High-resolution magnetic resonance blood vessel wall image segmentation method, analysis system, equipment and medium

Abstract

The invention relates to a high-resolution magnetic resonance blood vessel wall image segmentation method, an analysis system, equipment and a medium, and the method comprises the steps: S1, collecting a three-dimensional high-resolution magnetic resonance blood vessel wall imaging image and a PET image of a head and neck part, and carrying out the registration preprocessing; s2, inputting the image subjected to registration preprocessing into a blood vessel segmentation model to obtain a head and neck multi-segment blood vessel segmentation mask, and extracting an initial center line from each segmented blood vessel segment through skeletonization operation; s3, inputting the initial center line into a multi-modal center line correction unit to obtain a corrected center line point set; and S4, performing curved surface plane reconstruction according to the corrected center line point set, and inputting an obtained continuous two-dimensional curved surface reconstruction image into the blood vessel wall segmentation model to obtain a blood vessel segmentation result. Compared with the prior art, the method has the advantages of high segmentation accuracy and high analysis efficiency.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a high-resolution magnetic resonance imaging (MRI) method, analysis system, device, and medium for segmenting blood vessel walls. Background Technology

[0002] High-resolution magnetic resonance vessel wall imaging (HR-MRVWI) is a common diagnostic method for cerebrovascular diseases. Through dark-blood imaging and thin-slice scanning of magnetic resonance, the condition of the vessel walls, lumens, and plaques in the head and neck can be clearly displayed.

[0003] To display lesions more clearly and intuitively, traditional methods require manual reconstruction of images, which is time-consuming, labor-intensive, highly subjective, and results in poor consistency among doctors.

[0004] Currently, various methods exist for automated image reconstruction, but most only target single sites (head, neck, or aorta, etc.), lacking automated reconstruction methods for the entire head and carotid artery process. This is mainly because, in a single scan, there are significant scale differences and image feature variations from intracranial arteries in the head and neck (small diameter, low contrast) to carotid arteries (large diameter, variable morphology), posing a major challenge to integrated segmentation. Furthermore, existing methods largely remain at the image reconstruction stage, failing to form an effective closed loop with subsequent clinical quantitative tasks such as stenosis diagnosis and plaque parameter calculation, thus hindering true automation of the diagnostic and treatment process.

[0005] To address the aforementioned issues, there is an urgent need to design an automated method for segmenting the vessel wall throughout the entire process of the head and carotid arteries. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art by providing a high-resolution magnetic resonance vascular wall image segmentation method, analysis system, device and medium. Through multimodal image fusion and multimodal centerline correction unit, high-precision segmentation of the entire head and neck blood vessels is achieved.

[0007] The objective of this invention can be achieved through the following technical solutions: According to a first aspect of the present invention, a method for segmenting a high-resolution magnetic resonance imaging (MRI) blood vessel wall image is provided, comprising: S1. Acquire three-dimensional high-resolution magnetic resonance imaging (MRI) images and PET images of the blood vessel wall in the head and neck, and perform registration preprocessing; S2. Input the pre-processed image after registration into the three-dimensional blood vessel segmentation model to obtain a multi-segment blood vessel segmentation mask for the head and neck, and extract the initial center line from each segmented blood vessel segment through skeletonization operation. S3. Input the initial centerline into the multimodal centerline correction unit to obtain the corrected centerline point set; wherein, the multimodal centerline correction unit utilizes the characteristic that the vascular lumen is a metabolic resting area and the active inflammatory plaque is a high metabolic area on the PET image, and uses the PET image information as a biological spatial constraint to guide and correct the direction of the centerline. S4. Based on the corrected centerline point set, perform surface plane reconstruction, and input the obtained continuous two-dimensional surface reconstruction image into the blood vessel wall segmentation model to obtain the blood vessel segmentation result.

[0008] Preferably, the acquisition of three-dimensional high-resolution magnetic resonance imaging (MRI) images and PET images of the head and neck, followed by registration preprocessing, specifically includes: An integrated PET / MR device was used to simultaneously acquire three-dimensional high-resolution magnetic resonance imaging images of blood vessel walls and PET images of the head and neck. Intensity normalization was performed on three-dimensional high-resolution magnetic resonance imaging of blood vessel walls, attenuation correction and standardized uptake value (SUV) calculation were performed on PET images, and a standardized metabolic parameter map was generated. By leveraging the inherent hardware coordinate synchronization of integrated PET / MR equipment, spatial positioning information in the DICOM file headers of three-dimensional high-resolution magnetic resonance vascular wall imaging images and PET images is read. Through rigid transformation based on the DICOM coordinate system, the two modal images are aligned to achieve registration preprocessing.

[0009] Preferably, the three-dimensional blood vessel segmentation model is a 3D V-Net blood vessel segmentation model.

[0010] Preferably, in step S3, the initial centerline is input to the multimodal centerline correction unit to obtain the corrected centerline point set, specifically including: S301. Initialize multimodal parameters: Define the spherical search radius R, and set the metabolic activity threshold T according to the characteristics of PET images; S302, Point-by-point traversal and multimodal collaborative decision-making, specifically including: S3021. Initialize the centerline point set C_corrected to an empty list; S3022. For each point p_i in the initial centerline point set C_initial, execute the following loop: a) Simultaneous extraction of dual-modal features: Centered on point p_i, within a spherical neighborhood N(p_i) of radius R, the following operations are performed synchronously and in parallel: extracting local image patches from the pre-registered 3D high-resolution magnetic resonance vascular wall imaging image I_MR and calculating its gray-level gradient direction g_MR as an anatomical space guide; and extracting the uptake values ​​SUV of all voxels from the pre-registered PET image I_PET as functional metabolic distribution. b) Joint assessment of metabolic-anatomical consistency: Read the SUV value SUV(p_i) of point p_i on the pre-registered PET image I_PET. If SUV(p_i) > T, trigger the depth correction mechanism based on multimodal information. Otherwise, it is considered that the current point is located in a reliable vascular lumen at both the metabolic and anatomical levels, and it is directly added to the centerline point set C_corrected and then jumps to the next point. c) Calculation of the correction vector for multimodal information fusion, specifically including: Finding metabolic consensus regions: Within the neighborhood N(p_i), find all voxel locations that satisfy SUV(v) < T, forming a low-metabolic candidate point set V_low, which is used to characterize credible vascular lumen regions at the metabolic level; Calculate the fusion correction direction: Calculate the combined correction vector d_correction of fusion with PET metabolic attraction and MR anatomical guidance, expressed as: d_correction = λ * d_PET + (1 - λ) * g_MR, Where: d_PET is the vector pointing from point p_i to the center of the low-metabolism candidate point set V_low, used to characterize the spatial constraints and guidance of metabolic information; g_MR is the local gray-level gradient direction of the I_MR image, used to characterize the boundary constraints of anatomical structures; λ is a dynamic weighting factor, which is adjusted according to the degree to which SUV(p_i) exceeds the metabolic activity threshold T. The more significant the metabolic abnormality, the higher the weight λ of the PET image information. d) Dynamic step size correction and update: The correction step size α is positively correlated with (SUV(p_i) - T), and is dynamically adjusted. Then, the coordinates of the new point are calculated: p_corrected = p_i + α * normalize(d_correction), and the new point p_corrected is added to the centerline point set C_corrected. S303, Path Optimization: Smooth the centerline point set C_corrected obtained from the initial correction using spline curves; S304. Return result: Output the final corrected set of centerline points C_corrected, which is optimal in both anatomical and metabolic aspects.

[0011] Preferably, the metabolic activity threshold T is set as follows: T = μ_background + k * σ_background, Where μ_background and σ_background are the mean and standard deviation of the SUV in the image background region, respectively, and k is an empirical coefficient.

[0012] Preferably, in step S4, surface plane reconstruction is performed based on the corrected centerline point set, and the resulting continuous two-dimensional surface reconstruction image is input into the blood vessel wall segmentation model to obtain the blood vessel segmentation result, specifically including: Using the corrected centerline point set, curve plane reconstruction is automatically performed, and MR-CPR and PET-CPR image sequences with completely consistent spatial positions are generated simultaneously. The MR-CPR and PET-CPR image sequences with identical spatial locations are input into a vessel wall segmentation model based on an improved 2D U-Net to obtain vessel segmentation results.

[0013] Preferably, the blood vessel wall segmentation model based on the improved 2D U-Net is specifically as follows: Input: Multi-channel input is used, corresponding to MR-CPR and PET-CPR image sequences with completely consistent spatial locations; Improved 2D U-Net network structure: In the encoder part of U-Net, after initial convolution of the dual-channel input, feature concatenation is performed to fuse low-level features of different modalities; Training objective: To simultaneously learn the anatomical features of MR images and the metabolic function features of PET images. During training, high metabolic signals of plaques in PET images are used as auxiliary features. When the boundaries of MR images are blurred, the boundaries suggested by metabolically abnormal areas in PET images are given priority.

[0014] According to a second aspect of the present invention, a high-resolution magnetic resonance imaging system for analyzing blood vessel walls is provided, the system comprising: The segmentation module performs segmentation based on the high-resolution magnetic resonance blood vessel wall image segmentation method described above to obtain blood vessel segmentation results. The automatic plaque vulnerability assessment module is used to extract metabolic parameters of plaque regions from pre-registered PET images and determine the risk category of plaque regions through a multi-parameter fusion comprehensive vulnerability scoring model. The multi-modal parameters based on the comprehensive vulnerability scoring model include morphological parameters and functional parameters. An interpretable model is used to output a comprehensive vulnerability probability score between 0 and 1, and risk categories are classified according to the score threshold. The generation module is used to visualize multimodal image data of the head and neck, as well as the corresponding quantitative parameters and risk categories of the plaque areas output by the automatic plaque vulnerability assessment module.

[0015] According to a third aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement any of the methods described above.

[0016] According to a fourth aspect of the invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements any of the methods described herein.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) A multimodal centerline correction unit is introduced, which uses PET metabolic information as a biological spatial constraint to intelligently guide and correct the direction of the centerline, ensuring the robustness and accuracy of the whole-process vascular analysis from head to neck, and effectively solving the industry problem of inaccurate centerline extraction caused by vascular bifurcation, occlusion or low contrast in traditional MR images.

[0018] (2) The integrated PET / MR equipment hardware is used to simultaneously acquire three-dimensional high-resolution magnetic resonance vascular wall imaging images and PET images. The shared scanning aperture and positioning system ensure that the two modal images have accurate spatial consistency at the source of acquisition. A direct mapping registration scheme based on DICOM coordinates can be adopted to achieve pixel-level fast alignment without complex non-rigid deformation registration, laying the foundation for subsequent accurate quantitative fusion analysis.

[0019] (3) By reconstructing the curve plane, MR-CPR and PET-CPR image sequences with completely consistent spatial positions are generated simultaneously. The registered multimodal image pairs are then used as input to improve the 2D U-Net vascular segmentation model. The multi-channel input architecture can simultaneously learn the anatomical structural features of MR and the metabolic functional features of PET. During the training process, the model learns to use the high metabolic signal of plaques in PET as a powerful auxiliary feature, thereby enhancing the confidence of the segmentation boundary and identifying the biological properties of plaques in areas with poor contrast or blurred boundaries in MR images. Finally, it achieves accurate segmentation of the vascular lumen and outer wall with both anatomical accuracy and functional significance.

[0020] (4) The algorithm framework of “3D V-Net vessel segmentation + centerline extraction and correction unit + curve plane reconstruction + 2D U-Net vessel wall segmentation” constructed in this invention decouples and effectively integrates coarse-grained vessel localization and fine-grained vessel wall segmentation through cascaded network design and feature reuse mechanism. This allows 3D global spatial information and 2D local detail features to complement each other, successfully overcoming the cross-scale segmentation problem caused by the huge size difference between intracranial arteries and carotid arteries. It realizes integrated automatic segmentation and reconstruction of the entire blood vessels from head to neck. At the same time, by fully mining the anatomical localization and image pixel information of vessel inner diameter and outer diameter, the anatomical localization can provide additional useful information for vessel wall segmentation, improving the accuracy of vessel wall segmentation results. It can adapt to various actual scenarios of tortuous course and developmental variation of head and neck arteries, and achieve accurate and stable vessel wall segmentation results.

[0021] (5) Achieved full-process integration and high automation: This invention integrates three traditionally independent operations—image reconstruction, stenosis diagnosis, and plaque calculation—into a unified automated process for the first time. The system only needs to input the original HR-MRVWI image to automatically output auxiliary diagnostic results including CPR reconstructed images, stenosis location markers, and multiple quantitative parameters, completely changing the traditional mode that relies heavily on manual operations and reducing human intervention and subjective differences. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the overall system framework of the present invention; Figure 3 This is a schematic diagram illustrating the effects of the present invention; Figure 4 This is a schematic diagram illustrating the centerline extraction and correction process of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0024] Example like Figure 1 As shown in the figure, this embodiment provides a high-resolution magnetic resonance blood vessel wall image segmentation method, which includes the following steps: S1. Acquire three-dimensional high-resolution magnetic resonance imaging (MRI) images and PET images of the blood vessel wall in the head and neck, and perform registration preprocessing; S2. Input the pre-processed image after registration into the three-dimensional blood vessel segmentation model to obtain a multi-segment blood vessel segmentation mask for the head and neck, and extract the initial center line from each segmented blood vessel segment through skeletonization operation. S3. Input the initial centerline into the multimodal centerline correction unit to obtain the corrected centerline point set; wherein, the multimodal centerline correction unit utilizes the characteristic that the vascular lumen is a metabolic resting area and the active inflammatory plaque is a high metabolic area on the PET image, and uses the PET image information as a biological spatial constraint to guide and correct the direction of the centerline. S4. Based on the corrected centerline point set, perform surface plane reconstruction, and input the obtained continuous two-dimensional surface reconstruction image into the blood vessel wall segmentation model to obtain the blood vessel segmentation result.

[0025] The method of this embodiment will now be described in detail.

[0026] S1. Acquire three-dimensional high-resolution magnetic resonance imaging (MRI) images and PET images of the blood vessel walls in the head and neck region, and perform registration preprocessing, specifically including: S101. An integrated PET / MR device is used to simultaneously acquire three-dimensional high-resolution magnetic resonance vascular wall imaging images (3D HR-MRVWI, T1-weighted black blood sequence in this embodiment) and PET images (fluorodeoxyglucose positron emission tomography, FDG-PET in this embodiment). The shared scanning aperture and positioning system ensure that the HR-MRVWI and FDG-PET images have accurate spatial consistency at the source of acquisition.

[0027] S102. Normalize the intensity of the three-dimensional high-resolution magnetic resonance vascular wall imaging image, perform attenuation correction on the PET image and calculate the standardized uptake value (SUV) to generate a standardized metabolic parameter map. S103. Utilizing the inherent hardware coordinate synchronization of the integrated PET / MR equipment, the spatial positioning information in the DICOM file header of the three-dimensional high-resolution magnetic resonance vascular wall imaging image and the PET image is read. Through rigid transformation based on the DICOM coordinate system, the two modal images are aligned to achieve registration preprocessing.

[0028] Before proceeding with subsequent segmentation, the process also includes expert annotation and the establishment of a gold standard, specifically including: (1) Modal annotation platform: Load the registered HR-MRVWI and PET images simultaneously on a dedicated annotation platform (such as the open-source ITK-SNAP software).

[0029] (2) Labeling content and process: Primary annotation (vascular structure): The doctor delineates the boundaries of the vascular lumen and the outer wall layer by layer on the original axial image of HR-MRVWI.

[0030] Secondary annotation (plaque metabolism): On the registered PET image, the doctor delineates the region of interest (ROI) in the blood vessel wall area, records its maximum standardized uptake value (SUVmax), and calculates the target-to-sample ratio (TBR, i.e., the ratio of plaque SUVmax to the SUVmean of the contralateral or ipsilateral jugular vein blood pool).

[0031] Arbitration and Gold Standard Generation: An arbitration panel composed of at least two experienced radiologists and nuclear medicine physicians reviewed the annotation results based on the fused multimodal images. The final gold standard includes not only the segmentation mask but also a vulnerability level label (e.g., stable, vulnerable) for each plaque, which is determined based on a combination of morphological (e.g., presence of intraplaque hemorrhage, lipid necrosis nuclei) and functional (e.g., TBR value) information.

[0032] S2. Input the pre-processed image after registration into the three-dimensional blood vessel segmentation model to obtain a multi-segment blood vessel segmentation mask for the head and neck, and extract the initial center line from each segmented blood vessel segment through skeletonization operation.

[0033] In this embodiment, the 3D blood vessel segmentation model selected is the 3D V-Net blood vessel segmentation model, specifically: Input: Preprocessed 3D HR-MRVWI image.

[0034] Output: Segmentation mask of the seven major arteries in the head and neck (including the common carotid artery, internal carotid artery, carotid bifurcation, anterior cerebral artery, middle cerebral artery, basilar artery, vertebral artery, and posterior cerebral artery).

[0035] Training details: Training was performed using the Dice loss function and the Adam optimizer. The training objective was to achieve a blood vessel segmentation DSC ≥ 0.84 performance metric on the independent test set.

[0036] S3. Input the initial centerline into the multimodal centerline correction unit to obtain the corrected centerline point set; wherein, the multimodal centerline correction unit utilizes the characteristics that the vascular lumen is a metabolic resting area and the active inflammatory plaque is a high metabolic area on the PET image, and uses the PET image information as a biological spatial constraint to guide and correct the direction of the centerline.

[0037] In this embodiment, the multimodal centerline correction unit is essentially a lightweight graph neural network or rule-based reasoning system. The input includes an initial centerline (obtained from the skeletonization of the vessel segmentation mask) and registered PET / MR data. The algorithm iterates through each point on the initial centerline, simultaneously querying the grayscale gradient of HR-MRVWI and the SUV value of PET within its neighborhood space (e.g., 5x5x5 voxels). The algorithm sets a rule: when a centerline point deviates from a high-metabolic region in PET (e.g., SUV > twice the background standard deviation), it is determined that it may have deviated into the vessel wall or plaque, requiring coordinate correction towards a low-metabolic region (vascular lumen). Through iterative optimization, the centerline is positioned at the optimal path at both the anatomical and metabolic levels.

[0038] The specific implementation includes the following sub-steps: S301. Initialize multimodal parameters: Define the spherical search radius R (2~3 voxels), and set the metabolic activity threshold T according to the characteristics of PET images.

[0039] Specifically, the metabolic activity threshold T is set as follows: T = μ_background + k * σ_background, Where μ_background and σ_background are the mean and standard deviation of the SUV of the image background region, respectively, and k is an empirical coefficient, which is set to 2 in this embodiment. This threshold is used to reliably distinguish between vascular lumens (metabolic resting areas) and inflammatory plaques (hypermetabolic areas).

[0040] S302, Point-by-point traversal and multimodal collaborative decision-making, specifically including: S3021. Initialize the centerline point set C_corrected to an empty list; S3022. For each point p_i in the initial centerline point set C_initial, execute the following loop: a) Simultaneous extraction of dual-modal features: Centered on point p_i, within a spherical neighborhood N(p_i) of radius R, the following operations are performed synchronously and in parallel: extracting local image patches from the pre-registered 3D high-resolution magnetic resonance vascular wall imaging image I_MR and calculating their gray-level gradient direction g_MR as an anatomical space guide; and extracting the SUV values ​​of all voxels from the pre-registered PET image I_PET as a functional metabolic distribution. b) Joint assessment of metabolic-anatomical consistency: Read the SUV value SUV(p_i) of point p_i on the pre-registered PET image I_PET. If SUV(p_i) > T, trigger the depth correction mechanism based on multimodal information. Otherwise, it is considered that the current point is located in a reliable vascular lumen at both the metabolic and anatomical levels, and it is directly added to the centerline point set C_corrected and then jumps to the next point. c) Calculation of the correction vector for multimodal information fusion, specifically including: Finding metabolic consensus regions: Within the neighborhood N(p_i), find all voxel locations that satisfy SUV(v) < T, forming a low-metabolic candidate point set V_low, which is used to characterize credible vascular lumen regions at the metabolic level; The innovative aspect of computational fusion correction lies in its approach: instead of simply guiding the correction towards the geometric center, it computationally fuses a comprehensive correction vector d_correction that combines the metabolic attraction of PET and the anatomical guidance of MR. The expression is as follows: d_correction = λ * d_PET + (1 - λ) * g_MR, Where: d_PET is the vector pointing from point p_i to the center of the low-metabolism candidate point set V_low, used to characterize the spatial constraints and guidance of metabolic information; g_MR is the local gray-level gradient direction of the I_MR image, used to characterize the boundary constraints of anatomical structures; λ is a dynamic weighting factor, which is adjusted according to the degree to which SUV(p_i) exceeds the metabolic activity threshold T. The more significant the metabolic abnormality, the higher the weight λ of the PET image information. d) Dynamic step size correction and update: The correction step size α is positively correlated with (SUV(p_i) - T), and is dynamically adjusted. Then, the coordinates of the new point are calculated: p_corrected = p_i + α * normalize(d_correction), and the new point p_corrected is added to the centerline point set C_corrected.

[0041] S303, Path Optimization: Smooth the centerline point set C_corrected obtained from the initial correction using spline curves.

[0042] S304. Return result: Output the final corrected set of centerline points C_corrected, which is optimal in both anatomical and metabolic aspects.

[0043] S4. Reconstruct the surface plane based on the corrected centerline point set, and input the resulting continuous two-dimensional surface reconstruction image into the blood vessel wall segmentation model to obtain the blood vessel segmentation result, which specifically includes: Using the corrected centerline point set, curve plane reconstruction is automatically performed, and MR-CPR and PET-CPR image sequences with completely consistent spatial positions are generated simultaneously. The MR-CPR and PET-CPR image sequences with identical spatial locations are input into a vessel wall segmentation model based on an improved 2D U-Net to obtain vessel segmentation results.

[0044] Specifically, the blood vessel wall segmentation model based on the improved 2D U-Net in this embodiment is as follows: Input: Multi-channel input is used, corresponding to MR-CPR and PET-CPR image sequences with completely consistent spatial locations; Improved 2D U-Net network structure: In the encoder part of U-Net, after initial convolution of the dual-channel input, feature concatenation is performed to fuse low-level features of different modalities; Training objective: To simultaneously learn the anatomical features of MR images and the metabolic function features of PET images. During training, high metabolic signals of plaques in PET images are used as auxiliary features. When the boundaries of MR images are blurred, the boundaries suggested by metabolically abnormal areas in PET images are given priority.

[0045] This embodiment also provides a high-resolution magnetic resonance blood vessel wall image analysis system, the system comprising: The segmentation module performs segmentation based on the high-resolution magnetic resonance blood vessel wall image segmentation method described above to obtain blood vessel segmentation results. The automatic plaque vulnerability assessment module is used to extract metabolic parameters (such as SUVmax and TBR) of plaque regions from pre-registered PET images and determine the risk category of plaque regions through a multi-parameter fusion comprehensive vulnerability scoring model. The multi-modal parameters based on the comprehensive vulnerability scoring model include morphological parameters (standardized wall index, remodeling index, eccentricity index, and stenosis) and functional parameters (plaque region SUVmax, TBRmax). An interpretable model (using logistic regression or gradient boosting decision tree, etc.) is used to output a comprehensive vulnerability probability score between 0 and 1, and risk categories are classified according to the score threshold (plaques with a score > 0.7 are marked as high-risk). The generation module is used to visualize multimodal image data of the head and neck, as well as the corresponding quantitative parameters and risk categories of the plaque areas output by the automatic plaque vulnerability assessment module.

[0046] In this embodiment, the system can automatically generate a structured report that deeply integrates multimodal imaging evidence and quantitative parameters. The report not only displays fused CPR images from HR-MRVWI and PET side-by-side, achieving simultaneous visualization of anatomy and function, but also presents key morphological and metabolic parameters in a structured table format. For example, a plaque that simultaneously possesses a high standardized wall index (morphological thick wall), high TBR (functional inflammation), and positive remodeling characteristics will be automatically identified by the system as a high-risk vulnerable plaque, and a quantitative risk level indication will be provided, achieving a three-dimensional and quantitative assessment of plaque risk from "morphology" to "function."

[0047] Most importantly, the report outputs a comprehensive vulnerability score or risk level based on multi-dimensional information, and automatically highlights the locations of high-risk plaques on the image, providing clinicians with a precise decision support report that integrates qualitative, quantitative, and localization information. The specific data flow process is as follows: Figure 2 As shown.

[0048] Next, the effectiveness of the method of the present invention will be verified through experiments.

[0049] (1) Data preparation and preprocessing: High-quality HR-MRVWI images (DICOM format) covering the carotid artery to intracranial artery region were collected from three hospitals. Strict inclusion and exclusion criteria were established (e.g., excluding images with severe artifacts or incomplete data). A total of 472 scans were collected; after screening, data from 426 patients were used for model development, and data from another 117 patients were used as an independent prospective test set for performance evaluation.

[0050] (2) Expert annotation and gold standard construction: Professional image annotation software (such as ITK-SNAP) was used to perform fine pixel-level annotations on the vascular lumen and outer wall boundaries of all included images. A layered annotation process was adopted: first, multiple technicians performed preliminary annotations, and then two experienced physicians (8 years or more of experience) reviewed, corrected, and finally resolved any discrepancies, thus ensuring the accuracy of the annotation results and forming the "gold standard" for model training and performance evaluation.

[0051] (3) Model development and training: Blood vessel segmentation and centerline extraction model: The model was trained on a development set using 3D V-Net to learn vascular structures. After training, the Dice similarity coefficient (DSC) of the blood vessel segmentation model reached 0.84. After the model outputs a blood vessel segmentation mask, a skeletonization algorithm was applied to extract the centerlines of each segment, and correction units were used to process and smooth the bifurcation points to form complete centerlines, such as... Figure 4 As shown.

[0052] Blood vessel wall segmentation model: 2D images are generated through CPR reconstruction based on the extracted centerline. An improved 2D U-Net (incorporating residual units) is trained on these images to complete the final segmentation of the lumen and vessel wall. The blood vessel wall segmentation model achieves a DSC of 0.86.

[0053] The training process uses Dice similarity coefficient (DSC) as a loss function, employs optimizers such as Adam, and monitors performance on the validation set to prevent overfitting.

[0054] (4) System integration and deployment: It adopts a C / S or B / S architecture. The backend uses Python (such as a model service based on PyTorch / TensorFlow) and C++ (for high-performance image processing), while the frontend provides a graphical user interface.

[0055] The trained model sequences are packaged into standardized modules and integrated into image post-processing workstations or hospital PACS systems. The system can automatically capture protocol-compliant dual-modal data from integrated PET / MR equipment, triggering a fully automated analysis process.

[0056] (5) Performance verification and evaluation: The system was comprehensively evaluated using a pre-reserved, prospective, independent test set that was not used in the training. Evaluation metrics included: Image quality assessment: Multiple doctors blindly reviewed the CPR images automatically reconstructed by the system, scoring them according to criteria such as vascular continuity and boundary clarity, and calculated the image pass rate that meets clinical diagnostic requirements.

[0057] Diagnostic accuracy assessment: The system automatically outputs the diagnosis of vascular stenosis and compares it with the "gold standard" formed by the consensus of imaging experts. The consistency of diagnosis is assessed by calculating statistical indicators such as accuracy rate and Kappa value.

[0058] Multimodal parameter reliability assessment: Intra-group correlation coefficient (ICC) and Bland-Altman analysis were used to evaluate the consistency between the morphological and functional parameters automatically calculated by the system and the results of manual measurement.

[0059] Comprehensive efficacy assessment: The predictive efficacy is evaluated by analyzing the correlation between the plaque vulnerability results determined by the system and clinical follow-up data.

[0060] Work efficiency assessment: By comparing the operation time and number of clicks of doctors using the system-assisted mode with the traditional manual mode, the efficiency improvement of the system on the workflow is quantified.

[0061] Figure 3 This is a diagram illustrating the segmentation and evaluation of the results.

[0062] Validated by large-scale clinical data, the performance indicators of the proposed solution reached a high level. Specifically, the image reconstruction pass rate was as high as 92.3%; in the diagnosis of vascular stenosis (≥50%), the accuracy rate reached 90.5%, showing a high degree of consistency with the diagnostic consensus of senior radiologists (Kappa value of 0.84); in the calculation of plaque parameters, the accuracy rates for NWI, eccentricity index, and remodeling index were 0.94, 0.83, and 0.87, respectively, showing extremely high correlation with manual measurement results (correlation coefficients of 0.96, 0.86, and 0.83, respectively). Particularly noteworthy is that the NWI automatically calculated by the system was proven to be an independent predictor of plaque vulnerability (OR=18.8, P<0.001), possessing valuable clinical risk prediction value.

[0063] The clinical application of this embodiment can significantly shorten diagnostic time and reduce the workload of doctors. Experimental results show that after adopting this system, the overall processing time for doctors was significantly reduced from 32.0 ± 11.8 minutes to 12.9 ± 4.3 minutes (P<0.001). Among them, the time-consuming and tedious blood vessel wall analysis time was reduced from 6.3 ± 1.6 minutes to a negligible 0.3 ± 0.1 minutes. At the same time, the number of mouse clicks was sharply reduced from 104.5 ± 8.9 times to 12 times, which greatly optimized the workflow and improved diagnostic efficiency (from 5.4% to 13.4%).

[0064] The electronic device of this invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). The RAM may also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0065] Multiple components in the device are connected to the I / O interface, including: input units such as keyboards and mice; output units such as various types of displays and speakers; storage units such as disks and optical discs; and communication units such as network interface cards (NICs), modems, and wireless transceivers. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0066] The processing unit executes the various methods and processes described above, such as methods S1 to S4. For example, in some embodiments, methods S1 to S4 may be implemented as computer software programs tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of methods S1 to S4 described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute methods S1 to S4 by any other suitable means (e.g., by means of firmware).

[0067] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload programmable logic devices (CPLDs), and so on.

[0068] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0069] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0070] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for segmenting high-resolution magnetic resonance imaging of blood vessel walls, characterized in that, include: S1. Acquire three-dimensional high-resolution magnetic resonance imaging (MRI) images and PET images of the blood vessel wall in the head and neck, and perform registration preprocessing; S2. Input the pre-processed image after registration into the three-dimensional blood vessel segmentation model to obtain a multi-segment blood vessel segmentation mask for the head and neck, and extract the initial center line from each segmented blood vessel segment through skeletonization operation. S3. Input the initial centerline into the multimodal centerline correction unit to obtain the corrected centerline point set; wherein, the multimodal centerline correction unit utilizes the characteristic that the vascular lumen is a metabolic resting area and the active inflammatory plaque is a high metabolic area on the PET image, and uses the PET image information as a biological spatial constraint to guide and correct the direction of the centerline. S4. Based on the corrected centerline point set, perform surface plane reconstruction, and input the obtained continuous two-dimensional surface reconstruction image into the blood vessel wall segmentation model to obtain the blood vessel segmentation result.

2. The high-resolution magnetic resonance imaging blood vessel wall image segmentation method according to claim 1, characterized in that, The acquisition of three-dimensional high-resolution magnetic resonance imaging (MRI) images and PET images of the blood vessel wall in the head and neck region, followed by registration and preprocessing, specifically includes: An integrated PET / MR device was used to simultaneously acquire three-dimensional high-resolution magnetic resonance imaging images of blood vessel walls and PET images of the head and neck. Intensity normalization was performed on three-dimensional high-resolution magnetic resonance imaging of blood vessel walls, attenuation correction and standardized uptake value (SUV) calculation were performed on PET images, and a standardized metabolic parameter map was generated. By leveraging the inherent hardware coordinate synchronization of integrated PET / MR equipment, spatial positioning information in the DICOM file headers of three-dimensional high-resolution magnetic resonance vascular wall imaging images and PET images is read. Through rigid transformation based on the DICOM coordinate system, the two modal images are aligned to achieve registration preprocessing.

3. The high-resolution magnetic resonance imaging blood vessel wall image segmentation method according to claim 1, characterized in that, The three-dimensional blood vessel segmentation model is a 3D V-Net blood vessel segmentation model.

4. The high-resolution magnetic resonance imaging blood vessel wall image segmentation method according to claim 1, characterized in that, In step S3, the initial centerline is input to the multimodal centerline correction unit to obtain the corrected centerline point set, specifically including: S301. Initialize multimodal parameters: Define the spherical search radius R, and set the metabolic activity threshold T according to the characteristics of PET images; S302, Point-by-point traversal and multimodal collaborative decision-making, specifically including: S3021. Initialize the centerline point set C_corrected to an empty list; S3022. For each point p_i in the initial centerline point set C_initial, execute the following loop: a) Simultaneous extraction of dual-modal features: Centered on point p_i, within a spherical neighborhood N(p_i) of radius R, the following operations are performed synchronously and in parallel: extracting local image patches from the pre-registered 3D high-resolution magnetic resonance vascular wall imaging image I_MR and calculating its gray-level gradient direction g_MR as an anatomical space guide; and extracting the uptake values ​​SUV of all voxels from the pre-registered PET image I_PET as a functional metabolic distribution. b) Joint assessment of metabolic-anatomical consistency: Read the uptake value SUV SUV(p_i) of point p_i on the pre-registered PET image I_PET. If SUV(p_i) > T, trigger the depth correction mechanism based on multimodal information. Otherwise, it is considered that the current point is located in a reliable vascular lumen at both the metabolic and anatomical levels, and it is directly added to the centerline point set C_corrected and then jumps to the next point. c) Calculation of the correction vector for multimodal information fusion, specifically including: Finding metabolic consensus regions: Within the neighborhood N(p_i), find all voxel locations that satisfy SUV(v) < T, forming a low-metabolic candidate point set V_low, which is used to characterize credible vascular lumen regions at the metabolic level; Calculate the fusion correction direction: Calculate the combined correction vector d_correction of fusion with PET metabolic attraction and MR anatomical guidance, the expression is: d_correction = λ * d_PET + (1 - λ) * g_MR, Where: d_PET is the vector pointing from point p_i to the center of the low-metabolism candidate point set V_low, used to characterize the spatial constraints and guidance of metabolic information; g_MR is the local gray-level gradient direction of the I_MR image, used to characterize the boundary constraints of anatomical structures; λ is a dynamic weighting factor, which is adjusted according to the degree to which SUV(p_i) exceeds the metabolic activity threshold T. The more significant the metabolic abnormality, the higher the weight λ of the PET image information. d) Dynamic step size correction and update: The correction step size α is positively correlated with (SUV(p_i) - T), and is dynamically adjusted. Then, the coordinates of the new point are calculated: p_corrected = p_i + α * normalize(d_correction), and the new point p_corrected is added to the centerline point set C_corrected. S303, Path Optimization: Smooth the centerline point set C_corrected obtained from the initial correction using spline curves; S304. Return result: Output the final corrected set of centerline points C_corrected, which is optimal in both anatomical and metabolic aspects.

5. The high-resolution magnetic resonance imaging blood vessel wall image segmentation method according to claim 4, characterized in that, The metabolic activity threshold T is set by the following expression: T = μ_background + k * σ_background, Where μ_background and σ_background are the mean and standard deviation of the image background region capture value SUV, respectively, and k is an empirical coefficient.

6. The high-resolution magnetic resonance imaging blood vessel wall image segmentation method according to claim 1, characterized in that, In step S4, a curved plane reconstruction is performed based on the corrected centerline point set. The resulting continuous two-dimensional curved surface reconstruction image is then input into the blood vessel wall segmentation model to obtain the blood vessel segmentation result, which specifically includes: Using the corrected centerline point set, curve plane reconstruction is automatically performed, and MR-CPR and PET-CPR image sequences with completely consistent spatial positions are generated simultaneously. The MR-CPR and PET-CPR image sequences with identical spatial locations are input into a vessel wall segmentation model based on an improved 2D U-Net to obtain vessel segmentation results.

7. The high-resolution magnetic resonance imaging blood vessel wall image segmentation method according to claim 6, characterized in that, The blood vessel wall segmentation model based on the improved 2D U-Net is specifically as follows: Input: Multi-channel input is used, corresponding to MR-CPR and PET-CPR image sequences with completely consistent spatial locations; Improved 2D U-Net network structure: In the encoder part of U-Net, after initial convolution of the dual-channel input, feature concatenation is performed to fuse low-level features of different modalities; Training objective: To simultaneously learn the anatomical features of MR images and the metabolic function features of PET images. During training, high metabolic signals of plaques in PET images are used as auxiliary features. When the boundaries of MR images are blurred, the boundaries suggested by metabolically abnormal areas in PET images are given priority.

8. A high-resolution magnetic resonance imaging system for analyzing blood vessel walls, characterized in that, The system includes: The segmentation module performs segmentation according to the high-resolution magnetic resonance blood vessel wall image segmentation method described in any one of claims 1 to 7 to obtain blood vessel segmentation results; The automatic plaque vulnerability assessment module is used to extract metabolic parameters of plaque regions from pre-registered PET images and determine the risk category of plaque regions through a multi-parameter fusion comprehensive vulnerability scoring model. The multi-modal parameters based on the comprehensive vulnerability scoring model include morphological parameters and functional parameters. An interpretable model is used to output a comprehensive vulnerability probability score between 0 and 1, and risk categories are classified according to the score threshold. The generation module is used to visualize multimodal image data of the head and neck, as well as the corresponding quantitative parameters and risk categories of the plaque areas output by the automatic plaque vulnerability assessment module.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images