A Smart CT Image Analysis Method for Maxillofacial Fracture Classification

Abstract

This invention relates to the field of medical image processing technology, specifically to an intelligent CT image analysis method for maxillofacial fracture classification, solving the technical problem that existing technologies struggle to effectively restore the anatomical attribution of bone fragments. The method includes: acquiring a head CT voxel array of the target object and discretizing and extracting the head CT voxel array to obtain anatomical entity data; performing multidimensional constraint feature analysis on each free fracture fragment in the free fracture fragment set and each fracture section in the fracture section set to obtain a subset of fitting point pairs and a repositioning matching cost; assigning anatomical attribution based on the repositioning matching cost, and repositioning and reconstructing each free fracture fragment to its assigned fracture section on a stable skull reference block based on the fitting point pair subset, obtaining a virtual repositioning voxel array; and performing CT image analysis based on the virtual repositioning voxel array.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to an intelligent analysis method for CT images used in the classification of maxillofacial fractures. Background Technology

[0002] The classification of maxillofacial fractures is crucial for clinical decision-making regarding surgical treatment. Computed tomography (CT) imaging is the core imaging tool for diagnosing and classifying maxillofacial fractures. Currently, commonly used clinical fracture classification standards (such as the LeFort classification) primarily rely on whether the fracture line continuously traverses specific biomechanical supports such as the pterygoid process, orbital floor, or zygomaticofrontal suture. In clinical practice, accurate identification of the fracture line path and the fracture status of these anatomical supports via CT imaging is essential to provide reliable imaging evidence for surgical planning.

[0003] Current CT image analysis techniques for identifying and classifying maxillofacial fractures primarily rely on grayscale thresholding, edge detection, or simple geometric registration algorithms for fracture line identification and fragment reduction. While these techniques can effectively track fracture lines and meet basic classification requirements in conventional linear fracture scenarios, they are less effective in comminuted fractures caused by high-energy trauma. The disintegration of the anatomical support area, where stress is concentrated, results in the fragmentation of the previously continuous bone structure into numerous discrete, displaced, and rotating free bone fragments. This causes a break in the topological logic of the fracture line. Existing techniques, unable to track the broken fracture line path or prone to slippage and mismatch during reduction due to local geometric optima, struggle to effectively restore the anatomical classification of bone fragments, ultimately leading to classification errors and impacting the accuracy of clinical decision-making. Summary of the Invention

[0004] To address the technical problem of existing technologies' inability to effectively restore the anatomical classification of bone fragments, the present invention aims to provide an intelligent CT image analysis method for maxillofacial fracture classification. The specific technical solution adopted is as follows: This application provides an intelligent analysis method for CT images used in maxillofacial fracture classification, including: The head CT voxel array of the target object is obtained and discretized to extract anatomical entity data. The anatomical entity data includes a stable skull reference block, a set of free fracture fragments, and a set of fracture sections on the stable skull reference block. For each free fracture fragment in the set of free fracture fragments and each fracture section in the set of fracture sections, a multidimensional constraint feature analysis is performed to obtain a subset of contact point pairs between the free fracture fragments and the fracture sections, as well as the repositioning matching cost. The subset of contact point pairs includes the contact point pairs of geometrically fitted free fracture fragments and fracture sections. The repositioning matching cost is used to characterize the matching difference between the free fracture fragments and the fracture sections. Anatomical assignment is performed based on the repositioning matching cost between each free fracture fragment and each fracture section. Based on the subset of the fitting point pairs between each free fracture fragment and each fracture section, each free fracture fragment is repositioned and reconstructed to the fracture section to which it belongs on the stable skull reference block, thus obtaining a virtual repositioning voxel array. Intelligent analysis of CT images based on virtual reset voxel array.

[0005] In one possible implementation, the method includes: The original image data of the target object acquired by the scanner is analyzed, a global coordinate system is constructed, and the grayscale data of each three-dimensional point in the original image data is mapped to the global coordinate system to generate a head CT voxel array. Bone tissue segmentation was performed on the head CT voxel array, and stable skull reference blocks and free fracture fragment sets were extracted from the head CT voxel array. Fracture sections are identified at the vertices of the bone surface of the stable skull reference block to obtain a set of fracture sections on the stable skull reference block.

[0006] In one possible implementation, the method includes: Based on three-dimensional connected component analysis, the head CT voxel array is divided into multiple connected components; The connected component with the largest volume among multiple connected components is identified as the stable skull reference block, and the connected components with a volume greater than a preset volume threshold, excluding the stable skull reference block, are identified as free fracture fragments in the free fracture fragment set.

[0007] In one possible implementation, the method includes: For each bone surface vertex of the stable skull reference block, the discrete mean curvature corresponding to the bone surface vertex is calculated, and feature vertices are selected from each bone surface vertex based on the discrete mean curvature. Cluster analysis is performed on the feature vertices, and each cluster of points obtained from the clustering is identified as a fracture section in the fracture section set.

[0008] In one possible implementation, the method includes: For each free fracture fragment in the set of free fracture fragments and each fracture section in the set of fracture sections, spatial pose registration is performed on the free fracture fragments and fracture sections based on geometric morphology constraints to obtain the reset transformation matrix and the corresponding geometric residual between the free fracture fragments and fracture sections. The reset transformation matrix is ​​used to characterize the spatial pose transformation parameters of the free fracture fragments relative to the fracture sections. The geometric residual is used to characterize the geometric surface fitting error between the free fracture fragments and fracture sections. The coordinates of the vertices of the bone surface of the free fracture fragment are mapped based on the repositioning transformation matrix, and a subset of fitting points is generated according to the distance from each bone surface vertex to the fracture section after mapping. The repositioning matching cost between the free fracture fragment and the fracture section is obtained by evaluating the repositioning matching based on the repositioning transformation matrix, geometric residuals, and fitting point pairs subsets.

[0009] In one possible implementation, the method includes: Based on the normal vector distribution of each contact point pair in the set of contact point pairs, the concavity-convexity interlocking degree between the free fracture fragment and the fracture section is calculated; the concavity-convexity interlocking degree is used to characterize the degree of isotropy of the contact surface between the free fracture fragment and the fracture section in three-dimensional space. Based on the repositioning transformation matrix and the subset of fitting points, the grayscale information of the corresponding positions in the head CT voxel array is traced back to determine the trabecular texture difference between the free fracture fragment and the fracture section; the trabecular texture difference is used to characterize the degree of difference in the direction of trabecular texture on both sides of the fracture section. The cost of repositioning and matching is determined based on geometric residuals, occlusal symmetry, and trabecular texture differences.

[0010] In one possible implementation, the method includes: Extract the normal vectors of each contact point pair in the set of bonding point pairs, and construct the normal covariance matrix of the bonding surface based on the normal vectors of each contact point pair; The eigenvalue sequence is obtained by performing eigenvalue decomposition on the normal covariance matrix of the mating surface, and the concave-convex interlocking degree is determined based on the eigenvalue sequence.

[0011] In one possible implementation, the method includes: Calculate the geometric center and average normal vector of the subset of fitting points, and extend sampling along the positive and negative directions of the average normal vector starting from the geometric center to obtain the sampling points on the reference block side and the sampling points on the fracture fragment side. Based on the repositioning transformation matrix, the sampling points on the side of the fracture fragment are mapped back to the coordinate space of the free fracture fragment to obtain the index coordinates; The gray-level gradients in the neighborhood of the sampling point on the reference block side and the neighborhood of the index coordinate in the head CT voxel array are calculated respectively, and feature vectors are extracted based on the gray-level gradients to obtain the main texture direction on the reference block side and the main texture direction on the fracture fragment side. The difference in trabecular texture is determined based on the principal direction of the texture on the side of the reference block and the principal direction of the texture on the side of the fracture fragment.

[0012] In one possible implementation, the method includes: A matching cost matrix is ​​constructed based on the repositioning matching cost between each free fracture fragment and each fracture section. The matching cost matrix is ​​then optimized and assigned to obtain the anatomical assignment matrix. The anatomical assignment matrix is ​​used to characterize the correspondence between each free fracture fragment and the fracture section to which it belongs. The voxel data of the stable skull reference block is filled into the corresponding positions of the initial voxel array. Based on the anatomical assignment matrix, each free fracture fragment is transformed to the global coordinate system where the initial voxel array is located. The corresponding positions of the fitting point pairs are marked as fracture gap voxels, thereby generating a virtual repositioning voxel array.

[0013] In one possible implementation, the method includes: The standard anatomical atlas is registered to the virtual repositioned voxel array to generate an anatomical pillar region mask; the anatomical pillar region mask is used to identify the spatial extent of each anatomical pillar in the virtual repositioned voxel array. Identification of fracture voxels in each anatomical strut using anatomical strut region masking and virtual reset voxel array.

[0014] The present invention has the following beneficial effects: Based on the above technical solution, this application obtains the head CT voxel array of the target object and discretizes and extracts the head CT voxel array to obtain standardized anatomical entity data, providing accurate basic data for subsequent analysis. Then, for each free fracture fragment in the free fracture fragment set and each fracture section in the fracture section set, multidimensional constraint feature analysis is performed to realize the quantification of the matching relationship between free fracture fragments and fracture sections. Then, through anatomical assignment and virtual repositioning reconstruction, a virtual repositioning voxel array is obtained, and the complete skeletal structure is restored in virtual space, so that the anatomical assignment of bone fragments is effectively restored. In this way, this application can effectively improve the accuracy, efficiency and reliability of CT image analysis based on the image analysis of the virtual repositioning voxel array. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is one of the flowcharts illustrating an intelligent CT image analysis method for maxillofacial fracture classification according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the hardware structure of an intelligent CT image analysis device for maxillofacial fracture classification, provided in one embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a CT image intelligent analysis method for maxillofacial fracture classification proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] In all division and logarithmic operations covered in this application, a smoothing mechanism is employed to prevent computer program crashes or invalid values ​​from being generated due to a zero denominator or a zero input. Specifically, a positive correction factor is superimposed on the denominator term of the division operation or the argument term of the logarithmic function. For example, the value is This ensures the robustness and feasibility of the algorithm under extreme conditions.

[0020] The normalization function mentioned in this application Unless otherwise specified, all values ​​are normalized using maximum and minimum values. The maximum and minimum values ​​are preset empirical extreme values ​​derived from a large amount of historical experimental data. If the calculated result exceeds the [0,1] interval, it is restricted to the [0,1] range by a truncation function (i.e., if the result is less than 0, it is taken as 0, and if it is greater than 1, it is taken as 1) to eliminate the influence of outliers on the evaluation index.

[0021] The following describes in detail, with reference to the accompanying drawings, a specific scheme of the intelligent CT image analysis method for maxillofacial fracture classification provided by the present invention.

[0022] Please see Figure 1 The diagram illustrates a flowchart of an intelligent CT image analysis method for maxillofacial fracture classification according to an embodiment of the present invention. The method includes the following steps: Step 101: Obtain the head CT voxel array of the target object, and discretize and extract the head CT voxel array to obtain anatomical entity data.

[0023] The anatomical data includes a stable skull reference block, a collection of free fracture fragments, and a collection of fracture sections on the stable skull reference block. The head CT voxel array is a grayscale data array obtained from the parsing of raw digital imaging and communications in medicine (DICOM) sequences acquired by the CT scanner, used for subsequent texture sampling and geometric morphology analysis. Each voxel in the head CT voxel array contains a Hounsfield Unit (HU) value corresponding to its location, reflecting the degree of tissue attenuation of X-rays.

[0024] The stable skull reference block is the stable bone portion that remains relatively stationary and has not undergone displacement during maxillofacial trauma. It is typically the largest in volume and located at the skull base, serving as a static reference frame for spatial registration. The set of free fracture fragments refers to the collection of discrete bone fragments that have been displaced and rotated due to fractures, requiring subsequent reduction and matching analysis. The set of fracture sections is the collection of rough fracture areas exposed by fractures on the stable skull reference block, representing the matching area for free fracture fragments.

[0025] This application can transform continuous medical CT images into computer-operable discrete geometric entities, providing a clear operating object for subsequent registration, repositioning and other analysis steps, while eliminating interference from irrelevant information such as non-bone tissue and bone debris.

[0026] Step 102: Perform multidimensional constraint feature analysis on each free fracture fragment in the set of free fracture fragments and each fracture section in the set of fracture sections to obtain the subset of fitting points between the free fracture fragments and the fracture sections, as well as the repositioning matching cost.

[0027] The subset of fitting point pairs includes contact point pairs where the free fracture fragment and the fracture section are geometrically fitted. These are point pairs identified after spatial registration that have a matching relationship and serve as the spatial carriers of their physical contact. The repositioning matching cost is used to characterize the degree of matching difference between the free fracture fragment and the fracture section. The smaller the repositioning matching cost, the higher the degree of matching between the free fracture fragment and the fracture section, and the more it conforms to the true anatomical attribution.

[0028] This application can quantitatively analyze the matching between free fracture fragments and fracture sections through multidimensional constraint feature analysis, breaking through the limitations of simple geometric morphology matching, providing a core quantitative basis for subsequent anatomical assignment, and realizing the accurate quantification of the matching relationship between each free fracture fragment and each fracture section.

[0029] Step 103: Based on the repositioning matching cost between each free fracture fragment and each fracture section, perform anatomical assignment, and based on the subset of mating point pairs between each free fracture fragment and each fracture section, reposition and reconstruct each free fracture fragment to the fracture section to which it belongs on the stable skull reference block, thus obtaining a virtual repositioning voxel array.

[0030] This application can determine the fracture section to which each free fracture fragment belongs based on the quantification result of the repositioning matching cost, thereby achieving anatomical assignment. Then, through repositioning reconstruction, the spatial positions of the free fracture fragments, based on a subset of apposition point pairs, are mapped back to the corresponding fracture section positions of the stable skull reference block via coordinate transformation, achieving logical repositioning of comminuted fracture fragments. This virtual repositioning voxel array is generated through repositioning reconstruction based on an initial voxel array. The initial voxel array refers to a three-dimensional data array constructed in computer memory with the same dimensions and resolution as the head CT voxel array, used to carry the repositioned anatomical structure.

[0031] In this application, the virtual repositioning voxel array not only carries the repositioned skeletal anatomical structure but also marks the location information of the fracture gap, thus restoring the topological continuity of the anatomical support and solving the problem of topological disconnection of the fracture line in the head CT voxel array.

[0032] Step 104: Perform CT image analysis based on virtual reset voxel array.

[0033] In one possible implementation, this application can register a standard anatomical atlas to a virtual repositioned voxel array to generate an anatomical strut region mask.

[0034] The anatomical strut region mask is used to identify the spatial extent of each anatomical strut within the virtual repositioned voxel array. For example, the anatomical struts may include key mechanical transmission structures such as the pterygoid strut, piriform foramen strut, orbital floor strut, and zygomaticofrontal suture strut.

[0035] To define the spatial extent of key anatomical supports such as the pterygoid process and orbital floor, this application can utilize a stable skull reference block as an anchor point for rigid registration. Since the skull base structures typically remain morphologically stable during facial trauma, this application can load a pre-calibrated standard anatomical atlas containing the regions of interest (ROIs) of the aforementioned key biomechanical supports.

[0036] Thus, this application can use the Iterative Closest Point (ICP) algorithm to calculate the rigid transformation matrix between the skull base region in the standard anatomical atlas and the stable skull reference block. Based on this rigid transformation matrix, the four ROI masks in the standard atlas are projected onto the virtual repositioned voxel array space to generate anatomical support region masks, avoiding the complexity and deformation risk of whole-brain non-rigid registration and ensuring the anatomical rationality of the support region definition.

[0037] Subsequently, fracture voxels in each anatomical strut were identified based on the anatomical strut region mask and the virtual reset voxel array.

[0038] Among them, the fracture gap voxel is a voxel marked as representing a crack in the virtual reduction voxel array, which is a physical representation of the fracture line in the virtual reduction space.

[0039] This application can analyze each anatomical pillar corresponding to the mask in the anatomical pillar area mask one by one, extract the spatial coordinate range corresponding to each mask, and then process only the voxels within the spatial coordinate range to remove irrelevant voxel interference outside the analysis area, thereby improving the efficiency and accuracy of fracture gap voxel identification.

[0040] Next, the voxel coordinates of all voxels marked as fracture voxels in the virtual reset voxel array are extracted to construct a set of fracture voxel coordinates. The spatial intersection operation is performed between this set of coordinates and the spatial coordinate range of each anatomical support, and the voxel coordinates that belong to both the anatomical support analysis area and the set of fracture voxel coordinates are selected. The voxel corresponding to this voxel coordinate is the fracture voxel in each anatomical support.

[0041] Based on the above technical solution, this application obtains the head CT voxel array of the target object and discretizes and extracts the head CT voxel array to obtain standardized anatomical entity data, providing accurate basic data for subsequent analysis. Then, for each free fracture fragment in the free fracture fragment set and each fracture section in the fracture section set, multidimensional constraint feature analysis is performed to realize the quantification of the matching relationship between free fracture fragments and fracture sections. Then, through anatomical assignment and virtual repositioning reconstruction, a virtual repositioning voxel array is obtained, and the complete skeletal structure is restored in virtual space, so that the anatomical assignment of bone fragments is effectively restored. In this way, this application can effectively improve the accuracy, efficiency and reliability of CT image analysis based on the image analysis of the virtual repositioning voxel array.

[0042] As a possible embodiment of this application, step 101 above can be implemented through the following steps: Step 201: Analyze the original image data of the target object acquired by the scanner, construct a global coordinate system, and map the grayscale data of each three-dimensional point in the original image data to the global coordinate system to generate a head CT voxel array.

[0043] The original image data consists of a DICOM sequence file output by the scanner. This DICOM sequence file contains metadata such as pixel spacing, slice thickness, and scan origin, as well as grayscale scan data. The global coordinate system is a three-dimensional coordinate system constructed based on the CT scan metadata. This global coordinate system uses the physical origin defined by the scanner as its zero point, and the X, Y, and Z axes correspond to the left-right, front-back, and up-down directions of human anatomy, respectively. It serves as the unified reference for all subsequent spatial analysis operations. The grayscale data can be represented by the Heinz unit value of human tissue density. After mapping it to the global coordinate system, a head CT voxel array indexed by three-dimensional coordinates can be formed. The value of each element in the array is the grayscale value at the corresponding coordinate position.

[0044] For example, when parsing a DICOM sequence file, spatial parameters can be extracted to construct a global coordinate system, and interpolation operations can be used to achieve uniform mapping of grayscale data in the coordinate system, ensuring the spatial consistency of the head CT voxel array.

[0045] Step 202: Perform bone tissue segmentation on the head CT voxel array and extract the stable skull reference block and free fracture fragment set from the head CT voxel array.

[0046] In one possible implementation, this application can divide the head CT voxel array into multiple connected domains based on three-dimensional connected domain analysis. Then, the connected domain with the largest volume among the multiple connected domains is determined as the stable skull reference block, and the connected domains other than the stable skull reference block with a volume greater than a preset volume threshold are determined as free fracture fragments in the free fracture fragment set.

[0047] It should be noted that bone tissue and surrounding soft tissue have significantly different X-ray attenuation coefficients, and fractures caused by trauma can lead to the spatial splitting of originally continuous bone into independent, unconnected regions. To extract these independent regions as the target for manipulation, this application first applies Otsu's method to statistically analyze the grayscale histogram of the head CT voxel array to obtain the bone tissue segmentation threshold. Then, the head CT voxel array is traversed, and all voxels with grayscale values ​​greater than the bone tissue segmentation threshold are marked as foreground, generating a binarized bone mask. Subsequently, a three-dimensional connected component labeling algorithm is applied to this bone mask, using the 26-neighborhood connectivity rule to determine the connectivity between voxels. That is, if another bone voxel exists within the 26 spatial neighbors of a bone voxel, the two voxels are considered to belong to the same connected domain. Through the above three-dimensional connected domain analysis, the originally continuous bone voxels (actually discrete after trauma) can be divided into multiple spatially independent connected domains, each corresponding to an independent bone block structure.

[0048] Furthermore, based on anatomical knowledge, in maxillofacial trauma, the skull base or posterior portion of the skull typically remains relatively stable and has the largest volume, while free fracture fragments are relatively small. Therefore, this application can calculate the physical volume of each connected component, defining the connected component with the largest volume as a stable skull reference block, serving as a static reference system for subsequent spatial registration. For the remaining connected components, a preset volume threshold can be set (for example, a threshold value of...). (used to remove debris noise), and defined as free fracture fragments in the free fracture fragment set when the volume of the connected domain is greater than the preset volume threshold.

[0049] In this way, this application can extract isosurfaces and generate corresponding triangular mesh models for stable skull reference blocks and each free fracture fragment using the Marching Cubes algorithm. The vertex set of these triangular mesh models is defined as the set of vertices on the bone surface, and each vertex contains its coordinates in the global coordinate system and a unit normal vector calculated by fitting the mesh patches. It should be noted that each triangular mesh model not only contains a geometric mesh, but also retains the index range of its voxels in the head CT voxel array to ensure the alignment of the geometry with the internal density information.

[0050] Step 203: Identify the fracture sections at the vertices of the bone surface of the stable skull reference block to obtain a set of fracture sections on the stable skull reference block.

[0051] In one possible implementation, this application can calculate the discrete average curvature corresponding to each bone surface vertex of a stable skull reference block, and select feature vertices from each bone surface vertex based on the discrete average curvature. Then, cluster analysis is performed on the feature vertices, and each cluster of points obtained from the clustering is determined as a fracture section in the fracture section set.

[0052] It should be noted that the surface of the stable skull reference block includes both the originally smooth cortical bone surface and the rough fracture surface exposed by the fracture, and the free fracture fragment can only be physically classified as belonging to the rough fracture surface. In order to narrow down the matching search range and reduce the probability of mismatching bone fragments to smooth surfaces, this application can further segment the potential fracture area.

[0053] For example, this application can solve the geometric curvature characteristics of all vertices within a ring neighborhood of each bone surface vertex and perform averaging to obtain the discrete average curvature corresponding to the bone surface vertex. The discrete average curvature is a quantitative indicator characterizing the degree of local geometric curvature of the bone surface vertex. The fracture section on the stable skull reference block, formed by stress tearing, exhibits high-frequency geometric changes, and its discrete average curvature is significantly higher than that of the surrounding smooth cortical bone surface. This is the core geometric feature difference between the fracture section and the smooth surface. Based on this difference, this application can set a high curvature screening threshold (e.g., a threshold can be set...). All bone surface vertices with a discrete average curvature greater than the high curvature screening threshold are selected as feature vertices. These feature vertices are the constituent vertices of the fracture section. The preliminary location of the fracture section can be achieved by screening the feature vertices.

[0054] Then, using these feature vertices as seed points, Euclidean distance clustering analysis is performed, grouping nodes whose spatial distance is less than a preset neighborhood radius (e.g., taking...). The feature vertices of the image are merged into a single cluster. Each independent cluster of points constitutes a potential fracture region, defined as a fracture section in the fracture section set. For each fracture section, this application can calculate the geometric centroids of all its contained points and the average unit normal vector, which serve as guiding parameters for initial registration in subsequent steps. Through this step, the originally continuous reference block surface is structured into a series of discrete targets to be matched, completing the transformation from the original image to the physical object of operation.

[0055] Based on the above technical solutions, this application ensures the consistency of all subsequent processing steps on the spatial reference by constructing a global coordinate system. By segmenting bone tissue based on connected domain volume, it effectively distinguishes between stable reference structures and free fracture fragments. By identifying fracture sections based on geometric features, it accurately locates potential matching areas on stable skull reference blocks, eliminates interference from smooth cortical bone surfaces, narrows the search range for subsequent free fracture fragment matching, and provides effective anatomical entity data for subsequent multidimensional constraint feature analysis, significantly improving the search efficiency and accuracy of subsequent reduction and matching.

[0056] As a possible embodiment of this application, step 102 above can be implemented through the following steps: Step 301: For each free fracture fragment in the set of free fracture fragments and each fracture section in the set of fracture sections, perform spatial pose registration of the free fracture fragments and fracture sections based on geometric morphology constraints to obtain the reset transformation matrix between the free fracture fragments and fracture sections and the corresponding geometric residuals.

[0057] The repositioning transformation matrix characterizes the spatial orientation transformation parameters of the free fracture fragment relative to the fracture cross-section, while the geometric residual characterizes the geometrical fit error between the free fracture fragment and the fracture cross-section. For example, the repositioning transformation matrix includes translation and rotation components, which can map the free fracture fragment to a spatial position that geometrically fits the fracture cross-section. The geometric residual is typically represented by the root mean square error between point pairs; a smaller geometric residual indicates a higher degree of geometrical fit between the two.

[0058] This application enables constrained registration of the spatial relative position and orientation of free fracture fragments and fracture sections based on their geometric surface morphology, ensuring that the registration result meets the geometric fit requirements. Since fragments in a free fracture fragment set undergo significant random displacement and rotation after trauma, directly applying local optimization algorithms is prone to getting trapped in local maxima due to excessive initial value deviations. Therefore, this application divides spatial pose registration into two stages: coarse alignment and fine registration. First, coarse alignment between the bone fragment and the section is achieved through global features, pulling the bone fragment into the convergence region of fine registration. Then, fine registration is achieved through local iteration, ensuring the accuracy of the registration.

[0059] For example, for any free fracture fragment in the set of free fracture fragments and any fracture section in the set of fracture sections, the geometric centroid and covariance matrix of the set of vertices on the bone surface of the free fracture fragment and the fracture section are first calculated respectively. Eigenvalue decomposition is performed on the covariance matrix, and the eigenvector corresponding to the largest eigenvalue is extracted as the principal axis of inertia to construct the principal axis alignment transformation matrix.

[0060] The translation component ensures that the centroid of the free fracture fragment coincides with the centroid of the fracture section, while the rotation component aligns the principal axes of inertia of the free fracture fragment parallel to the principal axes of inertia of the fracture section. Given the 180° directional ambiguity of the principal axes of inertia, this application can generate two candidate initial values—forward alignment and reverse alignment—which are then independently input into the subsequent fine registration process for parallel iteration. Finally, the set with the smaller geometric residual is retained as the effective path for fragment-section pairing, thereby eliminating the uncertainty in pose initialization.

[0061] Subsequently, based on the initial pose determination, this application can further eliminate spatial overlap through fine iteration of the geometric surface. First, using the principal axis alignment transformation matrix as the initial value, the ICP algorithm is applied to rigidly register the free fracture fragment with the fracture section. In each iteration, the point on the surface of the free fracture fragment closest to the fracture section is retrieved to establish a correspondence. The transformation parameters are updated by minimizing the root mean square error between point pairs until convergence is obtained to obtain the final reset transformation matrix. At this point, the root mean square error between the corresponding point pairs is the final geometric residual.

[0062] Step 302: Based on the repositioning transformation matrix, perform coordinate mapping on the vertices of the bone surface of the free fracture fragment, and generate a subset of fitting points according to the distance from each bone surface vertex to the fracture section after mapping.

[0063] In some embodiments, this application can transform the set of vertices on the bone surface of the free fracture fragment to a global coordinate system using a repositioning transformation matrix. Then, iterates through the transformed bone surface vertices, calculates their Euclidean distances to the fracture surface, and selects all vertices with distances less than a preset contact threshold and their corresponding normal vectors to form a subset of contact point pairs. For example, the preset contact threshold can be taken as... This is approximately equal to the diagonal length of the voxel. To prevent false contact, if the number of point pairs in the subset of mating points is less than the minimum area threshold (e.g., taking 10 points), the subset of mating points can be directly determined to be invalid and not used for subsequent calculations, thus avoiding interference from false contact.

[0064] Step 303: Based on the repositioning transformation matrix, geometric residuals, and fitting point pairs, perform repositioning matching evaluation to obtain the repositioning matching cost between the free fracture fragment and the fracture section.

[0065] Among them, the repositioning and matching assessment refers to the comprehensive evaluation of the matching effectiveness and fit between the free fracture fragment and the fracture section from multiple dimensions. The assessment basis includes not only the geometric residual at the geometric morphology level, but also multi-dimensional indicators such as physical contact characteristics and internal material structure characteristics obtained based on the repositioning transformation matrix and the subset of fitting points. This application can obtain the final repositioning and matching cost by comprehensively quantifying the assessment indicators of each dimension, so that the repositioning and matching cost can fully characterize the matching difference between the two, and provide a reliable quantitative basis for subsequent anatomical assignment.

[0066] Based on the above technical solution, this application can perform spatial pose registration between each free fracture fragment in the free fracture fragment set and each fracture section in the fracture section set based on geometric morphological constraints. This yields the repositioning transformation matrix and the corresponding geometric residual between the free fracture fragment and the fracture section, ensuring the accuracy and stability of the registration results. Subsequently, coordinate mapping is performed on the bone surface vertices of the free fracture fragment based on the repositioning transformation matrix, and a subset of fitting point pairs is generated based on the distance from each bone surface vertex to the fracture section after mapping. This precisely defines the physical contact area between the fracture fragment and the section. Finally, repositioning matching evaluation is performed based on the repositioning transformation matrix, geometric residual, and the subset of fitting point pairs to obtain the repositioning matching cost between the free fracture fragment and the fracture section. This achieves comprehensive quantification of the matching relationship, making the calculation of the repositioning matching cost more realistic and providing a reliable basis for subsequent anatomical assignment. This further improves the accuracy and reliability of multidimensional constraint feature analysis.

[0067] As a possible embodiment of this application, step 303 above can be implemented through the following steps: Step 401: Calculate the degree of interlocking between the free fracture fragment and the fracture section based on the normal vector distribution of each contact point pair in the set of interlocking point pairs.

[0068] The concave-convex occlusion degree is used to characterize the degree of isotropy of the contact surface between the free fracture fragment and the fracture section in three-dimensional space, reflecting the kinematic interlocking ability of the contact surface. The real fracture surface is formed by stress tearing, presenting a complex morphology of interlocking canines. Its normal vector is discretely distributed in three-dimensional space and has significant isotropic characteristics. In contrast, the normal vector of the smooth cortical bone surface tends to be coplanarly distributed and has a low degree of isotropy. Therefore, this application can effectively distinguish the physical occlusion of the real fracture surface from the false occlusion of the smooth surface through the concave-convex occlusion degree.

[0069] In one possible implementation, this application can extract the normal vectors of each contact point pair in the set of bonding point pairs, and construct the normal covariance matrix of the bonding surface based on the normal vectors of each contact point pair.

[0070] For example, the normal covariance matrix of the mating surface satisfies the following formula: in, The normal covariance matrix of the mating surface. To fit the subset of points The first in One contact point pair, For the first The unit normal vector at each contact point pair for The transpose of . The constructed normal covariance matrix of the mating surface is The second-order moment matrix.

[0071] Then, the normal covariance matrix of the mating surface is decomposed into eigenvalues ​​to obtain the corresponding eigenvalue sequence, and the concave-convex interlocking degree is determined based on the eigenvalue sequence.

[0072] This application performs singular value decomposition on the above-mentioned mating surface normal covariance matrix, resulting in three eigenvalues ​​arranged in descending order. These three eigenvalues ​​represent the intensity of the projection components of the normal vector along the three orthogonal principal directions in space.

[0073] For example, the interlocking degree satisfies the following formula: in, For the degree of interlocking, , , The three eigenvalues ​​obtained after singular value decomposition are: 3 in the numerator is the normalization coefficient, which is used to limit the range of the concave-convex interlocking degree to between 0 and 1, and the denominator is the sum of the three eigenvalues.

[0074] This application uses the minimum eigenvalue With the core, through The ratio of the ratio to the sum of the eigenvalues ​​characterizes the degree of isotropy of the normal vector distribution. When the contact surface is an ideally smooth plane, all normal vectors are parallel. At this point, the degree of occlusion approaches 0, the lower the isotropy, the weaker the physical interlocking characteristics, resulting in a spurious match on a smooth surface. When the contact surface is a three-dimensional occlusal surface with interlocking canines, the normal vector is uniformly distributed in three-dimensional space. At this point, the degree of occlusion approaches 1, the higher the isotropy, the stronger the physical interlocking characteristics of the contact surface, resulting in a match on a true fracture surface.

[0075] Step 402: Based on the reset transformation array and the subset of fitting points, trace back the grayscale information of the corresponding positions in the head CT voxel array to determine the trabecular texture difference between the free fracture fragment and the fracture section.

[0076] Among them, the trabecular texture difference is used to characterize the degree of difference in the direction of trabecular texture on both sides of the fracture section, reflecting the continuity of the internal microstructure of the bone.

[0077] Geometric and interlocking analysis is mainly used to analyze surface morphology. In addition, the trabecular structure inside the skeleton forms a specific texture flow along the principal stress trajectory. The continuity of this microstructure is interrupted at the fracture site but still retains directional characteristics. Therefore, this application can further utilize structural tensor consistency to perform matching verification from the level of internal material properties.

[0078] In one possible implementation, this application can calculate the geometric center and average normal vector of the subset of fitting points, and extend sampling along the positive and negative directions of the average normal vector, starting from the geometric center, to obtain the sampling points on the reference block side and the sampling points on the fracture fragment side.

[0079] Here, the geometric center of the mating point pair subset can be the average spatial coordinates of all contact point pairs in the global coordinate system, and the average normal vector is the average unit normal vector of all contact point pairs. Both together characterize the center position and overall orientation of the mating area. Extended sampling refers to extending a preset sampling distance from the geometric center along the average normal vector and its opposite direction. For example, this sampling distance can be set to... To ensure that the sampling points are located in the cancellous bone region below the cortical bone, where the trabecular texture features are most prominent, sampling points along the direction of the average normal vector are the reference block side sampling points, located on the side of the stable skull reference block, and sampling points along the opposite direction of the average normal vector are the fracture fragment side sampling points, located on the side of the free fracture fragment.

[0080] It should be noted that, since the voxel data of the free fracture fragment is always stored in the original head CT voxel array and does not move with the geometric mesh, this application also needs to backtrack the grayscale information of the corresponding position in the head CT voxel array in order to obtain the true texture at the sampling point on the side of the fracture fragment.

[0081] This application can map the sampling points on the side of the fracture fragment back to the coordinate space of the free fracture fragment based on the reset transformation matrix to obtain the index coordinates. Then, the gray-level gradients in the neighborhood of the sampling points on the side of the reference block and in the neighborhood of the index coordinates in the head CT voxel array are calculated respectively. Based on the gray-level gradients, feature vectors are extracted to obtain the main direction of texture on the side of the reference block and the main direction of texture on the side of the fracture fragment. Among them, the sampling points on the side of the fracture fragment are the spatial coordinates of the free fracture fragment after repositioning and transformation, while the voxel data of the free fracture fragment is always stored in the original coordinate space of the head CT voxel array. This application can map the global coordinates of the sampling points on the side of the fracture fragment back to the original coordinate space of the free fracture fragment through the inverse transformation of the repositioning transformation matrix. The resulting mapped coordinates are the index coordinates. The index coordinates are the direct index for accessing the head CT voxel array. The grayscale information of the sampling points on the side of the fracture fragment can be obtained through the index coordinates.

[0082] In some embodiments, the sampling point neighborhood is a three-dimensional voxel neighborhood centered on the sampling point / index coordinates. For example, it can be set to a 3×3×3 three-dimensional voxel neighborhood. The gray-level gradient is a vector characterizing the direction and magnitude of the change in voxel gray-level values ​​within the sampling point neighborhood, obtained by solving the gradient of the gray-level data within the sampling point neighborhood. This application can construct a structure tensor using the gray-level gradient to characterize the texture direction features within the neighborhood. Then, the structure tensor is subjected to eigenvalue decomposition, and the eigenvector corresponding to the largest eigenvalue is the main texture direction. This main texture direction characterizes the main orientation of the trabecular texture at the sampling point.

[0083] Thus, this application can determine the difference in trabecular texture based on the principal direction of the texture on the side of the reference block and the principal direction of the texture on the side of the fracture fragment.

[0084] For example, the trabecular texture difference satisfies the following formula: in, For the trabecular texture difference, The main direction of the side texture of the reference block. The main direction of the texture on the side of the fracture fragment. and All are three-dimensional unit vectors. This represents the absolute value of the cosine of the angle between two directions. It is used when the two principal texture directions are nearly parallel. A value close to 1 indicates that the trabecular texture difference is close to 0, suggesting a more continuous trabecular texture orientation between the free fracture fragment and the fracture section, and a greater likelihood of a true anatomical match. When the two principal texture directions are nearly perpendicular... A value close to 0 indicates that the more discontinuous the trabecular texture direction is on both sides of the fracture surface, the greater the difference in trabecular texture direction, and the more likely the two are an incorrect matching combination.

[0085] Step 403: Determine the repositioning and matching cost based on the geometric residual, the degree of occlusion, and the difference in trabecular texture.

[0086] Among them, geometric residual reflects the degree of geometric fit between the fracture fragment and the cross section, concavity and convexity interlocking reflects the physical interlocking characteristics of the contact surface, and trabecular texture difference reflects the continuity of the material structure inside the bone. The three indicators characterize the fit of the matching from three dimensions: geometric shape, physical contact, and internal structure, and are independent and highly complementary to each other.

[0087] For example, the reset matching cost satisfies the following formula: in, To reset the matching cost, For geometric residuals, For the degree of interlocking, For the trabecular texture difference, , , For the weighting coefficients, satisfying For example, this application can obtain This emphasizes the dominant role of geometric fit and physical interlocking. This is a normalization function (e.g., maximum-minimum normalization) used to map geometric residuals to a range of 0 to 1.

[0088] Based on the above technical solution, this application calculates the concave-convex occlusion degree between the free fracture fragment and the fracture section according to the normal vector distribution of each contact point pair in the occlusion point pair subset. This effectively identifies the canine interlocking features of the real fracture section and eliminates sliding misfits on smooth surfaces. Based on the repositioning transformation matrix and the occlusion point pair subset, the grayscale information of the corresponding positions in the head CT voxel array is traced back to determine the trabecular texture difference between the free fracture fragment and the fracture section. The continuity of the internal microstructure of the bone solves the problem of distinguishing geometrically similar fragments. Finally, the repositioning matching cost determined by the geometric residual, concave-convex occlusion degree, and trabecular texture difference degree can comprehensively and accurately characterize the matching difference between the free fracture fragment and the fracture section, providing a more accurate and reliable quantitative basis for subsequent anatomical assignment and further improving the accuracy of multidimensional constraint feature analysis.

[0089] As a possible embodiment of this application, step 103 above can be implemented through the following steps: Step 501: Construct a matching cost matrix based on the repositioning matching cost between each free fracture fragment and each fracture section, and optimize the matching cost matrix for assignment matching to obtain the anatomical assignment matrix.

[0090] The anatomical attribution matrix, a two-dimensional matrix, characterizes the correspondence between each free fracture fragment and its assigned fracture section. For example, rows in the anatomical attribution matrix correspond to free fracture fragments in the set of free fracture fragments, and columns correspond to fracture sections in the set of fracture sections. Each element in the anatomical attribution matrix indicates the attribution relationship between the corresponding free fracture fragment and the fracture section. For instance, if the matrix element... , indicating the first The free fracture fragment belongs to the first For each fracture section, if the matrix elements , indicating the first The individual free fracture fragment does not belong to the first One fracture section.

[0091] In some embodiments, this application can construct a matching cost matrix based on the repositioning matching cost between each free fracture fragment and each fracture section. The matrix elements in the matching cost matrix are the repositioning matching costs between the corresponding free fracture fragment and the fracture section. If the number of free fracture fragments... With the number of fracture sections If they are not equal, the matching cost matrix can be completed by filling it with infinitely large value increments. For a square matrix, an infinitely large algebraic value indicates that the corresponding pairing has no actual matching meaning.

[0092] For example, this application can optimize the assignment matching based on the matching cost matrix, such as using the Hungarian algorithm. This algorithm is a globally optimal assignment algorithm that can find a set of element combinations in the matching cost matrix that minimizes the sum of the elements in the combination, and each row and column contains only one element, achieving a one-to-one exclusive assignment of free fracture fragments to fracture sections. This application can further filter the obtained element combinations, classifying free fracture fragments in the element combinations whose repositioning matching cost is greater than or equal to the acceptance threshold as unrepositionable free necrotic bone or noise, and retaining only element combinations whose repositioning matching cost is less than the acceptance threshold, thereby obtaining the final anatomical assignment matrix. This acceptance threshold can be based on a large amount of historical experimental data from maxillofacial fracture CT images, judging pairings with a matching degree of less than 40% as invalid pairings, and eliminating the interference of noise such as free necrotic bone and bone fragments on anatomical assignment, for example, it can be 0.5.

[0093] Step 502: Fill the voxel data of the stable skull reference block into the corresponding positions of the initial voxel array, and transform each free fracture fragment to the global coordinate system where the initial voxel array is located according to the anatomical assignment matrix, and mark the corresponding positions of the fitting point pairs as fracture gap voxels, thereby generating a virtual repositioning voxel array.

[0094] To analyze the distribution of fracture lines under a unified spatial reference, this application can first allocate a zero-value three-dimensional matrix in memory with the same dimensions and resolution as the head CT voxel array, defined as the initial voxel array. Then, the voxel data of the stable skull reference block are filled into the initial voxel array in situ according to their global coordinate system.

[0095] Then, traverse all the dissection assignment arrays. To ensure effective pairing, the original voxel coordinates of the free fracture fragments are transformed to the global coordinate system using the corresponding reduction transformation matrix, and the voxel values ​​at the corresponding positions of the transformed coordinates are marked as numerical values ​​representing bone (for example, marked as 1).

[0096] Finally, the corresponding subsets of mating points will be effectively matched. The voxel values ​​for the positions of all contact point pairs are labeled as numerical values ​​representing fracture gap voxels (for example, labeled as 2). These labeled voxels are the fracture gap voxels, used to explicitly characterize the location of the fracture line. The initial voxel array after the above filling and labeling is the virtual reduction voxel array.

[0097] Based on the above technical solution, this application can realize the matrix representation of the pairing relationship between each free fracture fragment and each fracture section based on the matching cost matrix constructed by the repositioning matching cost, which provides a standardized basis for global optimization assignment. Then, the matching cost matrix is ​​optimized and matched to obtain the anatomical assignment matrix, realizing the global optimal assignment of free fracture fragments and fracture sections. In this way, this application can fill the voxel data of the stable skull reference block into the corresponding position of the initial voxel array, and transform each free fracture fragment to the global coordinate system where the initial voxel array is located according to the anatomical assignment matrix, and mark the corresponding position of the fitting point pair subset as fracture gap voxels, thereby generating a virtual repositioning voxel array. The generated virtual repositioning voxel array has the same dimension and resolution as the head CT voxel array, and retains the precise position of the fracture line, providing a core carrier with topological determination and complete information for fracture classification analysis.

[0098] It should be noted that the various embodiments of this application can be referenced or learned from each other. For example, the same or similar steps, method embodiments, system embodiments and device embodiments can be referenced from each other without limitation.

[0099] This application also provides a hardware structure diagram of a CT image intelligent analysis device for maxillofacial fracture classification (referred to as CT image intelligent analysis device 20 for maxillofacial fracture classification), see [link to diagram]. Figure 2 The intelligent CT image analysis device 20 for maxillofacial fracture classification includes a processor 21, and optionally, a memory 22 connected to the processor 21.

[0100] In the first possible implementation, see Figure 2 The intelligent CT image analysis device 20 for maxillofacial fracture classification also includes a communication interface 23. The processor 21, memory 22, and communication interface 23 are connected via a bus. The communication interface 23 is used to communicate with other devices or communication networks. Optionally, the communication interface 23 may include a transmitter and a receiver. The device in the communication interface 23 that implements the receiving function can be considered as a receiver, which is used to perform the receiving steps in the embodiments of this application. The device in the communication interface 23 that implements the transmitting function can be considered as a transmitter, which is used to perform the transmitting steps in the embodiments of this application.

[0101] Based on the first possible implementation method Figure 2The schematic diagram shown can be used to illustrate the structure of the intelligent CT image analysis device for maxillofacial fracture classification involved in the above embodiments.

[0102] in, Figure 2 The diagram also illustrates the system chip in a CT image intelligent analysis device used for maxillofacial fracture classification. In this case, the actions performed by the aforementioned CT image intelligent analysis device for maxillofacial fracture classification can be implemented by this system chip; the specific actions performed are described above and will not be repeated here.

[0103] In implementation, each step of the method provided in this embodiment can be completed by integrated logic circuits in the processor or by instructions in software form. The steps of the method disclosed in the embodiments of this application can be directly manifested as being executed by a hardware processor, or being executed by a combination of hardware and software modules in the processor.

[0104] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0105] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A method for intelligent analysis of CT images for the classification of maxillofacial fractures, characterized in that, include: A head CT voxel array of the target object is acquired, and the head CT voxel array is discretized and extracted to obtain anatomical entity data; the anatomical entity data includes a stable skull reference block, a set of free fracture fragments, and a set of fracture sections on the stable skull reference block; For each free fracture fragment in the set of free fracture fragments and each fracture section in the set of fracture sections, a multidimensional constraint feature analysis is performed to obtain a subset of contact point pairs between the free fracture fragments and the fracture sections, as well as a repositioning and matching cost. The subset of contact point pairs includes pairs of contact points on the free fracture fragments and the fracture sections that are geometrically attached. The repositioning and matching cost is used to characterize the matching difference between the free fracture fragments and the fracture sections. Anatomical assignment is performed based on the repositioning matching cost between each free fracture fragment and each fracture section, and each free fracture fragment is repositioned and reconstructed to the fracture section to which it belongs on the stable skull reference block based on the subset of the mating point pairs between each free fracture fragment and each fracture section, thus obtaining a virtual repositioning voxel array. CT image analysis is performed based on the virtual reset voxel array.

2. The intelligent CT image analysis method for maxillofacial fracture classification according to claim 1, characterized in that, Obtain the head CT voxel array of the target object, and discretize and extract the head CT voxel array to obtain anatomical entity data, including: The original image data of the target object acquired by the scanner is analyzed, a global coordinate system is constructed, and the grayscale data of each three-dimensional point in the original image data is mapped to the global coordinate system to generate the head CT voxel array. Bone tissue segmentation was performed on the cranial CT voxel array to extract the stable skull reference block and the set of free fracture fragments from the cranial CT voxel array. Fracture sections are identified at the vertices of the bone surface of the stable skull reference block to obtain a set of fracture sections on the stable skull reference block.

3. The intelligent CT image analysis method for maxillofacial fracture classification according to claim 2, characterized in that, The cranial CT voxel array is subjected to bone tissue segmentation processing to extract the stable skull reference block and free fracture fragment set from the cranial CT voxel array, including: The head CT voxel array was divided into multiple connected components based on three-dimensional connected component analysis. The connected component with the largest volume among the plurality of connected components is determined as the stable skull reference block, and the connected components other than the stable skull reference block whose volume is greater than a preset volume threshold among the plurality of connected components are determined as free fracture fragments in the free fracture fragment set.

4. The intelligent CT image analysis method for maxillofacial fracture classification according to claim 2, characterized in that, Fracture sections are identified at the vertices of the bone surface of the stable skull reference block to obtain a set of fracture sections on the stable skull reference block, including: For each bone surface vertex of the stable skull reference block, the discrete mean curvature corresponding to the bone surface vertex is calculated, and feature vertices are selected from each bone surface vertex based on the discrete mean curvature. Cluster analysis is performed on the feature vertices, and each cluster of points obtained from the clustering is determined as a fracture section in the set of fracture sections.

5. The intelligent CT image analysis method for maxillofacial fracture classification according to claim 1, characterized in that, For each free fracture fragment in the set of free fracture fragments and each fracture section in the set of fracture sections, a multidimensional constraint feature analysis is performed to obtain a subset of fit point pairs between the free fracture fragments and the fracture sections, as well as the repositioning matching cost, including: For each free fracture fragment in the set of free fracture fragments and each fracture section in the set of fracture sections, spatial pose registration is performed on the free fracture fragments and the fracture sections based on geometric morphology constraints to obtain the reset transformation matrix and the corresponding geometric residual between the free fracture fragments and the fracture sections. The reset transformation matrix is ​​used to characterize the spatial pose transformation parameters of the free fracture fragments relative to the fracture sections. The geometric residual is used to characterize the geometric surface fitting error between the free fracture fragments and the fracture sections. Based on the repositioning transformation matrix, coordinate mapping is performed on the vertices of the bone surface of the free fracture fragment, and the set of fitting point pairs is generated according to the distance from each mapped bone surface vertex to the fracture section. The reset matching cost between the free fracture fragment and the fracture section is obtained by performing a reset transformation matrix, geometric residual, and fitting point pair subset.

6. The intelligent CT image analysis method for maxillofacial fracture classification according to claim 5, characterized in that, Based on the repositioning transformation matrix, the geometric residual, and the subset of fitting points, a repositioning matching evaluation is performed to obtain the repositioning matching cost between the free fracture fragment and the fracture section, including: Based on the normal vector distribution of each contact point pair in the set of contact point pairs, the concavity-convexity interlocking degree between the free fracture fragment and the fracture section is calculated; the concavity-convexity interlocking degree is used to characterize the isotropic degree of the contact surface between the free fracture fragment and the fracture section in three-dimensional space. Based on the reset transformation array and the fitting point pair subset, the grayscale information of the corresponding position in the head CT voxel array is traced back to determine the trabecular texture difference between the free fracture fragment and the fracture section; the trabecular texture difference is used to characterize the degree of difference in the direction of trabecular texture on both sides of the fracture section. The repositioning matching cost is determined based on the geometric residual, the occlusal degree, and the trabecular texture difference.

7. The intelligent CT image analysis method for maxillofacial fracture classification according to claim 6, characterized in that, Based on the normal vector distribution of each contact point pair in the subset of contact point pairs, the degree of concavity / convexity interlocking between the free fracture fragment and the fracture cross section is calculated, including: Extract the normal vector of each contact point pair in the set of bonding point pairs, and construct the normal covariance matrix of the bonding surface based on the normal vector of each contact point pair; The normal covariance matrix of the mating surface is decomposed into eigenvalues ​​to obtain the corresponding eigenvalue sequence, and the concave-convex interlocking degree is determined based on the eigenvalue sequence.

8. The intelligent CT image analysis method for maxillofacial fracture classification according to claim 6, characterized in that, Based on the reset transformation matrix and the subset of fitting points, the grayscale information of the corresponding positions in the cranial CT voxel array is traced back to determine the trabecular texture difference between the free fracture fragment and the fracture section, including: Calculate the geometric center and average normal vector of the subset of fitting points, and extend sampling along the positive and negative directions of the average normal vector starting from the geometric center to obtain the sampling points on the reference block side and the sampling points on the fracture fragment side. Based on the reset transformation matrix, the sampling points on the side of the fracture fragment are mapped back to the coordinate space of the free fracture fragment to obtain the index coordinates; The gray-level gradients in the neighborhood of the sampling point on the reference block side and the neighborhood of the index coordinates in the head CT voxel array are calculated respectively, and feature vectors are extracted based on the gray-level gradients to obtain the main texture direction on the reference block side and the main texture direction on the fracture fragment side. The trabecular texture difference is determined based on the main direction of the texture on the side of the reference block and the main direction of the texture on the side of the fracture fragment.

9. The intelligent CT image analysis method for maxillofacial fracture classification according to claim 1, characterized in that, Anatomical assignment is performed based on the reduction matching cost between each free fracture fragment and each fracture section. Then, based on a subset of the mating point pairs between each free fracture fragment and each fracture section, each free fracture fragment is repositioned and reconstructed onto the fracture section assigned to it on the stable skull reference block, resulting in a virtual reduction voxel array, including: A matching cost matrix is ​​constructed based on the repositioning matching cost between each free fracture fragment and each fracture section, and the matching cost matrix is ​​optimized and assigned to obtain an anatomical assignment matrix; the anatomical assignment matrix is ​​used to characterize the correspondence between each free fracture fragment and the fracture section to which it belongs. The voxel data of the stable skull reference block is filled into the corresponding positions of the initial voxel array, and each free fracture fragment is transformed to the global coordinate system where the initial voxel array is located according to the anatomical attribution matrix. The corresponding positions of the fitting point pairs are marked as fracture gap voxels, thereby generating the virtual repositioning voxel array.

10. The intelligent CT image analysis method for maxillofacial fracture classification according to claim 1, characterized in that, CT image analysis based on the virtual repositioning voxel array includes: The standard anatomical atlas is registered to the virtual repositioned voxel array to generate an anatomical pillar region mask; the anatomical pillar region mask is used to identify the spatial extent of each anatomical pillar in the virtual repositioned voxel array. Based on the anatomical strut region mask and the virtual reset voxel array, fracture voxels in each anatomical strut are identified.

Images