebral vascular CT lesion thrombus location; Measure the spatial diameter of the corresponding thrombus within the lesion area of the cerebrovascular CT image based on the thrombus localization position of the cerebrovascular CT lesion to obtain the size of the thrombus diameter in the cerebrovascular CT lesion; Estimate the thrombus area based on the size of the thrombus diameter in the cerebrovascular CT lesion to obtain the lesion thrombus area; Evaluate the lesion risk number of the corresponding cerebrovascular CT image lesion area using the lesion risk number calculation formula based on the cerebrovascular stenosis degree and the lesion thrombus area to obtain the size of the cerebrovascular CT lesion risk number.

[0012] Further, the obtaining of the corresponding cerebrovascular stenosis degree through the cerebrovascular CT image lesion area includes: Analyze the geometric shape of the blood vessel lumen in the cerebrovascular CT image lesion area to generate a cerebrovascular CT lumen contour map; Measure the distance between the inner and outer walls according to the cerebrovascular CT lumen contour map, calculate the diameter corresponding to the cerebrovascular lumen point by point for each cross-section, and map it to the corresponding CT image position to generate a cerebrovascular cross-sectional diameter map; Calculate the diameter difference between different cross-sections in the cerebrovascular cross-sectional diameter map to obtain the cerebrovascular diameter difference between different cross-sections within the lesion area; Evaluate and calculate the stenosis degree of the cerebrovascular CT image lesion area using the cerebrovascular stenosis evaluation calculation formula based on the cerebrovascular diameter difference between different cross-sections within the lesion area to obtain the cerebrovascular stenosis degree; Among them, the specific cerebrovascular stenosis evaluation calculation formula is: ; In the formula, is the cerebrovascular stenosis degree, is the reference diameter corresponding to the normal area of the cerebrovascular, is the length of the time interval, is the time variable parameter, is the total number of cross-sections corresponding to the cerebrovascular within the lesion area, is the th cross-section within the lesion area at time corresponding cerebrovascular diameter, is the th cross-section within the lesion area at time corresponding cerebrovascular diameter, is the cerebrovascular diameter difference adjustment coefficient, is the exponential function, is the size of the lesion area, is the length of the cerebrovascular corresponding to the lesion area.

[0013] Further, the specific lesion risk number calculation formula is: ; In the formula, is the critical number of cerebrovascular CT lesions, is the degree of cerebrovascular stenosis, is the weight of the lesion influence, is the thrombus area of the lesion.

[0014] Advantages of the present invention: The cerebrovascular image recognition and analysis system based on deep learning proposed by the present invention is generally composed of a cerebrovascular image registration module, a CT pixel connectivity module, a lesion area recognition and segmentation module, and a lesion risk assessment module. Compared with the prior art, the beneficial effect of this application is that through cerebrovascular CT angiography images, the morphology, blood flow changes, and potential lesion areas of blood vessels can be intuitively observed. However, due to the influence of different scanning devices, scanning angles, and image noise on CT images, there are certain noise or distortion problems in single-modal images. Image enhancement technology can effectively improve the contrast and details of images, making the edges of blood vessels, lesion areas, and other important features clearer, facilitating subsequent analysis, and thus being able to better reduce the corresponding noise and artifacts in cerebrovascular images. In addition, cerebrovascular CT images often require multi-modal image registration to accurately align images from different sources, different time series, or different imaging modalities to ensure the spatial consistency between images. The core of this step lies in using multi-modal image registration technology and the complementarity of different-modal images. By combining CT images and MRI images, detailed structural information of blood vessels and the state of soft tissues can be obtained simultaneously. Through image registration, images in different modalities will be merged into a unified coordinate system, thereby generating a sequence of cerebrovascular CT multi-modal registration images. Such a registered image sequence provides richer and more accurate information for subsequent analysis. Secondly, after obtaining the sequence of cerebrovascular CT multi-modal registration images, the next step is to extract the corresponding cerebrovascular CT pixel point cloud from it. By extracting the pixel point cloud, the spatial distribution and morphological characteristics of cerebrovascular can be effectively modeled. These point cloud data play an important role in identifying the distribution of blood vessel structures and potential abnormal areas within blood vessels. By analyzing the extracted cerebrovascular CT pixel point cloud, important information such as blood vessel branches, blood vessel walls, and lesion areas can be extracted and clustered through connectivity analysis. This clustering connectivity analysis not only helps to identify the overall morphology of blood vessels but also reveals abnormal changes in local areas, such as lesion characteristics like blood vessel stenosis, dilation, and thrombosis. By performing blood vessel clustering connectivity analysis on cerebrovascular CT images in different modalities, the structural differences inside and outside blood vessels can be better presented. Especially in multi-modal images, different types of lesion areas can be clearly identified.Then, a multi-branch network architecture based on deep learning can process data of multiple modalities simultaneously and fully extract feature information in different modalities. In the analysis of cerebral vascular CT images, deep learning can automatically learn the structure, morphology, and lesion characteristics of blood vessels from images through a large number of training samples, and then perform automatic recognition and segmentation of the lesion area. When a deep learning model with an attention mechanism processes cerebral vascular CT images, it can perform weighted processing based on the important regions of the image, thereby paying more attention to the characteristics of blood vessels and lesion areas. The attention mechanism enables the model to more flexibly select the key areas to focus on while ignoring unimportant areas, which can improve the accuracy and efficiency of image analysis, better adapt to the imaging characteristics of different patients, and thus improve the generalization ability of the deep learning model. Finally, once the lesion areas in cerebral vascular CT images are successfully identified and segmented, the next step is to perform quantitative analysis on these lesion areas, especially the assessment of stenosis degree and thrombus area. The cerebral vascular stenosis degree is an important indicator for evaluating the patency of blood vessels, while the thrombus area is directly related to the degree of blood vessel occlusion and the resulting cerebrovascular events (such as stroke). By calculating the stenosis degree and thrombus area of the lesion area, the severity of the lesion can be quantified, and through the comprehensive assessment of the cerebral vascular stenosis degree and thrombus area, the critical number of the lesion can be obtained, that is, the degree of impact of the lesion on the patient's health, which can more clearly understand the impact of the lesion and take corresponding intervention measures to reduce the corresponding health risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments read in conjunction with the accompanying drawings: Figure 1 It is a schematic diagram of the modules of the cerebral vascular image recognition and analysis system based on deep learning of the present invention; Figure 2 is Figure 1 a schematic diagram of the functional flow of the cerebral vascular image registration module in Figure 3 is Figure 1 a schematic diagram of the functional flow of the CT pixel connectivity module in DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] The following clearly and completely describes the technical system of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0017] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and thus repeated descriptions thereof will be omitted. Some of the block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor systems and / or microcontroller systems.

[0018] It should be understood that although terms such as "first" and "second" may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0019] To achieve the above object, please refer to Figures 1 to 3 , the present invention provides a cerebrovascular image recognition and analysis system based on deep learning, and the system includes the following modules: A cerebrovascular image registration module, configured to obtain a cerebrovascular CT angiography image, and perform image enhancement and multimodal image registration on the cerebrovascular CT angiography image to generate a cerebrovascular CT multimodal registration image sequence; A CT pixel connectivity module, configured to obtain a cerebrovascular CT pixel point cloud in the corresponding modality through the cerebrovascular CT multimodal registration image sequence, and perform vascular clustering connectivity analysis on the cerebrovascular CT image in the corresponding modality within the cerebrovascular CT multimodal registration image sequence based on the cerebrovascular CT pixel point cloud in the corresponding modality to generate a corresponding cerebrovascular CT pixel cluster connectivity map in each modality; A lesion area recognition and segmentation module, configured to perform lesion area recognition and segmentation on the corresponding cerebrovascular CT pixel cluster connectivity map in each modality based on a deep learning multi-branch network architecture and by introducing an attention mechanism to generate a cerebrovascular CT image lesion area; A lesion risk number assessment module, configured to obtain the corresponding cerebrovascular stenosis degree and the lesion thrombus area through the cerebrovascular CT image lesion area, and perform lesion risk number assessment on the corresponding cerebrovascular CT image lesion area based on the cerebrovascular stenosis degree and the lesion thrombus area to obtain the size of the cerebrovascular CT lesion risk number.

[0020] In an embodiment of the present invention, please refer to Figure 1As shown in the figure, it is a schematic diagram of the modules of the cerebrovascular image recognition and analysis system based on deep learning according to the present invention. In this example, the cerebrovascular image recognition and analysis system based on deep learning includes the following modules: The cerebrovascular image registration module is used to obtain cerebral vascular CT angiography images, and perform image enhancement and multi-modal image registration on the cerebral vascular CT angiography images to generate a sequence of cerebral vascular CT multi-modal registration images; In the embodiment of the present invention, when obtaining cerebral vascular CT angiography images, a high-precision CT scanning device is used to scan the patient's brain according to a specific scanning scheme to ensure that cerebral vascular images can be clearly captured. After the scanning is completed, image enhancement is performed on the cerebral vascular CT angiography images. The histogram equalization method is adopted to equalize the gray histogram of the image, expand the gray dynamic range of the image, and enhance the contrast of the image. For example, the gray level of an original CT angiography image mainly concentrates between 50-150. After histogram equalization, the gray level range is expanded to 0-255, and the distinction between cerebral blood vessels and surrounding tissues in the image is more obvious. In terms of multi-modal image registration, a registration algorithm based on mutual information is selected. Assuming that the CT translation modality image and the CT rotation modality image are registered, by maximizing the mutual information of the corresponding regions of the two modality images, the spatial position and angle of the image are continuously adjusted to make the cerebral vascular structures in the two modality images coincide as much as possible. After multiple iterative calculations, a sequence of cerebral vascular CT multi-modal registration images is generated, such as a series of images including different modalities such as translation and rotation and with good registration.

[0021] Preferably, the CT pixel connectivity module is used to obtain the cerebral vascular CT pixel point cloud in the corresponding modality through the sequence of cerebral vascular CT multi-modal registration images, and perform vascular clustering connectivity analysis on the cerebral vascular CT images in the corresponding modality within the sequence of cerebral vascular CT multi-modal registration images to generate the corresponding cerebral vascular CT pixel cluster connectivity map in each modality; In an embodiment of the present invention, a cerebrovascular CT voxel cloud corresponding to a corresponding modality is obtained from a cerebrovascular CT multi-modal registration image sequence. First, threshold segmentation is performed on the CT image. A suitable threshold is set, such as 120, and the pixel points with gray values greater than 120 in the image are extracted. These pixel points constitute the approximate area of the cerebrovascular. Then, using a three-dimensional reconstruction algorithm, the cerebrovascular pixel points in these two-dimensional images are mapped into a three-dimensional space to form a cerebrovascular CT voxel cloud. And through the voxel cloud, vascular cluster connectivity analysis is performed on the cerebrovascular CT image corresponding to the corresponding modality in the cerebrovascular CT multi-modal registration image sequence. The density-based spatial clustering algorithm DBSCAN is used, with the neighborhood radius set to 5 voxels and the minimum number of points set to 10. A point is selected from the voxel cloud, and the number of points in its neighborhood is checked. If the number of points is greater than or equal to 10, these points are used as the core points of a cluster, and the cluster is continuously expanded to include all the eligible points in the neighborhood. In this way, all vascular clusters are identified, and the cluster information is marked on the original CT image, and finally, a corresponding cerebrovascular CT pixel cluster connectivity map for each modality is generated. For example, in the CT translation modality, multiple vascular clusters are identified and different clusters are marked with different colors on the image.

[0022] Preferably, a lesion area identification and segmentation module is used to identify and segment the lesion area of the cerebrovascular CT pixel cluster connectivity map corresponding to each modality based on a deep learning multi-branch network architecture and by introducing an attention mechanism, and generate the lesion area of the cerebrovascular CT image. In an embodiment of the present invention, the lesion area of the cerebrovascular CT pixel cluster connectivity map corresponding to each modality is identified and segmented based on a deep learning multi-branch network architecture and by introducing an attention mechanism to construct a convolutional neural network (CNN) with multiple branches. Each branch corresponds to a pixel cluster connectivity map of a modality. Taking the CT translation modality branch as an example, the network structure includes multiple convolutional layers, pooling layers, and fully connected layers. The pixel cluster connectivity map of the CT translation modality is input into the branch network. In the convolutional layer, 3×3 and 5×5 convolutional kernels are used to extract image features, and the ReLU activation function is used to enhance the feature expression. And by introducing an attention mechanism into the network, an attention module is constructed. This module calculates the attention weights at each position in the feature map to weight and enhance the features containing the lesion area. For example, for a certain feature map area, the attention weight is 0.9, while for other areas it is 0.1, highlighting the lesion area features. After the learning and prediction of the network, the segmentation result of the lesion area is output, and finally, the lesion area of the cerebrovascular CT image is generated. During the training process, a large number of samples with lesion annotations are used for training, and the network parameters are continuously adjusted to improve the accuracy of identification and segmentation.

[0023] Preferably, a lesion risk assessment module is configured to obtain the corresponding cerebrovascular stenosis degree and the lesion thrombus area from the lesion region of the cerebrovascular CT image, and perform a lesion risk assessment on the lesion region of the corresponding cerebrovascular CT image based on the cerebrovascular stenosis degree and the lesion thrombus area, so as to obtain the size of the cerebrovascular CT lesion risk.

[0024] In an embodiment of the present invention, the corresponding cerebrovascular stenosis degree and the lesion thrombus area are obtained from the lesion region of the cerebrovascular CT image. When obtaining the cerebrovascular stenosis degree, first determine the center line of the blood vessel in the lesion region, use a thinning algorithm to thin the blood vessel structure in the lesion region to obtain the center line, select a cross-section every 3 voxels along the center line. On each cross-section, use an edge detection algorithm to determine the inner boundary and outer boundary of the blood vessel, calculate the area A1 enclosed by the inner boundary and the area A2 enclosed by the outer boundary, and calculate the blood vessel stenosis degree at this cross-section according to the formula (1 - A1 / A2) × 100%. Take the average value of the stenosis degrees of all cross-sections to obtain the cerebrovascular stenosis degree. For the estimation of the lesion thrombus area, assume that the thrombus is approximately an ellipsoid, use the previously located thrombus region, measure the diameters a, b, and c of the thrombus in the three coordinate axes directions, and calculate the thrombus area according to the ellipsoid surface area formula S = 4π((a^1.6075 * b^1.6075 + a^1.6075 * c^1.6075 + b^1.6075 * c^1.6075) / 3)^(1 / 1.6075). Perform a lesion risk assessment on the lesion region of the corresponding cerebrovascular CT image based on the cerebrovascular stenosis degree and the lesion thrombus area, so as to evaluate and calculate the size of the cerebrovascular CT lesion risk.

[0025] Further, the cerebrovascular image registration module includes the following functions: Obtain a cerebrovascular CT angiography image; Perform gray-scale processing on the cerebrovascular CT angiography image to obtain a cerebrovascular CT gray-scale image; Quantify the pixel blurriness of the cerebrovascular CT gray-scale image to obtain the pixel blurriness of the cerebrovascular CT image; perform pixel blurring denoising on the cerebrovascular CT gray-scale image based on the pixel blurriness of the cerebrovascular CT image to generate a cerebrovascular CT blurred denoised image; Perform image enhancement processing on the cerebrovascular CT blurred denoised image to adopt random rotation, flipping, scaling, and translation to perform modality augmentation on the cerebrovascular CT images at different time points, and generate a cerebrovascular CT multi-modal image sequence; Perform multi-modal image registration on the cerebrovascular CT images corresponding to different time points and different modalities in the cerebrovascular CT multi-modal image sequence to generate a cerebrovascular CT multi-modal registered image sequence.

[0026] As an embodiment of the present invention, refer to Figure 2 shown asFigure 1 Schematic diagram of the functional process of the cerebral vascular image registration module. In this embodiment, the cerebral vascular image registration module includes the following functions: S11: Obtain cerebral vascular CT angiography images; In the embodiment of the present invention, in the radiology department of the hospital, a Siemens SOMATOM DefinitionFlash CT scanner is used to perform cerebral vascular CT angiography on the patient. The patient lies flat on the examination bed of the CT scanner and remains stationary according to the doctor's instructions. The medical staff operates the console of the CT scanner to set the scanning parameters, such as the tube voltage is 120 kV, the tube current is 200 mA, the slice thickness is 0.625 mm, and the scanning range is from the base of the skull to the top of the skull. The CT scanner emits X-rays to penetrate the patient's head, and the detector receives the attenuated X-ray signal passing through the human body and converts it into an electrical signal. After analog-to-digital conversion and computer processing, cerebral vascular CT angiography images are finally generated, and these images are stored in the hospital's PACS (Picture Archiving and Communication System) system in DICOM (Digital Imaging and Communications in Medicine) format.

[0027] S12: Perform grayscale processing on the cerebral vascular CT angiography images to obtain cerebral vascular CT grayscale images; In the embodiment of the present invention, on the image processing workstation in the hospital, MATLAB software is used to perform grayscale processing on the obtained cerebral vascular CT angiography images. To read the corresponding file from the hospital's PACS system, import it into the workspace of MATLAB software, and by using built-in functions in MATLAB, such as the imread function to read the image data, and then use the rgb2gray function to convert the color cerebral vascular CT angiography images into grayscale images. Each pixel in the color image is represented by the color values of the red, green, and blue channels, while each pixel in the grayscale image is represented by only one grayscale value, ranging from 0 (black) to 255 (white). After grayscale processing, cerebral vascular CT grayscale images are finally obtained.

[0028] S13: Quantify the pixel blurriness of the cerebral vascular CT grayscale images to obtain the pixel blurriness of the cerebral vascular CT images; Based on the pixel blurriness of the cerebral vascular CT images, perform pixel blurring denoising on the cerebral vascular CT grayscale images to generate cerebral vascular CT blurred denoised images; In the embodiment of the present invention, on the image processing workstation in the hospital, the MATLAB software is continuously used to quantify the pixel blurriness of the cerebral vascular CT grayscale image, so as to import the corresponding cerebral vascular CT grayscale image into the MATLAB workspace, and by using the image processing toolbox of MATLAB, the pixel blurriness is quantified by calculating the gradient information of the pixels in the image. For example, the sobel operator is used to calculate the gradients of the image in the horizontal and vertical directions, and then the blurriness of the pixels is evaluated according to the magnitude and distribution of the gradients. Moreover, after obtaining the pixel blurriness of the cerebral vascular CT image, pixel blurring denoising is performed on the cerebral vascular CT grayscale image based on this blurriness. The median filtering algorithm is adopted, and the size of the filtering window is set to 3×3. The median filtering algorithm sorts the pixel values within the window according to their magnitudes and takes the median value as the new value of the pixel at the center of the window, so as to remove the noise and blurred pixels in the image. After processing, a cerebral vascular CT blurred denoised image is finally generated.

[0029] S14: Perform image enhancement processing on the cerebral vascular CT blurred denoised image, so as to generate a cerebral vascular CT multi-modal image sequence by performing corresponding modal augmentation such as random rotation, flipping, scaling, and translation on the cerebral vascular CT image at different time points; In the embodiment of the present invention, on the image processing workstation in the hospital, the Python language is used in combination with the OpenCV library to perform image enhancement processing on the cerebral vascular CT blurred denoised image, so as to read the image into the Python program by using the cv2.imread function in the OpenCV library, and perform corresponding modal augmentation such as random rotation, flipping, scaling, and translation on the cerebral vascular CT image at different time points. For example, the cv2.rotate function is used to rotate the image at a random angle, and the range is set to -10° to 10°; the cv2.flip function is used to perform random horizontal or vertical flipping on the image; the cv2.resize function is used to perform random scaling on the image, and the scaling ratio range is set to 0.8 to 1.2; the cv2.warpAffine function is used to perform random translation on the image, and the translation distance is within 10% of the width and height of the image. Through these operations, a series of cerebral vascular CT images with different modalities are finally generated, forming a cerebral vascular CT multi-modal image sequence.

[0030] S15: Perform multi-modal image registration on the cerebral vascular CT images corresponding to different time points and different modalities within the cerebral vascular CT multi-modal image sequence, so as to generate a cerebral vascular CT multi-modal registered image sequence.

[0031] In an embodiment of the present invention, on the image processing workstation in a hospital, the Elastix software is used to perform multi-modal image registration on cerebral vascular CT images corresponding to different time points and different modalities in a multi-modal image sequence of cerebral vascular CT, so as to open the Elastix software and set the registration parameters. For example, select the registration method based on mutual information, set the sampling step size to 4, and the maximum number of iterations to 100. And use the first image as the fixed image and the remaining images as floating images, and perform registration in sequence. The Elastix software finds the optimal transformation parameters between the floating image and the fixed image through an optimization algorithm to make the two as aligned as possible in space. After registration, a multi-modal registration image sequence of cerebral vascular CT is finally generated.

[0032] Further, the multi-modal image registration of cerebral vascular CT images corresponding to different time points and different modalities in the multi-modal image sequence of cerebral vascular CT includes: Performing temporal synchronization annotation on cerebral vascular CT images corresponding to different time points in the multi-modal image sequence of cerebral vascular CT to obtain a multi-modal temporal annotation image of cerebral vascular CT; In an embodiment of the present invention, on the image processing workstation in a hospital, a self-developed image annotation software is used to perform temporal synchronization annotation on cerebral vascular CT images corresponding to different time points in the multi-modal image sequence of cerebral vascular CT, so as to read all multi-modal image files of cerebral vascular CT and open the image annotation software. The time axis of the image sequence is displayed on the software interface. For each image, its corresponding time point is accurately marked on the time axis. For example, if the image acquisition time is the 5th second, the 10th second, etc. after the start of the examination, it is marked at the corresponding position on the time axis. At the same time, the modality information of the image is annotated, such as whether the image is obtained through rotation, flipping, scaling or translation processing. After annotation, a multi-modal temporal annotation image of cerebral vascular CT is finally generated.

[0033] Preferably, the contrast and resolution of the cerebral vascular CT images in the corresponding modality are obtained through the multi-modal temporal annotation image of cerebral vascular CT, and the modality-level sorting analysis is performed on the cerebral vascular CT images corresponding to different time points in the multi-modal temporal annotation image of cerebral vascular CT based on the contrast and resolution of the cerebral vascular CT images in the corresponding modality, so as to obtain the complementary modality-level relationship between the corresponding modalities according to the sorting of the corresponding contrast and resolution, and obtain the corresponding modality-level relationship between the cerebral vascular CT images at different time points; In the embodiment of the present invention, the cerebrovascular CT multi-modal temporal annotation images are analyzed by using the Python language in combination with the SimpleITK library. By using relevant functions in the SimpleITK library, such as the GetImageContrast function to obtain the contrast of the cerebrovascular CT image in the corresponding modality, and calculating the contrast index based on the gray value distribution of the image; using the GetSpacing function to obtain the resolution of the image, that is, the pixel spacing of the image in each direction. For cerebrovascular CT images at different time points, modal hierarchical sorting analysis is performed according to their contrast and resolution. For example, for a group of images processed in different modalities, the images with high contrast and high resolution are ranked at a higher level, and the images with low contrast and low resolution are ranked at a lower level. Through this sorting, the complementary hierarchical relationship between the corresponding modalities is obtained. For example, the rotated image may have an advantage in showing the overall morphology of the blood vessels, while the scaled image has an advantage in showing the details of the blood vessels. The complementary relationship between them is determined through hierarchical sorting, and finally the corresponding modal hierarchical relationship between cerebrovascular CT images at different time points is obtained.

[0034] Preferably, based on the corresponding modal hierarchical relationship between cerebrovascular CT images at different time points, the cerebrovascular CT images corresponding to the cerebrovascular CT multi-modal temporal annotation images at different time points are subjected to image space alignment to obtain the cerebrovascular CT image space transformation matrix between different time points and modalities; In an embodiment of the present invention, by using the ANTS (Advanced Normalization Tools) software to perform image spatial alignment on cerebral vascular CT images corresponding to different time points in a cerebral vascular CT multi-modal temporal annotation image based on the corresponding modal hierarchical relationship between cerebral vascular CT images at different time points, to open the ANTS software and set the spatial alignment parameters, such as selecting the Demons algorithm for image registration, setting the convergence condition as the maximum number of iterations of 200 times and the convergence threshold as 0.001, and by using the image at the higher level in the hierarchical relationship as the reference image and the image at the lower level as the floating image to perform spatial alignment in sequence, the ANTS software generates a spatial transformation matrix of cerebral vascular CT images between different time points and modalities through calculating the spatial transformation between images. For example, for a rotated image and a translated image, the ANTS software calculates a transformation matrix to transform the translated image to be spatially aligned with the rotated image. Taking the alignment of a translated cerebral vascular CT image with a rotated image as an example, the spatial transformation matrix needs to reposition each pixel in the translated image to match the corresponding position in the rotated image. Assuming that the coordinate of a certain vascular feature point in the rotated image is (x1, y1), the corresponding position of this feature point in the translated image needs to be moved to (x2, y2) after being calculated by the ANTS software. The spatial transformation matrix contains transformation parameters such as translation, rotation, and scaling. For example, in the two-dimensional case, a simple spatial transformation matrix , where , , , controls the rotation and scaling of the image, , controls the translation of the image. For cerebral vascular CT images, the ANTS software will accurately calculate the values of each parameter of this matrix by comprehensively considering the morphological and positional characteristics of cerebral blood vessels according to the modal hierarchical relationship of the images.

[0035] Preferably, based on the spatial transformation matrix of cerebral vascular CT images between different time points and modalities, a global optimization registration transformation is performed on the local regions of cerebral vascular CT images corresponding to different time points in a cerebral vascular CT multi-modal image sequence to correct the error corresponding to the spatial transformation matrix at the junction of cerebral blood vessels in cerebral vascular CT images at different time points, so as to generate a globally optimized cerebral vascular CT registration transformation function; In the embodiments of the present invention, by using MATLAB software, based on the spatial transformation matrix of cerebrovascular CT images at different time points and between modalities, global optimization registration transformation is performed on the local regions of corresponding cerebrovascular CT images at different time points in the cerebrovascular CT multi-modal image sequence. A program is written in MATLAB. For key local regions such as the cerebrovascular junction, the spatial transformation matrix is used for registration transformation. For example, by calculating the corresponding positions of pixel points at the cerebrovascular junction in different images, position adjustment is performed according to the spatial transformation matrix. During the adjustment process, the error corresponding to the transformation matrix is corrected. Through multiple iterations of optimization, such as setting the number of iterations to 50 times, and adjusting the transformation matrix parameters according to the error each time. For example, taking the cerebrovascular junction as an example, assuming that under the action of the initial spatial transformation matrix, the coordinates of a pixel point at a certain cerebrovascular junction in one image are (x1, y1), and the corresponding coordinates in another image are (x2, y2), but there is a deviation between the actual positions. The program will enter the iterative optimization process, with 50 calculations set for each iteration. In each iteration, by calculating the error between the actual position and the target position of the pixel point in different images, such as calculating the coordinate differences Delta x = x_actual - x_target, Delta y = y_actual - y_target, and then adjusting the spatial transformation matrix parameters according to these errors, such as adjusting the coefficients in the matrix that control translation, rotation, and scaling. After 50 iterations, a function that can accurately match the pixel points in key regions such as the cerebrovascular junction between different images is obtained, and finally, a globally optimized cerebrovascular CT registration transformation function is generated.

[0036] Preferably, based on the globally optimized cerebrovascular CT registration transformation function, full-modal synchronous alignment is performed on the cerebrovascular CT multi-modal image sequence to generate a cerebrovascular CT multi-modal registration image sequence.

[0037] In the embodiments of the present invention, by using the Python language in combination with the OpenCV library, based on the globally optimized cerebrovascular CT registration transformation function, full-modal synchronous alignment is performed on the cerebrovascular CT multi-modal image sequence to convert it into a form that can be called by Python (such as through MATLAB Engine for Python). All cerebrovascular CT multi-modal image files are read. In the Python program, all images are traversed. For each image, the globally optimized registration transformation function is used for transformation operations. For example, the function is called to perform coordinate transformation on each pixel point of the image, so that the images at different time points and different modalities are completely aligned in space. After full-modal synchronous alignment, a cerebrovascular CT multi-modal registration image sequence is finally generated.

[0038] Further, the CT pixel connectivity module includes the following functions: Perform pixel-level segmentation on each modal image in the cerebrovascular CT multi-modal registration image sequence to obtain the pixel spatial positions and structure points of the cerebrovascular structure in each modality through automatic threshold segmentation or region growing method, and form the corresponding pixel matrix, thereby obtaining the cerebrovascular CT region pixel matrix; Perform voxel density analysis on the corresponding cerebrovascular region voxels in each modal image based on the cerebrovascular CT region pixel matrix to generate a cerebrovascular modal feature volume cloud map; Perform multi-modal contrast enhancement on each voxel structure in the cerebrovascular modal feature volume cloud map to obtain a cerebrovascular multi-modal structure point cloud; Perform spatial point cloud fitting optimization on the cerebrovascular multi-modal structure point cloud to generate the cerebrovascular CT pixel point cloud in the corresponding modality; Perform vessel clustering connectivity analysis on the cerebrovascular CT image in the corresponding modality in the cerebrovascular CT multi-modal registration image sequence based on the cerebrovascular CT pixel point cloud in the corresponding modality to generate the corresponding cerebrovascular CT pixel cluster connectivity map in each modality.

[0039] As an embodiment of the present invention, refer to Figure 3 shown, which is Figure 1 the functional flow diagram of the CT pixel connectivity module in S21: Perform pixel-level segmentation on each modal image in the cerebrovascular CT multi-modal registration image sequence to obtain the pixel spatial positions and structure points of the cerebrovascular structure in each modality through automatic threshold segmentation or region growing method, and form the corresponding pixel matrix, thereby obtaining the cerebrovascular CT region pixel matrix; In the embodiment of the present invention, when performing pixel-level segmentation on each modal image in the cerebrovascular CT multi-modal registration image sequence, the Otsu algorithm is selected as the automatic threshold segmentation method. For an input modal image, its gray value range is divided into multiple intervals, the number of pixels and the sum of gray values in each interval are calculated, and an optimal threshold is determined by maximizing the between-class variance. This threshold can effectively distinguish the cerebrovascular structure in the image from the background. For the region growing method, first select a seed point located in the cerebrovascular region. Based on the set similarity criterion, such as the degree of gray value similarity, continuously merge the pixels adjacent to the seed point and meeting the criterion into the growing region until no further merging is possible. Record the spatial positions of the pixels of the segmented cerebrovascular structure and arrange them in row and column order to form the corresponding pixel matrix, and finally obtain the cerebrovascular CT region pixel matrix. For example, for a 512×512 modal image, the threshold determined by the Otsu algorithm is 120, and the pixels with a value greater than 120 are identified as cerebrovascular structure pixels, and then the pixel matrix is constructed.

[0040] S22: Based on the pixel matrix of the cerebrovascular CT region, perform voxel density analysis on the voxels in the corresponding cerebrovascular region within each modality image to generate a cerebrovascular modality feature volume cloud map; In the embodiment of the present invention, based on the obtained pixel matrix of the cerebrovascular CT region, perform voxel density analysis on the voxels in the corresponding cerebrovascular region within each modality image. Use the gray value of the voxel to characterize its density, view the cerebrovascular region from a three-dimensional spatial perspective. For each voxel, count its relative density relationship among adjacent voxels. By setting a density window, such as a 3×3×3 voxel window centered on the current voxel, calculate the average density of the voxels within the window. Map the density value of each voxel to a color space, assign bright colors to voxels with high density and dull colors to voxels with low density. When generating the cerebrovascular modality feature volume cloud map, use the voxel coordinates as the spatial position and the density-mapped color as the display attribute, and draw each voxel one by one, so as to intuitively present the density differences of different parts within the cerebrovascular region. For example, in a cerebrovascular region, the voxels at the blood vessel wall have high density and are displayed as bright white in the volume cloud map, while the voxels of the blood inside the blood vessel have relatively low density and are displayed as dark red.

[0041] S23: Perform multi-modal contrast enhancement on each voxel structure in the cerebrovascular modality feature volume cloud map to obtain a cerebrovascular multi-modal structure point cloud; In the embodiment of the present invention, by performing multi-modal contrast enhancement on each voxel structure in the cerebrovascular modality feature volume cloud map, first, determine the positional relationship of the corresponding voxels in different modality images. For example, map the voxels of the CT translation modality and the CT rotation modality through coordinate mapping. For each voxel, obtain its eigenvalue in different modalities, such as the density value in CT translation and the signal intensity value in CT rotation. Use the weighted fusion method for contrast enhancement, assign weights according to the importance of different modalities in displaying the cerebrovascular structure. Assume the weight of the CT translation modality is 0.6 and the weight of the CT rotation modality is 0.4. For a certain voxel, its enhanced eigenvalue is 0.6 times the eigenvalue of the CT translation modality plus 0.4 times the eigenvalue of the CT rotation modality. Remap the enhanced eigenvalue back to the voxel space and present it in the form of a point cloud. Each point represents a voxel, its coordinate is the position of the voxel in space, and its attribute is the enhanced eigenvalue, finally obtaining a cerebrovascular multi-modal structure point cloud.

[0042] S24: Perform spatial point cloud fitting optimization on the cerebrovascular multi-modal structure point cloud to generate a cerebrovascular CT pixel point cloud corresponding to the modality; In an embodiment of the present invention, through spatial point cloud fitting optimization of the multi-modal structural point cloud of cerebral blood vessels, the least squares method is selected for polynomial surface fitting. The three-dimensional coordinates of each point in the point cloud data are used as inputs. Assuming the fitted surface equation is in the form of a quadratic polynomial z = ax² + by² + cxy + dx + ey + f, by minimizing the sum of the squares of the distances from the point cloud data points to the fitted surface, the coefficients a, b, c, d, e, and f of the polynomial are solved. For the outlier points in the point cloud, the density-based spatial clustering algorithm DBSCAN is used for identification and elimination. Appropriate neighborhood radii and minimum numbers of points are set. Points with fewer points than the minimum number in the neighborhood are determined as outlier points. After fitting and outlier point processing, the optimized point cloud is rearranged according to the pixel coordinates of the cerebral blood vessel CT image, and finally the cerebral blood vessel CT pixel point cloud in the corresponding modality is generated. For example, in a point cloud containing 1000 points, 50 outlier points are identified and eliminated by the DBSCAN algorithm, and then polynomial surface fitting is performed to generate the pixel point cloud.

[0043] S25: Perform vascular cluster connectivity analysis on the cerebral blood vessel CT image in the corresponding modality in the cerebral blood vessel CT multi-modal registration image sequence based on the cerebral blood vessel CT pixel point cloud in the corresponding modality, so as to generate the cerebral blood vessel CT pixel cluster connectivity graph corresponding to each modality.

[0044] In an embodiment of the present invention, through vascular cluster connectivity analysis of the cerebral blood vessel CT image in the corresponding modality in the cerebral blood vessel CT multi-modal registration image sequence based on the cerebral blood vessel CT pixel point cloud in the corresponding modality, the connected component algorithm in graph theory is utilized. Each point in the pixel point cloud is regarded as a node in the graph, and the connection relationship between the nodes is determined based on the adjacency of the pixel points. If two pixel points are adjacent in space and the difference in gray values is within a certain range, it is considered that there is an edge connection between them. The graph is traversed through the depth-first search algorithm. Starting from a starting node, adjacent nodes are continuously accessed along the edges, and the accessed nodes are marked until no further access is possible. All the nodes accessed in one depth-first search are regarded as a connected component, that is, a vascular cluster. The entire pixel point cloud graph is traversed to identify all the vascular clusters, and the node position information of each vascular cluster is marked on the cerebral blood vessel CT image, with different colors or lines representing different clusters. Finally, the cerebral blood vessel CT pixel cluster connectivity graph corresponding to each modality is generated. For example, in the pixel point cloud graph corresponding to a cerebral blood vessel CT image, 10 vascular clusters are identified through depth-first search and clearly marked on the image.

[0045] Further, the spatial point cloud fitting optimization of the multi-modal structural point cloud of cerebral blood vessels includes: Perform pixel density statistical analysis on the cerebral vascular CT structure point cloud in the corresponding modality within the multi-modal structure point cloud of cerebral vessels to obtain the pixel density values corresponding to each cerebral vascular CT point cloud in the corresponding modality; In an embodiment of the present invention, when performing pixel density statistical analysis on the cerebral vascular CT structure point cloud in the corresponding modality within the multi-modal structure point cloud of cerebral vessels, the cerebral vascular CT structure point cloud is divided into multiple small spatial regions. For example, taking 10×10×10 voxels as a regional unit, for each regional unit, count the number of cerebral vascular CT point cloud pixels it contains. Assuming the volume of a certain regional unit is V and the number of pixels it contains is N, then the pixel density value corresponding to this regional unit is N / V. Calculate in this way for all regional units, so as to obtain the pixel density values of each cerebral vascular CT point cloud in different regions in the corresponding modality. Taking a cerebral vascular CT structure point cloud containing 1000×1000×1000 voxels as an example, it is divided into 100×100×100 regional units of 10×10×10, and calculate the pixel density value of each unit. For example, if there are 50 pixels in a certain unit and its volume is 10×10×10 = 1000 cubic voxels, then the pixel density value of this unit is 50 / 1000 = 0.05.

[0046] Preferably, perform pixel outlier removal on the cerebral vascular CT structure point cloud in the corresponding modality based on the pixel density values corresponding to each cerebral vascular CT point cloud in the corresponding modality to obtain the cerebral vascular CT pixel outlier removal point cloud in the corresponding modality; In an embodiment of the present invention, by performing pixel outlier removal on the cerebral vascular CT structure point cloud in the corresponding modality within the multi-modal structure point cloud of cerebral vessels based on the pixel density values corresponding to each cerebral vascular CT point cloud in the corresponding modality, set a density threshold range. For example, taking the overall pixel density average value μ as the benchmark, floating 1.5 times the standard deviation σ up and down as the threshold range (μ - 1.5σ, μ + 1.5σ). For the pixel density value of each regional unit where the cerebral vascular CT point cloud pixel is located, if it exceeds this threshold range, then determine that this pixel is an outlier, and remove these outliers from the cerebral vascular CT structure point cloud to obtain the cerebral vascular CT pixel outlier removal point cloud in the corresponding modality. For example, after calculation, the overall pixel density average value is 0.1 and the standard deviation is 0.02, then the threshold range is (0.1 - 1.5×0.02, 0.1 + 1.5×0.02), that is, (0.07, 0.13). If the pixel density value of a certain regional unit is 0.15, then the cerebral vascular CT point cloud pixel in this unit is determined as an outlier and removed.

[0047] Preferably, perform local smoothing fitting processing on the cerebral vascular CT pixel outlier removal point cloud in the corresponding modality to generate the cerebral vascular CT pixel point cloud in the corresponding modality.

[0048] In an embodiment of the present invention, local smoothing fitting processing is performed on the cerebrovascular CT pixel outlier removal point cloud in the corresponding modality. The moving least squares method is used for the operation. Taking each point as the center, a local neighborhood is set, such as a spherical neighborhood with a radius r = 5 voxels. Within this neighborhood, a local fitting surface is constructed using the moving least squares method. For each point within the neighborhood, the distance from the point to the fitting surface is calculated, and the position of the point is adjusted according to the distance to make the point cloud smoother. Suppose a point P is within its neighborhood, and the fitting surface equation obtained by the moving least squares method is z = a 0 + a 1 x + a 2 y + a 3 x² + a 4 xy + a 5 y². The perpendicular distance d from point P to the surface is calculated. If d is greater than the set deviation value (such as 0.5 voxel distance), the position of point P is corrected according to the direction and magnitude of d. Such local smoothing processing is performed on each point of the entire cerebrovascular CT pixel outlier removal point cloud, and finally, the cerebrovascular CT pixel point cloud in the corresponding modality is generated.

[0049] Furthermore, the vascular clustering and connectivity analysis of the cerebrovascular CT image in the corresponding modality in the cerebrovascular CT multimodal registration image sequence based on the cerebrovascular CT pixel point cloud in the corresponding modality includes: Performing spatial neighborhood division on the cerebrovascular CT pixel point cloud in the corresponding modality to divide each local spatial small branch of the cerebrovascular by using the local density distribution based on the point cloud, and obtaining the local spatial structure diagram of the cerebrovascular CT point cloud; In an embodiment of the present invention, when performing spatial neighborhood division on the cerebrovascular CT pixel point cloud in the corresponding modality, the KD - tree algorithm is used to construct a spatial index. Taking each pixel point as the query point, the neighborhood radius is set to r, for example, r = 3 voxel distances. All points within the neighborhood of each query point are quickly retrieved through the KD - tree. Based on the distribution of these neighborhood points, the local density is calculated. Suppose there are n points within the neighborhood of a certain point, and the neighborhood volume is V, then the local density is n / V. Using the density threshold D as the division basis, if the local density of a certain neighborhood is greater than D, such as D = 0.08, it is considered that this neighborhood belongs to a local spatial small branch of the cerebrovascular. By marking and connecting all eligible local spatial small branches, the local spatial structure diagram of the cerebrovascular CT point cloud is formed. For example, in a cerebrovascular CT pixel point cloud region, 20 points are retrieved within the neighborhood of a certain point through the KD - tree, the neighborhood volume is 27 cubic voxels (3×3×3), and its local density is 20 / 27≈0.74, which is greater than the threshold 0.08, and this neighborhood is identified as a local spatial small branch.

[0050] Preferably, vascular topology node connection analysis is performed according to the local spatial structure diagram of cerebral vascular CT point cloud to identify the branch points, intersection points and end points corresponding to the cerebral blood vessels, and their topological connections are made to form the corresponding vascular node connection structure, so as to generate the cerebral vascular CT point cloud node connection diagram; In the embodiment of the present invention, vascular topology node connection analysis is performed according to the local spatial structure diagram of cerebral vascular CT point cloud. By taking the end points of the local spatial small branches and the points at the branch points as candidate nodes, for each candidate node, the connection directions of the points in its neighborhood are analyzed. If a node has three or more neighborhood point connection branches in different directions, it is determined as a branch point; if two branches in different directions intersect at a point, this point is an intersection point; if a point has only one neighborhood point connection branch, it is an end point. The connection relationship between nodes is represented by the adjacency matrix in graph theory. For example, if node A is connected to nodes B and C, the values at the corresponding positions (A, B) and (A, C) in the adjacency matrix are set to 1, and the rest are 0. According to the adjacency matrix, the identified branch points, intersection points and end points are topologically connected according to their connection relationships, and finally the cerebral vascular CT point cloud node connection diagram is generated. For example, in a simple cerebral vascular structure, there are nodes a, b, and c, where a is connected to b and c, and the adjacency matrix is [[0, 1, 1], [1, 0, 0], [1, 0, 0]], and the node connection diagram is constructed accordingly.

[0051] Preferably, progressive connectivity analysis is performed on the cerebral vascular CT point cloud node connection diagram to calculate the connectivity paths between cerebral vascular nodes by using the shortest path recursion, and gradually merge the local vascular nodes into the corresponding vascular clusters to generate the cerebral vascular cluster connectivity sequence; In the embodiment of the present invention, through progressive connectivity analysis of the cerebral vascular CT point cloud node connection diagram, the Dijkstra algorithm is used to calculate the shortest path. Taking each node as the starting point, the shortest path from it to all other nodes is calculated. For example, for node i, all the edges in the graph are traversed by the Dijkstra algorithm, and the shortest path length d(i, j) from i to other nodes j is recorded. Starting from the node closest to the starting point, it is merged with the starting point into a vascular cluster. Then, taking the merged cluster as the new starting point, continue to calculate the shortest path to other nodes, and incorporate the node closest to this cluster into the cluster. Repeat this process. For example, starting from node A, the shortest path length to node B is 3, and to node C is 5. First, merge A and B into a cluster, then calculate the shortest path from this cluster to other nodes, and gradually merge the local vascular nodes into the corresponding vascular clusters, and finally generate the cerebral vascular cluster connectivity sequence.

[0052] Preferably, based on the cerebrovascular cluster connection sequence, the cerebrovascular CT pixel point cloud in the corresponding modality is optimized for the details of blood vessel clusters, so as to eliminate the tiny connection parts between cerebrovascular clusters according to the cerebrovascular cluster connection structures corresponding to different modalities, and ensure a high degree of connectivity in the cerebrovascular network connection of different modalities, so as to generate the corresponding cerebrovascular CT pixel cluster connection map in each modality.

[0053] In the embodiment of the present invention, by optimizing the details of blood vessel clusters for the cerebrovascular CT pixel point cloud in the corresponding modality based on the cerebrovascular cluster connection sequence, for the cerebrovascular cluster connection structures corresponding to different modalities, a tiny connection threshold T is set. For example, T = 2 voxel distances. The connection parts between cerebrovascular clusters are checked. If the length or width of the connection part is less than T, it is considered a tiny connection part and is eliminated. At the same time, to ensure a high degree of connectivity in the cerebrovascular network connection of different modalities, the connected component algorithm in graph theory is used to check the optimized structure. If there are unconnected parts, they are connected by adding virtual connection edges. The weight of the virtual connection edge is set to a relatively large value, such as 100, to indicate that this is an auxiliary connection added to ensure connectivity. After such processing, the corresponding cerebrovascular CT pixel cluster connection map in each modality is generated. For example, in the cerebrovascular cluster connection structure of a certain modality, it is found that there is a connection part with a length of 1 voxel between two clusters, which is less than the threshold of 2 and is eliminated. Then the connectivity is checked, and it is found that there is an isolated node, which is connected to other clusters by adding a virtual connection edge, and finally the pixel cluster connection map is generated.

[0054] Further, the lesion area recognition and segmentation module includes the following functions: Perform cerebrovascular region feature analysis on the corresponding cerebrovascular CT pixel cluster connection map in each modality to obtain the corresponding cerebrovascular region pixel differences and morphological features in each modality; In the embodiments of the present invention, for the connected graphs of cerebrovascular CT pixel clusters corresponding to their respective modalities, the gray-level co-occurrence matrix (GLCM) is used to analyze the characteristics of the cerebrovascular region. Taking the connected graph of pixel clusters of one modality as an example, the offsets are set to (1, 0) and (0, 1), and the frequencies of pixel pairs with different gray levels in these two offset directions are calculated to construct the GLCM matrix. Texture feature values such as contrast, correlation, energy, and entropy are extracted from the GLCM matrix. These values reflect the pixel difference situation in the cerebrovascular region. For the analysis of morphological features, the morphological skeleton extraction algorithm is used. Through erosion and thinning operations, the skeleton structure of the cerebrovascular is obtained, and morphological parameters such as the length and branch angle of the blood vessels are measured from the skeleton structure. For example, in a connected graph of pixel clusters in the CT modality, the contrast calculated by GLCM is 0.56 and the energy is 0.32. At the same time, the branch length of a certain blood vessel is measured as 15 voxels and the branch angle is 45 degrees from the skeleton structure, so as to obtain the pixel difference and morphological features of the cerebrovascular region in this modality.

[0055] Preferably, the pixel differences and morphological features of the cerebrovascular region corresponding to their respective modalities are input into the corresponding modality network branch in the deep learning multi-branch network architecture for modality feature fusion, so as to enhance and fuse the complementary information between different modality cerebrovascular images to generate a corresponding structural feature map of the cerebrovascular, and the attention mechanism is introduced to highlight and label the lesion regions included in the corresponding structural feature map of the cerebrovascular, so as to generate a cerebrovascular lesion region recognition model and output a corresponding cerebrovascular lesion sensitive region map; In an embodiment of the present invention, by inputting the pixel differences and morphological features of the corresponding cerebrovascular regions in each modality into the corresponding modality network branches within a deep learning multi-branch network architecture. Assuming that the multi-branch network architecture includes a CT translation modality branch, a CT rotation modality branch, etc., taking the CT translation modality branch as an example, the branch network structure adopts a convolutional neural network (CNN), which includes multiple convolutional layers, pooling layers, and fully connected layers. After normalizing the eigenvalue under the CT translation modality, it is input into the input layer of the CNN. In the convolutional layer, convolutional kernels of different sizes, such as 3×3 and 5×5 convolutional kernels, are used to extract features from the input features, and the feature expression is enhanced through the activation function ReLU. The features extracted by different modality branches are fused in the fusion layer, and a weighted average method is adopted. Weights are assigned according to the importance of different modalities for lesion recognition. For example, the weight of the CT translation modality is 0.6, and the weight of the CT rotation modality is 0.4. An attention mechanism is introduced during the fusion process to construct an attention module. By calculating the attention weights at each position in the feature map, the features containing the lesion area are weighted and enhanced to highlight the lesion area. For example, for a certain feature map area, the attention weight is 0.8, while for other areas it is 0.2. After weighting, the features of the lesion area are highlighted. Finally, a structural feature map corresponding to the cerebrovascular is generated, and a cerebrovascular lesion area recognition model is constructed to output a corresponding cerebrovascular lesion sensitive area map.

[0056] Preferably, region segmentation and contour extraction are performed on the corresponding lesion areas in the cerebrovascular lesion sensitive area map, so as to extract the lesion area boundary by combining edge detection and region growing, and form a clear lesion area contour in the cerebrovascular lesion sensitive area map to generate an accurate segmentation map of the cerebrovascular lesion area; In an embodiment of the present invention, by performing region segmentation and contour extraction on the corresponding lesion areas in the cerebrovascular lesion sensitive area map, first, the Canny edge detection algorithm is used, and the high and low thresholds are set to 100 and 200 respectively to perform edge detection on the lesion sensitive area map to obtain preliminary information on the edge contour of the lesion area. Then, the edge points obtained by edge detection are used as seed points, and the region growing algorithm is used for region segmentation. The similarity criterion is set that the gray value difference is within 10. Starting from the seed points, the pixels in the neighborhood that meet the similarity criterion are merged into the growing region until no further merging is possible. During the region growing process, the boundary of the growing region is continuously updated. For example, starting from a seed point, there are 3 pixels in its neighborhood whose gray values differ from the seed point by within 10, and these 3 pixels are incorporated into the growing region, and the region boundary is updated at the same time. By combining edge detection and region growing in this way, the lesion area boundary is extracted, and a clear lesion area contour is formed in the cerebrovascular lesion sensitive area map, and finally an accurate segmentation map of the cerebrovascular lesion area is generated.

[0057] Preferably, perform morphological optimization on the accurately segmented map of the cerebrovascular lesion area to repair the small breaks in the lesion area through erosion and dilation operations, smooth the segmentation boundary, and enhance the detail expression of the lesion area, thereby generating the lesion area of the cerebrovascular CT image.

[0058] In the embodiment of the present invention, by performing morphological optimization on the accurately segmented map of the cerebrovascular lesion area, an erosion operation is adopted. A 3×3 structuring element is selected to erode the lesion area in the segmented map to remove some isolated noise points and small burrs in the lesion area. For example, there are some isolated single-pixel points at the boundary of the lesion area, which are removed after the erosion operation. Then, a dilation operation is performed. Similarly, a 3×3 structuring element is used to dilate the eroded image to restore the main part of the lesion area, repair the small breaks caused by erosion, and smooth the segmentation boundary. During the dilation process, the boundary of the lesion area becomes more continuous and smooth. For example, there are some small breaks with a width of 1-2 voxels inside the lesion area, which are connected after the dilation operation. Through multiple alternating operations of erosion and dilation, the detail expression of the lesion area is enhanced, and finally the lesion area of the cerebrovascular CT image is generated.

[0059] Furthermore, the lesion risk number evaluation module includes the following functions: Obtain the corresponding cerebrovascular stenosis degree through the lesion area of the cerebrovascular CT image; In the embodiment of the present invention, to obtain the corresponding cerebrovascular stenosis degree by using the lesion area of the cerebrovascular CT image, first, in the generated map of the lesion area of the cerebrovascular CT image, determine the target blood vessel segment based on the previously extracted cerebrovascular skeleton structure. Adopt a blood vessel centerline extraction algorithm to refine the blood vessel structure in the lesion area to obtain the blood vessel centerline. Select multiple equally spaced cross-sections along the centerline. For example, select a cross-section every 2 voxels. On each cross-section, through an edge detection algorithm such as the Sobel operator, determine the inner boundary and outer boundary of the blood vessel, calculate the area A1 enclosed by the inner boundary and the area A2 enclosed by the outer boundary, and calculate the blood vessel stenosis degree at this cross-section according to the formula (1 - A1 / A2)×100%. For example, at a certain cross-section, the inner boundary area A1 is 10 square voxels, and the outer boundary area A2 is 20 square voxels, then the blood vessel stenosis degree at this cross-section is (1 - 10 / 20)×100% = 50%. Take the average value of the stenosis degrees of all selected cross-sections to finally obtain the cerebrovascular stenosis degree of the target blood vessel segment.

[0060] Preferably, locate the thrombus area in the lesion area of the cerebrovascular CT image to obtain the thrombus location position of the cerebrovascular CT lesion. In the embodiment of the present invention, by locating the thrombus area in the lesion area of the cerebrovascular CT image, the thrombus area is identified based on the density and texture feature differences of the pixels in the lesion area. The Gaussian mixture model (GMM) is used to model the pixels in the lesion area, and the pixels are divided into different categories. Since the density of the thrombus is different from that of the surrounding normal tissues and blood, its pixel features will be concentrated in specific categories. The number of categories of the GMM is set to 3, corresponding to normal tissues, blood, and thrombus respectively. The parameters of the GMM, including the mean, covariance, and weight, are estimated by the expectation-maximization (EM) algorithm. For each pixel in the lesion area, the probability of it belonging to each category is calculated, and the category with the highest probability is the category to which the pixel belongs. The pixels belonging to the thrombus category are marked, and the set of these marked pixels is the thrombus area. For example, in a lesion area, after GMM classification, it is found that the pixels in a certain area have an 80% probability of belonging to the thrombus category, and this area is located as the thrombus area, so as to obtain the thrombus localization position of the cerebrovascular CT lesion.

[0061] Preferably, based on the thrombus localization position of the cerebrovascular CT lesion, the spatial diameter of the corresponding thrombus in the lesion area of the cerebrovascular CT image is measured to obtain the diameter size of the cerebrovascular CT lesion thrombus; In the embodiment of the present invention, by measuring the spatial diameter of the corresponding thrombus in the lesion area of the cerebrovascular CT image based on the thrombus localization position of the cerebrovascular CT lesion, with the centroid of the thrombus area as the reference point, the thrombus area is transformed into a coordinate system with the centroid as the origin through three-dimensional coordinate transformation, and projected along the three coordinate axes (x, y, z) of the thrombus area. In each projection direction, the maximum inscribed sphere method is used to measure the diameter of the thrombus. For example, in the x-axis projection direction, starting from one end of the thrombus area, a virtual sphere is gradually moved with a fixed step size (such as 0.5 voxel), and the radius of the sphere is continuously increased until the sphere is tangent to the boundary of the thrombus area. At this time, the diameter of the sphere is the diameter of the thrombus in this projection direction. The diameters measured in the three coordinate axes are recorded. For example, the diameter in the x-axis direction is 5 voxels, the y-axis direction is 4 voxels, and the z-axis direction is 6 voxels, so as to obtain the diameter size of the cerebrovascular CT lesion thrombus.

[0062] Preferably, the thrombus area is estimated according to the diameter size of the cerebrovascular CT lesion thrombus to obtain the lesion thrombus area; In an embodiment of the present invention, by estimating the thrombus area according to the diameter of the thrombus in the cerebrovascular CT lesion, assuming that the thrombus is approximately an ellipsoid, and according to the diameters measured in the three coordinate axis directions (denoted as a, b, c), the surface area formula of the ellipsoid S = 4π((a^p*b^p+a^p*c^p+b^p*c^p) / 3)^(1 / p) is used to estimate the thrombus area, where p is taken as 1.6075 (this value is an empirical value and is applicable to the area estimation of most objects approximated as ellipsoids). For example, if a = 5 voxels, b = 4 voxels, and c = 6 voxels are measured and substituted into the formula, we can get S = 4π((5^1.6075*4^1.6075+5^1.6075*6^1.6075+4^1.6075*6^1.6075) / 3)^(1 / 1.6075). After calculation, the area of the lesion thrombus is obtained. If the shape of the thrombus is quite different from that of the ellipsoid, multiple small sub-regions are divided within the thrombus area, and each sub-region is approximated as an ellipsoid for area calculation, and finally the total thrombus area is obtained by accumulation.

[0063] Preferably, based on the cerebrovascular stenosis degree and the lesion thrombus area, the lesion risk number calculation formula is used to evaluate the lesion risk number of the corresponding cerebrovascular CT image lesion area, so as to obtain the size of the cerebrovascular CT lesion risk number.

[0064] In an embodiment of the present invention, by combining the cerebrovascular stenosis degree, the lesion influence weight, and the lesion thrombus area, a suitable lesion risk number calculation formula is formed to evaluate and calculate the lesion risk number of the corresponding cerebrovascular CT image lesion area, so as to quantitatively calculate the risk degree corresponding to the CT image lesion. This value can be used to evaluate the risk degree of cerebrovascular lesions. The higher the value, the higher the lesion risk degree. Finally, the size of the cerebrovascular CT lesion risk number is obtained. In addition, this lesion risk number calculation formula can also use any one of the risk detection algorithms in the field to replace the process of lesion risk number evaluation, and is not limited to this lesion risk number calculation formula. For example, the lesion risk number calculation formula is set as R = w1*S + w2*N, where R is the lesion risk number, S is the lesion thrombus area, N is the cerebrovascular stenosis degree, and w1 and w2 are weight coefficients.

[0065] Furthermore, the obtaining of the corresponding cerebrovascular stenosis degree through the cerebrovascular CT image lesion area includes: Performing vascular lumen geometric shape analysis on the cerebrovascular CT image lesion area to generate a cerebrovascular CT lumen contour map; In the embodiment of the present invention, by analyzing the geometric shape of the blood vessel lumen in the lesion area of the cerebral blood vessel CT image to generate a cerebral blood vessel CT lumen contour map. First, morphological operations are used to preprocess the lesion area image. An erosion operation is adopted, and a 3×3 structuring element is selected to remove some small noise points and irrelevant tiny structures in the lesion area, making the blood vessel lumen structure clearer. Then a dilation operation is carried out, also using a 3×3 structuring element, to restore the main part of the blood vessel lumen that has become smaller due to erosion. After that, an edge detection algorithm, such as the Canny edge detection algorithm, is used. The high and low thresholds are set to 80 and 160 respectively to extract the edges of the preprocessed image, obtaining the approximate edge contour of the blood vessel lumen. Then, through the contour tracking algorithm, starting from the edge points, adjacent edge points are sequentially connected in a clockwise or counterclockwise direction to form a closed blood vessel lumen contour. These contour information is drawn on a blank image in the form of lines to generate a cerebral blood vessel CT lumen contour map. For example, in a lesion area of a cerebral blood vessel CT image, after erosion, dilation, and Canny edge detection, a series of edge points are obtained, and these points are connected into a continuous contour through contour tracking to draw the lumen contour map.

[0066] Preferably, the inner and outer wall distances are measured according to the cerebral blood vessel CT lumen contour map, and the diameter corresponding to the cerebral blood vessel lumen is calculated point by point for each cross-section and mapped to the corresponding CT image position to generate a cerebral blood vessel cross-sectional diameter map. In the embodiment of the present invention, by measuring the inner and outer wall distances according to the cerebral blood vessel CT lumen contour map, the diameter corresponding to the cerebral blood vessel lumen is calculated point by point for each cross-section and mapped to the corresponding CT image position to generate a cerebral blood vessel cross-sectional diameter map. The cerebral blood vessel CT lumen contour map is sliced in a direction perpendicular to the blood vessel direction to obtain multiple cross-sectional images. For each cross-sectional image, with the centroid of the contour as the reference point, rays are emitted from the centroid to the contour edge, and the ray interval angle is set to 5 degrees. Along each ray, starting from the centroid, search towards the contour edge. When a contour edge point is encountered, record the distance at this time. This distance is the inner and outer wall distance at this point. Since the blood vessel is approximately circular, multiply this distance by 2 to obtain the diameter of the cerebral blood vessel lumen at this point. For example, in a certain cross-sectional image, a ray starting from the centroid encounters a contour edge point at a distance of 3 voxels from the centroid, then the diameter of the cerebral blood vessel lumen at this point is 6 voxels. Record the diameter values of all points on each cross-sectional image, and mark the diameter values at the corresponding positions on a new blank image according to their position information in the original CT image, finally generating a cerebral blood vessel cross-sectional diameter map.

[0067] Preferably, the diameter differences between different cross-sections within the cerebral blood vessel cross-sectional diameter map are calculated to obtain the cerebral blood vessel diameter differences between different cross-sections within the lesion area. In an embodiment of the present invention, by calculating the diameter difference between cross-sections in different regions within the cross-sectional diameter map of the cerebral blood vessels, the cross-sectional diameter map of the cerebral blood vessels is divided into multiple regions. For example, the blood vessel length direction is divided into 10 regions at equal intervals. For the cross-sections of adjacent regions, points at corresponding positions are selected, such as the points at the center positions of each cross-section, and the diameter differences of these corresponding points are calculated. Suppose the diameter of the point at the center position of the cross-section of region A at time t is d1, and the diameter of the point at the center position of the cross-section of adjacent region B is d2, then the diameter difference is |d1(t) - d2(t)|. Such calculations are performed for all corresponding points of the cross-sections of adjacent regions, obtaining a series of diameter differences. For example, in the cross-sections adjacent to region 1 and region 2, the diameter at the center of the cross-section of region 1 is 8 voxels, and the diameter at the center of the cross-section of region 2 is 6 voxels. Then the diameter difference between the points at the center positions of these two cross-sections is |8 - 6| = 2 voxels. These diameter differences are sorted and recorded, and finally the cerebral blood vessel diameter differences between different cross-sections within the lesion region are obtained.

[0068] Preferably, based on the cerebral blood vessel diameter differences between different cross-sections within the lesion region, the stenosis degree of the lesion region of the cerebral blood vessel CT image is evaluated and calculated using the cerebral blood vessel stenosis evaluation calculation formula to obtain the cerebral blood vessel stenosis degree. In an embodiment of the present invention, by obtaining the reference diameter corresponding to the normal region of the cerebral blood vessels, the total number of corresponding cross-sections of the cerebral blood vessels within the lesion region, the cerebral blood vessel diameters corresponding to the cross-sections within the lesion region at a certain time, the size of the lesion region, the length of the cerebral blood vessels corresponding within the lesion region, and combining the time interval length, time variable parameter, cerebral blood vessel diameter difference adjustment coefficient, and related parameters, a suitable cerebral blood vessel stenosis evaluation calculation formula is formed for quantitative calculation to quantitatively obtain the cerebral blood vessel stenosis degree. This value can be used to evaluate the stenosis condition of the lesion region of the cerebral blood vessels, and the higher the value, the higher the stenosis degree.

[0069] Among them, the cerebral blood vessel stenosis evaluation calculation formula is specifically: ; In the formula, is the cerebral blood vessel stenosis degree, is the reference diameter corresponding to the normal region of the cerebral blood vessels, is the time interval length, is the time variable parameter, is the total number of corresponding cross-sections of the cerebral blood vessels within the lesion region, is the th cross-section within the lesion region at time corresponding cerebral blood vessel diameter, is the The corresponding cerebrovascular diameter below, is the cerebrovascular diameter difference adjustment coefficient, is an exponential function, is the area size of the lesion area, is the length corresponding to the cerebrovascular in the lesion area.

[0070] The present invention obtains a cerebrovascular stenosis evaluation calculation formula through using a specific mathematical model and verification, which is used to evaluate and calculate the stenosis degree of the lesion area in the cerebrovascular CT image. This formula fully considers the cerebrovascular stenosis degree , the reference diameter corresponding to the normal area of the cerebrovascular , the length of the time interval , the time variable parameter , the total number of cross-sections corresponding to the cerebrovascular in the lesion area , the th cross-section in the lesion area at time The corresponding cerebrovascular diameter below , the th cross-section in the lesion area at time The corresponding cerebrovascular diameter below , the cerebrovascular diameter difference adjustment coefficient , the exponential function , the area size of the lesion area , the length corresponding to the cerebrovascular in the lesion area , according to the cerebrovascular stenosis degree And the mutual correlation relationship between the above parameters constitutes a functional relationship , this formula can achieve the calculation process of evaluating the stenosis degree of the lesion area in cerebral vascular CT images. At the same time, this formula incorporates multiple factors, including the diameter differences between different cross-sections within the lesion area, the geometric shape of the area, and the length of the lesion area, etc., and can more comprehensively reflect the actual situation of cerebral vascular stenosis. By introducing time parameters, the formula can dynamically evaluate the changing trend of cerebral vascular stenosis, which is crucial for observing the evolution of vascular stenosis at different time points, especially for tracking the intervention effects of chronic lesions. The diameter differences in this formula can reflect the morphological changes of cerebral blood vessels at different cross-sections. By calculating these differences and combining with other factors, the degree of vascular stenosis can be accurately evaluated, especially when there are obvious differences in each cross-section within the lesion area. The morphological changes and tortuosity of cerebral blood vessels will also affect blood flow. Therefore, this formula comprehensively evaluates the vascular stenosis situation by introducing the geometric features (such as length, area, etc.) of the lesion area, avoiding the limitations of traditional single diameter or morphology evaluation. The degree of cerebral vascular stenosis is not only related to its diameter, but also closely related to factors such as the thickness of the blood vessel wall and the morphological changes of the inner and outer walls. By combining these factors, this formula can accurately evaluate various lesion areas, especially applicable to cases where vascular stenosis is relatively complex or has irregular morphology. At the same time, the area of the lesion area is added to this formula, making the calculation of the stenosis degree more accurately reflect the actual situation of different lesion degree areas.

[0071] Personalized intervention: According to the evaluation results of the stenosis degree, it can guide doctors to decide whether to take interventional treatment, drug intervention or other treatment means. This helps to accurately formulate treatment plans and avoid over-treatment or under-treatment.

[0072] Furthermore, the specific formula for the lesion risk number is as follows: ; In the formula, is the size of the cerebral vascular CT lesion risk number, is the cerebral vascular stenosis degree, is the lesion influence weight, is the lesion thrombus area.

[0073] The present invention obtains a formula for calculating the lesion risk number through the use of a specific mathematical model and verification, and is used to evaluate the lesion risk number of the corresponding lesion area in cerebral vascular CT images. This formula fully considers the size of the cerebral vascular CT lesion risk number , the cerebral vascular stenosis degree , the lesion influence weight , the lesion thrombus area , and constitutes a functional relationship based on the mutual correlation relationship between the size of the cerebral vascular CT lesion risk number and the above parameters. , this formula can achieve the process of evaluating the lesion risk number of the corresponding lesion area in the cerebrovascular CT image. At the same time, by combining the cerebrovascular stenosis degree and the lesion thrombus area, it quantifies the harm degree of the lesion. The cerebrovascular stenosis degree is a key indicator for evaluating vascular patency, while the thrombus area directly affects the blood flow in the blood vessel. By combining these two factors, the formula can accurately reflect the lesion degree of the cerebrovascular. Secondly, the lesion influence weight in this formula can dynamically adjust the influence degree of the lesion, which means that in different clinical situations, the influence weight of the lesion can be adjusted according to the specific characteristics of the lesion (such as lesion type, location, patient condition, etc.), so as to make the evaluation of the lesion risk number more personalized and accurate. The exponential decay part in this formula can effectively reflect the non-linear influence of the lesion thrombus area. As the thrombus area increases, the harm degree of the lesion increases exponentially. This model can more accurately simulate the aggravating influence of thrombus expansion on the vascular stenosis degree, thus providing a more scientific basis for lesion evaluation. By calculating the lesion risk number, this formula provides a quantitative evaluation index for doctors, which can help doctors better understand the severity of vascular lesions, which is helpful for risk assessment of patients and timely formulation of more effective lesion intervention plans.

[0074] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, it is intended to cover all changes falling within the meaning and scope of the equivalent elements of the application documents within the present invention.

[0075] The above are only the specific implementation manners of the present invention, enabling those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A cerebrovascular image recognition and analysis system based on deep learning, characterized in that: Includes the following modules: A cerebrovascular image registration module is used to obtain cerebrovascular CT angiography images, and perform image enhancement and multimodal image registration on the cerebrovascular CT angiography images to generate a cerebrovascular CT multimodal registration image sequence; A CT pixel connectivity module is used to obtain a cerebrovascular CT pixel point cloud under a corresponding modality through a cerebrovascular CT multimodal registration image sequence, and perform a vascular cluster connectivity analysis on the cerebrovascular CT images under a corresponding modality in the cerebrovascular CT multimodal registration image sequence based on the cerebrovascular CT pixel point cloud under the corresponding modality, so as to generate a cerebrovascular CT pixel cluster connectivity map corresponding to each modality; The lesion area recognition and segmentation module is used to identify and segment the lesion area of ​​the corresponding cerebrovascular CT pixel cluster connectivity map under each modality based on the deep learning multi-branch network architecture and by introducing the attention mechanism to generate the lesion area of ​​the cerebrovascular CT image; The lesion risk assessment module is used to obtain the corresponding cerebrovascular stenosis and lesion thrombus area through the lesion area of ​​the cerebrovascular CT image, and to perform lesion risk assessment on the corresponding lesion area of ​​the cerebrovascular CT image based on the cerebrovascular stenosis and lesion thrombus area to obtain the size of the cerebrovascular CT lesion risk.

2. The cerebrovascular image recognition and analysis system based on deep learning according to claim 1, characterized in that: The cerebrovascular image registration module includes the following functions: Acquire cerebral vascular CT angiography images; Grayscale processing is performed on the cerebral vascular CT angiography image to obtain a cerebral vascular CT grayscale image; Quantify the pixel fuzziness of the cerebrovascular CT grayscale image to obtain the pixel fuzziness of the cerebrovascular CT image; Based on the pixel fuzziness of the cerebrovascular CT image, pixel fuzzy denoising is performed on the cerebrovascular CT grayscale image to generate a cerebrovascular CT fuzzy denoised image; Perform image enhancement processing on the fuzzy denoised cerebrovascular CT images, and use the corresponding modes of random rotation, flipping, scaling and translation to expand the cerebrovascular CT images at different time points to generate a cerebrovascular CT multimodal image sequence; Multimodal image registration is performed on cerebrovascular CT images corresponding to different modalities at different time points in the cerebrovascular CT multimodal image sequence to generate a cerebrovascular CT multimodal registered image sequence.

3. The cerebrovascular image recognition and analysis system based on deep learning according to claim 2, characterized in that: The multimodal image registration of cerebrovascular CT images corresponding to different modalities at different time points in the cerebrovascular CT multimodal image sequence includes: Performing time-series synchronous annotation on the cerebrovascular CT images corresponding to different time points in the cerebrovascular CT multimodal image sequence to obtain cerebrovascular CT multimodal time-series annotation images; The contrast and resolution of the cerebrovascular CT images under the corresponding modality are obtained through the cerebrovascular CT multimodal time-series annotated images, and the cerebrovascular CT images corresponding to different time points of the cerebrovascular CT multimodal time-series annotated images are subjected to modality hierarchy sorting analysis based on the contrast and resolution of the cerebrovascular CT images under the corresponding modality, so as to obtain the complementary hierarchy of modal characteristics between the corresponding modalities according to the corresponding contrast and resolution sorting, and obtain the corresponding modality hierarchy relationship between the cerebrovascular CT images at different time points; Based on the corresponding modality hierarchical relationship between cerebrovascular CT images at different time points, the cerebrovascular CT images corresponding to different time points of the cerebrovascular CT multimodal time series annotated images are spatially aligned to obtain the spatial transformation matrix of cerebrovascular CT images at different time points and between modalities; Based on the spatial transformation matrix of cerebrovascular CT images at different time points and between modalities, a global optimization registration transformation is performed on the local areas corresponding to the cerebrovascular CT images at different time points in the cerebrovascular CT multimodal image sequence, so as to align the cerebrovascular junctions of the cerebrovascular CT images at different time points and correct the errors corresponding to the spatial transformation matrix, so as to generate a globally optimized cerebrovascular CT registration transformation function; Based on the globally optimized cerebrovascular CT registration transformation function, the cerebrovascular CT multimodal image sequence is fully modally synchronously aligned to generate a cerebrovascular CT multimodal registered image sequence.

4. The cerebrovascular image recognition and analysis system based on deep learning according to claim 1, characterized in that: The CT pixel connectivity module includes the following functions: Perform pixel-level segmentation on each modality image in the cerebrovascular CT multimodal registration image sequence, so as to obtain the pixel spatial position and structural points of the cerebrovascular structure in each modality by automatic threshold segmentation or region growing method and form a corresponding pixel matrix to obtain a cerebrovascular CT regional pixel matrix; Based on the pixel matrix of the cerebrovascular CT region, the voxel density of the corresponding cerebrovascular region voxels in each modality image is analyzed to generate a cerebrovascular modality characteristic volume cloud map; Perform multimodal contrast enhancement on each voxel structure in the cerebrovascular modal characteristic volume cloud image to obtain a cerebrovascular multimodal structure point cloud; Perform spatial point cloud fitting optimization on the cerebrovascular multimodal structure point cloud to generate cerebrovascular CT pixel point cloud under the corresponding modality; Based on the cerebrovascular CT pixel point cloud under the corresponding modality, the vascular clustering connectivity analysis is performed on the cerebrovascular CT images under the corresponding modality in the cerebrovascular CT multimodal registration image sequence to generate the corresponding cerebrovascular CT pixel cluster connectivity map under each modality.

5. The cerebrovascular image recognition and analysis system based on deep learning according to claim 4, characterized in that: The spatial point cloud fitting optimization of the cerebrovascular multimodal structure point cloud comprises: Perform pixel density statistical analysis on the cerebrovascular CT structure point cloud under the corresponding mode in the cerebrovascular multimodal structure point cloud to obtain the pixel density value corresponding to each cerebrovascular CT point cloud under the corresponding mode; Based on the pixel density values ​​corresponding to each cerebrovascular CT point cloud under the corresponding modality, pixel outliers are removed from the cerebrovascular CT structure point cloud under the corresponding modality in the cerebrovascular multimodal structure point cloud to obtain the cerebrovascular CT pixel outlier removal point cloud under the corresponding modality; The cerebrovascular CT pixel outlier removal point cloud under the corresponding modality is subjected to local smoothing fitting processing to generate the cerebrovascular CT pixel point cloud under the corresponding modality.

6. The cerebrovascular image recognition and analysis system based on deep learning according to claim 4, characterized in that: The performing of vascular cluster connectivity analysis on the cerebrovascular CT images in the corresponding mode in the cerebrovascular CT multimodal registration image sequence based on the cerebrovascular CT pixel point cloud in the corresponding mode comprises: The cerebrovascular CT pixel point cloud under the corresponding modality is divided into spatial neighborhoods, so as to divide the local spatial small branches of the cerebrovascular vessels based on the local density distribution of the point cloud, and obtain the local spatial structure map of the cerebrovascular CT point cloud; According to the local spatial structure diagram of cerebrovascular CT point cloud, the vascular topological node connection analysis is performed to identify the branch points, intersection points and terminal points corresponding to the cerebrovascular vessels, and the corresponding vascular node connection structure is topologically connected to generate the cerebrovascular CT point cloud node connection diagram; A progressive connectivity analysis is performed on the cerebrovascular CT point cloud node connection diagram to recursively calculate the connectivity paths between cerebrovascular nodes using the shortest path, and the local vascular nodes are gradually merged into corresponding vascular clusters to generate a cerebrovascular cluster connectivity sequence. Based on the cerebrovascular cluster connectivity sequence, the cerebrovascular CT pixel point cloud under the corresponding modality is optimized for vascular cluster details, so as to eliminate the tiny connection parts between cerebrovascular clusters according to the cerebrovascular cluster connectivity structure corresponding to different modalities, and ensure the high connectivity of different modalities in the cerebrovascular network connection, so as to generate the corresponding cerebrovascular CT pixel cluster connectivity map under each modality.

7. The cerebrovascular image recognition and analysis system based on deep learning according to claim 1, characterized in that: The lesion area recognition and segmentation module includes the following functions: The cerebrovascular regional characteristics are analyzed on the cerebrovascular CT pixel cluster connectivity map corresponding to each modality, and the pixel differences and morphological characteristics of the cerebrovascular regions corresponding to each modality are obtained; The pixel differences and morphological features of the corresponding cerebrovascular regions under each modality are input into the corresponding modality network branches in the deep learning multi-branch network architecture for modality feature fusion, so as to enhance and fuse the complementary information between cerebrovascular images of different modalities to generate the structural feature map corresponding to the cerebrovascular vessels, and introduce the attention mechanism to highlight the lesion areas in the structural feature map corresponding to the cerebrovascular vessels to generate a cerebrovascular lesion area recognition model, and output the corresponding cerebrovascular lesion sensitive area map; Performing regional segmentation and contour extraction on the corresponding lesion area in the cerebrovascular lesion sensitive area map, extracting the lesion area boundary by combining edge detection and region growing, and forming a clear lesion area contour in the cerebrovascular lesion sensitive area map, generating an accurate segmentation map of the cerebrovascular lesion area; The precise segmentation map of the cerebrovascular lesion area is morphologically optimized to repair small fractures in the lesion area through corrosion and dilation operations, smooth the segmentation boundary, and enhance the detail expression of the lesion area to generate the lesion area of ​​the cerebrovascular CT image.

8. The cerebrovascular image recognition and analysis system based on deep learning according to claim 1, characterized in that: The lesion risk assessment module includes the following functions: Obtain the corresponding cerebrovascular stenosis through the lesion area of ​​cerebrovascular CT images; The thrombus area is located in the lesion area of ​​the cerebrovascular CT image to obtain the thrombus location of the lesion in the cerebrovascular CT image; Based on the location of the thrombus on the cerebrovascular CT lesion, the spatial diameter of the thrombus corresponding to the lesion area on the cerebrovascular CT image is measured to obtain the diameter of the thrombus on the cerebrovascular CT lesion; The thrombus area is estimated based on the diameter of the thrombus in the cerebral vascular CT lesions to obtain the thrombus area of ​​the lesions; Based on the degree of cerebral vascular stenosis and the area of ​​lesion thrombus, the lesion risk calculation formula is used to evaluate the lesion risk of the corresponding lesion area of ​​cerebrovascular CT images to obtain the size of the cerebrovascular CT lesion risk.

9. The deep learning-based cerebrovascular image recognition and analysis system according to claim 8, characterized in that: The step of obtaining the corresponding cerebrovascular stenosis through the lesion area of ​​the cerebrovascular CT image includes: Performing vascular lumen geometry analysis on the lesion area of ​​cerebral vascular CT images to generate cerebral vascular CT lumen contour map; The inner and outer wall distances are measured based on the CT lumen contour map of the cerebral blood vessels, so as to calculate the corresponding diameter of the cerebral blood vessel lumen point by point for each cross section, and map it to the corresponding CT image position to generate a cerebral blood vessel cross-sectional diameter map; Calculate the diameter difference between cross sections in different regions of the cerebral vascular cross-sectional diameter map to obtain the diameter difference between different cross sections in the lesion area; Based on the difference in cerebral blood vessel diameters between different cross sections in the lesion area, the stenosis degree of the lesion area in the cerebral blood vessel CT image is evaluated and calculated using the cerebral blood vessel stenosis evaluation calculation formula to obtain the stenosis degree of the cerebral blood vessel; Among them, the calculation formula for cerebral vascular stenosis assessment is as follows: ； In the formula, The degree of cerebral vascular stenosis, is the reference diameter corresponding to the normal area of ​​cerebral blood vessels, is the length of the time interval, is the time variable parameter, is the total number of corresponding cross-sections of cerebral vessels in the lesion area, In the lesion area cross section in time The corresponding cerebral blood vessel diameter is In the lesion area cross section in time The corresponding cerebral blood vessel diameter is is the adjustment coefficient for cerebral vascular diameter difference, is an exponential function, is the size of the lesion area, is the corresponding length of the cerebral blood vessels in the lesion area.

10. The deep learning-based cerebrovascular image recognition and analysis system according to claim 8, characterized in that: The specific calculation formula of the lesion risk number is: ； In the formula, It is the criticality of cerebral vascular CT lesions. The degree of cerebral vascular stenosis, is the lesion impact weight, is the area of ​​thrombus lesion.

Images