Vascular calcification analysis method and device based on non-enhanced CT image

Abstract

The invention provides a vascular calcification analysis method and device based on a non-enhanced CT image, and relates to the technical field of artificial intelligence. The method comprises the following steps: analyzing scanning parameter information of a non-contrast enhanced scanning image set; standardizing the non-contrast enhanced scanning image set based on the scanning parameter information to obtain a preprocessed image sequence; inputting the preprocessed image sequence into a multi-modal registration network, and generating a fused image based on the obtained key anatomical mark points; inputting the fused image into an attention-enhanced segmentation network, and determining blood vessel segmentation masks of the thoracic aorta, the common carotid artery and the intracranial artery; a candidate voxel set is extracted from the fused image based on the blood vessel segmentation mask, three-dimensional connected domain analysis is carried out on the candidate voxel set, and a blood vessel calcification three-dimensional area is determined; and inputting the vascular calcification three-dimensional region into the multi-task prediction model, determining a quantitative score of the vascular calcification degree, and improving the accuracy and reliability of a vascular calcification evaluation result.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and in particular to a method and device for analyzing vascular calcification based on non-enhanced CT images. Background Art

[0002] The assessment of vascular calcification levels in existing technologies primarily relies on single-modality computed tomography (CT) imaging, typically targeting specific anatomical locations, such as the coronary arteries or common carotid arteries. While this method can provide information on localized vascular calcification, it is limited to a single location and may not fully reflect the calcification status of the entire vascular system.

[0003] In recent years, the development of non-contrast enhanced scanning imaging technology has provided a new approach for the assessment of vascular calcification. By combining chest, neck, and head scans, CT images of different parts of the body, such as plain CT images of the lungs, neck, and brain, can be combined to obtain more comprehensive vascular information. However, related multimodal image fusion methods still have some shortcomings in practical applications. First, during the acquisition process of CT images of different parts of the body, there may be spatial inconsistencies between images due to factors such as patient position and respiratory movement. This inconsistency increases the difficulty of image registration and may affect the accuracy of subsequent analysis. Second, when processing multimodal images, related image registration algorithms may not fully consider the specific characteristics of each modality, resulting in insufficient registration accuracy. In addition, during the segmentation of calcified areas, high-density tissues such as bones may produce artifacts, interfering with the accurate identification of vascular calcification areas and affecting the accuracy and reliability of calcification assessment results.

[0004] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute prior art known to ordinary technicians in the field. Summary of the Invention

[0005] The purpose of the embodiments of the present disclosure is to provide a vascular calcification analysis method based on non-enhanced CT images and a vascular calcification analysis device based on non-enhanced CT images, thereby improving the accuracy and reliability of vascular calcification assessment results.

[0006] According to a first aspect of an embodiment of the present disclosure, a method for analyzing vascular calcification based on non-enhanced CT images is provided, comprising:

[0007] Acquiring a non-contrast enhanced scan image set, the non-contrast enhanced scan image set including a lung CT plain scan image, a neck CT plain scan image, and a brain CT plain scan image, and parsing scan parameter information in a metadata field of the non-contrast enhanced scan image set;

[0008] performing spatial resolution normalization processing on the non-contrast enhanced scan image set based on the scan parameter information to generate a preprocessed image sequence;

[0009] Inputting the preprocessed image sequence into a pre-trained multimodal registration network to extract key anatomical landmarks in the images of each part, and generating a spatially aligned fused image based on the key anatomical landmarks;

[0010] Inputting the fused image into a pre-trained attention-enhanced segmentation network to determine vessel segmentation masks of the thoracic aorta, common carotid artery, and intracranial artery in the fused image;

[0011] extracting a set of candidate voxels from the fused image based on the vascular segmentation mask, and performing a three-dimensional connected domain analysis on the set of candidate voxels to determine a three-dimensional region of vascular calcification in the fused image;

[0012] The three-dimensional vascular calcification region is input into a pre-trained multi-task prediction model to determine a quantitative score of the vascular calcification degree and a risk assessment level corresponding to the three-dimensional vascular calcification region.

[0013] According to a second aspect of an embodiment of the present disclosure, a device for analyzing vascular calcification based on non-enhanced CT images is provided, comprising:

[0014] an image data acquisition module, configured to acquire a non-contrast enhanced scan image set, the non-contrast enhanced scan image set including a lung CT plain scan image, a neck CT plain scan image, and a brain CT plain scan image, and parse scan parameter information in a metadata field of the non-contrast enhanced scan image set;

[0015] an image standardization module, configured to perform spatial resolution standardization processing on the non-contrast enhanced scan image set based on the scan parameter information to generate a preprocessed image sequence;

[0016] a fused image generation module, configured to input the pre-processed image sequence into a pre-trained multimodal registration network to extract key anatomical landmarks from the images of each part, and generate a spatially aligned fused image based on the key anatomical landmarks;

[0017] a vascular mask determination module, configured to input the fused image into a pre-trained attention-enhanced segmentation network to determine vascular segmentation masks of the thoracic aorta, common carotid artery, and intracranial artery in the fused image;

[0018] a calcification region identification module, configured to extract a candidate voxel set from the fused image based on the vascular segmentation mask, and perform a three-dimensional connected domain analysis on the candidate voxel set to determine a three-dimensional region of vascular calcification in the fused image;

[0019] The vascular calcification inference module is used to input the three-dimensional vascular calcification area into a pre-trained multi-task prediction model to determine the quantitative score of the vascular calcification degree and the risk assessment level corresponding to the three-dimensional vascular calcification area.

[0020] According to a third aspect of an embodiment of the present disclosure, an electronic device is provided, comprising: a processor; and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, any one of the above-mentioned methods for analyzing vascular calcification based on non-enhanced CT images is implemented.

[0021] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the method for analyzing vascular calcification based on non-enhanced CT images according to any one of the above items is implemented.

[0022] The technical solutions provided by the embodiments of the present disclosure may have the following beneficial effects:

[0023] The vascular calcification analysis method based on non-enhanced CT images in the exemplary embodiments of the present disclosure, on the one hand, can extract key anatomical landmarks from the images of each part by inputting a set of non-contrast enhanced scan images into a multimodal registration network, and generate spatially aligned fused images based on these landmarks. This can establish precise spatial correspondences between images of different modalities, ensuring the accuracy and consistency of subsequent analysis. Furthermore, the fused images are input into an attention-enhanced segmentation network to effectively identify and segment the vascular structures of the thoracic aorta, common carotid artery, and intracranial arteries, effectively improving the accuracy of vascular structure extraction and providing a reliable basis for subsequent calcification detection and assessment. Furthermore, based on the vascular segmentation mask, candidate voxel sets are extracted from the fused images and subjected to three-dimensional connected domain analysis, accurately identifying three-dimensional regions of vascular calcification and comprehensively assessing the spatial distribution characteristics of calcification, providing important reference for clinical diagnosis and treatment decisions. Furthermore, the three-dimensional vascular calcification regions are input into a pre-trained multi-task prediction model to determine the corresponding quantitative scores of vascular calcification levels and risk assessment levels, effectively improving the accuracy and reliability of vascular calcification assessment results.

[0024] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the specification, are used to explain the principles of the present disclosure. Obviously, the drawings described below are only some embodiments of the present disclosure, and those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0026] Figure 1 A schematic diagram of a system architecture of an exemplary application environment in which a method and apparatus for analyzing vascular calcification based on non-enhanced CT images according to an embodiment of the present disclosure can be applied is shown.

[0027] Figure 2 The flowchart of the method for analyzing vascular calcification based on non-enhanced CT images according to some embodiments of the present disclosure is schematically shown.

[0028] Figure 3 The figure schematically shows a flow chart of generating a fused image according to some embodiments of the present disclosure.

[0029] Figure 4 The following schematically illustrates a flow chart of screening a candidate voxel set according to some embodiments of the present disclosure.

[0030] Figure 5 The flowchart of determining the three-dimensional area of vascular calcification according to some embodiments of the present disclosure is schematically shown.

[0031] Figure 6 A schematic diagram of a vascular calcification analysis device based on non-enhanced CT images according to some embodiments of the present disclosure is schematically shown.

[0032] Figure 7 A schematic structural diagram of a computer system of an electronic device according to some embodiments of the present disclosure is schematically shown.

[0033] Figure 8 A schematic diagram of a computer-readable storage medium according to some embodiments of the present disclosure is schematically shown.

[0034] In the drawings, the same or corresponding reference numerals denote the same or corresponding parts. DETAILED DESCRIPTION

[0035] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. In the following description, when referring to the drawings, identical numerals in different figures represent identical or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with this specification. Rather, they are merely examples of apparatus and methods consistent with certain aspects of this specification, as detailed in the appended claims.

[0036] Furthermore, the drawings are schematic illustrations only and are not necessarily drawn to scale. The block diagrams shown in the drawings are merely functional entities and do not necessarily correspond to physically separate entities. In other words, these functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0037] Figure 1 A schematic diagram of a system architecture of an exemplary application environment in which a method and apparatus for analyzing vascular calcification based on non-enhanced CT images according to an embodiment of the present disclosure can be applied is shown.

[0038] like Figure 1 As shown, the system architecture 100 may include one or more terminal devices 101, 102, 103, a network 104, and a server 105. The network 104 is used to provide a medium for communication links between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables, etc. The terminal devices 101, 102, 103 may be various electronic devices with artificial intelligence (AI) computing capabilities, including but not limited to desktop computers, portable computers, smart phones, tablet computers, and medical diagnostic terminals, etc. It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is merely illustrative. Any number of terminal devices, networks and servers may be provided as needed. For example, the server 105 may be a server cluster consisting of multiple servers.

[0039] The vascular calcification analysis method based on non-enhanced CT images provided in the embodiments of the present disclosure is generally executed by the terminal devices 101, 102, and 103. Accordingly, the vascular calcification analysis device based on non-enhanced CT images is generally disposed in the terminal devices 101, 102, and 103. However, it will be readily understood by those skilled in the art that the vascular calcification analysis method based on non-enhanced CT images provided in the embodiments of the present disclosure may also be executed by the server 105. Accordingly, the vascular calcification analysis device based on non-enhanced CT images may also be disposed in the server 105. This is not particularly limited in the present exemplary embodiment.

[0040] In this example embodiment, a vascular calcification analysis method based on non-enhanced CT images is first provided. This vascular calcification analysis method based on non-enhanced CT images can be applied to a terminal device or a server. This example embodiment is not limited to this, and the subsequent description will take the execution of this method by a server as an example. Figure 2The flowchart of the method for analyzing vascular calcification based on non-enhanced CT images according to some embodiments of the present disclosure is schematically shown. Figure 2 As shown, the vascular calcification analysis method based on non-enhanced CT images may include the following steps:

[0041] Step S210: Acquire a non-contrast enhanced scan image set, the non-contrast enhanced scan image set including a lung CT plain scan image, a neck CT plain scan image, and a brain CT plain scan image, and parse scan parameter information in a metadata field of the non-contrast enhanced scan image set;

[0042] Step S220, performing spatial resolution normalization processing on the non-contrast enhanced scan image set based on the scan parameter information to generate a pre-processed image sequence;

[0043] Step S230 , inputting the pre-processed image sequence into a pre-trained multimodal registration network to extract key anatomical landmarks in the images of each part, and generating a spatially aligned fused image based on the key anatomical landmarks;

[0044] Step S240: inputting the fused image into a pre-trained attention-enhanced segmentation network to determine the vessel segmentation masks of the thoracic aorta, common carotid artery, and intracranial artery in the fused image;

[0045] Step S250 , extracting a candidate voxel set from the fused image based on the vascular segmentation mask, and performing a three-dimensional connected domain analysis on the candidate voxel set to determine a three-dimensional region of vascular calcification in the fused image;

[0046] Step S260 : Input the three-dimensional vascular calcification region into a pre-trained multi-task prediction model to determine a quantitative score of the vascular calcification degree and a risk assessment level corresponding to the three-dimensional vascular calcification region.

[0047] According to the vascular calcification analysis method based on non-enhanced CT images in this exemplary embodiment, on the one hand, by inputting a set of non-contrast enhanced scanned images into a multimodal registration network, key anatomical landmarks in the images of each part can be extracted, and spatially aligned fused images can be generated based on these landmarks. This can establish precise spatial correspondences between images of different modalities, ensuring the accuracy and consistency of subsequent analysis. Furthermore, by inputting the fused images into an attention-enhanced segmentation network, the vascular structures of the thoracic aorta, common carotid artery, and intracranial arteries can be effectively identified and segmented, effectively improving the accuracy of vascular structure extraction and providing a reliable basis for subsequent calcification detection and assessment. On the other hand, based on the vascular segmentation mask, candidate voxel sets are extracted from the fused images and subjected to three-dimensional connected domain analysis. This can accurately identify three-dimensional regions of vascular calcification and comprehensively assess the spatial distribution characteristics of calcification, providing an important reference for clinical diagnosis and treatment decisions. Furthermore, by inputting the three-dimensional vascular calcification regions into a pre-trained multi-task prediction model, the corresponding quantitative scores of vascular calcification levels and risk assessment levels can be determined, effectively improving the accuracy and reliability of vascular calcification assessment results.

[0048] The following further describes the vascular calcification analysis method based on non-enhanced CT images in this exemplary embodiment.

[0049] In step S210, a non-contrast enhanced scan image set is acquired, wherein the non-contrast enhanced scan image set includes lung CT plain scan images, neck CT plain scan images, and brain CT plain scan images, and scan parameter information in the metadata field of the non-contrast enhanced scan image set is parsed.

[0050] In an example embodiment of the present disclosure, a non-contrast enhanced scanning image set may include lung CT plain scan images, neck CT plain scan images, and brain CT plain scan images, wherein the lung CT images can be used to extract coronary artery vessels and thoracic aorta images, the neck CT images can cover the vessels of the common carotid artery and the vascular direction of most internal carotid arteries, and the brain CT images can provide anatomical information of part of the internal carotid artery and the intracranial arteries.

[0051] Image data from non-contrast-enhanced scanned images can be stored in the Digital Imaging and Communications in Medicine (DICOM) format. The metadata fields in the DICOM format can contain parameters related to the scanning device and scanning process, namely, scanning parameter information. For example, scanning parameter information can include slice thickness, scanning voltage, reconstruction filter, voxel size, etc. The process of parsing metadata can be implemented using medical image processing software (such as 3D Slicer and ITK-SNAP) or programming tools (such as the Python pydicom library) to extract key scanning parameters and store them in a standardized format.

[0052] In step S220, spatial resolution normalization processing is performed on the non-contrast enhanced scan image set based on the scan parameter information to generate a pre-processed image sequence.

[0053] In an exemplary embodiment of the present disclosure, spatial resolution normalization refers to unifying CT images of different resolutions and belonging to different parts of the body through interpolation operations so that the voxel sizes of all image data are consistent, thereby reducing the subsequent analysis deviation caused by resolution differences. For example, the original resolution of each image can be calculated, based on a preset target voxel size (such as 1mm 3 ) calculates the resampling ratio and uses an interpolation method for reconstruction. For example, the interpolation method can be selected from linear interpolation, cubic spline interpolation, or nearest neighbor interpolation. Linear interpolation can be used for most CT image processing and has high computational efficiency; cubic spline interpolation can provide smooth resampling results and reduce interpolation errors; nearest neighbor interpolation can be used for classification tasks to maintain the integrity of discrete values of the image.

[0054] Optionally, the interpolation strategy can be dynamically adjusted according to the complexity of the tissue structure in the CT image. For example, high-order interpolation can be used for high-contrast areas such as blood vessels, while low-order interpolation can be used for uniform areas to optimize computational efficiency and image quality. By standardizing spatial resolution, the comparability between different images can be improved, and the accuracy of subsequent image registration and feature extraction can be optimized.

[0055] In step S230, the pre-processed image sequence is input into a pre-trained multimodal registration network to extract key anatomical landmarks in the images of each part, and a spatially aligned fused image is generated based on the key anatomical landmarks.

[0056] In an example embodiment of the present disclosure, a multimodal registration network refers to a deep learning model used for image space transformation. For example, the multimodal registration network can adopt a convolutional neural network (CNN). Of course, a deformable network (Deformable Network) structure can also be adopted. This example embodiment does not specifically limit the network type of the multimodal registration network.

[0057] Key anatomical landmarks refer to characteristic points with anatomical stability in CT images of various parts of the body. For example, a key anatomical landmark may be the tracheal carina in a lung CT image, or the anterior commissure apex and the posterior commissure bottom in a brain CT image. Of course, a key anatomical landmark may also be the common carotid artery bifurcation in a neck CT image. Key anatomical landmarks may also be a combination of the above different anatomical landmarks, and this embodiment is not limited thereto.

[0058] The method for detecting key anatomical landmarks can be based on a regression network or a key point detection model based on an attention mechanism. For example, the regression network can locate the key points by predicting their three-dimensional coordinates, while the attention mechanism model can enhance the network's attention to key anatomical areas and improve detection accuracy.

[0059] After the key points are extracted, the joint deformation matrix between the images can be further calculated. The joint deformation matrix can adopt rigid transformation (Rigid Transformation), affine transformation (Affine Transformation) or free-form deformation (FFD). Among them, rigid transformation only considers rotation and translation, which is suitable for the case where the anatomical structure is relatively stable; affine transformation can further adapt to scale changes, and free deformation can perform nonlinear transformation based on the B-spline interpolation model to optimize the alignment accuracy. As for which transformation method to use, it can be customized according to actual usage, and this embodiment is not limited to this.

[0060] The CT images of each part are spatially aligned by combining the deformation matrix and resampled using voxel interpolation to generate a fused image. In some optional embodiments, a registration method based on image features can also be used. For example, image similarity can be calculated using mutual information (MI) or normalized cross correlation (NCC) to optimize the registration results. The generation of fused images can improve the alignment of images of different anatomical parts, enhance the overall visualization of vascular structures, and optimize the accuracy of subsequent vascular segmentation and calcification detection.

[0061] In step S240, the fused image is input into a pre-trained attention-enhanced segmentation network to determine the vessel segmentation masks of the thoracic aorta, common carotid artery, and intracranial artery in the fused image.

[0062] In an exemplary embodiment of the present disclosure, vessel segmentation is based on pixel-by-pixel classification of images using a deep learning model to distinguish vascular tissue from background tissue and generate a vessel segmentation mask. The attention-enhanced segmentation network is an improved convolutional neural network that incorporates an attention mechanism to improve the accuracy of target region recognition.

[0063] Specifically, the attention-enhanced segmentation network may include a feature extraction layer, an attention mechanism layer, and a region segmentation layer, wherein the feature extraction layer can be used to extract multi-scale features of the image, the attention mechanism layer can be used to enhance the feature expression of the vascular region, and the region segmentation layer can achieve the final pixel-level classification. Among them, the attention mechanism of the attention mechanism layer can adopt channel attention, spatial attention, or self-attention methods, wherein channel attention can adjust the importance of different feature channels; spatial attention is used to suppress background noise and improve the clarity of vascular boundaries; the self-attention mechanism can be used for long-range dependency modeling to improve the global perception ability of the segmentation network, which can be customized according to actual usage, and this embodiment is not limited to this. In some optional embodiments, multi-scale pyramid features (Feature Pyramid Networks, FPN) can also be combined in the segmentation process to enhance the recognition ability of vascular structures of different scales. Finally, after processing by the attention-enhanced segmentation network, vascular segmentation masks of the thoracic aorta, common carotid artery, and intracranial artery are generated, and the vascular segmentation masks can be used for subsequent calcification area extraction and quantitative analysis.

[0064] In step S250 , a candidate voxel set is extracted from the fused image based on the vascular segmentation mask, and a three-dimensional connected domain analysis is performed on the candidate voxel set to determine a three-dimensional region of vascular calcification in the fused image.

[0065] In an exemplary embodiment of the present disclosure, when screening voxels for vascular calcification, the CT value distribution characteristics of the image voxels can be used. Generally, vascular calcification areas have higher CT values in non-contrast CT images. For example, the CT value of vascular calcification areas is generally greater than or equal to 130 HU (Hounsfield Units), while the CT values of non-calcified vascular tissue and surrounding soft tissue are relatively low. The basic principle of voxel screening can be to set a dynamic CT value threshold to distinguish possible calcified voxels from background tissue voxels.

[0066] First, all voxels can be extracted within the region of interest (ROI) covered by the vascular segmentation mask, and their CT value histogram distribution can be calculated to analyze the CT value range of calcified tissue. Then, based on the image noise level and individual differences, an adaptive threshold method or statistical method (such as Otsu threshold segmentation) can be applied to determine the final CT value threshold, and voxels that meet the calcification characteristics can be screened to form a candidate voxel set.

[0067] Three-dimensional connected domain analysis refers to an analysis method that uses a connectivity algorithm to cluster adjacent candidate voxels to identify independent vascular calcification areas. For example, connected domain analysis can be based on three-dimensional connectivity of six-neighborhood, eighteen-neighborhood, or twenty-six-neighborhood. Six-neighborhood connectivity only considers direct axial connections between voxels and is suitable for more regular calcified areas. Eighteen-neighborhood connectivity adds diagonal connectivity to the six-neighborhood connectivity, improving the detection capability of connected domains. Twenty-six-neighborhood connectivity covers all adjacent voxels and can more comprehensively identify densely distributed calcified areas in space, but may increase computational complexity. In this embodiment, twenty-six-neighborhood connectivity can be used to implement three-dimensional connected domain analysis of a set of candidate voxels.

[0068] In some optional implementations, the volume of each connected domain can be calculated, and initial connected regions with volumes less than a preset threshold can be filtered out to remove small areas of pseudo-calcification. Of course, morphological optimization can also be performed on the connected regions. For example, morphological dilation and erosion can be performed on the three-dimensional calcified regions to smooth the boundaries and remove isolated noise points. Ultimately, after connected domain analysis and morphological optimization, the three-dimensional region of vascular calcification in the fused image is obtained, which can be used for subsequent quantitative scoring and risk assessment.

[0069] In step S260, the three-dimensional vascular calcification region is input into a pre-trained multi-task prediction model to determine a quantitative score of the vascular calcification degree and a risk assessment level corresponding to the three-dimensional vascular calcification region.

[0070] In an exemplary embodiment of the present disclosure, a multi-task prediction model refers to a pre-trained deep learning model that outputs multiple prediction results. For example, the multi-task prediction model can be a deep learning model that predicts the volume, shape characteristics, and clinical risk grading of vascular calcification areas. The multi-task prediction model can employ a multi-branch network structure, where one branch can be used for regression prediction, for example, to quantify the volume and morphological parameters of vascular calcification areas, and another branch can be used for classification prediction, for example, to assess risk levels.

[0071] Vascular calcification quantitative scoring involves numerically calculating the volume, density, and spatial distribution of vascular calcification areas based on imaging data to quantify the degree of calcification. The degree of vascular calcification directly impacts cardiovascular disease risk assessment, and therefore quantitative scoring provides important diagnostic evidence for clinicians. The basic principle of quantitative scoring is to calculate the three-dimensional volume of vascular calcification areas in CT images and, combined with the CT value distribution, employ a standardized scoring method. For example, calcification scoring methods may include, but are not limited to, the Agatston score, volumetric score, and mass score, though this embodiment is not limited thereto.

[0072] The risk assessment stratification is a quantitative scoring method based on the degree of vascular calcification, combined with epidemiological data and clinical guidelines, to categorize a patient's cardiovascular disease risk. A statistical correlation can be established between the degree of vascular calcification and the probability of developing cardiovascular disease, and risk levels can be divided according to different scoring ranges. For example, in the risk assessment of coronary heart disease, the risk assessment levels can be divided into low risk, medium risk, high risk and extremely high risk. Among them, low risk can be the level corresponding to a coronary vascular calcification score less than 100, indicating a mild degree of calcification and a low risk of cardiovascular disease in patients. Usually, no special intervention is required, and only regular follow-up is required; medium risk can be the level corresponding to a coronary vascular calcification score greater than or equal to 100 and less than 300, indicating a moderate degree of vascular calcification. Patients may have a higher risk of cardiovascular disease and further cardiovascular health assessment is recommended. Lifestyle adjustments or drug interventions may be required; high risk can be the level corresponding to a coronary vascular calcification score greater than or equal to 300, indicating severe vascular calcification. Patients have a significantly increased risk of cardiovascular disease (such as myocardial infarction and stroke) and require detailed clinical examinations and consideration of drugs or other interventions; extremely high risk can be the level corresponding to a coronary vascular calcification score greater than or equal to 1000, indicating an extremely high risk of cardiovascular events. Patients need close monitoring and may need intensive lipid-lowering therapy or interventional therapy. Of course, the above risk assessment levels are merely illustrative examples, and this exemplary embodiment does not impose any special limitations thereto.

[0073] The contents of steps S210 to S260 are described in detail below.

[0074] In an exemplary embodiment of the present disclosure, step S220 can be implemented by performing spatial resolution normalization processing on the non-contrast enhanced scan image set based on the scan parameter information to generate a pre-processed image sequence through the following steps:

[0075] The spatial resolution between the CT images of each part in the non-contrast enhanced scan image set can be determined by the scanning parameter information; the CT images of each part in the non-contrast enhanced scan image set are resampled based on the spatial resolution to generate a preprocessed image sequence.

[0076] CT images of different parts of the body often have significant differences in spatial resolution due to differences in acquisition equipment and imaging protocols, leading to inconsistencies in voxel size, spatial scale, and image quality. The core goal of spatial resolution standardization is to unify images through mathematical transformations to ensure accurate cross-modal comparison and fusion during subsequent image registration, segmentation, and analysis. The basic principle is to use image resampling technology to adjust CT images of various parts to the same spatial resolution, thereby eliminating scale differences caused by different imaging conditions.

[0077] The spatial resolution between CT images of various parts of a non-contrast-enhanced scan set can be determined using scanning parameter information. Scanning parameter information refers to the equipment parameters and imaging settings related to image acquisition, as recorded in the CT image metadata. These include, but are not limited to, slice thickness, voxel size, scanning voltage, reconstruction filter, and field of view (FOV). Slice thickness and voxel size are key factors in determining spatial resolution. Slice thickness represents the thickness of each slice during a CT scan and is typically measured in millimeters. Smaller slice thicknesses provide higher longitudinal resolution, while larger slice thicknesses reduce scanning noise. Voxel size refers to the physical size of a single voxel in a CT image and is typically determined by both the transverse pixel pitch and slice thickness. The transverse pixel pitch is typically measured in millimeters and depends on the spatial resolution of the scanner and the size of the acquisition matrix. For example, a 512×512 pixel image matrix with a 250 mm FOV has a pixel pitch of approximately 0.49 mm. Different CT scanners may have different sampling intervals, reconstruction algorithms, and FOVs, resulting in differences in the spatial resolution of CT images of the same anatomical region across different modalities. When performing spatial resolution normalization, the slice thickness and voxel size information in the DICOM image metadata fields can be parsed, and the spatial resolution differences between different modality images can be calculated to determine the target scale for resampling.

[0078] The CT images of each part in the non-contrast enhanced scan image set can be resampled based on the spatial resolution to generate a preprocessed image sequence. Resampling refers to the use of an interpolation algorithm to scale the image so that it meets the target spatial resolution requirements. Specifically, resampling can recalculate the original image data through a three-dimensional interpolation method to match the preset standard voxel size. Common three-dimensional interpolation methods include nearest neighbor interpolation (Nearest Neighbor Interpolation), linear interpolation (Linear Interpolation) and cubic spline interpolation (Cubic Spline Interpolation). Of course, a deep learning-driven super-resolution reconstruction (Super-Resolution Reconstruction, SRR) method can also be used to perform high-quality reconstruction of low-resolution images through a convolutional neural network or a generative adversarial network (Generative Adversarial Network, GAN) to improve the image detail retention capability, and this example embodiment does not specifically limit this.

[0079] During implementation, a uniform target spatial resolution can be pre-selected. For example, the standardized voxel size of CT images can be set to 1 mm. 3, in order to ensure high spatial resolution and adapt to different CT scanning protocols; then, the original voxel size of each modality CT image can be analyzed and the scaling factor can be calculated, that is, the ratio of the target voxel size to the current voxel size. For example, if the voxel size of the lung CT image is 0.8×0.8×1.5mm 3 , and the target voxel size is 1×1×1mm 3 , it can be scaled 0.8 times in the horizontal direction and 1.5 times in the vertical direction; then, the CT image can be resampled using a three-dimensional interpolation method and the data can be reconstructed to ensure the consistency of the image structure.

[0080] In an optional embodiment, a resampling method based on regularization can also be used to reduce the distortion that may occur during the spatial transformation process. For example, a total variation (TV) regularization method can be used to impose constraints on areas with large gradient changes during the interpolation calculation process to preserve the edge structure of the image. In addition, an image reconstruction method based on deep learning can be combined. For example, a variational autoencoder (VAE) can be used to reconstruct images of different modalities to further improve the spatial consistency of the image.

[0081] By normalizing the spatial resolution of CT images across modalities, we can effectively eliminate resolution differences between them and improve the comparability of images during subsequent analysis. Furthermore, the standardized image sequences provide more stable input data for registration, segmentation, and calcification analysis, improving the accuracy and robustness of the overall algorithm.

[0082] In an exemplary embodiment of the present disclosure, the following steps may be performed to implement step S230 of inputting the pre-processed image sequence into a pre-trained multimodal registration network to extract key anatomical landmarks from images of various parts, and generating a spatially aligned fused image based on the key anatomical landmarks. Specifically:

[0083] The preprocessed image sequence can be input into the input layer of the multimodal registration network to extract the CT image features corresponding to the CT images of each part in the preprocessed image sequence; through the feature extraction layer of the multimodal registration network, the key anatomical landmarks in the CT images of each part are identified and extracted from the CT image features; based on the key anatomical landmarks, the joint deformation matrix between the CT images of each part is determined, and the preprocessed image sequence is spatially aligned and fused through the joint deformation matrix to generate a fused image.

[0084] Among them, in the input layer, the CT image data of each part can be pre-processed by standardization to ensure that the grayscale value range of different modal images is consistent. For example, standardization to [0, 1] or Z-score normalization can be adopted to reduce the impact of brightness differences on alignment. The CT images can also be edge enhanced or histogram matched to improve the feature alignment between different modal images.

[0085] Key anatomical landmarks refer to feature points in an image that have a stable anatomical structure and can be used for registration. The extraction of key anatomical landmarks can use a regression network based on deep learning or a feature extraction layer based on an attention mechanism. The regression network can use supervised learning to train the network to predict the three-dimensional coordinates of key anatomical landmarks. For example, a ResNet regression model can be used to perform feature mapping on the input CT image, and the coordinates of the key points can be output through a fully connected layer; and the method based on the attention mechanism can enhance the anatomical structure information in the CT image through spatial attention or channel attention to improve the robustness of key point detection. In some optional embodiments, the Hessian matrix can be used to detect vascular bifurcation points or the SIFT (Scale-Invariant Feature Transform) algorithm can be used to extract multi-scale feature points to improve the positioning accuracy of anatomical landmarks.

[0086] In the case of rigid registration, the joint deformation matrix can be calculated by the least squares method or the iterative closest point algorithm (ICP). In the case of affine registration, the joint deformation matrix can be solved by homography transformation. In the case of non-rigid registration, a deformation model based on B-spline or thin-plate spline (TPS) can be used for optimization calculation. Optionally, when calculating the joint deformation matrix, the coordinates of key anatomical landmarks can be matched. For example, the position of the tracheal carina in three-dimensional space is calculated in the lung CT image, and the anterior commissure vertex and posterior commissure bottom are matched in the brain CT image, and the common carotid artery bifurcation point is matched in the neck CT image. Then, the least squares method or robust estimation method (such as RANSAC algorithm) can be used to calculate the optimal registration parameters to optimize the stability of the joint deformation matrix.

[0087] After the deformation matrix is calculated, the pre-processed image sequence can be spatially transformed to ensure that all images are aligned in the same coordinate system. Spatial transformation methods can include rigid transformation, affine transformation, and free deformation transformation. In the case of rigid transformation and affine transformation, the spatial transformation of CT images can be performed through standard matrix operations. For example, homogeneous coordinates can be used to represent the image coordinates and transform them through matrix multiplication. In the case of free deformation, a B-spline free-form deformation model (FFD) can be used to deform the image by controlling the grid to adapt to non-rigid changes in local anatomical structures.

[0088] Fusion images can be fused at the voxel level based on the registered images to optimize image quality and enhance the visualization of vascular structures in the images. Fusion methods can include weighted average fusion, maximum fusion, and deep learning-based fusion models. Weighted average fusion refers to the weighted summation of corresponding voxels of different modal images to retain high-confidence image information. The weights can be adaptively adjusted based on image quality, registration error, or tissue contrast. Maximum fusion can select the maximum intensity value in different images to enhance the contrast of high-brightness areas. Deep learning-based fusion models can use generative adversarial networks or autoencoders to fuse features of multimodal images to generate high-quality fused images.

[0089] Through this method, fused images can accurately align images of different anatomical locations within the same coordinate system, improving the integrity and visualization of vascular structures and optimizing the accuracy of subsequent vascular segmentation and calcification detection. Compared to traditional image registration methods, this embodiment utilizes a multimodal registration network based on deep learning to improve the stability and automation of image registration, reduce the need for manual adjustments, and is applicable to image data from different CT scanning devices and imaging protocols.

[0090] In an exemplary embodiment of the present disclosure, key anatomical landmarks may include the tracheal carina in a lung CT plain scan image, the anterior commissure apex and the posterior commissure bottom in a cranial CT plain scan image, and the common carotid artery bifurcation in a neck CT plain scan image. The tracheal carina refers to the anatomical structure located at the junction of the trachea and the main bronchus. Because of its high contrast and clear morphology in chest CT images, it is suitable as a reference point for registration; the anterior commissure apex and the posterior commissure bottom refer to two key anatomical points in a cranial CT image, corresponding to the highest point and the lowest point of the anterior commissure and the posterior commissure, respectively. Their positions are relatively stable and easy to identify in cranial images; the common carotid artery bifurcation refers to the bifurcation position of the common carotid artery in a neck CT image. Since the bifurcation of the common carotid artery usually has obvious vascular structural characteristics, it can be used for spatial alignment of images.

[0091] Can be achieved through Figure 3 The steps in the paper are used to determine the joint deformation matrix between the CT images of each part based on the key anatomical landmarks, and to perform spatial alignment and fusion of the pre-processed image sequence through the joint deformation matrix to generate a fused image. Figure 3 Specifically, it may include:

[0092] Step S310, establishing a first three-dimensional Cartesian coordinate system with the tracheal carina as the center, constructing a second three-dimensional Cartesian coordinate system with the anterior commissure line between the anterior commissure vertex and the posterior commissure base as the coordinate axis, and establishing a third three-dimensional Cartesian coordinate system with the common carotid artery bifurcation as the center;

[0093] Step S320: determining an axial tensile deformation component based on the first three-dimensional Cartesian coordinate system and the third three-dimensional Cartesian coordinate system;

[0094] Step S330, performing singular value decomposition on the spatial rotation matrix between the second three-dimensional Cartesian coordinate system and the third three-dimensional Cartesian coordinate system to extract a pitch angle deviation compensation parameter;

[0095] Step S340: fusing the axial tensile deformation component and the pitch angle deviation compensation parameter into a preset B-spline free deformation model to generate a joint deformation matrix;

[0096] Step S350 , performing voxel space coordinate transformation on the CT images of each part in the pre-processed image sequence using the joint deformation matrix to generate a fused image.

[0097] Among them, the Cartesian coordinate system refers to a coordinate system used for spatial geometric transformation. Its basic principle is to establish a standardized reference frame based on points or objects in three-dimensional space to facilitate the registration of image data. The tracheal carina is located at the tracheal bifurcation and is usually clearly visible in lung CT images. It is suitable as the center point of the chest image. Therefore, the tracheal carina can be used as a reference to establish a first three-dimensional Cartesian coordinate system, whose X-axis, Y-axis, and Z-axis correspond to the left-right, front-back, and top-bottom directions of the image respectively; the anterior commissure vertex and the posterior commissure base are located at the midline position of the cranial image. The second three-dimensional Cartesian coordinate system can be constructed by connecting the anterior and posterior commissures to ensure the spatial alignment of the brain image; the common carotid artery bifurcation is located in the neck CT image. Its anatomical position is relatively stable. Therefore, the third three-dimensional Cartesian coordinate system is established with this point as a reference to perform spatial standardization of the neck image.

[0098] Axial stretching deformation refers to the process of image registration, in which axial scale transformation is performed to address possible longitudinal scale differences between different images to ensure spatial consistency of the images in the longitudinal direction. Since different CT scans may use different layer thicknesses, resulting in inconsistent longitudinal resolution, the images can be scaled using stretching transformation. Specifically, the longitudinal scale difference between the first 3D Cartesian coordinate system and the third 3D Cartesian coordinate system can be calculated and adjusted using a scale transformation function. For example, bilinear interpolation or cubic spline interpolation can be used to resample the images to optimize the accuracy of the longitudinal stretching transformation.

[0099] The spatial rotation matrix refers to a mathematical tool that describes the rotation of a three-dimensional object. Its basic principle is to rotate and align the image through matrix transformation. Since different CT scans may have slight changes in the patient's head or neck posture, resulting in deviations in the pitch angle of the image in three-dimensional space, the spatial rotation matrix can be calculated and compensated. Singular value decomposition is a mathematical method for matrix decomposition, which can be used to calculate the optimal rotation parameters in image registration to minimize the error between different images. Specifically, the spatial rotation matrix can be calculated based on the three-dimensional coordinates of the anterior commissure vertex, the posterior commissure bottom point and the common carotid artery bifurcation point, and then the singular value decomposition algorithm can be used to decompose the matrix and extract the pitch angle deviation parameters to perform rotation adjustment of the image.

[0100] The B-spline free-form deformation model is a deformation method used for non-rigid registration. It can achieve flexible image registration by defining multiple control points and continuously deforming the image using a B-spline interpolation function. The advantage of the B-spline free-form deformation model is that it can fine-tune local areas of the image to accommodate anatomical variations between individuals. For example, a set of regular grids can be defined in the image space, and control points can be set at the grid intersections. The positions of the control points can then be adjusted based on the calculated deformation components, and the deformation variables of all voxels can be calculated using B-spline interpolation, ultimately generating a continuous deformation field.

[0101] A joint deformation matrix can be used to perform voxel-space coordinate transformation on the CT images of each part in the preprocessed image sequence to generate a fused image. Voxel-space coordinate transformation refers to mapping the registered image data to a unified coordinate system to ensure spatial consistency between the images. For example, methods such as nearest neighbor interpolation, bilinear interpolation, and cubic spline interpolation can be used in conjunction with the joint deformation matrix to achieve voxel-space coordinate transformation of the CT images of each part in the preprocessed image sequence. This example embodiment does not specifically limit this.

[0102] By determining the joint deformation matrix between the CT images of each part based on key anatomical landmarks, and spatially aligning and fusing the preprocessed image sequence through the joint deformation matrix, a fused image is generated, which improves the degree of automation of the registration of the preprocessed image sequence, reduces the need for manual adjustment, and improves the efficiency of generating fused images, thereby improving the accuracy and stability of image analysis.

[0103] In an exemplary embodiment of the present disclosure, the following steps may be performed to implement step S240 of inputting the fused image into a pre-trained attention-enhanced segmentation network to determine the vessel segmentation masks of the thoracic aorta, intracranial artery, and common carotid artery in the fused image. Specifically,

[0104] The fused image can be input into the input layer of the attention-enhanced segmentation network to extract the fused image features corresponding to the fused image; the bone artifact area in the fused image features can be suppressed through the attention mechanism layer of the attention-enhanced segmentation network to obtain enhanced image features; the enhanced image features can be identified using the regional segmentation layer of the attention-enhanced segmentation network to extract the vascular segmentation masks corresponding to the thoracic aorta, intracranial artery and common carotid artery in the fused image.

[0105] Among them, the main function of the input layer is to pre-process the fused image and extract preliminary low-level features so that the subsequent deep network can perform high-level semantic analysis. In the input layer, the input fused image can be normalized to ensure that the grayscale value distribution between each region in the fused image is consistent. Commonly used normalization methods can be minimum-maximum normalization and Z-score normalization, which are not specifically limited in this example embodiment. In addition, histogram equalization or adaptive contrast enhancement (Adaptive Histogram Equalization, AHE) methods can also be used in the input layer to enhance the contrast of blood vessel edges to optimize the separability of vascular tissue and background tissue.

[0106] The attention mechanism layer in the attention enhanced segmentation network is used to assign different weights to different areas of the fused image to enhance the feature expression of the vascular area and suppress the interference of the background area. Bone artifacts are common interference factors in CT images, especially in head and neck and chest CT images, the skull, clavicle and vertebrae may produce high-density artifacts, affecting the boundary identification of blood vessels. In optional implementations, the attention mechanism can adopt channel attention, spatial attention or self-attention methods. For example, the attention mechanism in this embodiment can adopt spatial attention, which can be used to calculate the local response of the fused image and enhance the spatial feature expression of the vascular area. Specifically, it can be achieved through attention mapping based on convolutional neural networks or self-attention mechanism based on Transformer. The self-attention mechanism based on Transformer can improve the clarity of the vascular boundary by calculating the global dependency between pixels. In addition, it can also be combined with multi-scale feature extraction methods to extract vascular features of different scales through pyramid pooling (PPM) or void convolution to enhance the adaptability of the network to complex vascular structures.

[0107] The core goal of the regional segmentation layer is to classify the vascular regions in the image pixel by pixel and generate an accurate vascular segmentation mask. For example, the regional segmentation layer can adopt a fully convolutional network structure, and restore the spatial resolution of the image through multi-scale feature extraction and layer-by-layer upsampling to generate pixel-level segmentation results. Specifically, a U-Net structure can be used. The U-Net structure network can extract features from the fused image layer by layer through a symmetric encoding-decoding structure, and retain high-resolution features in combination with jump connections to optimize the segmentation accuracy of the vascular boundary. Of course, the regional segmentation layer can also adopt the DeepLabV3+ network, which combines the Atrous Spatial Pyramid Pooling (ASPP) technology to enhance the network's adaptability to vascular structures of different scales. It is understandable that in actual use, the network architecture of the attention-enhanced segmentation network can also be adjusted according to different CT image segmentation requirements. For example, for complex vascular networks, a graph neural network can also be used for topological structure modeling to optimize the connectivity of vascular branches. This example embodiment does not specifically limit this.

[0108] By enhancing the segmentation network with attention, the accuracy of vessel segmentation is improved, the interference of CT image noise and bone artifacts is reduced, and the integrity of vessel boundaries is optimized. In addition, the use of the attention mechanism can adapt to the needs of vessel segmentation in different anatomical locations and maintain high segmentation stability under different CT scan parameters, thereby providing more reliable input data for vascular calcification detection and improving the accuracy and robustness of the overall analysis.

[0109] In an exemplary embodiment of the present disclosure, Figure 4 The steps in step S250 are used to extract a candidate voxel set from the fused image based on the blood vessel segmentation mask, referring to Figure 4 Specifically, it may include:

[0110] Step S410, determining an image region of interest in the fused image based on the blood vessel segmentation mask;

[0111] Step S420 : In each of the image regions of interest, voxels that meet the characteristics of vascular calcification are screened according to the determined dynamic CT value threshold to form a candidate voxel set.

[0112] The image region of interest refers to a specific spatial range determined by the vascular segmentation results. This range can include vascular structures such as the thoracic aorta, common carotid artery, and intracranial artery in the fused image, and is used for subsequent vascular calcification extraction and analysis. The determination of the image region of interest is based on the vascular segmentation mask, that is, the vascular segmentation mask generated by the attention-enhanced segmentation network is used to screen the affected area of the fused image, retaining only the vascular area and removing other background tissues, such as soft tissue, bone, and air areas. The determination of the image region of interest can be achieved through mathematical morphological operations or methods based on connected domain analysis to extract the valid area within the vascular segmentation mask and map it to the original fused image to construct an analysis space for vascular calcification detection.

[0113] In the specific implementation process, the vessel segmentation mask can be first morphologically optimized. For example, dilation can be used to fill small missing areas to improve the integrity of the vessel region, and erosion can be used to remove isolated small false positive areas. Subsequently, the image region of interest can be cropped in the fused image to reduce the impact of non-vascular tissue on subsequent calculations.

[0114] Dynamic CT thresholding refers to an automated screening method that dynamically adjusts the threshold for vascular calcification identification based on the statistical characteristics of different fused image data, adaptive segmentation methods, and machine learning algorithms to adapt to varying scanning conditions, individual anatomical differences, and image quality variations. Compared to fixed threshold methods (such as the traditional 130 HU), dynamic CT thresholding intelligently adjusts the CT threshold based on information such as image background, local contrast, and statistical distribution. This improves the ability to identify low-density calcifications and high-density artifacts, while reducing false positives and missed detections.

[0115] The determination of dynamic CT value thresholds can be based on statistical analysis, adaptive algorithms, or deep learning models to model the distribution of CT values within the image region of interest and automatically select the optimal segmentation threshold. For example, in a specific implementation, all CT values within the image region of interest can be initially screened to remove abnormally high-density areas (such as bone artifacts) and low-density background areas (such as air or soft tissue). Then, a Gaussian mixture model (GMM) or K-means clustering method can be used to cluster the CT values within the image region of interest and automatically divide the CT value range belonging to the calcified area. For example, the Gaussian mixture model can fit multiple Gaussian distributions based on maximum likelihood estimation (MLE) to identify the CT value distribution of different tissue types and automatically select the optimal threshold range for calcified tissue. The K-means clustering method can use an iterative optimization algorithm to divide the voxels in the image region of interest into multiple categories and automatically select the category closest to the known calcification CT value range to set the dynamic threshold. Of course, the above is only an illustrative example, and this embodiment does not impose any special limitation on this.

[0116] Dynamic CT thresholds can be adaptively adjusted to suit the characteristics of different imaging data, thereby improving the accuracy of identifying vascular calcification areas and reducing segmentation errors caused by image quality, scanning parameters, or individual differences. Furthermore, adaptive adjustment of the dynamic CT threshold optimizes the identification of calcified areas, increasing the sensitivity of detecting low-density calcifications and reducing the impact of false positives due to bone artifacts or high-density vessel walls, ensuring the accuracy and stability of subsequent 3D connected domain analysis and quantitative assessment of vascular calcification.

[0117] In an exemplary embodiment of the present disclosure, Figure 5 The steps in step S250 are used to perform a three-dimensional connected domain analysis on the candidate voxel set to generate a three-dimensional vascular calcification region. Figure 5 Specifically, it may include:

[0118] Step S510 , performing connectivity marking on each voxel in the candidate voxel set to determine an initial connected voxel group;

[0119] Step S520, calculating the volume of each of the initial connected voxel groups, and filtering out the initial connected voxel groups whose volumes are smaller than a preset volume threshold, to obtain a target connected voxel group;

[0120] Step S530 : performing morphological optimization on the target connected voxel group to obtain a three-dimensional region of vascular calcification.

[0121] Among them, connectivity labeling refers to traversing all voxels in the candidate voxel set and labeling voxels belonging to the same connected area with the same label according to the preset neighborhood connection rules. In the specific implementation process, a three-dimensional voxel index data structure can be constructed to store the coordinate information of each voxel and its adjacency relationship. Then, a breadth-first search (BFS) or depth-first search (DFS) algorithm can be used to traverse all candidate voxels and, based on the neighborhood connection standard, voxels that are spatially adjacent and whose CT values meet the threshold requirements are classified as the same initial connected voxel group. In addition, a connectivity labeling method based on parallel computing can also be used. For example, a parallel union-find algorithm or a GPU-accelerated region growing method can be used to improve the processing efficiency of large-scale voxel data. This example embodiment does not specifically limit the method of connectivity labeling.

[0122] Volume calculation can be based on three-dimensional voxel counting, counting the total number of voxels in each connected voxel group and multiplying it by the physical volume of a single voxel to obtain the total volume of the connected area. Since vascular calcification usually shows a certain degree of spatial connectivity, and isolated noise points or misdetected voxels are often small in volume, it is possible to set a volume threshold (such as 0.5 mm 3 or 1mm 3 ) to filter out small pseudo-calcification areas to improve analysis accuracy. In a specific implementation, voxels in each initial connected voxel cluster can be counted and their volume calculated. Then, based on a preset volume threshold, voxel clusters with volumes smaller than the threshold can be eliminated, retaining only the target connected voxel clusters that meet the characteristics of vascular calcification.

[0123] Morphological optimization refers to an image processing method used to improve the quality of image segmentation. It can adjust the morphology of a voxel set through mathematical morphological operations to remove isolated points, fill small holes, and optimize the boundaries of calcified areas. Specifically, a morphological dilation operation can be performed on the target connected voxel group to enhance the connectivity of the voxel group and fill small missing areas caused by CT scan noise or image artifacts. A morphological erosion operation can then be applied to remove isolated noise points in the image and optimize the boundary morphology of the calcified area. In addition, morphological opening and closing operations can be combined to further smooth the morphology of the vascular calcification area and improve spatial consistency.

[0124] Through three-dimensional connected domain analysis, a complete three-dimensional area of vascular calcification is obtained, providing accurate data input for subsequent vascular calcification quantitative analysis and risk assessment. Then, through spatial connectivity analysis and morphological optimization, the integrity and segmentation accuracy of the vascular calcification area are improved, and false detections caused by imaging artifacts or noise are reduced. At the same time, by combining dynamic volume thresholding and mathematical morphological optimization, it can adapt to different scanning conditions and individual anatomical differences, improving the stability and reliability of vascular calcification detection.

[0125] In an exemplary embodiment of the present disclosure, the following steps may be performed to input the three-dimensional vascular calcification region into a pre-trained multi-task prediction model to determine a quantitative score of the vascular calcification degree and a risk assessment level corresponding to the three-dimensional vascular calcification region. Specifically,

[0126] The three-dimensional area of vascular calcification can be input into the input layer of the multi-task prediction model to extract the vascular calcification features corresponding to the three-dimensional area of vascular calcification. The vascular calcification features include the volume of the calcified area, the shape of the calcified area, and the CT value of the calcified area. The multi-dimensional feature vector of the vascular calcification features can be extracted through the feature extraction layer of the multi-task prediction model. The multi-dimensional feature vector is input into the regression and classification layers of the multi-task prediction model to output the quantitative score of the vascular calcification degree and the risk assessment level corresponding to the three-dimensional area of vascular calcification.

[0127] The input layer is primarily used to extract features from the three-dimensional regions of vascular calcification and construct feature vectors suitable for regression and classification tasks. In the input layer, the volume of the calcified region can be calculated by voxel counting, which counts the number of voxels belonging to the calcified region and multiplies it by the physical volume of a single voxel to obtain the total calcified volume. The shape of the calcified region can be measured by calculating morphological parameters such as compactness, sphericity, and flatness. For example, compactness is used to measure the shape complexity of the calcified region, while sphericity is used to assess the regularity of the calcified region. The CT value characteristics of the calcified region can be obtained through statistical analysis, including calculating the mean, standard deviation, maximum value, and quantile of the CT value to reflect the density distribution of the calcified tissue.

[0128] The feature extraction layer can use convolutional neural networks or variational autoencoders for feature learning. Convolutional neural networks can extract spatial features through multi-layer convolution operations, while variational autoencoders can learn the latent representation of calcification features through probability distribution modeling. Specifically, the variational autoencoder can use multiple convolutional layers to extract multi-scale features from calcified voxel data, and use pooling layers to reduce the data dimension while retaining key information. A fully connected layer can then be used to linearly transform the features and generate a multi-dimensional feature vector corresponding to the vascular calcification characteristics.

[0129] The regression layer is primarily used to predict a quantitative score for vascular calcification based on the input feature vector. The regression method employed by the regression layer can include linear regression, support vector regression, or a neural network-based regression model. In specific implementations, a linear transformation can be performed on the input multidimensional feature vector, and a ReLU activation function can be used for nonlinear mapping to enhance the model's expressive power. The mean squared error (MSE) or Huber loss can then be used as the regression objective function to optimize the quantitative score for vascular calcification.

[0130] The classification layer's primary task is to predict the vascular calcification risk assessment level based on the regression results and additional classification features. This layer can employ a Softmax classifier, random forest, or gradient boosting decision tree (GBDT) for classification. Specifically, the input multidimensional feature vector is mapped to categories, and a cross-entropy loss is optimized to maximize classification accuracy. A multi-layer perceptron (MLP) is then used for classification, incorporating clinical characteristics such as patient age, gender, and medical history to improve the accuracy of risk assessment.

[0131] Through a multi-task prediction model based on deep learning, the automation level of vascular calcification assessment is improved and human error is reduced. At the same time, by combining regression and classification tasks, a more comprehensive vascular calcification analysis can be provided, improving the accuracy and interpretability of clinical diagnosis. The model can be applied to imaging data from different populations and with different CT scan parameters, further enhancing the universality and reliability of vascular calcification assessment.

[0132] In an exemplary embodiment of the present disclosure, an assessment report including a three-dimensional coordinate mapping can be generated by combining the quantitative scoring of the degree of vascular calcification and the risk assessment level with the coordinate information of the three-dimensional area of vascular calcification. The assessment report is converted into a DICOM structured report that complies with the DICOM standard, and the DICOM structured report is uploaded to the electronic medical record associated with the non-contrast enhanced scan image set through the interface of the medical imaging system.

[0133] Among them, the vascular calcification assessment report can be based on multi-dimensional data fusion, including quantitative scores, risk levels and spatial distribution information, to provide accurate vascular calcification analysis results. The core goal of three-dimensional coordinate mapping is to visualize the vascular calcification area and provide accurate anatomical positioning to facilitate doctors to make clinical decisions. The acquisition of three-dimensional coordinate information can rely on the image registration results. Through the coordinate transformation matrix of the fused image, the vascular calcification area is positioned in the standard anatomical coordinate system to ensure spatial consistency between different imaging modalities. Specifically, the center coordinates of the calcified area can be extracted based on the vascular segmentation mask, and the spatial distribution of the calcified area in the fused image can be calculated. Secondly, a three-dimensional reconstruction algorithm can be used to generate a three-dimensional model of vascular calcification, and the model can be spatially transformed to match the standard medical coordinate system. In an optional embodiment, a spatial mapping method based on image registration can be used. For example, an affine transformation or a B-spline free deformation model can be used to convert the coordinates of the calcified area to a unified anatomical reference frame to improve the comparability of the imaging data.

[0134] DICOM Structured Report (DICOM-SR) is a standardized medical imaging report format that can store structured text, image annotations, and analysis results to ensure the interoperability and standardized storage of medical data. The basic principle of DICOM-SR is to store image analysis information based on a hierarchical structure and encode different types of data through standard DICOM tags. In the specific implementation process, a DICOM-SR data structure can be constructed based on the results of vascular calcification analysis, including fields such as patient basic information, image source, analysis method, and quantitative score. Then, a DICOM tool library (such as DCMTK or dcm4che) can be used to generate a DICOM-SR file and perform data format verification to ensure compliance with the DICOM standard.

[0135] Through the interface of the medical imaging system, DICOM structured reports are uploaded to the electronic medical record (EMR) associated with the non-contrast enhanced scan image set. The EMR system is an important component of modern medical imaging management, storing patients' historical imaging data and related diagnostic information to facilitate physicians' medical history review and longitudinal analysis. Uploading DICOM-SR files relies on a medical picture archiving and communication system (PACS) or a radiology information system (RIS), which transmits the report data to the hospital database via standard DICOM network protocols (such as DICOM C-STORE).

[0136] The generated vascular calcification assessment report provides standardized quantitative analysis results and supports storage and sharing in the DICOM-SR format, optimizing the management and clinical application of medical imaging data. Compared to traditional manual input methods, the automatic generation of structured assessment reports improves the accuracy and consistency of data records and reduces the workload of physicians. Furthermore, through seamless integration with PACS and EMR systems, vascular calcification analysis results can be rapidly stored and remotely accessed, improving clinical efficiency and providing standardized data support for multi-center medical research.

[0137] It should be noted that although the steps of the method disclosed herein are depicted in a particular order in the accompanying drawings, this does not require or imply that the steps must be performed in that particular order, or that all steps must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one, and / or one step may be decomposed into multiple steps.

[0138] In addition, in this exemplary embodiment, a vascular calcification analysis device based on non-enhanced CT images is also provided. Figure 6 As shown, the non-enhanced CT image-based vascular calcification analysis device 600 includes: an image data acquisition module 610, an image standardization module 620, a fusion image generation module 630, a vascular mask determination module 640, a calcified region identification module 650, and a vascular calcification inference module 660.

[0139] An image data acquisition module 610 is configured to acquire a non-contrast enhanced scan image set, the non-contrast enhanced scan image set including a lung CT plain scan image, a neck CT plain scan image, and a brain CT plain scan image, and parse scan parameter information in a metadata field of the non-contrast enhanced scan image set;

[0140] An image normalization module 620 is configured to perform spatial resolution normalization processing on the non-contrast enhanced scan image set based on the scan parameter information to generate a pre-processed image sequence;

[0141] A fused image generation module 630 is configured to input the pre-processed image sequence into a pre-trained multimodal registration network to extract key anatomical landmarks from the images of each part, and generate a spatially aligned fused image based on the key anatomical landmarks;

[0142] a vessel mask determination module 640 , configured to input the fused image into a pre-trained attention-enhanced segmentation network to determine vessel segmentation masks for the thoracic aorta, common carotid artery, and intracranial artery in the fused image;

[0143] a calcification region identification module 650 for extracting a candidate voxel set from the fused image based on the vascular segmentation mask, and performing a three-dimensional connected domain analysis on the candidate voxel set to determine a three-dimensional region of vascular calcification in the fused image;

[0144] The vascular calcification inference module 660 is used to input the three-dimensional vascular calcification region into a pre-trained multi-task prediction model to determine the quantitative score of the vascular calcification degree and the risk assessment level corresponding to the three-dimensional vascular calcification region.

[0145] In an exemplary embodiment of the present disclosure, based on the above solution, the image standardization module 620 is configured as follows:

[0146] Determining the spatial resolution between CT images of various parts in the non-contrast enhanced scan image set by using the scanning parameter information;

[0147] The CT images of each part in the non-contrast enhanced scan image set are resampled based on the spatial resolution to generate a preprocessed image sequence.

[0148] In an exemplary embodiment of the present disclosure, based on the above solution, the fused image generation module 630 is configured as follows:

[0149] Inputting the preprocessed image sequence into the input layer of the multimodal registration network, and extracting CT image features corresponding to the CT images of each part in the preprocessed image sequence;

[0150] Identifying and extracting key anatomical landmarks in the CT images of various parts from the CT image features through the feature extraction layer of the multimodal registration network;

[0151] A joint deformation matrix between CT images of various parts is determined based on the key anatomical landmarks, and the pre-processed image sequence is spatially aligned and fused using the joint deformation matrix to generate a fused image.

[0152] In an exemplary embodiment of the present disclosure, based on the aforementioned solution, the key anatomical landmarks include the tracheal carina in the lung CT plain scan image, the anterior commissure apex and the posterior commissure base in the brain CT plain scan image, and the common carotid artery bifurcation in the neck CT plain scan image; the fusion image generation module 630 is configured to:

[0153] A first three-dimensional Cartesian coordinate system is established with the tracheal carina as the center, a second three-dimensional Cartesian coordinate system is established with the anterior and posterior commissure connecting the apex of the anterior commissure and the base of the posterior commissure as the coordinate axis, and a third three-dimensional Cartesian coordinate system is established with the common carotid artery bifurcation as the center;

[0154] determining an axial tensile deformation component based on the first three-dimensional Cartesian coordinate system and the third three-dimensional Cartesian coordinate system;

[0155] performing singular value decomposition on a spatial rotation matrix between the second three-dimensional Cartesian coordinate system and the third three-dimensional Cartesian coordinate system to extract a pitch angle deviation compensation parameter;

[0156] fusing the axial tensile deformation component and the pitch angle deviation compensation parameter into a preset B-spline free deformation model to generate a joint deformation matrix;

[0157] The voxel space coordinates of the CT images of each part in the pre-processed image sequence are transformed by using the joint deformation matrix to generate a fused image.

[0158] In an exemplary embodiment of the present disclosure, based on the above solution, the blood vessel mask determination module 640 is configured as follows:

[0159] Inputting the fused image into the input layer of the attention-enhanced segmentation network to extract fused image features corresponding to the fused image;

[0160] Suppressing the bone artifact region in the fused image feature through the attention mechanism layer of the attention enhanced segmentation network to obtain enhanced image features;

[0161] The enhanced image features are identified using the regional segmentation layer of the attention-enhanced segmentation network to extract the vascular segmentation masks corresponding to the thoracic aorta, common carotid artery and intracranial artery in the fused image.

[0162] In an exemplary embodiment of the present disclosure, based on the above solution, the calcified region identification module 650 is configured to:

[0163] determining an image region of interest in the fused image based on the blood vessel segmentation mask;

[0164] In each of the image regions of interest, voxels that meet the characteristics of vascular calcification are screened out according to the determined dynamic CT value threshold to form a candidate voxel set.

[0165] In an exemplary embodiment of the present disclosure, based on the above solution, the calcified region identification module 650 is configured to:

[0166] Performing connectivity marking on each voxel in the candidate voxel set to determine an initial connected voxel group;

[0167] Calculating the volume of each of the initial connected voxel groups, and filtering out the initial connected voxel groups whose volumes are smaller than a preset volume threshold, to obtain a target connected voxel group;

[0168] Morphological optimization is performed on the target connected voxel group to obtain a three-dimensional region of vascular calcification.

[0169] In an exemplary embodiment of the present disclosure, based on the aforementioned solution, the vascular calcification reasoning module 660 is configured as follows:

[0170] Inputting the three-dimensional vascular calcification region into the input layer of the multi-task prediction model, extracting vascular calcification features corresponding to the three-dimensional vascular calcification region, wherein the vascular calcification features include the volume of the calcified region, the shape of the calcified region, and the CT value of the calcified region;

[0171] Extracting a multidimensional feature vector of the vascular calcification feature through a feature extraction layer of the multi-task prediction model;

[0172] The multidimensional feature vector is input into the regression and classification layers of the multi-task prediction model, and the quantitative score of the vascular calcification degree and the risk assessment level corresponding to the three-dimensional vascular calcification area are output respectively.

[0173] In an exemplary embodiment of the present disclosure, based on the aforementioned solution, the vascular calcification analysis device 600 based on non-enhanced CT images further includes an evaluation report generation module, which is configured to:

[0174] Generate an assessment report including a three-dimensional coordinate mapping by combining the quantitative score of the degree of vascular calcification and the risk assessment level with the coordinate information of the three-dimensional area of vascular calcification;

[0175] The evaluation report is converted into a DICOM structured report that complies with the DICOM standard, and the DICOM structured report is uploaded to the electronic medical record associated with the non-contrast enhanced scan image set through an interface of a medical imaging system.

[0176] The specific details of each module of the above-mentioned vascular calcification analysis device based on non-enhanced CT images have been described in detail in the corresponding vascular calcification analysis method based on non-enhanced CT images, and therefore will not be repeated here.

[0177] It should be noted that while the detailed description above mentions several modules or units of the non-enhanced CT image-based vascular calcification analysis device, this division is not mandatory. In fact, depending on the embodiments of the present disclosure, the features and functions of two or more modules or units described above can be embodied in a single module or unit. Conversely, the features and functions of a single module or unit described above can be further divided and embodied by multiple modules or units.

[0178] In addition, in an exemplary embodiment of the present disclosure, an electronic device capable of implementing the above-mentioned vascular calcification analysis method based on non-enhanced CT images is also provided.

[0179] Those skilled in the art will appreciate that various aspects of the present disclosure may be implemented as systems, methods, or program products. Therefore, various aspects of the present disclosure may be implemented in the following forms: a complete hardware embodiment, a complete software embodiment (including firmware, microcode, etc.), or an embodiment combining hardware and software aspects, which may be collectively referred to herein as a "circuit," "module," or "system."

[0180] Refer to the following Figure 7 7 to describe an electronic device 700 according to such an embodiment of the present disclosure. Figure 7 The electronic device 700 shown is merely an example and should not limit the functions and scope of use of the embodiments of the present disclosure.

[0181] like Figure 7 As shown, electronic device 700 is implemented as a general-purpose computing device. Components of electronic device 700 may include, but are not limited to, the aforementioned at least one processing unit 710, the aforementioned at least one storage unit 720, a bus 730 connecting various system components (including storage unit 720 and processing unit 710), and a display unit 740.

[0182] The storage unit stores program codes, which can be executed by the processing unit 710, so that the processing unit 710 performs the steps described in the "Exemplary Method" section of the present disclosure according to various exemplary embodiments. For example, the processing unit 710 can perform the following steps: Figure 2 In step S210, a non-contrast enhanced scan image set is obtained, wherein the non-contrast enhanced scan image set includes a lung CT plain scan image, a neck CT plain scan image, and a brain CT plain scan image, and the scan parameter information in the metadata field of the non-contrast enhanced scan image set is parsed; in step S220, a spatial resolution normalization process is performed on the non-contrast enhanced scan image set based on the scan parameter information to generate a pre-processed image sequence; in step S230, the pre-processed image sequence is input into a pre-trained multimodal registration network to extract key anatomical landmarks in the images of each part, and a spatial image sequence is generated based on the key anatomical landmarks. In step S240, the fused image is input into a pre-trained attention-enhanced segmentation network to determine a vascular segmentation mask for the thoracic aorta, common carotid artery, and intracranial artery in the fused image; in step S250, a candidate voxel set is extracted from the fused image based on the vascular segmentation mask, and a three-dimensional connected domain analysis is performed on the candidate voxel set to determine a three-dimensional region of vascular calcification in the fused image; in step S260, the three-dimensional region of vascular calcification is input into a pre-trained multi-task prediction model to determine a quantitative score of the degree of vascular calcification and a risk assessment level corresponding to the three-dimensional region of vascular calcification.

[0183] The storage unit 720 may include a readable medium in the form of a volatile storage unit, such as a random access memory unit (RAM) 721 and / or a cache memory unit 722 , and may further include a read-only memory unit (ROM) 723 .

[0184] The storage unit 720 may also include a program / utility 724 having a set (at least one) of program modules 725, such program modules 725 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0185] Bus 730 may represent one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of a variety of bus architectures.

[0186] The electronic device 700 can also communicate with one or more external devices 770 (e.g., a keyboard, a pointing device, a Bluetooth device, etc.), one or more devices that enable a user to interact with the electronic device 700, and / or any device that enables the electronic device 700 to communicate with one or more other computing devices (e.g., a router, a modem, etc.). Such communication can occur via an input / output (I / O) interface 750. Furthermore, the electronic device 700 can also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network such as the Internet) via a network adapter 760. As shown, the network adapter 760 communicates with other modules of the electronic device 700 via a bus 730. It should be understood that, although not shown in the figure, other hardware and / or software modules can be used in conjunction with the electronic device 700, including but not limited to microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0187] Through the description of the above embodiments, it is easy for those skilled in the art to understand that the example embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solution according to the embodiments of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes a number of instructions to enable a computing device (which can be a personal computer, a server, a terminal device, or a network device, etc.) to execute the method according to the embodiments of the present disclosure.

[0188] In exemplary embodiments of the present disclosure, a computer-readable storage medium is also provided, on which is stored a program product capable of implementing the aforementioned methods of this specification. In some possible embodiments, various aspects of the present disclosure may also be implemented in the form of a program product comprising program code. When the program product is executed on a terminal device, the program code is configured to cause the terminal device to execute the steps described in the "Exemplary Methods" section of this specification according to various exemplary embodiments of the present disclosure.

[0189] refer to Figure 8 As shown, a program product 800 for implementing the above-described method for analyzing vascular calcification based on non-enhanced CT images according to an embodiment of the present disclosure is described. This program product can be implemented in a portable compact disk read-only memory (CD-ROM) and include program code, and can be executed on a terminal device, such as a personal computer. However, the program product of the present disclosure is not limited thereto. In this document, a readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0190] The program product may be implemented in any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or component, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0191] A computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, which carries readable program code. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium that can transmit, propagate, or transfer a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0192] The program code embodied on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0193] The program code for performing the operations of the present disclosure may be written in any combination of one or more programming languages, including object-oriented programming languages such as Java, C++, and the like, as well as conventional procedural programming languages such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a stand-alone software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving a remote computing device, the remote computing device may be connected to the user computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0194] Furthermore, the figures above are merely illustrative of the processes included in the methods according to exemplary embodiments of the present disclosure and are not intended to be limiting. It is readily understood that the processes illustrated in the figures above do not indicate or limit the temporal order of these processes. Furthermore, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0195] Through the description of the above embodiments, it is easy for those skilled in the art to understand that the example embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solution according to the embodiments of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes several instructions to enable a computing device (which can be a personal computer, a server, a touch terminal, or a network device, etc.) to execute the method according to the embodiments of the present disclosure.

[0196] Other embodiments of the present disclosure will readily occur to those skilled in the art after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow from the general principles of the present disclosure and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, with the true scope and spirit of the present disclosure being indicated by the claims.

[0197] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes can be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A method for analyzing vascular calcification based on non-enhanced CT images, characterized in that: include: Acquiring a non-contrast enhanced scan image set, the non-contrast enhanced scan image set including a lung CT plain scan image, a neck CT plain scan image, and a brain CT plain scan image, and parsing scan parameter information in a metadata field of the non-contrast enhanced scan image set; performing spatial resolution normalization processing on the non-contrast enhanced scan image set based on the scan parameter information to generate a preprocessed image sequence; Inputting the preprocessed image sequence into a pre-trained multimodal registration network to extract key anatomical landmarks in the images of each part, and generating a spatially aligned fused image based on the key anatomical landmarks; Inputting the fused image into a pre-trained attention-enhanced segmentation network to determine vessel segmentation masks of the thoracic aorta, common carotid artery, and intracranial artery in the fused image; extracting a set of candidate voxels from the fused image based on the vascular segmentation mask, and performing a three-dimensional connected domain analysis on the set of candidate voxels to determine a three-dimensional region of vascular calcification in the fused image; The three-dimensional vascular calcification region is input into a pre-trained multi-task prediction model to determine a quantitative score of the vascular calcification degree and a risk assessment level corresponding to the three-dimensional vascular calcification region.

2. The vascular calcification analysis method according to claim 1, characterized in that: The performing spatial resolution normalization processing on the non-contrast enhanced scan image set based on the scan parameter information to generate a pre-processed image sequence includes: Determining the spatial resolution between CT images of various parts in the non-contrast enhanced scan image set by using the scanning parameter information; The CT images of each part in the non-contrast enhanced scan image set are resampled based on the spatial resolution to generate a preprocessed image sequence.

3. The vascular calcification analysis method according to claim 1, characterized in that: Inputting the pre-processed image sequence into a pre-trained multimodal registration network to extract key anatomical landmarks in the images of each part, and generating a spatially aligned fused image based on the key anatomical landmarks, includes: Inputting the preprocessed image sequence into the input layer of the multimodal registration network, and extracting CT image features corresponding to the CT images of each part in the preprocessed image sequence; Identifying and extracting key anatomical landmarks in the CT images of various parts from the CT image features through the feature extraction layer of the multimodal registration network; A joint deformation matrix between CT images of various parts is determined based on the key anatomical landmarks, and the pre-processed image sequence is spatially aligned and fused using the joint deformation matrix to generate a fused image.

4. The vascular calcification analysis method according to claim 3, characterized in that: The key anatomical landmarks include the tracheal carina in the lung CT plain scan image, the anterior commissure apex and the posterior commissure base in the brain CT plain scan image, and the common carotid artery bifurcation in the neck CT plain scan image; The step of determining a joint deformation matrix between CT images of various parts based on the key anatomical landmarks, and performing spatial alignment and fusion on the pre-processed image sequence using the joint deformation matrix to generate a fused image includes: A first three-dimensional Cartesian coordinate system is established with the tracheal carina as the center, a second three-dimensional Cartesian coordinate system is established with the anterior and posterior commissure connecting the apex of the anterior commissure and the base of the posterior commissure as the coordinate axis, and a third three-dimensional Cartesian coordinate system is established with the common carotid artery bifurcation as the center; determining an axial tensile deformation component based on the first three-dimensional Cartesian coordinate system and the third three-dimensional Cartesian coordinate system; performing singular value decomposition on a spatial rotation matrix between the second three-dimensional Cartesian coordinate system and the third three-dimensional Cartesian coordinate system to extract a pitch angle deviation compensation parameter; fusing the axial tensile deformation component and the pitch angle deviation compensation parameter into a preset B-spline free deformation model to generate a joint deformation matrix; The voxel space coordinates of the CT images of each part in the pre-processed image sequence are transformed by using the joint deformation matrix to generate a fused image.

5. The vascular calcification analysis method according to claim 1, characterized in that: Inputting the fused image into a pre-trained attention-enhanced segmentation network to determine the vessel segmentation masks of the thoracic aorta, the common carotid artery, and the intracranial artery in the fused image includes: Inputting the fused image into the input layer of the attention-enhanced segmentation network to extract fused image features corresponding to the fused image; Suppressing the bone artifact region in the fused image feature through the attention mechanism layer of the attention enhanced segmentation network to obtain enhanced image features; The enhanced image features are identified using the regional segmentation layer of the attention-enhanced segmentation network to extract the vascular segmentation masks corresponding to the thoracic aorta, common carotid artery and intracranial artery in the fused image.

6. The vascular calcification analysis method according to claim 1, characterized in that: Extracting a candidate voxel set from the fused image based on the blood vessel segmentation mask includes: determining an image region of interest in the fused image based on the blood vessel segmentation mask; In each of the image regions of interest, voxels that meet the characteristics of vascular calcification are screened out according to the determined dynamic CT value threshold to form a candidate voxel set.

7. The vascular calcification analysis method according to claim 1 or 6, characterized in that: Performing a three-dimensional connected domain analysis on the candidate voxel set to generate a three-dimensional vascular calcification region includes: Performing connectivity marking on each voxel in the candidate voxel set to determine an initial connected voxel group; Calculating the volume of each of the initial connected voxel groups, and filtering out the initial connected voxel groups whose volumes are smaller than a preset volume threshold, to obtain a target connected voxel group; Morphological optimization is performed on the target connected voxel group to obtain a three-dimensional region of vascular calcification.

8. The vascular calcification analysis method according to claim 1, characterized in that: Inputting the three-dimensional vascular calcification region into a pre-trained multi-task prediction model to determine a quantitative score of the degree of vascular calcification and a risk assessment level corresponding to the three-dimensional vascular calcification region includes: Inputting the three-dimensional vascular calcification region into the input layer of the multi-task prediction model, extracting vascular calcification features corresponding to the three-dimensional vascular calcification region, wherein the vascular calcification features include the volume of the calcified region, the shape of the calcified region, and the CT value of the calcified region; Extracting a multidimensional feature vector of the vascular calcification feature through a feature extraction layer of the multi-task prediction model; The multidimensional feature vector is input into the regression and classification layers of the multi-task prediction model, and the quantitative score of the vascular calcification degree and the risk assessment level corresponding to the three-dimensional vascular calcification area are output respectively.

9. The vascular calcification analysis method according to claim 1, characterized in that: The method further comprises: Generate an assessment report including a three-dimensional coordinate mapping by combining the quantitative score of the degree of vascular calcification and the risk assessment level with the coordinate information of the three-dimensional area of vascular calcification; The evaluation report is converted into a DICOM structured report that complies with the DICOM standard, and the DICOM structured report is uploaded to the electronic medical record associated with the non-contrast enhanced scan image set through an interface of a medical imaging system.

10. A vascular calcification analysis device based on non-enhanced CT images, characterized in that: include: an image data acquisition module, configured to acquire a non-contrast enhanced scan image set, the non-contrast enhanced scan image set including a lung CT plain scan image, a neck CT plain scan image, and a brain CT plain scan image, and parse scan parameter information in a metadata field of the non-contrast enhanced scan image set; an image standardization module, configured to perform spatial resolution standardization processing on the non-contrast enhanced scan image set based on the scan parameter information to generate a preprocessed image sequence; a fused image generation module, configured to input the pre-processed image sequence into a pre-trained multimodal registration network to extract key anatomical landmarks from the images of each part, and generate a spatially aligned fused image based on the key anatomical landmarks; a vascular mask determination module, configured to input the fused image into a pre-trained attention-enhanced segmentation network to determine vascular segmentation masks of the thoracic aorta, common carotid artery, and intracranial artery in the fused image; a calcification region identification module, configured to extract a candidate voxel set from the fused image based on the vascular segmentation mask, and perform a three-dimensional connected domain analysis on the candidate voxel set to determine a three-dimensional region of vascular calcification in the fused image; The vascular calcification inference module is used to input the three-dimensional vascular calcification area into a pre-trained multi-task prediction model to determine the quantitative score of the vascular calcification degree and the risk assessment level corresponding to the three-dimensional vascular calcification area.

Images