Maxillary anterior tooth alveolar bone defect identification method and device

Abstract

The invention discloses a maxillary anterior tooth alveolar bone defect identification method and device. The method comprises the following steps: S1, acquiring a global voxel matrix and voxel size information; segmenting the global voxel matrix to obtain N pieces of tooth information; s2, based on the voxel size information, processing each piece of tooth information to obtain a vector set corresponding to each piece of tooth information; s3, processing each piece of tooth information and the corresponding vector set to obtain a classification image pair set corresponding to each piece of tooth information; s4, processing the classification image pair set corresponding to each piece of tooth information by using a pre-trained multi-scale classification recognition model to obtain a probability vector set corresponding to each piece of tooth information; and S5, processing the probability vector set corresponding to each piece of tooth information to obtain a classification identification result corresponding to each piece of tooth information. According to the method, high-precision classification and identification of alveolar bone defect conditions can be automatically completed.

Description

Technical Field

[0001] This invention belongs to the field of dental damage detection technology, specifically a method and device for identifying alveolar bone defects in the maxillary anterior teeth. Background Technology

[0002] Alveolar bone defects (ABD) are widespread in individuals without a history of orthodontic treatment. For patients requiring anterior tooth retraction after extraction, a thorough examination of the anterior tooth root bone morphology before orthodontic treatment is crucial. Orthodontists need to develop personalized orthodontic treatment plans based on the patient's anterior tooth root bone morphology to ensure safe root movement within the alveolar bone and minimize alveolar bone defects and root resorption. For anterior alveolar bone defects that occur after orthodontic treatment, orthodontists should diagnose them promptly. Mild to moderate alveolar bone defects require close long-term observation, while severe alveolar bone defects necessitate periodontal bone grafting. Therefore, accurately detecting and identifying alveolar bone fenestrations, fissures, and normal areas in maxillary anterior teeth CBCT images is essential for ensuring the safe range of alveolar bone movement during orthodontic surgery and minimizing alveolar bone damage.

[0003] However, existing recognition methods have the following problems: First, they are highly dependent on manual intervention: traditional CBCT image analysis requires experienced physicians to manually reconstruct sagittal and coronal slices, which is time-consuming and inefficient. Second, the recognition accuracy is limited: two-dimensional images cannot fully assess the three-dimensional morphology of the alveolar bone, and existing convolutional neural network (CNN) models are limited by local receptive fields, making it difficult to effectively integrate multi-view features. Third, the degree of automation is insufficient: existing recognition methods require manual pre-definition of regions of interest (ROIs), which cannot achieve end-to-end fully automated recognition. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and device for identifying alveolar bone defects in maxillary anterior teeth, which can automatically complete high-precision classification and identification of alveolar bone defects.

[0005] To address the aforementioned technical problems, a first aspect of the present invention discloses a method for identifying alveolar bone defects in maxillary anterior teeth, the method comprising:

[0006] S1. Obtain the global voxel matrix and voxel size information; perform segmentation processing on the global voxel matrix to obtain N tooth information; N is an integer from 1 to 4;

[0007] The voxel size information includes the first voxel size, the second voxel size, and the third voxel size;

[0008] The tooth information includes tooth position number, local voxel matrix, and mask matrix;

[0009] The global voxel matrix, the local voxel matrix, and the mask matrix are all three-dimensional matrices, and the local voxel matrix and the mask matrix have the same size; the values ​​of the elements in the mask matrix are 0 or 1.

[0010] S2. Based on the voxel size information, process each piece of tooth information to obtain a vector set corresponding to each piece of tooth information;

[0011] The vector set includes a first vector, a second vector, and a third vector;

[0012] S3. Process each piece of tooth information and the corresponding vector set to obtain a set of classification image pairs corresponding to each piece of tooth information;

[0013] The set of classified image pairs includes several classified image pairs; each classified image pair includes a first classified image and a second classified image.

[0014] S4. Using a pre-trained multi-scale classification and recognition model, process the set of classification images corresponding to each piece of tooth information to obtain a set of probability vectors corresponding to each piece of tooth information.

[0015] The probability vector set includes the probability vector corresponding to each classification image pair in the corresponding classification image pair set; the dimension of the probability vector is 5.

[0016] S5. Process the probability vector set corresponding to each piece of tooth information to obtain the classification and recognition result corresponding to each piece of tooth information; the value of the classification and recognition result is an integer from 1 to 5.

[0017] As an optional implementation, in the first aspect of the present invention, the step of processing each piece of tooth information based on the voxel size information to obtain a vector set corresponding to each piece of tooth information includes:

[0018] S21. A preset template information set; the template information set includes M template information items; the template information includes the tooth position number and a two-dimensional template point set;

[0019] S22. Based on the voxel size information, process the mask matrix of each tooth information to obtain the compressed index matrix and the first vector corresponding to each tooth information;

[0020] S23. Based on the voxel size information, process the compressed index matrix and the two-dimensional template point set corresponding to each tooth information to obtain the three-dimensional feature points corresponding to each tooth information.

[0021] S24. Process the first vector and the three-dimensional feature point corresponding to each tooth information to obtain the second vector and the third vector corresponding to each tooth information; the first vector, the second vector and the third vector corresponding to each tooth information constitute the corresponding vector set.

[0022] As an optional implementation, in the first aspect of the present invention, processing the mask matrix of each tooth information based on the voxel size information to obtain a compressed index matrix and the first vector corresponding to each tooth information includes:

[0023] S221. Compress the mask matrix to obtain the corresponding compressed index matrix;

[0024] S222. Construct a model using a standard mask matrix, process the compressed index matrix and the voxel size information, and construct the corresponding standard mask matrix.

[0025] The standard mask matrix construction model is as follows:

[0026]

[0027] In the formula, b i,j and a i,j These are the elements in the i-th row and j-th column of the standard mask matrix and the compressed index matrix, respectively; i is an integer from 1 to 1, and 1 is the total number of rows in the compressed index matrix; j is 1, 2, or 3; Δ1, Δ2, and Δ3 are the first voxel size, the second voxel size, and the third voxel size, respectively;

[0028] S223. Using the feature matrix calculation model, the standard mask moments are processed to obtain the corresponding feature matrix;

[0029] The feature matrix calculation model is as follows:

[0030]

[0031] In the formula, C is the feature matrix, and B is the standard mask matrix;

[0032] S224. Perform eigenvalue decomposition on the feature matrix to obtain three feature pairs; each feature pair includes an eigenvalue and a corresponding eigenvector.

[0033] S225. Normalize the feature vector corresponding to the largest feature value among the three feature pairs to obtain the corresponding first vector.

[0034] As an optional implementation, in the first aspect of the present invention, the step of processing the compressed index matrix and the two-dimensional template point set corresponding to each tooth information based on the voxel size information to obtain the three-dimensional feature points corresponding to each tooth information includes:

[0035] S231. Process the compressed index matrix and the voxel size information corresponding to the tooth information to obtain the corresponding transformed coordinate information and two-dimensional coordinate point set;

[0036] The transformed coordinate information includes a selected origin, a first unit vector, and a second unit vector; the two-dimensional coordinate point set includes J two-dimensional coordinate points; J is an integer greater than 1;

[0037] S232. Register the J two-dimensional coordinate points to the corresponding two-dimensional template point set to obtain the corresponding rotation matrix;

[0038] S233. Multiply the J two-dimensional coordinate points by the rotation matrix respectively to obtain the corresponding J two-dimensional registration coordinate points;

[0039] S234. Perform convex hull detection on the J two-dimensional registration coordinate points to obtain the corresponding two-dimensional vertex set; the two-dimensional vertex set includes several two-dimensional vertices;

[0040] S235. Set the two-dimensional feature point as the lowest two-dimensional vertex in the set of two-dimensional vertices;

[0041] S236. Using a three-dimensional feature point calculation model, process the rotation matrix, the two-dimensional feature points, and the transformation coordinate information to obtain the corresponding three-dimensional feature points;

[0042] The three-dimensional feature point calculation model is as follows:

[0043] [u e ,v e ] T =R -1 [u c ,v c ] T

[0044] Q = u e V1+v e V2+O

[0045] In the formula, Q is a vector composed of the three-dimensional coordinates of the three-dimensional feature points; R -1 [u] is the inverse of the rotation matrix; c ,v c ] TV1 and V2 are two-dimensional column vectors formed by the two-dimensional coordinate values ​​of the two-dimensional feature points; V1 and V2 are column vectors formed by the three-dimensional coordinate values ​​of the selected origin, the first unit vector, and the second unit vector, respectively.

[0046] As an optional implementation, in the first aspect of the present invention, processing the compressed index matrix and the voxel size information corresponding to the tooth information to obtain the corresponding transformed coordinate information and two-dimensional coordinate point set includes:

[0047] S2311. The compressed index matrix and the voxel size information are processed to obtain a mask point set; the mask point set includes I mask points;

[0048] S2312. Project the I mask points onto the line containing the first vector to obtain I first projection points;

[0049] S2313. Set the cutting point as the midpoint of the two farthest first projection points among the I first projection points;

[0050] S2314. Take J mask points from the I mask points whose corresponding voxels intersect with the cutting plane and form a cutting point set; the cutting plane is a plane that passes through the cutting points and whose normal vector is the first vector.

[0051] S2315. Project the J mask points in the cutting point set onto the cutting surface to obtain the corresponding second projection points;

[0052] S2316. Based on the cutting plane, perform coordinate transformation processing on the J second projection points respectively to obtain the corresponding transformed coordinate information and two-dimensional coordinate point set.

[0053] As an optional implementation, in the first aspect of the present invention, processing each piece of tooth information and the corresponding vector set to obtain a set of classification image pairs corresponding to each piece of tooth information includes:

[0054] S31. Process the local voxel matrix and voxel size information of the tooth information to obtain the three-dimensional center point;

[0055] S32. Set N1 sampling points at equal intervals on the main axis; the main axis is a straight line along the first vector and passing through the three-dimensional center point; N1 is an integer greater than 1;

[0056] S33. Based on the N1 sampling points, the second vector, and the third vector, set up N1 sampling plane pairs; the sampling plane pair includes a first sampling plane and a second sampling plane;

[0057] S34. Using N1 sampling plane pairs, sample the three-dimensional voxel region corresponding to the local voxel matrix to obtain N1 initial image pairs; the initial image pairs include a first initial image and a second initial image.

[0058] S35. Using an image enhancement model, enhance the N1 initial image pairs to obtain N1 enhanced image pairs; the enhanced image pairs include a first enhanced image and a second enhanced image;

[0059] S36. The size of each of the N1 enhanced image pairs is adjusted to obtain N1 classified image pairs; the N1 classified image pairs constitute the set of classified image pairs corresponding to the tooth information.

[0060] As an optional implementation, in the first aspect of the present invention, the multi-scale classification and recognition model includes an image preprocessing module, a global feature generation module, a local feature generation module, a feature fusion module, and a classification module;

[0061] The image preprocessing module is data-connected to the global feature generation module and the local feature generation module, and is used to preprocess the classified image pairs to obtain a preprocessed image.

[0062] The global feature generation module is data-connected to the feature fusion module and is used to process the preprocessed image to obtain a global feature map;

[0063] The local feature generation module is data-connected to the feature fusion module and is used to process the preprocessed image to obtain a local feature map;

[0064] The feature fusion module is data-connected to the classification module and is used to perform feature fusion processing on the global feature map and the local feature map to obtain a fused feature map.

[0065] The classification module is used to classify the fused feature map to obtain the probability vector.

[0066] As an optional implementation, in the first aspect of the present invention, the feature fusion module includes an initial fusion unit and R intermediate fusion units; R is an integer greater than 2.

[0067] The initial fusion unit is used to process the global feature map and the local feature map to obtain an initial fused feature map;

[0068] The intermediate fusion unit is used to process the global feature map, the local feature map, and the previous feature map to obtain an intermediate fused feature map;

[0069] The preceding feature map of the first intermediate fusion unit is the initial fusion feature map; the preceding feature maps of the subsequent R-1 intermediate fusion units are the intermediate fusion feature maps output by the previous intermediate fusion unit.

[0070] As an optional implementation, in the first aspect of the present invention, the intermediate fusion unit includes a channel attention subunit, a first multiplication subunit, a spatial attention subunit, a second multiplication subunit, a first splicing subunit, a first convolution subunit, an average pooling subunit, a second splicing subunit, a second convolution subunit, a residual subunit, and an addition subunit.

[0071] The global feature map is input to the input terminal of the channel attention subunit; the output terminal of the channel attention subunit is connected to the first input terminal of the first multiplication subunit; the global feature map is input to the second input terminal of the first multiplication subunit; and the output terminal of the first multiplication subunit is connected to the first input terminal of the first splicing subunit.

[0072] The spatial attention subunit receives the local feature map as its input; the spatial attention subunit outputs the first input of the second multiplication subunit; the second input of the second multiplication subunit receives the local feature map; and the second output of the second multiplication subunit outputs the second input of the first splicing subunit.

[0073] The input terminal of the first convolutional subunit of the first intermediate fusion unit is connected to the output terminal of the initial fusion unit; the input terminals of the first convolutional subunits of the subsequent R-1 intermediate fusion units are connected to the output terminal of the previous intermediate fusion unit.

[0074] The output of the first convolutional subunit is connected to the input of the average pooling subunit; the output of the average pooling subunit is connected to the first input of the second concatenation unit and the first input of the addition subunit; the second and third inputs of the second concatenation subunit are respectively input to the global feature map and the local feature map.

[0075] The output of the second stitching unit is connected to the input of the second convolution subunit; the output of the second convolution subunit is connected to the third input of the first stitching subunit; the output of the first stitching subunit is connected to the input of the residual subunit; the output of the residual subunit is connected to the second input of the addition subunit; and the output of the addition subunit serves as the output of the feature fusion module.

[0076] The second aspect of the present invention discloses a computer storage medium storing computer instructions, which, when invoked, are used to execute some or all of the steps in the maxillary anterior alveolar bone defect identification method disclosed in the first aspect of the present invention.

[0077] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0078] This invention processes each tooth's information based on voxel size information to obtain a corresponding vector set, and further obtains a corresponding set of classified image pairs. Using a multi-scale classification and recognition model, each set of classified image pairs is processed to obtain a corresponding set of probability vectors, and further obtains the corresponding classification and recognition results. This eliminates the need for manual intervention and automatically completes high-precision classification and recognition of alveolar bone defects. Attached Figure Description

[0079] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings that need to be activated in the description of the embodiments 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.

[0080] Figure 1 This is a flowchart illustrating a method for identifying alveolar bone defects in the maxillary anterior teeth, as disclosed in an embodiment of the present invention.

[0081] Figure 2 This is a schematic diagram of the structure of a multi-scale classification and recognition model for a method for identifying alveolar bone defects in maxillary anterior teeth disclosed in an embodiment of the present invention.

[0082] Figure 3 This is a schematic diagram of the global feature generation module of a multi-scale classification recognition model for a maxillary anterior alveolar bone defect recognition method disclosed in an embodiment of the present invention.

[0083] Figure 4 This is a schematic diagram of the local feature generation module of a multi-scale classification recognition model for a maxillary anterior alveolar bone defect recognition method disclosed in an embodiment of the present invention.

[0084] Figure 5 This is a schematic diagram of the feature fusion module of a multi-scale classification and recognition model for a maxillary anterior alveolar bone defect identification method disclosed in an embodiment of the present invention.

[0085] Figure 6 This is a schematic diagram of the intermediate fusion unit of a multi-scale classification and recognition model for a method for identifying alveolar bone defects in maxillary anterior teeth, as disclosed in an embodiment of the present invention.

[0086] Figure 7 This is a schematic diagram of the structure of a maxillary anterior alveolar bone defect identification device disclosed in an embodiment of the present invention.

[0087] Figure 8 This is a schematic diagram of another maxillary anterior alveolar bone defect identification device disclosed in an embodiment of the present invention. Detailed Implementation

[0088] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0089] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0090] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0091] Example 1

[0092] Please see Figure 1-6 . Figure 1 This is a flowchart illustrating a method for identifying alveolar bone defects in the maxillary anterior teeth, as disclosed in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a multi-scale classification and recognition model for a method for identifying alveolar bone defects in maxillary anterior teeth disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of the global feature generation module of the multi-scale classification and recognition model of a method for identifying alveolar bone defects in maxillary anterior teeth disclosed in an embodiment of the present invention; Figure 4This is a schematic diagram of the local feature generation module of the multi-scale classification recognition model of a method for identifying alveolar bone defects in maxillary anterior teeth disclosed in an embodiment of the present invention; Figure 5 This is a schematic diagram of the feature fusion module of a multi-scale classification and recognition model for a method for identifying alveolar bone defects in maxillary anterior teeth, as disclosed in an embodiment of the present invention. Figure 6 This is a schematic diagram of the intermediate fusion unit of a multi-scale classification and recognition model for a method for identifying alveolar bone defects in the maxillary anterior teeth, as disclosed in an embodiment of the present invention. Figure 1 The described method for identifying alveolar bone defects in the maxillary anterior teeth is applied in the field of dental damage detection, such as the identification of alveolar bone defects. This invention does not limit its application to specific applications. Figure 1 As shown, the method includes:

[0093] S1. Obtain the global voxel matrix and voxel size information; perform segmentation processing on the global voxel matrix to obtain N tooth information; N is an integer from 1 to 4.

[0094] It should be noted that, in order to obtain the global voxel matrix and voxel size information, the patient's maxillary anterior teeth were first subjected to CBCT scanning to obtain a set of two-dimensional grayscale tomographic images (such as axial, coronal, and sagittal views), which were stored in DICOM format. Then, the above two-dimensional tomographic images were converted into a global voxel matrix using a reconstruction algorithm. Each element in the matrix corresponds to a cuboid voxel, and its element value represents the X-ray attenuation coefficient of the human tissue in that cuboid voxel.

[0095] The above voxel size information includes the first voxel size, the second voxel size, and the third voxel size.

[0096] It should be noted that the first voxel size, the second voxel size, and the third voxel size mentioned above represent the length of each cuboid voxel in the X, Y, and Z axes of the preset global coordinate system, respectively.

[0097] The aforementioned tooth information includes tooth position number, local voxel matrix, and mask matrix.

[0098] It should be noted that the tooth position numbering mentioned above is a standardized numbering system used in dentistry to accurately identify the position of teeth. Its core function is to clearly and consistently identify each tooth in scenarios such as medical records, doctor-patient communication, and academic exchanges, avoiding confusion. The tooth position numbering above can adopt the ISO system, ADA system, or Palmer system, etc., and this embodiment of the invention does not limit it.

[0099] The global voxel matrix, the local voxel matrix, and the mask matrix are all three-dimensional matrices, and the size of the local voxel matrix and the mask matrix are exactly the same; the value of the element in the mask matrix is ​​0 or 1.

[0100] It should be noted that the dimensions of the local voxel matrix and the mask matrix are exactly the same, meaning that the number of rows, columns and depths of the two are equal.

[0101] It should be noted that the local voxel matrix in the above tooth information is the submatrix in the global voxel matrix corresponding to the region where the tooth position number of the tooth information is located; when the element value of the above mask matrix is ​​1 or 0, it means that the cuboid voxel corresponding to the corresponding element in the local voxel matrix of the same tooth information represents tooth tissue or does not represent tooth tissue.

[0102] Optionally, the above segmentation process can be implemented by processing the global voxel matrix using a neural network based on the 3D U-Net architecture. This method is known and will not be described in detail in this embodiment of the invention.

[0103] S2. Based on the above voxel size information, process each of the above tooth information to obtain the vector set corresponding to each of the above tooth information.

[0104] It should be noted that there is a one-to-one correspondence between the information of N teeth and the N sets of vectors.

[0105] The aforementioned vector set includes the first vector, the second vector, and the third vector.

[0106] It should be noted that the first vector mentioned above is used as the normal vector of the cross section of the corresponding tooth; the second vector mentioned above is used as the normal vector of the coronal plane of the corresponding tooth; and the third vector mentioned above is used as the normal vector of the sagittal plane of the corresponding tooth.

[0107] S3. Process each of the above-mentioned tooth information and the corresponding vector set to obtain a set of classification image pairs corresponding to each of the above-mentioned tooth information.

[0108] The aforementioned set of classified image pairs includes several classified image pairs; the aforementioned classified image pairs include a first classified image and a second classified image.

[0109] S4. Using a pre-trained multi-scale classification and recognition model, process the set of classification images corresponding to each of the above tooth information to obtain a set of probability vectors corresponding to each of the above tooth information.

[0110] The aforementioned set of probability vectors includes the probability vector corresponding to each of the aforementioned classified image pairs in the aforementioned set of classified image pairs; the dimension of the aforementioned probability vectors is 5.

[0111] It should be noted that the above multi-scale classification and recognition model processes each class image pair in the set of class image pairs separately to obtain the probability vector corresponding to each class image pair.

[0112] S5. Process the probability vector set corresponding to each of the above tooth information to obtain the classification and recognition result corresponding to each of the above tooth information; the value of the classification and recognition result is an integer from 1 to 5.

[0113] It should be noted that the above classification and identification results are used to indicate the degree of cracking of the maxillary anterior teeth, thereby reflecting the alveolar bone defect. When the above identification result is 1, 2, 3, 4 or 5, it means that the tooth corresponding to the tooth position number in the corresponding tooth information has normal, mild bone cracking, moderate bone cracking, severe bone cracking and bone fenestration, respectively.

[0114] In an optional embodiment, the above-mentioned processing of each of the above-mentioned tooth information based on the voxel size information to obtain a vector set corresponding to each of the above-mentioned tooth information includes:

[0115] S21. Preset template information set; the above template information set includes M template information; the above template information includes the above tooth position number and two-dimensional template point set.

[0116] Preferably, M is 4. The tooth position numbers in the above 4 template information correspond to the 4 maxillary anterior teeth respectively.

[0117] It should be noted that the two-dimensional template point set in each template information is a set of several two-dimensional points on the cross-sectional (i.e., horizontal section) image of the tooth corresponding to the tooth position number of the template information at 1 / 2 distance from the root apex for any patient.

[0118] It should be noted that the two-dimensional coordinate system (defined by mutually perpendicular u-axis and v-axis) corresponding to each two-dimensional template point set is different. The v-axis is the left and right axis of symmetry of the corresponding cross-sectional image, and it points away from the lingual side of the corresponding tooth.

[0119] S22. Based on the above voxel size information, the mask matrix of each of the above tooth information is processed to obtain the compressed index matrix and the first vector corresponding to each of the above tooth information.

[0120] S23. Based on the above voxel size information, process the above compressed index matrix and the above two-dimensional template point set corresponding to each of the above tooth information to obtain the three-dimensional feature points corresponding to each of the above tooth information.

[0121] It should be noted that when a certain tooth information corresponds to a certain two-dimensional template point set, it means that the two contain the same tooth position number.

[0122] S24. Process the first vector and the three-dimensional feature point corresponding to each of the above tooth information to obtain the second vector and the third vector corresponding to each of the above tooth information; the first vector, the second vector and the third vector corresponding to each of the above tooth information constitute the corresponding vector set.

[0123] In another optional embodiment, based on the voxel size information, the mask matrix for each of the tooth pieces of information is processed to obtain a compressed index matrix and the first vector corresponding to each of the tooth pieces of information, including:

[0124] S221. Compress the above mask matrix to obtain the corresponding compressed index matrix.

[0125] It should be noted that the above compression of the mask matrix is ​​to take the indices (ix, iy, iz) of all elements with a value of 1 in the mask matrix in the X, Y, and Z axes as row vectors and concatenate them to obtain a compressed index matrix.

[0126] S222. Using the standard mask matrix, construct a model by processing the above compressed index matrix and the above voxel size information to construct the corresponding standard mask matrix.

[0127] The standard mask matrix construction model is as follows:

[0128]

[0129] In the formula, b i,j and a i,j These are the elements in the i-th row and j-th column of the standard mask matrix and the compressed index matrix, respectively; i is an integer from 1 to I, and I is the total number of rows in the compressed index matrix; j is 1, 2, or 3; Δ1, Δ2, and Δ3 are the first voxel size, the second voxel size, and the third voxel size, respectively.

[0130] S223. Using the feature matrix calculation model, process the above standard mask moments to obtain the corresponding feature matrix.

[0131] The above feature matrix calculation model is as follows:

[0132]

[0133] In the formula, C is the aforementioned feature matrix, and B is the aforementioned standard mask matrix.

[0134] S224. Perform eigenvalue decomposition on the above feature matrix to obtain 3 feature pairs; each feature pair includes an eigenvalue and its corresponding eigenvector.

[0135] S225. Normalize the eigenvector corresponding to the largest eigenvalue among the three eigenpairs to obtain the corresponding first vector.

[0136] In another optional embodiment, based on the aforementioned voxel size information, the aforementioned compressed index matrix and the aforementioned two-dimensional template point set corresponding to each of the aforementioned tooth information are processed to obtain three-dimensional feature points corresponding to each of the aforementioned tooth information, including:

[0137] S231. Process the compressed index matrix and voxel size information corresponding to the above tooth information to obtain the corresponding transformed coordinate information and two-dimensional coordinate point set.

[0138] The aforementioned transformed coordinate information includes the selected origin, the first unit vector, and the second unit vector; the aforementioned two-dimensional coordinate point set includes J two-dimensional coordinate points; J is an integer greater than 1.

[0139] It should be noted that the selected origin is a three-dimensional coordinate point in the global coordinate system; the first unit vector and the second unit vector are both three-dimensional vectors in the global coordinate system.

[0140] S232. Register the J two-dimensional coordinate points mentioned above to the corresponding two-dimensional template point set to obtain the corresponding rotation matrix.

[0141] Optionally, the J two-dimensional coordinate points mentioned above are registered to the corresponding two-dimensional template point set using the 2D-ICP algorithm to obtain the rotation matrix.

[0142] It should be noted that the rotation matrix mentioned above is a 2×2 matrix.

[0143] S233. Multiply the J two-dimensional coordinate points mentioned above with the rotation matrix respectively to obtain the corresponding J two-dimensional registration coordinate points.

[0144] It should be noted that the multiplication of the two-dimensional coordinate points with the rotation matrix refers to the multiplication of the two-dimensional column vector formed by the two-dimensional coordinate values ​​of the two-dimensional coordinate points with the rotation matrix. After rotation, the coordinate system of the two-dimensional registration coordinate points is the same as that of the two-dimensional template point set.

[0145] S234. Perform convex hull detection on the J two-dimensional registration coordinate points mentioned above to obtain the corresponding two-dimensional vertex set; the two-dimensional vertex set mentioned above includes several two-dimensional vertices.

[0146] It should be noted that the Graham algorithm can be used for the convex hull detection described above, and this embodiment of the invention is not limited thereto. The vertex set described above is the set of vertices of the smallest polygon (i.e., the convex hull) that surrounds the J two-dimensional registration coordinate points.

[0147] S235. Set the two-dimensional feature point to the lowest two-dimensional vertex in the above two-dimensional vertex set.

[0148] It should be noted that the lowest vertex in the above set of two-dimensional vertices refers to the two-dimensional vertex with the smallest v-axis coordinate.

[0149] S236. Using the three-dimensional feature point calculation model, the above rotation matrix, the above two-dimensional feature points, and the above transformation coordinate information are processed to obtain the corresponding three-dimensional feature points.

[0150] The above three-dimensional feature point calculation model is as follows:

[0151] [u e ,v e ] T =R -1 [u c ,v c ] T

[0152] Q = u e V1+v e V2+O

[0153] In the formula, Q is a vector composed of the three-dimensional coordinates of the aforementioned three-dimensional feature points; R -1 This is the inverse of the rotation matrix mentioned above; [u c ,v c ] T V1 and V2 are the two-dimensional column vectors formed by the two-dimensional coordinate values ​​of the two-dimensional feature points mentioned above; V1 and V2 are the column vectors formed by the three-dimensional coordinate values ​​of the selected origin, the first unit vector, and the second unit vector, respectively.

[0154] In another optional embodiment, the above-mentioned processing of the compressed index matrix and voxel size information corresponding to the tooth information to obtain the corresponding transformed coordinate information and two-dimensional coordinate point set includes:

[0155] S2311. Process the above compressed index matrix and the above voxel size information to obtain a mask point set; the above mask point set includes I mask points.

[0156] It should be noted that the above processing of the compressed index matrix and voxel size information involves multiplying the first, second, and third columns of the compressed index matrix by the first, second, and third voxel sizes, respectively, to obtain the local coordinate matrix D. Each row of D (d... x ,d y ,d z () corresponds to a mask point in the mask point set.

[0157] S2312. Project the I mask points mentioned above onto the line containing the first vector to obtain I first projection points.

[0158] S2313. Set the cutting point as the midpoint of the two farthest first projection points among the I first projection points.

[0159] S2314. Take J mask points from the I mask points whose corresponding voxels intersect with the cutting plane and form a cutting point set; the cutting plane is a plane that passes through the cutting points and whose normal vector is the first vector.

[0160] It should be noted that the voxel corresponding to the mask point refers to the cuboid voxel with the mask point as its centroid.

[0161] S2315. Project the J mask points in the above cutting point set onto the above cutting surface to obtain the corresponding second projection points.

[0162] S2316. Based on the above cutting plane, perform coordinate transformation on the J second projection points respectively to obtain the corresponding transformed coordinate information and two-dimensional coordinate point set.

[0163] In another optional embodiment, based on the above-mentioned cutting plane, coordinate transformation processing is performed on J of the above-mentioned second projection points to obtain corresponding transformed coordinate information and a two-dimensional coordinate point set, including:

[0164] S23161. Set the selected origin point as any point on the cutting surface that is different from all J of the second projection points.

[0165] S23162. Select any two perpendicular unit vectors V1 and V2 on the above-mentioned cutting plane, and use them as the first unit vector and the second unit vector, respectively.

[0166] S23163. Using a two-dimensional coordinate calculation model, the selected origin, the first unit vector, the second unit vector, and each of the second projection points are processed to obtain the two-dimensional coordinate points corresponding to each of the second projection points.

[0167] The above two-dimensional coordinate calculation model is as follows:

[0168] (u ii ,v ii )=((PO)·V1,(PO)·V2)

[0169] In the formula, (u ii ,v ii) represents the two-dimensional coordinate value of the i-th two-dimensional coordinate point; P represents the vector composed of the three-dimensional coordinate values ​​of the i-th second projection point; O represents the vector composed of the three-dimensional coordinate values ​​of the selected origin; · represents the vector inner product operation.

[0170] In another optional embodiment, the above-described processing of the first vector and the three-dimensional feature points corresponding to each of the above-described tooth information to obtain the second vector and the third vector corresponding to each of the above-described tooth information includes:

[0171] S241. Set the construction point as any point on the line containing the first vector mentioned above.

[0172] S242. Using the second vector calculation model, the above-mentioned construction points and the above-mentioned three-dimensional feature points are processed to obtain the corresponding second vector.

[0173] The second vector calculation model mentioned above is as follows:

[0174]

[0175] In the formula, V b and V a Q1 and Q2 are the second vector and the first vector corresponding to the tooth information, respectively; Q2 and Q1 are column vectors composed of the three-dimensional coordinate values ​​of the three-dimensional feature points and the three-dimensional construction points corresponding to the tooth information, respectively.

[0176] S243. Calculate the cross product of the first vector and the corresponding second vector to obtain the corresponding third vector.

[0177] In another optional embodiment, the above processing of each of the above-mentioned tooth information and the corresponding vector set to obtain a set of classification image pairs corresponding to each of the above-mentioned tooth information includes:

[0178] S31. Process the local voxel matrix and voxel size information of the tooth information to obtain the three-dimensional center point.

[0179] Optionally, the coordinates of the three-dimensional center point are (Δ1×II / 2, Δ2×JJ / 2, Δ3×KK / 2), where II, JJ, and KK are the number of elements in the local voxel matrix along the X, Y, and Z axes, respectively.

[0180] S32. Set N1 sampling points at equal intervals on the main axis; the main axis is a straight line along the first vector and passing through the three-dimensional center point; N1 is an integer greater than 1.

[0181] Preferably, the interval between the adjacent sampling points is 0.25 mm.

[0182] Preferably, the above-mentioned N1 sampling points include the above-mentioned three-dimensional center point, and N2 sampling points on both sides of the three-dimensional center point, where N2 is an integer greater than 1, and N1 = 2 × N2 + 1.

[0183] S33. Based on the N1 sampling points, the second vector, and the third vector, set up N1 sampling plane pairs; the sampling plane pairs include the first sampling plane and the second sampling plane.

[0184] It should be noted that the first and second sampling planes of the above-mentioned nth sampling plane pair both pass through the nth sampling point, and take the corresponding second and third vectors as normal vectors respectively.

[0185] S34. Using N1 sampling plane pairs, sample the three-dimensional voxel regions corresponding to the local voxel matrix to obtain N1 initial image pairs; the initial image pairs include a first initial image and a second initial image.

[0186] It should be noted that the three-dimensional voxel region corresponding to the aforementioned local voxel matrix is ​​a cuboid region, including cuboid voxels centered at each element in the local voxel matrix, and the center coordinates of the cuboid voxel corresponding to the element with index (ia, ib, ic) in the local voxel matrix are (ia×Δ1, ib×Δ1). 2, (ic×Δ3). In addition, the voxel value of each cuboid voxel is the corresponding element value in the local voxel matrix.

[0187] It should be noted that the above sampling process utilizes trilinear interpolation, which slices the three-dimensional voxel region corresponding to the local voxel matrix based on the first and second sampling planes of each sampling plane pair, respectively, to obtain an initial image pair corresponding to each sampling plane pair. The first initial image is obtained by slicing from the first sampling plane, and the second initial image is obtained by slicing from the second sampling plane. The above trilinear interpolation method is a known method, and will not be described further in this embodiment.

[0188] S35. Using an image enhancement model, enhance the N1 initial image pairs to obtain N1 enhanced image pairs; the enhanced image pairs include a first enhanced image and a second enhanced image.

[0189] S36. The size of each of the N1 enhanced image pairs is adjusted to obtain N1 classified image pairs; the N1 classified image pairs constitute the set of classified image pairs corresponding to the above-mentioned tooth information.

[0190] It should be noted that the above-described resizing process involves scaling all the first and second enhanced images to obtain first and second classification images of the same size. For example, the first enhanced image, after scaling, yields a first classification image with a size of 150×224 pixels; the second enhanced image, after scaling, yields a second classification image with the same size of 150×224 pixels.

[0191] In yet another optional embodiment, the expression for the above image enhancement model is:

[0192] V 1,l =αW 1,l-1 +βW 1,l +γW 1,l+1

[0193] V 1,1 =(α+β)W 1,1 +γW 1,2

[0194] V 1,L =αW 1,L-1 +(β+γ)W 1,L

[0195] V 2,l =αW 2,l-1 +βW 2,l +γW 2,l+1

[0196] V 2,1 =(α+β)W 2,1 +γW 2,2

[0197] V 2,L =αW 2,L-1 +(β+γ)W 2,L

[0198] In the formula, l is an integer from 2 to L-1, and L = N1; W 1,ll and W 2,ll These are the first initial image and the second initial image of the ll-th initial image pair, respectively; V 1,ll and V 2,ll These are the first enhanced image and the second enhanced image of the ll-th enhanced image pair, respectively; ll is an integer from 1 to L; α, β and γ are the preset first coefficient, second coefficient and third coefficient, respectively, and α+β+γ=1.

[0199] Preferably, α, β and γ are 0.3, 0.4 and 0.3, respectively.

[0200] In yet another alternative embodiment, such as Figure 2As shown, the above multi-scale classification and recognition model includes an image preprocessing module, a global feature generation module, a local feature generation module, a feature fusion module, and a classification module.

[0201] The image preprocessing module described above is data-connected to the global feature generation module and the local feature generation module, and is used to preprocess the classified image pairs to obtain preprocessed images.

[0202] It should be noted that the above preprocessing of the classified image pairs involves first stitching the first and second classification images of the pair together horizontally to obtain a stitched image, and then scaling the stitched image to obtain a preprocessed image. For example, the first and second classification images, both with a size of 150×224 pixels, are stitched together to obtain a stitched image with a size of 300×224 pixels, which is then further scaled to obtain a preprocessed image of 224×224 pixels.

[0203] The aforementioned global feature generation module is data-connected to the aforementioned feature fusion module and is used to process the aforementioned preprocessed image to obtain a global feature map.

[0204] The aforementioned local feature generation module is data-connected to the aforementioned feature fusion module and is used to process the aforementioned preprocessed image to obtain a local feature map.

[0205] The aforementioned feature fusion module is connected to the aforementioned classification module and is used to perform feature fusion processing on the aforementioned global feature map and the aforementioned local feature map to obtain a fused feature map.

[0206] The classification module described above is used to classify the fused feature map to obtain the probability vector.

[0207] It should be noted that the above multi-scale classification and recognition model achieved an accuracy of 92.9% and an F1 score of 89.8% on the test set, which is 5%-8% higher than traditional models (CNN, SwinTransformer, etc.).

[0208] In yet another alternative embodiment, such as Figure 3 As shown, the global feature generation module includes a first downsampling unit and a global feature extraction unit.

[0209] The first downsampling unit is connected to the global feature extraction unit and is used to perform a first downsampling process on the stitched image to obtain a first downsampled image.

[0210] It should be noted that the aforementioned first downsampling is achieved through image patch merging. For example, by performing a 4-fold first downsampling on a stitched image of size 224×224 pixels, a first downsampled image of size 56×56 pixels is obtained.

[0211] The aforementioned global feature extraction unit is used to process the aforementioned first downsampled image to obtain the aforementioned global feature map.

[0212] Optionally, the aforementioned global feature extraction unit is built based on the Transformer architecture.

[0213] It should be noted that the aforementioned global feature extraction unit can dynamically model the global interaction between different image patches, capture the dependencies between distant pixels / regions, and achieve global semantic extraction (such as the boundary continuity between the sagittal root morphology and the coronal alveolar bone).

[0214] In yet another alternative embodiment, such as Figure 4 As shown, the aforementioned local feature generation module includes a second downsampling unit and a local feature extraction unit.

[0215] The second downsampling unit is connected to the local feature extraction unit and is used to perform a second downsampling process on the stitched image to obtain a second downsampling image.

[0216] It should be noted that the second downsampling mentioned above is achieved through a convolution operation (kernel size 4×4, stride 2). For example, by performing the second downsampling on a stitched image of size 224×224 pixels, a local feature map of size 112×112 pixels is obtained.

[0217] The aforementioned local feature extraction unit is used to process the aforementioned second downsampled image to obtain the aforementioned local feature map.

[0218] Optionally, the aforementioned local feature extraction unit is built on the ResNet architecture, thereby efficiently extracting multi-scale local features while maintaining gradient flow, which enhances detail capture and prevents degradation.

[0219] In yet another alternative embodiment, such as Figure 5 As shown, the feature fusion module includes an initial fusion unit and R intermediate fusion units; R is an integer greater than 2.

[0220] The aforementioned initial fusion unit is used to process the aforementioned global feature map and the aforementioned local feature map to obtain an initial fused feature map.

[0221] The aforementioned intermediate fusion unit is used to process the aforementioned global feature map, the aforementioned local feature map, and the previous-level feature map to obtain an intermediate fused feature map.

[0222] The preceding feature map of the first intermediate fusion unit is the initial fusion feature map; the preceding feature maps of the subsequent R-1 intermediate fusion units are the intermediate fusion feature maps output by the previous intermediate fusion unit.

[0223] Preferably, R is 3.

[0224] As can be seen, the feature fusion module described above improves the expressive power of the output features by fusioning the global feature map and the local feature map in a hierarchical manner multiple times.

[0225] In yet another alternative embodiment, such as Figure 6 As shown, the aforementioned intermediate fusion unit includes a channel attention subunit, a first multiplication subunit, a spatial attention subunit, a second multiplication subunit, a first splicing subunit, a first convolution subunit, an average pooling subunit, a second splicing subunit, a second convolution subunit, a residual subunit, and an addition subunit.

[0226] The global feature map is input to the input terminal of the channel attention subunit; the output terminal of the channel attention subunit is connected to the first input terminal of the first multiplication subunit; the global feature map is input to the second input terminal of the first multiplication subunit; and the output terminal of the first multiplication subunit is connected to the first input terminal of the first splicing subunit.

[0227] It should be noted that the aforementioned channel attention subunit can be constructed using the Squeeze-and-Excitation module, and this embodiment of the invention is not limited thereto. The aforementioned channel attention subunit generates channel attention weights by weighting the channel dimensions of the input feature map through global average pooling and fully connected layers, thereby enhancing the feature correlation between channels.

[0228] The spatial attention subunit receives the local feature map as input; the spatial attention subunit outputs the first input of the second multiplication subunit; the second input of the second multiplication subunit receives the local feature map; and the second output of the second multiplication subunit outputs the second input of the first splicing subunit.

[0229] It should be noted that the aforementioned spatial attention subunit can be constructed using the SAM structure in CBAM (Convolutional Block Attention Module), and this embodiment of the invention is not limited thereto. The aforementioned spatial attention subunit generates a spatial attention map by performing average pooling and max pooling operations along the channel dimension, which is used to highlight the importance of key spatial locations in the feature map.

[0230] The input terminal of the first convolutional subunit of the first intermediate fusion unit is connected to the output terminal of the initial fusion unit; the input terminals of the first convolutional subunits of the subsequent R-1 intermediate fusion units are connected to the output terminal of the previous intermediate fusion unit.

[0231] It should be noted that the first convolutional subunit mentioned above is used to perform 1×1 convolution operations, which can adjust the number of channels of the input early features, reduce the amount of computation, and increase the feature representation ability.

[0232] The output of the first convolutional subunit is connected to the input of the average pooling subunit; the output of the average pooling subunit is connected to the first input of the second concatenation subunit and the first input of the addition subunit; the second and third inputs of the second concatenation subunit are respectively input to the global feature map and the local feature map.

[0233] It should be noted that the above average pooling subunit is used to perform average pooling operation on the feature map after the above 1×1 convolution, reduce the spatial resolution of the feature map, and further extract global information.

[0234] The output of the second stitching unit is connected to the input of the second convolution subunit; the output of the second convolution subunit is connected to the third input of the first stitching subunit; the output of the first stitching subunit is connected to the input of the residual subunit; the output of the residual subunit is connected to the second input of the addition subunit; and the output of the addition subunit serves as the output of the feature fusion module.

[0235] It should be noted that the above residual sub-units are constructed based on Inverted Residual MLP (IRMLP) to enhance feature representation capabilities and mitigate gradient vanishing or exploding. Their input-output relationship is as follows, where LN represents the layer normalization operation, and Conv1 and Conv3 represent 1×1 and 3×3 convolution operations, respectively:

[0236] Output(x)=Conv1(Conv1(Conv3(LN(Input)+LN(Input))))

[0237] It should be noted that the second convolutional subunit mentioned above is used to perform 3×3 convolution operations.

[0238] As can be seen, the aforementioned intermediate fusion unit, through global (channel attention) and local (spatial attention) branches, accurately mines important information from different dimensions of features, thereby improving feature discriminative power. Furthermore, through a hierarchical fusion mechanism, the original multi-source features are first concatenated (via the second concatenation sub-unit) and refined via convolution; then, weighted enhanced features are concatenated (via the first concatenation sub-unit), and deep fusion is performed using IRMLP to ensure feature richness and expressiveness.

[0239] In another alternative embodiment, the initial fusion unit is obtained by removing the first convolution sub-unit, the average pooling sub-unit, the second splicing sub-unit, and the second convolution sub-unit from the intermediate fusion unit.

[0240] In another optional embodiment, the classification module includes a pooling layer, a layer normalization layer, a fully connected layer, and a Softmax layer connected in sequence.

[0241] The pooling layer described above is used to perform global average pooling (GAP) on the fused feature map to obtain a compressed vector. For example, after GAP processing, the dimension of the compressed vector can be 1×1×512.

[0242] The aforementioned layer normalization layer is used to perform layer normalization on the compressed vector to obtain a layer normalized vector.

[0243] The fully connected layer described above is used to perform category mapping on the normalized vectors of the above layers to obtain a recognition vector with a dimension of 5.

[0244] The aforementioned Softmax layer is used to process the aforementioned recognition vector to obtain a probability vector with a dimension of 5.

[0245] It should be noted that the Softmax layer described above processes the recognition vector using the following expression, where p jj Let z be the j-th element of the probability vector, representing the probability value that the recognition result is jj. jj and z kk These are the jj-th and kk-th elements of the recognition vector, respectively.

[0246]

[0247] In yet another optional embodiment, the loss function expression used for pre-training the above-mentioned multi-scale classification and recognition model is:

[0248]

[0249] In the formula, L is the total loss value; P k N represents the value of the k-th element of the probability vector; kThis represents the number of classification and recognition results that are actually k among all training samples.

[0250] As can be seen, by using the above loss function, when the number of samples for the five classification results is unbalanced (for example, there are few samples of moderate bone fracture), the weight of the recognition result with fewer training samples can be increased, thereby increasing its contribution to the total loss value and thus improving the classification ability of the multi-scale classification recognition model after pre-training.

[0251] In another optional embodiment, the above-mentioned processing of the probability vector set corresponding to each of the above-mentioned tooth information to obtain the classification and recognition result corresponding to each of the above-mentioned tooth information includes:

[0252] S51. For each of the above-mentioned tooth information corresponding to the above-mentioned probability vector set, execute S52 to S53.

[0253] S52. Take each of the above probability vectors in the above probability vector set as a row vector and concatenate them to obtain a probability matrix.

[0254] S53. Set the above classification and recognition result as the column number of the largest element in the above probability matrix.

[0255] It is evident that the method for identifying alveolar bone defects in maxillary anterior teeth described in the embodiments of the present invention can eliminate the need for manual intervention and automatically complete the high-precision classification and identification of alveolar bone defects.

[0256] Example 2

[0257] Please see Figure 7 , Figure 7 This is a schematic diagram of the structure of a maxillary anterior alveolar bone defect identification device disclosed in an embodiment of the present invention. Figure 7 The described device can be applied to the field of tooth damage detection, such as the identification of alveolar bone defects; however, the embodiments of this invention are not limited to this. Figure 7 As shown, the device may include:

[0258] The information acquisition module 201 is used to acquire the global voxel matrix and voxel size information; the global voxel matrix is ​​segmented to obtain N tooth information.

[0259] The vector generation module 202 is used to process each of the above-mentioned tooth information based on the above-mentioned voxel size information to obtain a vector set corresponding to each of the above-mentioned tooth information.

[0260] The classification image generation module 203 is used to process each of the above-mentioned tooth information and the corresponding vector set to obtain a set of classification image pairs corresponding to each of the above-mentioned tooth information.

[0261] The probability vector generation module 204 is used to process the set of classification images corresponding to each of the above-mentioned tooth information using a pre-trained multi-scale classification and recognition model to obtain a set of probability vectors corresponding to each of the above-mentioned tooth information.

[0262] The classification result generation module 205 is used to process the probability vector set corresponding to each of the above tooth information to obtain the classification recognition result corresponding to each of the above tooth information.

[0263] As can be seen, the maxillary anterior alveolar bone defect identification device described in the embodiments of the present invention can eliminate the need for manual intervention and automatically complete the high-precision classification and identification of alveolar bone defects.

[0264] Example 3

[0265] Please see Figure 8 , Figure 8 This is a schematic diagram of another maxillary anterior alveolar bone defect identification device disclosed in an embodiment of the present invention. Figure 8 The described maxillary anterior alveolar bone defect identification device can be applied to the field of tooth damage detection, such as the identification of alveolar bone defects. This invention does not limit its application to specific applications. Figure 8 As shown, the maxillary anterior alveolar bone defect identification device may include the following parts:

[0266] Memory 301 that stores executable program code.

[0267] Processor 302 coupled to memory 301.

[0268] The processor 302 calls the executable program code stored in the memory 301 to execute the steps in the maxillary anterior tooth alveolar bone defect identification method described in Embodiment 1.

[0269] Example 4

[0270] This invention discloses a computer read storage medium that stores a computer program for electronic data exchange, wherein the computer program causes a computer to execute the steps in the maxillary anterior tooth alveolar bone defect identification method described in Embodiment 1.

[0271] The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0272] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0273] Finally, it should be noted that the method and apparatus for identifying alveolar bone defects in maxillary anterior teeth disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention and are only used to illustrate the technical solutions of the present invention, not to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying alveolar bone defects in the maxillary anterior teeth, characterized in that, include: S1. Obtain the global voxel matrix and voxel size information; The global voxel matrix is ​​segmented to obtain N tooth information; N is an integer from 1 to 4; The voxel size information includes the first voxel size, the second voxel size, and the third voxel size; The tooth information includes tooth position number, local voxel matrix, and mask matrix; The global voxel matrix, the local voxel matrix, and the mask matrix are all three-dimensional matrices, and the local voxel matrix and the mask matrix have the same size. The values ​​of the elements in the mask matrix are 0 or 1; S2. Based on the voxel size information, process each piece of tooth information to obtain a vector set corresponding to each piece of tooth information; The vector set includes a first vector, a second vector, and a third vector; S3. Process each piece of tooth information and the corresponding vector set to obtain a set of classification image pairs corresponding to each piece of tooth information; The set of classified image pairs includes several classified image pairs; each classified image pair includes a first classified image and a second classified image. S4. Using a pre-trained multi-scale classification and recognition model, process the set of classification images corresponding to each piece of tooth information to obtain a set of probability vectors corresponding to each piece of tooth information. The probability vector set includes the probability vector corresponding to each classification image pair in the corresponding classification image pair set; the dimension of the probability vector is 5. S5. Process the probability vector set corresponding to each piece of tooth information to obtain the classification and recognition result corresponding to each piece of tooth information; the value of the classification and recognition result is an integer from 1 to 5.

2. The method for identifying alveolar bone defects in the maxillary anterior teeth according to claim 1, characterized in that, The process of processing each tooth information based on the voxel size information to obtain a vector set corresponding to each tooth information includes: S21. A preset template information set; the template information set includes M template information items; the template information includes the tooth position number and a two-dimensional template point set; S22. Based on the voxel size information, process the mask matrix of each tooth information to obtain the compressed index matrix and the first vector corresponding to each tooth information; S23. Based on the voxel size information, process the compressed index matrix and the two-dimensional template point set corresponding to each tooth information to obtain the three-dimensional feature points corresponding to each tooth information. S24. Process the first vector and the three-dimensional feature point corresponding to each tooth information to obtain the second vector and the third vector corresponding to each tooth information; the first vector, the second vector and the third vector corresponding to each tooth information constitute the corresponding vector set.

3. The method for identifying alveolar bone defects in the maxillary anterior teeth according to claim 2, characterized in that, The process of processing the mask matrix for each tooth based on the voxel size information to obtain the compressed index matrix and the first vector corresponding to each tooth information includes: S221. Compress the mask matrix to obtain the corresponding compressed index matrix; S222. Construct a model using a standard mask matrix, process the compressed index matrix and the voxel size information, and construct the corresponding standard mask matrix. The standard mask matrix construction model is as follows: In the formula, b i,j and a i,j These are the elements in the i-th row and j-th column of the standard mask matrix and the compressed index matrix, respectively; i is an integer from 1 to 1, and 1 is the total number of rows in the compressed index matrix; j is 1, 2, or 3; Δ1, Δ2, and Δ3 are the first voxel size, the second voxel size, and the third voxel size, respectively; S223. Using the feature matrix calculation model, the standard mask moments are processed to obtain the corresponding feature matrix; The feature matrix calculation model is as follows: In the formula, C is the feature matrix, and B is the standard mask matrix; S224. Perform eigenvalue decomposition on the feature matrix to obtain three feature pairs; each feature pair includes an eigenvalue and a corresponding eigenvector. S225. Normalize the feature vector corresponding to the largest feature value among the three feature pairs to obtain the corresponding first vector.

4. The method for identifying alveolar bone defects in the maxillary anterior teeth according to claim 2, characterized in that, The process of processing the compressed index matrix and the two-dimensional template point set corresponding to each tooth information based on the voxel size information to obtain the three-dimensional feature points corresponding to each tooth information includes: S231. Process the compressed index matrix and the voxel size information corresponding to the tooth information to obtain the corresponding transformed coordinate information and two-dimensional coordinate point set; The transformed coordinate information includes a selected origin, a first unit vector, and a second unit vector; the two-dimensional coordinate point set includes J two-dimensional coordinate points; J is an integer greater than 1; S232. Register the J two-dimensional coordinate points to the corresponding two-dimensional template point set to obtain the corresponding rotation matrix; S233. Multiply the J two-dimensional coordinate points by the rotation matrix respectively to obtain the corresponding J two-dimensional registration coordinate points; S234. Perform convex hull detection on the J two-dimensional registration coordinate points to obtain the corresponding two-dimensional vertex set; the two-dimensional vertex set includes several two-dimensional vertices; S235. Set the two-dimensional feature point as the lowest two-dimensional vertex in the set of two-dimensional vertices; S236. Using a three-dimensional feature point calculation model, process the rotation matrix, the two-dimensional feature points, and the transformation coordinate information to obtain the corresponding three-dimensional feature points; The three-dimensional feature point calculation model is as follows: [u e ,v e ] T =R -1 [u c ,v c ] T Q=u e V1+v e V2+O In the formula, Q is a vector composed of the three-dimensional coordinates of the three-dimensional feature points; R -1 [u] is the inverse of the rotation matrix; c ,v c ] T V1 and V2 are two-dimensional column vectors formed by the two-dimensional coordinate values ​​of the two-dimensional feature points; V1 and V2 are column vectors formed by the three-dimensional coordinate values ​​of the selected origin, the first unit vector, and the second unit vector, respectively.

5. The method for identifying alveolar bone defects in the maxillary anterior teeth according to claim 4, characterized in that, The process of processing the compressed index matrix and voxel size information corresponding to the tooth information to obtain the corresponding transformed coordinate information and two-dimensional coordinate point set includes: S2311. The compressed index matrix and the voxel size information are processed to obtain a mask point set; the mask point set includes I mask points; S2312. Project the I mask points onto the line containing the first vector to obtain I first projection points; S2313. Set the cutting point as the midpoint of the two farthest first projection points among the I first projection points; S2314. Take J mask points from the I mask points whose corresponding voxels intersect with the cutting plane and form a cutting point set; the cutting plane is a plane that passes through the cutting points and whose normal vector is the first vector. S2315. Project the J mask points in the cutting point set onto the cutting surface to obtain the corresponding second projection points; S2316. Based on the cutting plane, perform coordinate transformation processing on the J second projection points respectively to obtain the corresponding transformed coordinate information and two-dimensional coordinate point set.

6. The method for identifying alveolar bone defects in the maxillary anterior teeth according to claim 1, characterized in that, The process of processing each piece of tooth information and the corresponding vector set to obtain a set of classification image pairs corresponding to each piece of tooth information includes: S31. Process the local voxel matrix and voxel size information of the tooth information to obtain the three-dimensional center point; S32. Set N1 sampling points at equal intervals on the main axis; the main axis is a straight line along the first vector and passing through the three-dimensional center point; N1 is an integer greater than 1; S33. Based on the N1 sampling points, the second vector, and the third vector, set up N1 sampling plane pairs; the sampling plane pairs include a first sampling plane and a second sampling plane; S34. Using N1 sampling plane pairs, sample the three-dimensional voxel region corresponding to the local voxel matrix to obtain N1 initial image pairs; the initial image pairs include a first initial image and a second initial image. S35. Using an image enhancement model, enhance the N1 initial image pairs to obtain N1 enhanced image pairs; the enhanced image pairs include a first enhanced image and a second enhanced image; S36. The size of each of the N1 enhanced image pairs is adjusted to obtain N1 classified image pairs; the N1 classified image pairs constitute the set of classified image pairs corresponding to the tooth information.

7. The method for identifying alveolar bone defects in the maxillary anterior teeth according to claim 1, characterized in that, The multi-scale classification and recognition model includes an image preprocessing module, a global feature generation module, a local feature generation module, a feature fusion module, and a classification module; The image preprocessing module is data-connected to the global feature generation module and the local feature generation module, and is used to preprocess the classified image pairs to obtain a preprocessed image. The global feature generation module is data-connected to the feature fusion module and is used to process the preprocessed image to obtain a global feature map; The local feature generation module is data-connected to the feature fusion module and is used to process the preprocessed image to obtain a local feature map; The feature fusion module is data-connected to the classification module and is used to perform feature fusion processing on the global feature map and the local feature map to obtain a fused feature map. The classification module is used to classify the fused feature map to obtain the probability vector.

8. The method for identifying alveolar bone defects in the maxillary anterior teeth according to claim 7, characterized in that, The feature fusion module includes an initial fusion unit and R intermediate fusion units; R is an integer greater than 2. The initial fusion unit is used to process the global feature map and the local feature map to obtain an initial fused feature map; The intermediate fusion unit is used to process the global feature map, the local feature map, and the previous feature map to obtain an intermediate fused feature map; The preceding feature map of the first intermediate fusion unit is the initial fusion feature map; the preceding feature maps of the subsequent R-1 intermediate fusion units are the intermediate fusion feature maps output by the previous intermediate fusion unit.

9. The method for identifying alveolar bone defects in the maxillary anterior teeth according to claim 8, characterized in that, The intermediate fusion unit includes a channel attention subunit, a first multiplication subunit, a spatial attention subunit, a second multiplication subunit, a first splicing subunit, a first convolution subunit, an average pooling subunit, a second splicing subunit, a second convolution subunit, a residual subunit, and an addition subunit. The global feature map is input to the input terminal of the channel attention subunit; the output terminal of the channel attention subunit is connected to the first input terminal of the first multiplication subunit; the global feature map is input to the second input terminal of the first multiplication subunit; and the output terminal of the first multiplication subunit is connected to the first input terminal of the first splicing subunit. The spatial attention subunit receives the local feature map as its input; the spatial attention subunit is connected to the first input of the second multiplication subunit; and the second input of the second multiplication subunit receives the local feature map. The output of the second multiplication subunit is connected to the second input of the first splicing subunit; The input terminal of the first convolutional subunit of the first intermediate fusion unit is connected to the output terminal of the initial fusion unit; the input terminals of the first convolutional subunits of the subsequent R-1 intermediate fusion units are connected to the output terminal of the previous intermediate fusion unit. The output of the first convolutional subunit is connected to the input of the average pooling subunit; the output of the average pooling subunit is connected to the first input of the second concatenation unit and the first input of the addition subunit; the second and third inputs of the second concatenation subunit are respectively input to the global feature map and the local feature map. The output of the second stitching unit is connected to the input of the second convolution subunit; the output of the second convolution subunit is connected to the third input of the first stitching subunit; the output of the first stitching subunit is connected to the input of the residual subunit; the output of the residual subunit is connected to the second input of the addition subunit; and the output of the addition subunit serves as the output of the feature fusion module.

10. A computer storage medium, characterized in that, The computer storage medium stores computer instructions, which, when invoked, are used to execute the maxillary anterior alveolar bone defect identification method as described in any one of claims 1-9.

Images