Automatic naming method and system for head and neck blood vessels based on CTA image

Abstract

The invention belongs to the field of blood vessel naming, and provides a head and neck blood vessel automatic naming method and system based on a CTA image, and the method comprises the steps: obtaining a head and neck CTA image, carrying out the segmentation through a pre-trained blood vessel segmentation model, generating a blood vessel mask image, and generating multi-dimensional point cloud data according to the blood vessel mask image; classifying each point by using a pre-trained head and neck branch point cloud segmentation model to obtain a point cloud classification result; constructing a feature vector based on the multi-dimensional point cloud data and the point cloud classification result, performing classification decision by using a pre-trained decision tree model, and determining a point cloud naming tag; and constructing a comprehensive feature vector according to the multi-dimensional point cloud data and the point cloud naming tag, clustering the comprehensive feature vector by adopting a constraint clustering algorithm, and correcting the point cloud naming tag according to a clustering result to obtain a head and neck vessel name. According to the method, local geometric features and spatial information of the point cloud are comprehensively considered as multi-dimensional point cloud data to correct the naming tag, and the classification precision of vessel naming is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of blood vessel naming, and particularly relates to a method and system for automatically naming head and neck blood vessels based on CTA images. Background Art

[0002] The statements in this part merely provide background technical information related to the present invention and do not necessarily constitute prior art.

[0003] With the development of computed tomography technology and angiography, CT angiography (CTA) has become an important tool for diagnosing and treating cardiovascular diseases. It can provide high-resolution three-dimensional blood vessel structure information to help doctors perform lesion detection, surgical planning, and efficacy evaluation. However, the complexity of the head and neck blood vessel structure poses challenges to the interpretation of CTA images. There are numerous blood vessel branches in this area, and their shapes are diverse. Traditional manual annotation and naming methods are not only extremely time-consuming but also prone to errors, especially in cases with dense blood vessel branches or complex lesions. Such limitations not only affect the accuracy of diagnosis but also limit the efficiency of clinical decision-making. Therefore, developing a method and system for automatically naming head and neck blood vessels based on CTA images and improving accuracy aims to improve the accuracy and efficiency of diagnosing cardiovascular diseases through technological innovation, reduce the dependence on manual annotation, and promote the intelligent development of medical imaging technology. This research not only has important clinical significance but also provides a new direction for the future development of medical imaging technology.

[0004] In the prior art, for blood vessel naming, deep learning is mostly used for image segmentation and then point cloud features are used for fine classification. However, when using point clouds for classification, it ignores the influence of spatial position information and geometric features on classification accuracy, resulting in a decrease in the accuracy of blood vessel classification and naming. Summary of the Invention

[0005] To solve the above problems, the present invention proposes a method and system for automatically naming head and neck blood vessels based on CTA images. The present invention adopts a pre-trained decision tree model introducing a new loss function. This model comprehensively considers the probability output of point cloud segmentation and multi-dimensional point cloud data, and through a dynamic adjustment mechanism, obtains the preliminary naming labels of blood vessels and the K-nearest neighbor labels of each point to enhance the comprehensiveness and relevance of label information. Finally, a clustering algorithm with constraints is introduced as an error correction mechanism. This algorithm can accurately identify and correct potential naming errors by re-evaluating the similarity between point clouds, ensuring that the accuracy and consistency of blood vessel naming reach a new level.

[0006] According to some embodiments, the first aspect of the present invention provides a method for automatically naming head and neck blood vessels based on CTA images, adopting the following technical solution:

[0007] A method for automatically naming head and neck blood vessels based on CTA images, comprising:

[0008] Obtain head and neck CTA images and perform preprocessing, use a pre-trained blood vessel segmentation model to perform segmentation to generate a blood vessel mask image, and generate multi-dimensional point cloud data based on the blood vessel mask image;

[0009] Use a pre-trained point cloud segmentation model for each branch of the head and neck to classify each point in the multi-dimensional point cloud data to obtain a point cloud classification result;

[0010] Construct a feature vector based on the multi-dimensional point cloud data and the point cloud classification result, use a pre-trained decision tree model to perform classification decision on the feature vector, and determine the point cloud naming label;

[0011] Construct a comprehensive feature vector based on the multi-dimensional point cloud data and the point cloud naming label, use a constrained clustering algorithm to cluster the comprehensive feature vector, and correct the point cloud naming label according to the clustering result to obtain the head and neck blood vessel naming.

[0012] Further, the multi-dimensional point cloud data includes point cloud local geometric features and point cloud spatial position information;

[0013] The point cloud local geometric features include normal vectors, plane curvatures, and connected domain features; wherein, the connected domain feature is obtained by performing connected domain analysis on the blood vessel mask image to identify and label all points belonging to the same blood vessel branch;

[0014] For each point in the point cloud, fit through its K nearest neighboring points, calculate the K-nearest neighbor normal and curvature of the point as the normal vector and plane curvature;

[0015] The point cloud spatial position information is the spatial coordinates of the point cloud after normalization processing.

[0016] Further, the use of a pre-trained point cloud segmentation model for each branch of the head and neck to classify each point in the multi-dimensional point cloud data to obtain a point cloud classification result is specifically:

[0017] Use a pre-trained point cloud segmentation model for each branch of the head and neck to classify each point;

[0018] Obtain the probability distribution of each point belonging to each category to generate the point cloud classification result of each point.

[0019] Further, the construction of a feature vector based on the multi-dimensional point cloud data and the point cloud classification result, and the use of a pre-trained decision tree model to perform classification decision on the feature vector to determine the point cloud naming label is specifically:

[0020] Construct a feature vector based on multi-dimensional point cloud data and the point cloud classification results;

[0021] According to the feature vector, starting from the root node of the decision tree model, at each node, compare each eigenvalue in the feature vector with the splitting threshold in turn based on a preset splitting rule;

[0022] According to the comparison results, transfer the point cloud data down along the branches of the decision tree, continuously repeat the splitting process until reaching the leaf node, and use the class label of the leaf node as the preliminary naming label of the point cloud data;

[0023] Use the naming information of the K nearest points around each point as the K-nearest neighbor label of each point, and form the point cloud naming label by combining the K-nearest neighbor label and the preliminary naming label of each point.

[0024] Further, construct a comprehensive feature vector according to the multi-dimensional point cloud data and the point cloud naming label, use a constrained clustering algorithm to cluster the comprehensive feature vector, and correct the point cloud naming label according to the clustering results to obtain the head and neck blood vessel naming, specifically:

[0025] Construct a comprehensive feature vector from the multi-dimensional point cloud data and the point cloud naming label;

[0026] Use a constrained clustering algorithm to construct an objective function containing label information, and cluster the comprehensive feature vector based on the objective function;

[0027] Judge according to the consistency of the point cloud naming label information within each cluster in the clustering results. If there are points with inconsistent naming within the cluster, correct them according to the point cloud label information and the comprehensive feature vector of the surrounding points of the points with inconsistent naming to obtain the head and neck blood vessel naming.

[0028] Further, the correction according to the point cloud label information and the comprehensive feature vector of the surrounding points of the points with inconsistent naming is specifically:

[0029] For a point p with inconsistent naming i , based on the comprehensive feature vector, use a weighted distance metric method to calculate the similarity between the point with inconsistent naming and other points in the cluster;

[0030] Find several points most similar to point p according to the similarity results i , and determine the new naming label of point p through multi-point voting on the point cloud naming label information of these points i .

[0031] According to some embodiments, the second solution of the present invention provides a head and neck blood vessel automatic naming system based on CTA images, and adopts the following technical solutions:

[0032] A head and neck blood vessel automatic naming system based on CTA images, comprising:

[0033] A blood vessel segmentation unit, configured to obtain head and neck CTA images, segment them using a pre-trained blood vessel segmentation model to generate a blood vessel mask image, and generate multi-dimensional point cloud data based on the blood vessel mask image;

[0034] A point cloud segmentation unit, configured to classify each point in the multi-dimensional point cloud data using a pre-trained head and neck branch point cloud segmentation model to obtain a point cloud classification result;

[0035] A decision tree naming unit, configured to construct a feature vector based on the multi-dimensional point cloud data and the point cloud classification result, and classify and make a decision on the feature vector using a pre-trained decision tree model to determine a point cloud naming label;

[0036] A clustering optimization unit, configured to construct a comprehensive feature vector based on the multi-dimensional point cloud data and the point cloud naming label, perform clustering on the comprehensive feature vector using a constrained clustering algorithm, and correct the point cloud naming label according to the clustering result to obtain the head and neck blood vessel naming.

[0037] According to some embodiments, the third aspect of the present invention provides a computer-readable storage medium.

[0038] A computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements the steps in a method for automatically naming head and neck blood vessels based on CTA images as described in the first aspect above.

[0039] According to some embodiments, the fourth aspect of the present invention provides a computer device.

[0040] A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, it implements the steps in a method for automatically naming head and neck blood vessels based on CTA images as described in the first aspect above.

[0041] According to some embodiments, the fifth aspect of the present invention provides a computer program product or a computer program.

[0042] The present invention provides a computer program product or a computer program, which includes computer instructions stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the steps in a method for automatically naming head and neck blood vessels based on CTA images as described in the first aspect above.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] The present invention proposes multi-modal feature fusion decision tree classification. During the decision tree classification process, it not only fuses the probability output of point cloud segmentation, but also innovatively combines spatial position information and various local geometric features (normal vector, plane curvature, connected domain feature) to construct a feature vector, and creates a new loss function to guide classification. This multi-modal feature fusion method enables the decision tree model to more comprehensively capture the features of point cloud data and improve the classification accuracy of complex vascular structures. Compared with traditional decision tree classification methods based only on single features or simple feature combinations, this method can better handle complex situations such as dense vascular branches and diverse morphologies.

[0045] The present invention proposes an adaptive optimization of the constrained clustering algorithm. In the clustering analysis correction stage, by designing an objective function containing label information and adaptively adjusting the weights according to the characteristics of the actual data, the clustering algorithm can flexibly optimize the clustering process based on the preliminary naming results and point cloud features. At the same time, it can effectively utilize the preliminary naming label information as prior knowledge to guide the clustering direction and improve the efficiency and accuracy of correcting misnaming. In addition, introducing context information such as anatomical standards and historical data further enhances the adaptability and reliability of the clustering algorithm in dealing with complex vascular structures and special situations.

[0046] The present invention can ensure the robustness and reliability of the naming results to meet high-standard requirements. During the vascular naming process, strict quality control measures are implemented to ensure the robustness and reliability of the naming results. This meets the high-standard requirements for accurate anatomical information in the clinical and research fields, providing reliable and accurate anatomical information for clinical diagnosis and research; doctors can more confidently rely on these automatically generated naming results, thus making more confident diagnostic and treatment decisions and improving patient satisfaction and treatment effects.

[0047] The present invention integrates advanced segmentation models, point cloud processing technologies, and various machine learning algorithms, achieving a breakthrough in the field of vascular naming. This method has not only achieved remarkable results in improving the accuracy and consistency of naming, but also ensured the robustness and reliability of the naming results through strict quality control, meeting the high-standard requirements for accurate anatomical information in the clinical and research fields. The high-quality vascular naming results generated by this solution provide reliable and accurate anatomical information for clinical diagnosis and research, helping doctors more accurately evaluate vascular lesions, optimize treatment plans, and improve the overall medical quality and patient satisfaction; the application value and market prospect of this solution cannot be underestimated, and it is expected to play an important role in the treatment and diagnosis of head and neck vascular diseases, promoting the continuous progress of medical technology. Brief Description of the Drawings

[0048] The accompanying drawings forming a part of this invention are used to provide a further understanding of the invention. The schematic embodiments of the invention and their descriptions are used to explain the invention and do not unduly limit the invention.

[0049] Figure 1 is a flowchart of an automatic naming method for head and neck blood vessels based on CTA images in the first embodiment of the present invention;

[0050] Figure 2 is a comparison diagram of the segmentation results of traditional blood vessel segmentation + point cloud segmentation and the segmentation scheme of this embodiment in the first embodiment of the present invention. Among them, Figure 2 (a) in is the result diagram of traditional blood vessel segmentation + point cloud segmentation, Figure 2 (b) in is the segmentation result diagram of the scheme of this embodiment;

[0051] Figure 3 is a structural diagram of an automatic naming system for head and neck blood vessels based on CTA images in the second embodiment of the present invention. Detailed implementation manners

[0052] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0053] It should be noted that the following detailed descriptions are all illustrative and are intended to provide further descriptions of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0054] It should be noted that the terms used herein are only for describing specific implementation manners and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular forms are also intended to include the plural forms. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0055] In the case of no conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0056] Embodiment 1

[0057] This embodiment provides a method for automatically naming head and neck blood vessels based on CTA images. In this embodiment, taking the application of this method to a server as an example, it can be understood that this method can also be applied to a terminal, or to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, web servers, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited thereto. The terminal and the server can be directly or indirectly connected through wired or wireless communication methods, and this application does not limit this. In this embodiment, the method includes the following steps:

[0058] Obtain a head and neck CTA image and perform preprocessing, use a pre-trained blood vessel segmentation model to perform segmentation to generate a blood vessel mask image, and generate multi-dimensional point cloud data based on the blood vessel mask image;

[0059] Use a pre-trained point cloud segmentation model for each branch of the head and neck to classify each point in the multi-dimensional point cloud data to obtain a point cloud classification result;

[0060] Construct a feature vector based on the multi-dimensional point cloud data and the point cloud classification result, use a pre-trained decision tree model to perform classification decision on the feature vector, and determine the point cloud naming label;

[0061] Construct a comprehensive feature vector based on the multi-dimensional point cloud data and the point cloud naming label, use a constrained clustering algorithm to cluster the comprehensive feature vector, and correct the point cloud naming label according to the clustering result to obtain the head and neck blood vessel naming.

[0062] As Figure 1 shown, the method described in this embodiment specifically includes:

[0063] First, perform blood vessel segmentation on the head and neck CTA image based on a pre-trained blood vessel segmentation model to generate a high-quality blood vessel mask; then, generate multi-dimensional point cloud data based on this blood vessel mask. Next, based on a pre-trained point cloud segmentation model for each branch of the head and neck, classify each point and retain the probability distribution of all categories to provide richer information. Subsequently, based on a pre-trained decision tree model, combine the probability output of point cloud segmentation and multi-dimensional point cloud data to dynamically adjust the final category of each point and output point cloud naming label information. Finally, introduce a clustering algorithm as an error correction mechanism to re-evaluate the similarity between point clouds, correct naming errors, and further improve the accuracy and consistency of blood vessel naming.

[0064] Step S101, precise blood vessel segmentation and high-quality point cloud generation: Based on a blood vessel segmentation model trained with a large amount of data, the head and neck CTA images are deeply processed. This model can accurately identify and extract the blood vessel structures in the images, generate a high-quality blood vessel mask, and further obtain multi-dimensional point cloud data. The blood vessel mask is crucial for subsequent processing, ensuring the integrity and clarity of the blood vessel structure; each point not only retains its precise spatial position information but also additional rich local geometric features.

[0065] Step S102, detailed point cloud classification and probability distribution retention: Based on a pre-trained point cloud segmentation model for each branch of the head and neck, each point in the point cloud is classified, and the probability distributions of all categories are retained, which can provide additional information support for the subsequent decision tree dynamic adjustment mechanism. An example of the blood vessel naming result generated at this step is shown in Figure 2 (a).

[0066] Step S103, decision tree dynamic adjustment and naming label output: Based on the pre-trained decision tree model, on the basis of Step S102, the point cloud is classified and adjusted. The decision tree model comprehensively considers the spatial position, geometric features, and classification probability distribution of the point cloud. Through a series of rules and conditional judgments, the final category of each point is dynamically adjusted, and the point cloud naming label information is output, including the preliminary naming label of the blood vessel and the K-nearest neighbor label of each point. The introduction of the decision tree model makes the classification process more intelligent and adaptive. It can flexibly adjust the classification strategy according to the actual situation of the point cloud data, thus generating more accurate and consistent point cloud naming label information. At the same time, the decision tree model can also handle complex and ambiguous situations, improving the robustness and reliability of naming;

[0067] Step S104, clustering algorithm error correction and high-quality naming result generation: The constrained clustering algorithm is introduced as an error correction mechanism to further optimize the point cloud naming label information output in Step S103. This algorithm re-evaluates the similarity between point clouds, detects, and corrects possible naming errors. The application of the constrained clustering algorithm makes the naming result more accurate and consistent. It can identify and correct naming inconsistencies caused by classification errors or noise interference. At the same time, the constrained clustering algorithm can also consider additional context information such as anatomical knowledge and historical data to further improve the reliability and applicability of naming. Through this step, a high-quality blood vessel naming result is finally generated, providing reliable anatomical information for clinical diagnosis and research. An example of its blood vessel naming result is shown in Figure 2 (b).

[0068] In this embodiment, the vascular model, the head and neck branch point cloud segmentation model, and the decision tree model can be pre-trained. Based on the trained vascular segmentation model, the vascular region is extracted from the head and neck vascular CTA image. Based on the trained head and neck branch point cloud segmentation model, the probability distribution of each point is obtained from the vascular region. Based on the trained decision tree model, the point cloud naming label information is obtained. Among them, the vascular segmentation model can be but is not limited to common segmentation networks such as nnUnet, the head and neck branch point cloud segmentation model can be but is not limited to common point cloud segmentation networks such as PointNet, and the decision tree model can be but is not limited to models such as LightGBM. On the basis of the solution of this embodiment, the network model needed can be selected.

[0069] In this embodiment, the head and neck CTA images carefully labeled by professionals are used as training sample images to ensure that each vascular region and its branches have clear class labels. This step not only provides high-quality training data but also lays a solid foundation for subsequent model training.

[0070] Among them, the trained vascular segmentation model is obtained by the following steps:

[0071] Data input and integration: The head and neck CTA image after manually labeling the vascular region is input into the pre-constructed vascular segmentation model. During this process, the corresponding manually labeled tags are integrated into a single vascular tag so that the model can accurately identify and segment the vascular region.

[0072] Model parameter adjustment: Based on the input data, the relevant parameters in the vascular segmentation model are carefully adjusted to ensure that the model can accurately capture the characteristics of the blood vessels.

[0073] Cross-validation and hyperparameter optimization: In order to further improve the performance of the model, the cross-validation method is used to verify the model, and through hyperparameter optimization technology, the hyperparameters of the model are finely adjusted, so as to obtain a vascular segmentation model with better performance. This model can accurately segment the vascular region and provide reliable input for subsequent point cloud processing and naming.

[0074] Among them, the pre-trained head and neck branch point cloud segmentation model is obtained by the following steps:

[0075] Data input: The manually labeled head and neck branch vascular regions (i.e., manually labeled tag data) are input into the pre-constructed head and neck branch point cloud segmentation model.

[0076] Model parameter adjustment: Based on the input data, the relevant parameters in the head and neck branch point cloud segmentation model are carefully adjusted to ensure that the model can accurately identify and segment each branch blood vessel.

[0077] Cross-validation and hyperparameter optimization: Similarly, the cross-validation method is used to validate the model, and the hyperparameters of the model are finely tuned through hyperparameter optimization techniques, so as to obtain a head and neck branch point cloud segmentation model with better performance. This model can accurately segment the blood vessels of each branch of the head and neck, providing rich information for subsequent point cloud classification and naming.

[0078] Among them, the pre-trained decision tree model is obtained by the following steps:

[0079] Feature extraction: The manually labeled blood vessel regions of each branch of the head and neck are input into the pre-trained head and neck branch point cloud segmentation model to extract the probability distribution result p of each point i , spatial position information x i and local geometric features (such as normal vector n i , plane curvature k i and connected domain feature c i ). These features are combined into a feature vector f i , which is used for subsequent model training.

[0080] Design a new loss function formula:

[0081]

[0082] Among them, N is the number of sample points, C is the total number of categories, y ij is the true class label of the sample point (if the sample q i belongs to the j-th class, then y ij = 1, otherwise y ij = 0), is the probability that the sample point q i is predicted to be the j-th class (from the probability output p of the point cloud segmentation i ).

[0083] The term in the formula is used to penalize the degree of drastic change of the normal vector. represents the gradient of the normal vector, calculate its norm, and α is a tuning coefficient used to control the importance of normal vector smoothness in the loss function.

[0084] The i term penalizes the plane curvature. The plane curvature k

[0085] This item considers the consistency of the connected component features. c i is the connected component feature of the sample point, is the average value of the connected component features of its surrounding neighborhood points (for example, the average value of the connected component features of the surrounding K neighborhood points can be taken). γ is a regulation coefficient. This item encourages the connected component feature of the sample point to be consistent with its neighborhood points to ensure that, during the blood vessel naming process, the points belonging to the same blood vessel branch have consistent connected component features and enhance the consistency of naming.

[0086] Model training and loss function optimization: For the regulation coefficients α, β, and γ, they can be tuned through methods such as cross-validation. The training data is divided into a training set, a validation set, and a test set. The model is trained on the training set using different combinations of coefficients, and the model performance (such as accuracy, F1 value, etc.) is evaluated on the validation set. The coefficient combination with the best performance is selected. Then, a final test is conducted on the test set to ensure the effectiveness of the model in practical applications.

[0087] During the training process, continuously monitor the change of the loss function and the performance of the model on the validation set. If it is found that the loss function no longer decreases or the model performance no longer improves, the training can be stopped in advance to prevent overfitting. At the same time, according to the actual situation, the regulation coefficients can be adjusted in a timely manner or the loss function can be further improved to adapt to different data characteristics and task requirements.

[0088] Cross-validation and hyperparameter optimization: Finally, the cross-validation method is used to validate the model, and the hyperparameters of the model are finely adjusted through hyperparameter optimization techniques to obtain a decision tree model with better performance. This model can comprehensively consider various features of the point cloud and achieve accurate classification and naming.

[0089] In this embodiment, after obtaining the pre-trained blood vessel segmentation model, the head and neck CTA image of a certain patient can be recognized based on this blood vessel segmentation model to obtain the blood vessel region. Further, multi-dimensional point cloud data is generated based on this blood vessel mask. The spatial position features of the blood vessel region point cloud are input into the trained head and neck branch point cloud segmentation model, and the probability distribution of all categories of each point is retained. At this time, the above process can be implemented by the following steps:

[0090] S1011, In the blood vessel region obtained based on the blood vessel segmentation model, extract the spatial coordinates of each point in the blood vessel region and normalize them as the spatial position information.

[0091] In point cloud data processing, the normal vector is obtained by fitting the K nearest points around each point. Assume that there is a point q in the point cloud data i , whose coordinate is x i , in this point cloud, q iThe K nearest neighbor points in the vicinity are {x i1 , x i2 , ……, x iK}. Then, for the point q i , the normal vector n i can be obtained through principal component analysis (PCA).

[0092] n i = PCA({x i1 - x i , x i2 - x i , ……, x iK - x i})

[0093] The planar curvature is estimated by fitting a quadratic surface to the local point cloud. Assume that the local point cloud conforms to a quadratic polynomial z = ax 2 + by 2 + cxy + d x + ey + f. Then, the curvature k i can be calculated through the derivative of this polynomial.

[0094]

[0095] The connected component feature is obtained by performing a connected component analysis on the blood vessel region. Assume a blood vessel mask M, where M(i, j, k) = 1 indicates that the voxel belongs to the blood vessel, and M(i, j, k) = 0 indicates that it does not belong to the blood vessel (i.e., the background). The connected component matrix L is a matrix of the same size as M, used to store the connected component feature id to which each voxel belongs. Traverse the mask M. For each unlabeled blood vessel voxel (i, j, k), start a breadth-first search (BFS) or depth-first search (DFS) to label all adjacent blood vessel voxels with the same id.

[0096]

[0097] Among them, id is the unique identifier of the current connected component, that is, the connected component feature, ensuring that all voxels within the connected component share the same connected component feature.

[0098] For each voxel (i, j, k), its connected component feature c i is the value of L(i, j, k).

[0099] c i = L(i, j, k)

[0100] S1021. Based on the head and neck branch point cloud segmentation model, for each point qi , the probability distribution p output by the model i contains the probabilities of all classes:

[0101] p i = [p i1 , p i2 ,..., p iC

[0102] where C is the total number of classes, and p in represents the probability that the point q i belongs to the nth class, where n = 1, 2,..., C.

[0103] In this embodiment, based on the pre-trained decision tree model, combined with the probability output of point cloud segmentation, spatial position, and geometric features, the final class of each point is dynamically adjusted to output the point cloud naming label information. At this time, the above process can be achieved through the following steps:

[0104] S1031, the input of the decision tree model is to combine the probability output p i of point cloud segmentation, the spatial position information x i and the geometric features (n i , k i , c i ) to construct the feature vector f i :

[0105] f i = [p i , x i , n i , k i , c i

[0106] S1032, the decision tree model makes a classification decision starting from the root node according to the constructed feature vector f i . The root node is the starting point of the decision tree, which contains all the point cloud data to be classified. At each node, the decision tree model compares each feature value in the feature vector f i with the splitting threshold according to the pre-set splitting rules. The splitting rules can be determined based on metrics such as information gain and Gini index. For example, for information gain, calculate the information gain of each feature at the current node, select the feature with the largest information gain as the splitting feature, and determine the best splitting value of this feature as the splitting threshold.

[0107] S1033, according to the comparison results, the point cloud data is passed down along the branches of the decision tree, and the above splitting process is continuously repeated until the leaf node is reached. The leaf node is the end node of the decision tree, and it no longer performs splitting operations.

[0108] ​​S1034. The class label on the leaf node is the preliminary naming label of the point cloud data. This class label is determined according to the true labels and feature distributions of the training data during the decision tree training process. For example, if in the training data, most of the point cloud data with similar features are labeled as "left common carotid artery", then when the point cloud data to be classified reaches the corresponding leaf node, its preliminary naming label is determined as "left common carotid artery".

[0109] S1035. During the classification process, the decision tree model also outputs the K-nearest neighbor labels of each point, that is, introducing the naming information of the K nearest points around it. The K-nearest neighbor labels can strengthen the local consistency and help subsequent clustering algorithms correct the misclassification of isolated points. For example, if the preliminary naming label of a point is different from the nearest neighbor labels of most of the surrounding points, then in the subsequent clustering algorithm, this point may be regarded as an outlier and its naming label needs to be further checked and corrected.

[0110] In this embodiment, finally, a constrained clustering algorithm is introduced as an error correction mechanism. Based on the point cloud naming label information, the similarity between point clouds is re-evaluated, naming errors are corrected, and the accuracy and consistency of blood vessel naming are further improved. At this time, the above process can be achieved through the following steps:

[0111] S1041. The constrained clustering algorithm takes the point cloud naming label information as a feature input to guide the generation of the final label. This approach can help the clustering algorithm better understand the structure of the data and, in some cases, improve the classification accuracy. By combining the spatial position information, geometric features, and the point cloud naming label l i , constructing a comprehensive feature vector g i , can make the clustering results more reasonable and consistent.

[0112] g i = [x i , n i , k i , l i

[0113] This construction method enables the clustering algorithm to comprehensively consider various information of the point cloud, better understand the data structure, and improve the classification accuracy.

[0114] S1042. Design a constrained clustering, adjust the weights in the objective function, so that the clustering results tend to retain the consistency of the preliminary naming labels and correct those results that are significantly inconsistent with the preliminary naming labels. For the constrained clustering, an objective function J containing label information can be defined:

[0115]

[0116] Among them, E is the set of clusters, d(qi , q j ) is the point q i and q j The distance between them, λ is the balance coefficient that controls the influence degree of label constraints, S is the set of all point pairs, is the weight of the point pair (q i , q j ), which can be set according to the preliminary named labels. I(·) is the indicator function that returns 1 when the condition holds and 0 otherwise.

[0117] The objective function J defined here serves both to judge the classification and clustering methods and to perform clustering based on specific principles. Its main purposes are as follows:

[0118] 1. Guide the clustering process: The objective function J is the core basis for performing constrained clustering. It is defined based on label information and the distance between points. During the clustering process, the algorithm adjusts the clustering method according to this function. Among them, d(q i , q j ) in the formula is the distance between point q i and q j . Through this distance metric, the algorithm can measure the similarity between points and thus judge which points are more suitable to be clustered together. This determines the clustering principle based on distance similarity, and points with closer distances are more likely to be assigned to the same cluster.

[0119] 2. Correct the named labels: λ in the function is the balance coefficient that controls the influence degree of label constraints; is the weight of the point pair (q i , q j ), which is set according to the preliminary named labels; I(·) is the indicator function. Through these parameters, the objective function J incorporates the preliminary named label information into the clustering process. When there are obvious inconsistencies with the preliminary named labels in the clustering results, the algorithm adjusts according to the objective function J. If the named label of a point is inconsistent with the named labels of most surrounding points, the objective function will tend to reclassify this point into the cluster with the same named label as the surrounding points by adjusting the weights and calculating distances, etc., thus correcting those results that are significantly inconsistent with the preliminary named labels and making the clustering results retain the consistency of the preliminary named labels.

[0120] S1043, after the clustering algorithm converges, the final clustering result is obtained, that is, the data points are divided into K clusters. For each cluster, check the naming consistency of the points within the cluster. If a point with a naming inconsistent with other points (judged according to the point cloud naming label l i ) is found, then correct its naming label according to the naming information of the surrounding points and the comprehensive feature vector.

[0121] Suppose that in the processing of head and neck vascular CTA images, after the clustering algorithm converges, there is a cluster containing 20 points.

[0122] 1. Check naming consistency: In this cluster, the naming labels of 16 points are "right vertebral artery", and the naming labels of the other 4 points are "left vertebral artery". It is initially judged that these 4 points with the label "left vertebral artery" may be points with inconsistent naming.

[0123] 2. Determine points with inconsistent naming: For a point p with inconsistent naming i , based on the comprehensive feature vector, use the weighted distance metric method to calculate the similarity between the point with inconsistent naming and other points in the cluster;

[0124] Select one of the points with the label "left vertebral artery" as point p i , whose comprehensive feature vector is g i . Suppose that according to experience, the importance of spatial position information, normal vector, plane curvature, and naming label information is different, and weights of 0.3, 0.2, 0.2, and 0.3 are assigned respectively. Calculate the similarity between p i and the other 19 points in the cluster through the weighted distance metric formula. Suppose that the average similarity between p i and the 16 "right vertebral artery" points is 0.4, and the average similarity with the other 3 "left vertebral artery" points is 0.8. Set the similarity threshold to 0.6. The similarity between p i and the majority of points (16 "right vertebral artery" points) is lower than the threshold, so it is determined that p i is a point with inconsistent naming.

[0125] 3. Correct the naming label: According to the similarity result, find several points that are most similar to point p i , and determine the new naming label of point p i through multi-point voting on the naming label information of these points;

[0126] Set the number of the most similar points to be selected as 5. According to the similarity calculation, find the 5 points that are most similar to p i . Among them, the naming labels of 4 points are "right vertebral artery", and the naming label of 1 point is "left vertebral artery". Through multi-point voting, change the naming label of p i from "left vertebral artery" to "right vertebral artery".

[0127] 4. Process special regions: Suppose the region where this cluster is located is in a relatively blurred position in the vascular image, and there is still a certain degree of uncertainty in the naming determined only by the above steps. At this time, referring to the anatomical atlas, it is found that this region should belong to the branch region of the "right vertebral artery" under normal circumstances. Combining with the historical data of dealing with similar cases before, it is further confirmed that p iIt is reasonable to name it as the "right vertebral artery". At the same time, record information such as the weights, thresholds, and processing methods for special regions determined during this processing, and feedback it to the algorithm. For example, when processing similar regions next time, appropriately increase the weight of spatial position information in the comprehensive feature vector, or adjust the threshold of the clustering algorithm, so that the algorithm is more accurate and stable when processing similar situations in the future.

[0128] Combined with the above description, in the embodiment of the present application, through the head and neck vascular CTA image, based on the vascular segmentation model, the mask of the vascular region is extracted, and the preliminary naming of the blood vessels is generated based on the head and neck branch point cloud segmentation model and the decision tree. Finally, the constrained clustering algorithm is introduced to correct the preliminary naming label to obtain a high-precision vascular naming result. The high-quality vascular naming result generated by this solution not only significantly improves the efficiency and quality of naming, but also provides reliable anatomical support for clinical diagnosis and research, helps doctors more accurately evaluate the vascular health status, and thus improves the accuracy and credibility of diagnosis.

[0129] Based on the pre-trained high-precision vascular segmentation model, this embodiment performs a detailed vascular segmentation task on the head and neck CTA image to generate a high-quality vascular mask. Based on this vascular mask, multi-dimensional point cloud data (spatial position information, normal vector, plane curvature, and connected domain features) is further generated. Based on the pre-trained head and neck vascular branch point cloud segmentation model, the class probability distribution of all point clouds is obtained. On this basis, a pre-trained decision tree model with a new loss function is adopted. This model comprehensively considers the probability output of point cloud segmentation and multi-dimensional point cloud data, and through a dynamic adjustment mechanism, obtains the preliminary naming label of the blood vessel and the K-nearest neighbor label of each point to enhance the comprehensiveness and relevance of the label information. Finally, a constrained clustering algorithm is introduced as an error correction mechanism. This algorithm can accurately identify and correct potential naming errors by re-evaluating the similarity between point clouds, ensuring that the accuracy and consistency of vascular naming reach a new height. The method and system proposed by the present invention are not only applicable to head and neck vascular CTA images of various complexities, showing strong generalization ability, but also the accurate vascular naming and structural information generated can be widely applied to multiple clinical fields such as early diagnosis of vascular diseases, surgical plan planning, and treatment effect evaluation, significantly improving the accuracy and efficiency of clinical work.

[0130] Embodiment 2

[0131] As Figure 3 shown, this embodiment provides a head and neck vascular automatic naming system based on CTA images, including:

[0132] A vascular segmentation unit, configured to obtain head and neck CTA images, perform segmentation using a pre-trained vascular segmentation model to generate a vascular mask image, and generate multi-dimensional point cloud data based on the vascular mask image; this unit is based on a deeply trained vascular segmentation model and can accurately extract a high-quality vascular mask from the CTA images of head and neck blood vessels. This mask not only retains the fine structure of the blood vessels but also greatly reduces noise interference. Subsequently, based on this high-quality vascular mask, the system further generates detailed and accurate multi-dimensional point cloud data, providing a solid foundation for subsequent steps.

[0133] A point cloud segmentation unit, configured to classify each point in the multi-dimensional point cloud data using a pre-trained point cloud segmentation model for each branch of the head and neck to obtain a point cloud classification result; this unit incorporates a point cloud segmentation model for each branch of the head and neck blood vessels that has been rigorously pre-trained. The acquisition process of this model is based on the manually and carefully labeled regions of each branch of the head and neck blood vessels (i.e., manually labeled tag data), and is gradually constructed through meticulous model parameter adjustment as well as cross-validation and hyperparameter optimization techniques. Its main function is to perform preliminary processing on the multi-dimensional point cloud data transmitted by the high-precision vascular segmentation unit, and can calculate the probability distribution of each point in the point cloud belonging to each category, thereby capturing the subtle category difference information in the point cloud data and providing key data support for the classification work of the subsequent decision tree model.

[0134] A decision tree naming unit, configured to construct a feature vector based on the multi-dimensional point cloud data and the point cloud classification result, and use a pre-trained decision tree model to perform classification decision on the feature vector to determine the point cloud naming label; this unit incorporates a decision tree model that is trained using the probability distribution results, spatial position information, and local geometric features of each point extracted by the point cloud segmentation model for each branch of the head and neck, and has also undergone a rigorous pre-training process. When receiving the multi-dimensional point cloud data and related feature information processed by the point cloud segmentation unit, the decision tree naming unit can, based on the decision tree model within it, comprehensively consider these input information, efficiently and accurately identify each blood vessel, assign an appropriate name to it, and finally generate detailed and accurate point cloud naming label information.

[0135] The clustering optimization unit is configured to construct a comprehensive feature vector based on multi-dimensional point cloud data and point cloud naming labels, perform clustering on the comprehensive feature vector using a constrained clustering algorithm, and correct the point cloud naming labels according to the clustering results to obtain head and neck blood vessel naming. To further improve the accuracy and reliability of blood vessel naming, the system introduces a constrained clustering algorithm. After receiving multi-dimensional point cloud data and preliminary point cloud naming label information, this algorithm can intelligently analyze and adjust the naming results to ensure that the finally output blood vessel naming not only conforms to medical standards but also closely matches the actual blood vessel structure. This step effectively reduces the misnaming rate and improves the overall performance of the system.

[0136] The examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed in the first embodiment above. It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0137] In the above embodiments, the descriptions of each embodiment have their own focuses. For parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0138] The proposed system can be implemented in other ways. For example, the above-described system embodiments are merely illustrative. For example, the above module division is only a logical function division. In actual implementation, there can be other division methods. For example, multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0139] Embodiment Three

[0140] This embodiment provides a computer-readable storage medium with a computer program stored thereon. When the program is executed by a processor, it implements the steps in a method for automatically naming head and neck blood vessels based on CTA images as described in the first embodiment above.

[0141] Embodiment Four

[0142] This embodiment provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps in a method for automatically naming head and neck blood vessels based on CTA images as described in the first embodiment above.

[0143] Embodiment Five

[0144] This embodiment provides a computer program product or a computer program. The computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the steps in the method for automatically naming head and neck blood vessels based on CTA images described in Embodiment 1 above.

[0145] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories and optical memories, etc.) containing computer-usable program code.

[0146] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in one Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0147] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the functions specified in one Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0148] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0149] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0150] Although the specific implementation manners of the present invention have been described above in conjunction with the accompanying drawings, it is not a limitation to the protection scope of the present invention. Those skilled in the art should understand that, based on the technical solutions of the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present invention.

Claims

1. A method for automatically naming head and neck blood vessels based on CTA images, characterized in that: include: Acquire head and neck CTA images and perform preprocessing, use the pre-trained vascular segmentation model to segment and generate vascular mask images, and generate multi-dimensional point cloud data based on the vascular mask images; Using the pre-trained point cloud segmentation model of each branch of the head and neck, each point in the multi-dimensional point cloud data is classified to obtain the point cloud classification result; Construct feature vectors based on multi-dimensional point cloud data and point cloud classification results, use pre-trained decision tree models to make classification decisions on feature vectors, and determine point cloud naming labels; A comprehensive feature vector is constructed based on the multidimensional point cloud data and point cloud naming labels. The constrained clustering algorithm is used to cluster the comprehensive feature vector. The point cloud naming labels are modified according to the clustering results to obtain the naming of head and neck blood vessels.

2. The method for automatically naming head and neck blood vessels based on CTA images as claimed in claim 1, characterized in that: The multi-dimensional point cloud data includes local geometric features of the point cloud and spatial position information of the point cloud; The local geometric features of the point cloud include normal vectors, plane curvatures, and connected domain features; wherein the connected domain features are obtained by performing connected domain analysis on the vascular mask image to identify and mark all points belonging to the same vascular branch; For each point in the point cloud, fit it through the K nearest points around it, and calculate the K nearest neighbor normals and curvatures of the point as the normal vector and plane curvature; The point cloud spatial position information is the spatial coordinates of the point cloud after normalization.

3. The method for automatically naming head and neck blood vessels based on CTA images as claimed in claim 1, characterized in that: The pre-trained head and neck branch point cloud segmentation model is used to classify each point in the multi-dimensional point cloud data to obtain the point cloud classification result, which is specifically: Use the pre-trained point cloud segmentation model of each branch of the head and neck to classify each point; The probability distribution of each point belonging to each category is obtained to generate the point cloud classification result of each point.

4. The method for automatically naming head and neck blood vessels based on CTA images as claimed in claim 1, characterized in that: The feature vector is constructed based on the multi-dimensional point cloud data and the point cloud classification result, and the pre-trained decision tree model is used to classify the feature vector and determine the point cloud naming label, specifically: Construct feature vectors based on multi-dimensional point cloud data and point cloud classification results; According to the feature vector, starting from the root node of the decision tree model, each feature value in the feature vector is compared with the splitting threshold in turn based on the pre-set splitting rule at each node; According to the comparison results, the point cloud data is passed down along the branches of the decision tree, and the splitting process is repeated until it reaches the leaf node. The category label of the leaf node is used as the preliminary naming label of the point cloud data; The naming information of the K nearest points around each point is used as the K nearest neighbor label of each point, and the K nearest neighbor label of each point and the preliminary naming label constitute the point cloud naming label.

5. The method for automatically naming head and neck blood vessels based on CTA images as claimed in claim 1, characterized in that: The method constructs a comprehensive feature vector based on the multidimensional point cloud data and the point cloud naming label, clusters the comprehensive feature vector using a constrained clustering algorithm, and modifies the point cloud naming label according to the clustering result to obtain the naming of the head and neck blood vessels, specifically: Construct a comprehensive feature vector from multi-dimensional point cloud data and point cloud naming labels; The constrained clustering algorithm is used to construct the objective function containing label information, and the comprehensive feature vectors are clustered based on the objective function; The consistency of the point cloud naming label information in each cluster in the clustering results is used for judgment. If there are inconsistently named points in other points in the cluster, the naming is corrected based on the point cloud label information and comprehensive feature vector of the surrounding points of the inconsistently named points to obtain the naming of the head and neck blood vessels.

6. The method for automatically naming head and neck blood vessels based on CTA images as claimed in claim 5, characterized in that: The correction is performed based on the point cloud label information and comprehensive feature vector of the surrounding points of the inconsistently named points, specifically: For an inconsistently named point p i ,Based on the comprehensive feature vector, a weighted distance metric is used to calculate the similarity between the points with inconsistent naming and other points in the cluster; According to the similarity results, find the point p i The most similar points are determined by multi-point voting on the point cloud naming label information of these points. i New naming label.

7. An automatic naming system for head and neck blood vessels based on CTA images, characterized in that: include: A blood vessel segmentation unit is configured to acquire a head and neck CTA image, perform segmentation using a pre-trained blood vessel segmentation model to generate a blood vessel mask image, and generate multi-dimensional point cloud data based on the blood vessel mask image; The point cloud segmentation unit is configured to classify each point in the multi-dimensional point cloud data using the pre-trained point cloud segmentation models of the head and neck branches to obtain a point cloud classification result; A decision tree naming unit is configured to construct a feature vector based on the multi-dimensional point cloud data and the point cloud classification result, and use a pre-trained decision tree model to make a classification decision on the feature vector to determine the point cloud naming label; The clustering optimization unit is configured to construct a comprehensive feature vector based on the multi-dimensional point cloud data and the point cloud naming labels, cluster the comprehensive feature vector using a constrained clustering algorithm, and modify the point cloud naming labels according to the clustering results to obtain the head and neck blood vessel naming.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the method for automatic naming of head and neck blood vessels based on CTA images as described in any one of claims 1 to 6 are implemented.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the method for automatic naming of head and neck blood vessels based on CTA images as described in any one of claims 1 to 6 are implemented.

10. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by a processor, the steps in the method for automatic naming of head and neck blood vessels based on CTA images as claimed in any one of claims 1 to 6 are implemented.

Images