A dental image processing system based on CBCT and multi-scale wavelet attention network

Abstract

The application discloses a kind of based on CBCT and multi-scale wavelet attention network's dental image processing system, comprising: image acquisition module, for collecting CBCT image, wherein CBCT image includes panorama, cross section, sagittal view and coronal view;Data preprocessing module is used to locate the central axis of the panorama, cross section, sagittal view and coronal view slice of each tooth in CBCT image by principal component analysis method, and according to the central axis, the image of panorama, cross section, sagittal view and coronal view perspective is intercepted and spliced, and tooth data image is obtained;Image feature acquisition module is used to extract features to tooth data image, and obtain tooth image feature.The application significantly improves the efficiency and quality of dental image processing.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of dentistry, and particularly relates to a tooth image processing system based on CBCT and a multi-scale wavelet attention network. BACKGROUND

[0002] With the continuous progress of dental imaging technology, cone beam computed tomography (CBCT) has gradually become the core technology of dental three-dimensional imaging. CBCT emits a cone-shaped X-ray beam and obtains multiple tomographic images in one scan, and uses computer algorithms to reconstruct high-resolution three-dimensional images. In contrast, traditional CT systems use a fan-shaped X-ray beam to collect data by layer-by-layer scanning, and then reconstruct each layer of image into a three-dimensional image. Although traditional CT can generate complete three-dimensional images through multi-layer scanning, its imaging efficiency is low. Compared with traditional CT, CBCT has significant advantages in imaging principle and technical implementation, can generate high-precision three-dimensional images, and has the characteristics of small device volume, low radiation dose and short imaging time. These advantages make CBCT more suitable for daily use in dental clinics, significantly improving clinical diagnosis efficiency and improving patient comfort.

[0003] Although CBCT has obvious advantages in the field of dentistry, its application still has certain limitations. Compared with traditional CT, especially in low-dose scanning mode, the image noise of CBCT is higher. In addition, metal restorations in patients' oral cavities can cause artifacts and image distortion, thereby affecting the accuracy of diagnosis. At the same time, the large amount of three-dimensional image information generated by CBCT puts higher requirements on doctors' interpretation ability. In complex cases, doctors need to analyze three-dimensional image slices one by one, accurately label and diagnose lesions, which not only consumes time, but also increases the workload of doctors and reduces diagnosis efficiency.

[0004] Therefore, how to effectively combine the multi-view image information of CBCT has become a key problem in current research. The present application aims to propose a tooth image processing system based on CBCT and a multi-scale wavelet attention network, which solves the problems existing in the prior art through multi-view image fusion, accurate positioning and efficient feature extraction. SUMMARY

[0005] The present application proposes a tooth image processing system based on CBCT and a multi-scale wavelet attention network to solve the problems existing in the prior art.

[0006] To achieve the above purpose, the present application provides a tooth image processing system based on CBCT and a multi-scale wavelet attention network, comprising:

[0007] An image acquisition module is used to acquire CBCT images, wherein the CBCT images include panoramic images, cross-sectional images, sagittal images, and coronal images;

[0008] a data preprocessing module for locating the central axis of the panoramic, cross-sectional, sagittal, and coronal slices of each tooth in the CBCT image by a principal component analysis method, and intercepting and splicing the images of the panoramic, cross-sectional, sagittal, and coronal view angles according to the central axis to obtain a tooth data image;

[0009] The image feature acquisition module is used to extract features from the tooth data image to obtain tooth image features.

[0010] Preferably, the image acquisition module includes:

[0011] A scanning unit, used for performing a circumferential scan on the head and acquiring scanning data;

[0012] The reconstruction unit is used to perform three-dimensional reconstruction based on the scanning data to obtain a CBCT image.

[0013] Preferably, the data preprocessing module includes:

[0014] Multi-view slice extraction unit for extracting panoramic, cross-sectional, sagittal, and coronal slices from CBCT images;

[0015] The central axis positioning unit is used to locate the central axis of each tooth through the principal component analysis method, and to intercept and splice slices of each perspective based on the central axis to obtain a multi-perspective fusion image.

[0016] Preferably, the central axis positioning unit includes:

[0017] A centroid calculation unit, used to extract the area of ​​all teeth in the panoramic image through a segmentation algorithm and calculate the centroid of each tooth;

[0018] The medial axis direction determination unit is used to use the principal component analysis method to take the contour point set as input to find the direction with the largest variation in the data, that is, the medial axis direction of the tooth;

[0019] The stitching unit is used to locate the coronal view, sagittal view and cross-sectional view according to the central axis direction, and stitch the images of the four perspectives to obtain a multi-perspective fusion image.

[0020] Preferably, the calculation expression of the tooth axis is:

[0021]

[0022] In the formula, (x i ,y i) represents the coordinates of the contour points, N is the total number of contour points; (x c ,y c ) is the coordinate of the center of mass; (v 1x ,v 1y ) is the principal direction vector obtained by PCA; t is an adjustable parameter used to control the length of the central axis; (x, y) is the two-dimensional coordinate of a point on the central axis of the tooth contour; (x, y) is located on a line extending along the principal direction with the centroid as the reference; X is the contour point set matrix; and μ is the mean vector.

[0023] Preferably, the image feature acquisition module includes:

[0024] Multi-scale wavelet attention unit is used to extract features from multi-view fusion images and fuse the extracted features to obtain a fusion feature map;

[0025] The cross-stage feature fusion unit is used to extract local features, global features and semantic features of the fused feature map, and fuse the local features, global features and semantic features to obtain tooth image features.

[0026] Preferably, the multi-scale wavelet attention unit comprises:

[0027] The high-frequency attention branch is used to extract information of different frequency bands through wavelet convolution, and reorganize and combine the extracted multi-band information through channel shuffling and downsampling operations;

[0028] Low-frequency attention branch, used to extract global features through global average pooling;

[0029] The splicing unit is used to splice the output features of the high-frequency attention branch and the low-frequency attention branch in the channel dimension to form a fused feature map.

[0030] Preferably, the cross-stage feature fusion unit includes:

[0031] IBBneck module, which is used to extract local and global features by adopting different convolution strategies through multiple branches;

[0032] EGConv module, used to extract semantic features and improve the network's computational efficiency and feature expression capabilities;

[0033] The feature integration unit is used to integrate the features extracted by the IBBneck module and the EGConv module to generate the final tooth image feature representation.

[0034] The present invention also provides a dental image classification system based on the dental image processing system, comprising:

[0035] Model building module, used to build the initial classification model;

[0036] A training module, used to train the initial classification model using tooth image features to obtain a classification model;

[0037] The classification module is used to analyze the real-time collected dental data through the classification model to obtain classification results.

[0038] Preferably, the optimized loss function expression when training the initial classification model is:

[0039]

[0040] Where, ρ 2 (b pred ,b gt ) represents the Euclidean distance between the center point of the predicted box and the center point of the real box; c is the diagonal length of the bounding rectangle of the predicted box and the real box; α is the weight factor used to balance the influence of IoU loss and aspect ratio loss; ν is the aspect ratio difference, which is used to measure the aspect ratio difference between the predicted box and the real box; ∑(p i ,g i ) represents the intersection of the predicted box and the true box, Indicates the total size of the predicted area and the true area.

[0041] Compared with the prior art, the present invention has the following advantages and technical effects:

[0042] The dental image processing system based on CBCT and multi-scale wavelet attention network of the present invention collects multi-view CBCT images including panoramic images, cross-sections, sagittal images and coronal images through the image acquisition module, which can fully capture the three-dimensional structural information of the teeth. The data preprocessing module uses the principal component analysis method to accurately locate the central axis of each tooth, and accordingly intercepts and splices multi-view images to form high-quality dental data images, effectively solving the problems of image misalignment and incompleteness. The image feature acquisition module further performs deep feature extraction on the preprocessed image, and uses the advanced multi-scale wavelet attention mechanism to accurately identify the characteristics of dental lesions and improve the robustness and accuracy of feature extraction. This series of innovative designs significantly improves the efficiency and quality of dental image processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0044] Figure 1 A schematic diagram of tooth image processing and classification according to an embodiment of the present invention;

[0045] Figure 2Flowchart of data preprocessing according to an embodiment of the present invention;

[0046] Figure 3 This is a model framework diagram of an embodiment of the present invention;

[0047] Figure 4 This is a structural diagram of a multi-scale wavelet attention module according to an embodiment of the present invention;

[0048] Figure 5 This is a structural diagram of the cross-stage feature fusion module of an embodiment of the present invention. DETAILED DESCRIPTION

[0049] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0050] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0051] Example 1

[0052] like Figure 1 As shown, this embodiment provides a dental image processing system based on CBCT and a multi-scale wavelet attention network, including:

[0053] An image acquisition module is used to acquire CBCT images, wherein the CBCT images include panoramic images, cross-sectional images, sagittal images, and coronal images;

[0054] a data preprocessing module for locating the central axis of the panoramic, cross-sectional, sagittal, and coronal slices of each tooth in the CBCT image by a principal component analysis method, and intercepting and splicing the images of the panoramic, cross-sectional, sagittal, and coronal view angles according to the central axis to obtain a tooth data image;

[0055] The image feature acquisition module is used to extract features from the tooth data image to obtain tooth image features.

[0056] Furthermore, the image acquisition module includes:

[0057] A scanning unit, used for performing a circumferential scan on the head and acquiring scanning data;

[0058] The reconstruction unit is used to perform three-dimensional reconstruction based on the scanning data to obtain a CBCT image.

[0059] Furthermore, the data preprocessing module includes:

[0060] Multi-view slice extraction unit for extracting panoramic, cross-sectional, sagittal, and coronal slices from CBCT images;

[0061] The central axis positioning unit is used to locate the central axis of each tooth through the principal component analysis method, and to intercept and splice slices of each perspective based on the central axis to obtain a multi-perspective fusion image.

[0062] The specific implementation methods are as follows:

[0063] This method uses principal component analysis to locate the central axis of the panoramic, cross-sectional, sagittal, and coronal slices of a single tooth in CBCT images, and then intercepts and splices the images from these four perspectives to form the final data image. The specific steps are:

[0064] Since YOLO performs end-to-end target detection based on a single image, the collected CBCT images are first intercepted and spliced. The data collection process is as follows: Figure 2 . First, the maximum intensity projection method is used to process the slices of different viewing angles in CBCT image processing. For panoramic images, the maximum intensity projection is completed within the volume data range with a thickness of 0.2mm; while for coronal images, sagittal images and cross-sections, the projection is performed within the volume data range of 1.0mm. Secondly, the segmentation algorithm is used to extract the areas of all teeth in the panoramic image, and the axis of each tooth is calculated. Based on the segmented image, the corresponding area is cropped with each tooth as the center. The edge information of the tooth area is extracted by the segmentation algorithm, and the contour point set C is generated. The center of mass of the tooth area is calculated based on this set to determine the position and direction of the tooth. At the same time, the principal component analysis (PCA) method is used to extract the contour point set C={(x1,y1),(x2,y2),…,(x N ,y N )} is input, and the direction with the largest variation in the data is found, that is, the tooth's central axis direction. The specific formula for calculating the tooth's central axis is as follows:

[0065]

[0066] Among them, (x i ,y i ) represents the coordinates of the contour points, N is the total number of contour points; (x c ,y c ) is the coordinate of the center of mass; (v 1x ,v 1y) is the principal direction vector obtained by PCA; t is an adjustable parameter, (x, y) is the two-dimensional coordinate of a point on the central axis of the tooth contour, located on a line extending along the principal direction with the centroid as the reference. By adjusting the parameter t, the length of the central axis can be flexibly controlled. In the PCA calculation process, the data needs to be centered on its mean, so the contour point set C is first organized into a matrix X, and then the mean vector μ of the point set is calculated = (x c ,y c ) and subtract the mean vector from the data matrix to obtain a centered data matrix. Next, based on this matrix, the covariance matrix Σ is calculated and the eigenvalue decomposition of the covariance matrix is ​​performed to obtain eigenvalues ​​and eigenvectors. The eigenvector represents the main direction of the data, while the eigenvalue indicates the degree of variation in that direction. The eigenvector corresponding to the largest eigenvalue represents the eigenvector with the greatest variation in the data (i.e., the main direction of the tooth contour).

[0067] This method allows CBCT slices to be precisely positioned relative to the tooth's central axis. Coronal, sagittal, and cross-sectional views are also positioned relative to the central axis. The resulting images from these four perspectives are then stitched together to generate a multi-view image for subsequent label training.

[0068] Furthermore, the central axis positioning unit includes:

[0069] A centroid calculation unit, used to extract the area of ​​all teeth in the panoramic image through a segmentation algorithm and calculate the centroid of each tooth;

[0070] The medial axis direction determination unit is used to use the principal component analysis method to take the contour point set as input to find the direction with the largest variation in the data, that is, the medial axis direction of the tooth;

[0071] The stitching unit is used to locate the coronal view, sagittal view and cross-sectional view according to the central axis direction, and stitch the images of the four perspectives to obtain a multi-perspective fusion image.

[0072] Furthermore, the image feature acquisition module includes:

[0073] Multi-scale wavelet attention unit is used to extract features from multi-view fusion images and fuse the extracted features to obtain a fusion feature map;

[0074] The cross-stage feature fusion unit is used to extract local features, global features and semantic features of the fused feature map, and fuse the local features, global features and semantic features to obtain tooth image features.

[0075] Furthermore, the multi-scale wavelet attention unit includes:

[0076] The high-frequency attention branch is used to extract information of different frequency bands through wavelet convolution, and reorganize and combine the extracted multi-band information through channel shuffling and downsampling operations;

[0077] Low-frequency attention branch, used to extract global features through global average pooling;

[0078] The splicing unit is used to splice the output features of the high-frequency attention branch and the low-frequency attention branch in the channel dimension to form a fused feature map.

[0079] The specific implementation methods are as follows:

[0080] To further improve the feature expression capability, this method proposes a multi-scale wavelet attention module. This module is designed based on the self-attention mechanism framework and contains two independent paths: a high-frequency attention branch and a low-frequency attention branch. It realizes effective interaction of features from the cross-channel and cross-space dimensions respectively. Its specific structure is shown in Figure 4 Specifically, given the input feature F∈R c×H×W , where C represents the number of channels, H and W represent the height and width of the feature map, respectively. First, the feature map F is expanded through the convolution layer, increasing the number of channels from C to 2C, and obtaining the expanded feature map S. Then, the feature map S is split into S1 and S2 along the channel dimension, and S1, S2∈R c×H×W These two branches are respectively input into the high-frequency attention branch and the low-frequency attention branch for processing.

[0081] In the high-frequency attention branch, the feature map S1 is divided into three parts: the query feature q1, the key feature k1, and the value feature v1. These features are further processed through wavelet convolution (WTConv) to extract information from different frequency bands. During WTConv processing, the input feature map undergoes wavelet decomposition to extract information from four frequency bands: low frequency (LL) and high frequency (LH, HL, and HH). The extracted multi-band information is then reorganized and combined through operations such as channel shuffling and downsampling, and reconstructed using an inverse wavelet transform to output the final feature map.

[0082] s 1out =σ(Conv(WTConv(q1))⊙WTConv(k1))·v1

[0083] s 2out =Softmax(Pool(k2)·q2)·Pool(v2)

[0084] Among them, S 1out and S 2outThey represent the output feature maps after processing by the high-frequency attention branch and the low-frequency attention branch, σ represents the activation function Sigmod, Softmax represents the activation function, Pool represents the global average pooling, q1, k1, v1 and q2, k2, v2 represent the query features, key features and value features after the high-frequency attention branch and the low-frequency attention branch divide the input feature maps S1 and S2 respectively.

[0085] The low-frequency attention branch focuses on modeling global information. The input feature S2 is similarly split into query feature q2, key feature k2, and value feature v2. The low-frequency branch extracts global features from k2 and v2 through global average pooling (GAP), effectively capturing global contextual information, suppressing local noise, and strengthening the modeling of long-range dependencies.

[0086] F out =Conv(Concat(s 1out ,s 2out ))+F

[0087] Among them, F and F out Represent the input features and the final output features respectively, and the convolution kernel size of Conv is 3×3.

[0088] The output features of the high-frequency and low-frequency branches are first concatenated along the channel dimension to form a fused feature map. The multi-scale wavelet attention module integrates high- and low-frequency feature modeling, enhancing the richness and robustness of feature representation through cross-channel and cross-spatial information interaction. The high-frequency branch effectively enhances the ability to capture details and local information, while the low-frequency branch strengthens global dependency modeling and suppresses noise interference on the model. This overall design achieves efficient modeling of multi-scale information and strong adaptability in feature representation.

[0089] Furthermore, the cross-stage feature fusion unit includes:

[0090] IBBneck module, which is used to extract local and global features by adopting different convolution strategies through multiple branches;

[0091] EGConv module, used to extract semantic features and improve the network's computational efficiency and feature expression capabilities;

[0092] The feature integration unit is used to integrate the features extracted by the IBBneck module and the EGConv module to generate the final tooth image feature representation.

[0093] The specific implementation methods are as follows:

[0094] To further optimize the computational efficiency and feature extraction ability, a cross-stage feature fusion module is proposed to improve the computational efficiency and feature expression ability of the target classification task in the Head part. The specific structure is shown in Figure 5

[0095] The CSPFB receives the input feature map F and divides it into two branches. The first branch reduces the channel number to half through dimension reduction convolution, thereby reducing the computational cost. After dimension reduction, the feature map is input to the IBBneck module. The IBBneck is designed based on the Inception module, which uses four branches with different convolution strategies: 3x3 depth separable convolution is used to extract local features, 1x11 and 11x1 depth separable convolution is used to capture long strip features in different directions, and the combination of EGConv and ordinary convolution layer is used to further extract more delicate semantic features. The features extracted by each branch are finally spliced in the channel dimension to complete multi-path feature fusion.

[0096] S3 = Concat(DWConv(x1), DWConv(x2), DWConv(x3),

[0097] EGConv(EGConv(x4)) + Conv(x4)

[0098] EGConv = Shuffle Concat(Conat(Conv(x), x), Conv(Conv(x)))

[0099] where S3 represents the output feature after processing by the IBBneck module, x1, x2, x3, and x4 represent the four features after splitting the input features of the IBBneck module, Shuffle represents channel shuffling. DWConv represents depth separable convolution, and the convolution kernel size is 3x3, 1x11, and 11x1 in turn. Conv represents 3x3 convolution.

[0100] The second branch directly processes the input feature through the EGConv module.

[0101] S4 = EGConv(x)

[0102] F 1out = Concat(Conv(S3), S4)

[0103] where S4 represents the output feature of the second branch. F 1out represents the final output after processing by the CSPFB module.

[0104] ​The cross-stage feature fusion module is designed to balance efficiency and sophistication. It provides an efficient, robust, and adaptable solution for object classification tasks, improving computational efficiency while enhancing classification performance.

[0105] This embodiment further provides a dental image classification system based on a dental image processing system, comprising:

[0106] Model building module, used to build the initial classification model;

[0107] A training module, used to train the initial classification model using tooth image features to obtain a classification model;

[0108] The classification module is used to analyze the real-time collected dental data through the classification model to obtain classification results.

[0109] Furthermore, the model framework is as follows:

[0110] The overall structure of MWA-YOLO is mainly composed of two parts: the backbone network (Backbone) and the head network (Head). Each part plays a different role in the target detection process. The model framework is as follows Figure 3 As shown in the figure. The backbone network adopts CSPDarknet53 and combines it with a multi-scale wavelet attention module for feature extraction. The feature maps extracted by the multi-scale wavelet attention module and multiple convolution operations are further enhanced by the SPPF module and then passed to the Head part. In the Head part, feature fusion modules such as PANet and FPN are used to improve the detection ability of targets of different scales and sizes. Through multi-scale feature fusion, contextual information enhancement and optimized convolution operations, the Head part further refines the feature representation provided by the backbone network. In addition, a cross-stage feature fusion module is introduced to effectively integrate channel and spatial features to improve feature expression capabilities. At the same time, the design of efficient group convolution and cross-stage feature fusion modules is combined to reduce the computational complexity of the model. Finally, the optimized features are passed to the detection head part for accurate target detection, thereby improving detection accuracy and speed.

[0111] Furthermore, the optimized loss function expression when training the initial classification model is:

[0112]

[0113] Where, ρ 2 (b pred ,b gt) represents the Euclidean distance between the center point of the predicted box and the center point of the real box; c is the diagonal length of the bounding rectangle of the predicted box and the real box; α is the weight factor used to balance the influence of IoU loss and aspect ratio loss; ν is the aspect ratio difference, which is used to measure the aspect ratio difference between the predicted box and the real box; ∑(p i ,g i ) represents the intersection of the predicted box and the true box, Indicates the total size of the predicted area and the true area.

[0114] The hyperparameter settings are shown in Table 1:

[0115] Table 1

[0116]

[0117]

[0118] Performance evaluation indicators:

[0119] In this paper, mAP@0.5, mAP@0.5:0.95, parameter count (Parameters) and computational complexity (GFLOPs) are selected as classification indicators of positioning performance.

[0120] This example performs model verification as follows:

[0121] Finally, the model's performance was verified on the collected dataset. Table 2 shows the simulation results. Compared to the baseline model YOLOv8n (baseline), the proposed method improves caries classification by 1.3% and 1.9% in mAP@0.5 and mAP@0.5:0.95, respectively, while maintaining comparable parameter count and computational complexity. This method demonstrates excellent performance in practical applications.

[0122] Table 2

[0123]

[0124] The beneficial effects brought about by the present invention are as follows:

[0125] (1) A dental image classification method based on CBCT images is proposed. By integrating multiple advanced technologies, this method effectively addresses the challenge of automatically analyzing multi-view slices in CBCT images. By comprehensively analyzing slices from multiple viewpoints, the model improves the accuracy of CBCT caries detection, reaching a detection accuracy of 98%.

[0126] (2) Based on the strategy of wavelet convolution combined with self-attention mechanism, by improving the data features of different frequency bands, the characteristic information of teeth can be better extracted. At the same time, the interference of noise on model detection is effectively suppressed, and the adaptability and robustness of the model to the real environment are improved.

[0127] (3) Based on the cross-stage feature fusion module and efficient convolution, the complexity of the model is reduced while maintaining the model accuracy, making the model more efficient.

[0128] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A dental image processing system based on CBCT and multi-scale wavelet attention network, characterized in that: include: An image acquisition module is used to acquire CBCT images, wherein the CBCT images include panoramic images, cross-sectional images, sagittal images, and coronal images; a data preprocessing module for locating the central axis of the panoramic, cross-sectional, sagittal, and coronal slices of each tooth in the CBCT image by a principal component analysis method, and intercepting and splicing the images of the panoramic, cross-sectional, sagittal, and coronal view angles according to the central axis to obtain a tooth data image; An image feature acquisition module is used to extract features from the tooth data image to obtain tooth image features; The image feature acquisition module includes: Multi-scale wavelet attention unit is used to extract features from multi-view fusion images and fuse the extracted features to obtain a fusion feature map; a cross-stage feature fusion unit, configured to extract local features, global features, and semantic features from the fused feature map, and fuse the local features, global features, and semantic features to obtain tooth image features; The multi-scale wavelet attention unit includes: The high-frequency attention branch is used to extract information of different frequency bands through wavelet convolution, and reorganize and combine the extracted multi-band information through channel shuffling and downsampling operations; Low-frequency attention branch, used to extract global features through global average pooling; The splicing unit is used to splice the output features of the high-frequency attention branch and the low-frequency attention branch in the channel dimension to form a fused feature map.

2. The system according to claim 1, wherein: The image acquisition module includes: A scanning unit, used for performing a circumferential scan on the head and acquiring scanning data; The reconstruction unit is used to perform three-dimensional reconstruction based on the scanning data to obtain a CBCT image.

3. The system according to claim 1, wherein: The data preprocessing module includes: Multi-view slice extraction unit for extracting panoramic, cross-sectional, sagittal, and coronal slices from CBCT images; The central axis positioning unit is used to locate the central axis of each tooth through the principal component analysis method, and to intercept and splice slices of each perspective based on the central axis to obtain a multi-perspective fusion image.

4. The system according to claim 3, characterized in that The central axis positioning unit includes: A centroid calculation unit, used to extract the area of ​​all teeth in the panoramic image through a segmentation algorithm and calculate the centroid of each tooth; The medial axis direction determination unit is used to use the principal component analysis method to take the contour point set as input to find the direction with the largest variation in the data, that is, the medial axis direction of the tooth; The stitching unit is used to locate the coronal view, sagittal view and cross-sectional view according to the central axis direction, and stitch the images of the four perspectives to obtain a multi-perspective fusion image.

5. The system according to claim 3, wherein: The calculation expression of the tooth axis is: In the formula, (x i ,y i ) represents the coordinates of the contour points, N is the total number of contour points; (x c ,y c ) is the coordinate of the center of mass; (v 1x ,v 1y ) is the principal direction vector obtained by PCA; t is an adjustable parameter used to control the length of the central axis; (x, y) is the two-dimensional coordinate of a point on the central axis of the tooth contour; (x, y) is located on a line extending along the principal direction with the centroid as the reference; X is the contour point set matrix; and μ is the mean vector.

6. The system according to claim 1, wherein: The cross-stage feature fusion unit includes: IBBneck module, which is used to extract local and global features by adopting different convolution strategies through multiple branches; EGConv module, used to extract semantic features; The feature integration unit is used to integrate the features extracted by the IBBneck module and the EGConv module to generate the final tooth image feature representation.

7. A dental image classification system based on the dental image processing system according to claim 1, characterized in that: include: Model building module, used to build the initial classification model; A training module, used to train the initial classification model using tooth image features to obtain a classification model; The classification module is used to analyze the real-time collected dental data through the classification model to obtain classification results.

8. The system according to claim 7, characterized in that The optimized loss function expression when training the initial classification model is: Where, ρ 2 (b pred ,b gt ) represents the Euclidean distance between the center point of the predicted box and the center point of the real box; c is the diagonal length of the bounding rectangle of the predicted box and the real box; α is the weight factor used to balance the influence of IoU loss and aspect ratio loss; ν is the aspect ratio difference, which is used to measure the aspect ratio difference between the predicted box and the real box; ∑(p i ,g i ) represents the intersection of the predicted box and the true box, Indicates the total size of the predicted area and the true area.

Images