Tooth segmentation method of CBCT image

Abstract

The invention discloses a tooth segmentation method of a CBCT (Cone Beam Computed Tomography) image in the technical field of medical image segmentation. The method comprises the following steps: S1, sequentially carrying out standardization and data enhancement processing on a CBCT image data set; s2, inputting the CBCT image preprocessed in the S1 into an encoder part of a network structure, and completing step-by-step extraction of image features to obtain multi-scale features; s3, further fusing the multi-scale feature maps generated by each downsampling layer of the encoder network structure to obtain richer and more effective feature expressions; and S4, performing step-by-step spatial resolution recovery on the encoder features, and finally realizing fine segmentation of the tooth image, the tooth segmentation method of the CBCT image solves the problem that the teeth and surrounding tissue boundary details are fuzzy and the teeth are not clear due to the influence of metal artifacts in CBCT segmentation.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image segmentation, and in particular to a method for tooth segmentation in CBCT images. Background Art

[0002] Cone-beam computed tomography (CBCT) technology, a core imaging tool in modern dentistry, overcomes the limitations of traditional two-dimensional X-ray images in terms of structural overlap and lack of depth information with its high-resolution three-dimensional imaging capabilities. This technology significantly improves the accuracy of three-dimensional tooth positioning and measurement, and has been widely used in oral clinical diagnosis. However, accurate segmentation of teeth in CBCT images still faces many challenges. First, the high grayscale similarity between teeth and alveolar bone leads to blurred boundaries, which is prone to leakage errors. Second, streak artifacts generated by metal restorations cover multiple slice regions, severely distorting the local anatomy and further complicating segmentation. Furthermore, the complexity of the three-dimensional structure of teeth (such as impacted teeth, overlapping adjacent teeth, and missing teeth) significantly exacerbates the challenges of segmentation. Furthermore, the inherent noise and structural complexity of CBCT images make manual segmentation not only time-consuming and labor-intensive, but also susceptible to operator subjective factors, which can lead to missed or misdiagnosed diagnoses, thereby affecting the formulation and implementation of subsequent treatment plans. Therefore, the development of efficient and accurate automated tooth segmentation methods is of great significance.

[0003] Early tooth segmentation methods primarily relied on traditional image processing techniques, such as morphological operations and the watershed algorithm. Some researchers have proposed a segmentation method based on morphological operations. This method combines morphological operations with a region growing algorithm, utilizing the grayscale and morphological features of teeth for segmentation. Morphological filters are then used to remove image noise, thereby improving the accuracy of tooth region recognition. However, this method has significant limitations when processing complex images. In particular, when the image is noisy or grayscale differences are small, morphological operations struggle to effectively distinguish teeth from the background, resulting in decreased segmentation accuracy. To overcome the shortcomings of traditional morphological operations, some researchers have proposed an improved watershed algorithm. This method, based on the traditional watershed algorithm, incorporates image smoothing to improve segmentation accuracy by removing high-frequency noise and enhancing the contrast of the tooth region. Although the improved watershed algorithm alleviates the noise interference problem to some extent, its performance remains limited when processing complex scenes, with severe metal artifacts, or when the tooth boundaries are blurred. This is because the watershed algorithm inherently relies on image gradient information, which is easily distorted in the presence of noise, leading to over-segmentation or under-segmentation.

[0004] Traditional tooth segmentation methods primarily rely on threshold segmentation and image smoothing. While these methods have improved accuracy to some extent, they rely heavily on artificially designed features and are sensitive to noise, making it difficult to effectively address the problems of bleed-through and over-segmentation caused by low-contrast boundaries. Furthermore, traditional methods place high demands on image quality and often struggle to achieve reliable segmentation results in images with significant artifacts or noise. In recent years, with the rise of deep learning techniques in the field of medical imaging, methods represented by convolutional neural networks (CNNs) have gradually become the mainstream technology for tooth segmentation. Among them, encoder-decoder architectures, such as U-Net and its variants (3D U-Net, V-Net, and UNet++), have been widely used in medical image segmentation due to their outstanding performance in extracting local features and recovering spatial information. However, CNNs still have inherent limitations in handling complex backgrounds and long-range dependencies, making them particularly difficult to accurately locate and distinguish three-dimensional tooth structures (such as impacted teeth). Furthermore, CNN models generally lack dynamic feature fusion mechanisms, making it difficult to adaptively optimize features at different scales and in different regions, thus limiting further improvements in segmentation accuracy.

[0005] To improve the ability to model long-range dependencies, Transformer-based medical image segmentation methods, such as TransUNet and Swin Transformer, have gradually emerged. Their global modeling significantly improves segmentation accuracy in high-noise and complex backgrounds. However, the computational complexity of the Transformer architecture is much higher than that of CNNs, and it faces severe computational and memory bottlenecks, especially when processing three-dimensional medical images. This limits its widespread deployment and application in real-world scenarios. To alleviate the computational burden of the Transformer model, researchers have proposed improved models such as VM-UNet and LightM-UNet. These models effectively reduce computational complexity and enhance the ability to model long-range spatial dependencies through parallel processing, reduced model parameters, and the introduction of the residual Mamba method. However, these methods still have shortcomings in multi-scale feature fusion in the presence of metal artifacts and uneven noise distribution, and their dynamic feature interaction capabilities are limited. In addition, existing methods generally fail to effectively integrate spatial information across scales and channels, which also restricts their performance in complex image environments. Summary of the Invention

[0006] The purpose of this section is to summarize some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of this application to avoid blurring the purpose of this section, the abstract and the title of the invention, and such simplifications or omissions should not be used to limit the scope of the present invention.

[0007] Therefore, the object of the present invention is to provide a method for teeth segmentation in CBCT images to solve the problem of blurred boundary details between teeth and surrounding tissues and unclear teeth due to the influence of metal artifacts in CBCT segmentation.

[0008] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:

[0009] A method for tooth segmentation of a CBCT image, comprising the following steps:

[0010] S1. Preprocessing: The CBCT image dataset is standardized and data enhanced in sequence;

[0011] S2, encoding: The CBCT image preprocessed in S1 is input into the encoder part of the network structure to complete the step-by-step extraction of image features and obtain multi-scale features {X1, X2, X3};

[0012] S3, feature fusion and interaction: further fuse the multi-scale feature maps generated by each downsampling layer of the encoder network structure to obtain richer and more effective feature expression;

[0013] S4. Decoding: The encoder features are restored to a certain spatial resolution level by level, ultimately achieving fine segmentation of the tooth image.

[0014] As a preferred solution of the method for tooth segmentation of CBCT images described in the present invention, step S1 further includes the following step: using a 3D patch extraction method to divide the large-size CBCT data into blocks.

[0015] As a preferred embodiment of the method for tooth segmentation of CBCT images of the present invention, in step S2, the encoding includes three consecutive downsampling layers, and each downsampling layer sequentially performs the following operations:

[0016] First, the three-dimensional convolution operation is used to extract local features to improve the feature expression ability of the network. The input feature of the first layer is set to X l-1 , then the convolution operation formula is defined as:

[0017]

[0018] in, is the output feature of the lth layer, Conv3D represents 3×3 convolution, and They represent the weight matrix and bias term of the convolution kernel of the lth layer respectively, and σ is a nonlinear activation function used to enhance the nonlinear feature learning ability of the model;

[0019] Secondly, the features output by the convolution operation are input into the feature extraction method. The specific implementation of the feature extraction method is as follows:

[0020] Let the output of the kth sublayer in the feature extraction method be Its mathematical expression is:

[0021]

[0022] In the above formula, [] represents the feature concatenation operation in the channel dimension. The learning parameters for the feature extraction method Mamba is a state space model;

[0023] The Mamba operation is defined as:

[0024] Mamba(X)=BN(X+f SSM (LN(X)))

[0025] Among them, X is the input feature, BN is the layer normalization function, f SSM Represents the calculation of the state space model. The specific formula is as follows:

[0026] f SSM (X) = Conv1D(X; W in )*h SSM (X)

[0027] Among them, W in is the convolution kernel weight, * represents element-by-element multiplication, and the recursive filter function h SSM The specific expression is:

[0028] h t =A·h t-1 +B·X t ,Y t =C·h t +D·X t

[0029] Among them, A, B, C, and D are the state space model parameters of the feature extraction method, and h t is the hidden state, h t-1 is the hidden state of the previous time step, which retains the past information, X t and Y t They are input and output features respectively;

[0030] Finally, the maximum pooling operation is used to reduce the spatial resolution of the feature map to extract higher-level and more abstract semantic features. The pooling operation formula is:

[0031]

[0032] in, Represents the output value at position (i, j, k) in the lth layer, Represents the value of position (p,q,r) in the original feature map, Ω i,j,k Represents the local area window used for pooling operation in the feature map.

[0033] The above process is repeated three times in sequence, and each execution produces a higher level of semantic features, thus forming a multi-scale feature set output by the encoder: {X1, X2, X3}.

[0034] As a preferred solution of the method for tooth segmentation of CBCT images described in the present invention, in step S3, the specific steps of feature fusion and interaction are as follows:

[0035] First, perform global average pooling on each level feature map to obtain the global feature vector. For the i-th level feature map, the calculation formula is:

[0036]

[0037] in:

[0038]

[0039] Among them, z i Represents the feature map calculated by the i-th layer, H i ×W i ×D i Represent the height, width and depth of the feature map of the i-th layer respectively, Represents the eigenvalue of the feature map of layer i at channel c, located at coordinates (h, w, d). Subsequently, the global eigenvectors of all levels are concatenated into a comprehensive eigenvector Z:

[0040]

[0041] Among them, C1+C2+C3 is the total number of channels after splicing, z 1 、z 2 、z 3 They are features from different layers, and then the comprehensive feature vector is mapped to the weight value W through a 1×1 convolution;

[0042]

[0043] Where W=[W 1, W 2, W3], respectively represent different weights of different feature maps, W fuse is the convolution weight matrix, b fuseis a bias term, both of which are learnable parameters. To ensure the validity of the weight value, W is normalized using the Softmax function to obtain the normalized weight value:

[0044]

[0045] Finally, W i are different weights of different feature maps. The fused feature map is obtained by weighted summation of the feature maps of each level in the scale dimension and their corresponding weights:

[0046]

[0047] Then, the feature interaction method is used to further enhance the interaction ability between fusion features. First, the fusion feature map output by the feature fusion method is transformed into Flatten into 2D features:

[0048]

[0049] Next, channel attention and spatial attention calculations are performed separately. The specific process is as follows:

[0050] Through the learnable projection matrix W Q ,W K ,W V Input features X flat Mapped to a matrix of query Q, key K and value V:

[0051] Q=W Q X flat , K=W K X flat , V=W V X flat

[0052] Then, we calculate the scaled dot product attention:

[0053]

[0054] Among them, d k is the feature dimension of each attention head, QK T Indicates the matching degree of each query with all keys. Softmax is used to normalize the attention weight. After the output of all attention heads is spliced ​​in the channel dimension, the interleaved channel attention output X is obtained through linear mapping. CA ;

[0055] X CA =W O [head1;head2;…;head n ],W O is a learnable parameter;

[0056] Then, the gating mechanism is used to dynamically adjust the channel attention strength and calculate the gating factor:

[0057] g CA =σ(W g1 X CA +b g1 )

[0058] Among them, W g1 、b g1 is a learnable parameter, σ is the Softmax activation function, which compresses the result into the interval [0,1];

[0059] The interleaved spatial attention method captures the dependencies between spatial positions and also calculates the gating factor gSA:

[0060] g SA =σ(W g2 X SA +b g2 )

[0061] Among them, X SA is the spatial dimension feature output obtained by the spatial attention method, W g2 、b g2 is another set of gated learnable parameters;

[0062] The final feature interaction method outputs features by multiplying the interleaved channel attention output and the spatial attention output by the corresponding gating factor g CA and g SA , and summing over the channel and spatial dimensions to obtain:

[0063] X ICSA =g CA ·X CA +g SA ·X SA

[0064] in, The final output feature map of the feature interaction method integrates the channel and spatial attention information, making the feature expression more accurate and robust.

[0065] As a preferred embodiment of the method for tooth segmentation of CBCT images of the present invention, in step S4, decoding includes step-by-step upsampling, skip connection, and final pixel-level classification. The specific process is as follows:

[0066] The decoding uses three consecutive interpolation upsampling layers to gradually restore the spatial resolution of the feature map. Each upsampling layer performs the following operations in sequence: Set the input feature map of the lth upsampling layer to U l, then perform interpolation upsampling operation to obtain a feature map U with higher spatial resolution l+1 , the specific formula is:

[0067]

[0068] Among them, the Interpolation function uses trilinear interpolation, that is, the feature map is interpolated in three dimensions: depth, height, and width, so that the feature size is enlarged to gradually restore the resolution of the original image.

[0069] To preserve the details of the tooth anatomy extracted during the encoding phase and maintain semantic consistency of the features, after upsampling each interpolation, it is fused with the corresponding encoder features through a multi-level skip connection. The specific process is as follows:

[0070] Assume that the feature after interpolation upsampling of the lth decoding layer is The corresponding encoder level l feature is The fusion method of the two is splicing in the feature channel dimension:

[0071]

[0072] Then the fused features Further fusion and optimization through a convolution operation:

[0073]

[0074] are the learnable weights and biases of the l+1th layer decoder convolution;

[0075] The above process is repeated three times to gradually restore the feature map from low resolution to the original input resolution;

[0076] After the feature recovery process is completed, the final feature map with the same size as the input CBCT image is obtained. Then, the convolution is used to map the pixel-level segmentation prediction results, and the Sigmoid function is used:

[0077]

[0078] P h,w,d is the probability of belonging to the target category at position (h,w,d);

[0079] The pixel-level mask generated by the above process is the final tooth segmentation result of the CBCT image.

[0080] Compared with existing technologies, the present invention offers the following advantages: It addresses the challenges of CBCT tooth segmentation caused by blurred boundaries and high-noise metal artifacts by proposing an optimization solution. This solution leverages deep learning technology to enhance tooth boundary clarity, suppress metal artifact interference, and improve segmentation accuracy. This enables more accurate extraction of tooth morphology information, providing reliable data support for orthodontic analysis, implant planning, and personalized restoration, thereby enhancing the accuracy and efficiency of dental diagnosis and treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort. Among them:

[0082] Figure 1 This is a schematic diagram of the overall network architecture of the present invention;

[0083] Figure 2 Schematic diagram of feature extraction in the present invention;

[0084] Figure 3 Schematic diagram of feature fusion and interaction in the present invention;

[0085] Figure 4 This is the tooth segmentation flow chart of CBCT images in the present invention. DETAILED DESCRIPTION

[0086] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0087] like Figure 4 As shown, the present invention provides a method for tooth segmentation of CBCT images, the steps are as follows:

[0088] S1: During the preprocessing phase, the CBCT image dataset is first standardized to ensure input data consistency. Subsequently, data augmentation techniques, including rotation, flipping, contrast adjustment, and artifact simulation, are employed to improve the model's generalization capabilities. Furthermore, to improve computational efficiency and reduce video memory usage, 3D patch extraction is employed to segment large CBCT data into blocks, optimizing subsequent analysis and model training.

[0089] S2: Encoding stage: The CBCT image normalized in S1 is input into the encoder part of the network structure to complete the step-by-step extraction of image features. The network structure is as follows Figure 1As shown in Figure 1, it includes an encoder, a feature fusion and interaction module, and a decoder. Specifically, the encoding stage contains three consecutive downsampling layers, each of which performs the following three operations in sequence: First, a three-dimensional convolution (Conv3D) operation is used to extract local features to improve the feature expression ability of the network. Assume that the input feature of the first layer is X l-1 , then the convolution operation formula is defined as:

[0090]

[0091] in, is the output feature of the lth layer, Conv3D represents 3×3 convolution, and They represent the weight matrix and bias term of the l-th layer convolution kernel respectively, and σ is a nonlinear activation function, which is used to enhance the nonlinear feature learning ability of the model.

[0092] Secondly, the features output by the convolution operation are fed into the feature extraction method, such as Figure 2 As shown in Figure 2, the core of the feature extraction method is a cross-layer global context memory pool constructed by dense connection and recursive filtering. By closely connecting features at different levels, the semantic expression ability of features is effectively improved. The specific implementation of the feature extraction method is as follows:

[0093] Let the output of the kth sublayer in the feature extraction method be Its mathematical expression is:

[0094]

[0095] In the above formula, [] represents the feature concatenation operation in the channel dimension. are the learning parameters of the feature extraction method.

[0096] Mamba is a state-space model, and the operation is defined as:

[0097] Mamba(X)=BN(X+f SSM (LN(X)))

[0098] Among them, X is the input feature, BN is the layer normalization function, and f SSM Represents the calculation of the state space model. The specific formula is as follows:

[0099] f SSM (X) = Conv1D(X; W in )*h SSM (X)W in is the convolution kernel weight, * represents element-by-element multiplication, and the recursive filter function h SSM The specific expression is:

[0100] ht =A·h t-1 +B·X t ,Y t =C·h t +D·X t

[0101] Among them, A, B, C, and D are the state space model parameters of the feature extraction method, and h t is the hidden state, h t-1 is the hidden state of the previous time step, which retains the past information, X t and Y t They are input and output features respectively. Through this recursive operation, the global context information of the image is effectively captured, forming a global context memory pool of cross-layer features, which significantly improves the semantic representation ability of the encoder to extract features.

[0102] Finally, the maximum pooling operation is used to reduce the spatial resolution of the feature map to extract higher-level and more abstract semantic features. The pooling operation formula is:

[0103]

[0104] in,, Represents the output value at position (i, j, k) in the lth layer, Represents the value of position (p,q,r) in the original feature map, Ω i,j,k Represents the local area window used for pooling operation in the feature map.

[0105] The above process is repeated three times in sequence, and each execution produces a higher level of semantic features, thus forming a multi-scale feature set output by the encoder: {X1, X2, X3}. After completing the above encoding stage, the deep feature expression of CBCT images in the multi-scale semantic space can be realized.

[0106] S3: Feature fusion and interaction stage: further fuse the multi-scale feature maps generated by each downsampling layer of the encoder network structure to obtain richer and more effective feature expression. Specifically, feature fusion method and feature interaction method are used to achieve feature enhancement in this stage, such as Figure 3 As shown in Figure 2, the specific process and formula are described as follows: To effectively utilize the multi-scale features {X1, X2, X3} obtained from S2, a dynamic receptive field weighting mechanism is proposed. This mechanism first performs global average pooling (GAP) on each level feature map to obtain a global feature vector. For the i-th level feature map, the calculation formula is:

[0107]

[0108] in:

[0109]

[0110] Among them, z i Represents the feature map calculated by the i-th layer, H i ×W i ×D i Represent the height, width and depth of the feature map of the i-th layer respectively, Represents the eigenvalue of the feature map of layer i at channel c, located at coordinates (h, w, d). Subsequently, the global eigenvectors of all levels are concatenated into a comprehensive eigenvector Z:

[0111]

[0112] Among them, C1+C2+C3 is the total number of channels after splicing, z 1 、z 2 、z 3 They are features from different layers, and then the comprehensive feature vector is mapped to the weight value W through a 1×1 convolution:

[0113]

[0114] Where W=[W 1, W 2, W3], respectively represent different weights of different feature maps, W fuse is the convolution weight matrix, b fuse is the bias term, and both are learnable parameters. To ensure the validity of the weight value, W is normalized using the Softmax function to obtain the normalized weight value:

[0115]

[0116] W i are different weights for different feature maps. Finally, the fused feature map is obtained by weighted summation of the feature maps at each level in the scale dimension and their corresponding weights:

[0117]

[0118] Through the above process, the feature fusion method realizes the dynamic fusion of adaptive weights of multi-scale feature maps, effectively enhancing the multi-scale adaptability of feature expression.

[0119] Then, the feature interaction method is used to further enhance the interaction ability between fusion features. First, the fusion feature map output by the feature fusion method is converted into Flatten into 2D features:

[0120]

[0121] Next, channel attention and spatial attention calculations are performed separately. The specific process is as follows:

[0122] Interleaved channel attention, via a learnable projection matrix W Q ,W K ,W V Input features X flat Mapped into query (Query, Q), key (Key, K) and value (Value, V) matrices:

[0123] Q=W Q X flat ,K=W K X flat ,V=W V X flat

[0124] Then, we calculate the scaled dot product attention

[0125]

[0126] Among them, d k is the feature dimension of each attention head, QK T Indicates the matching degree of each query with all keys, and Softmax is used to normalize the attention weight. After the output of all attention heads is spliced ​​in the channel dimension, the interleaved channel attention output X is obtained through linear mapping. CA

[0127] X CA =W O [head1;head2;…;head n ],W O is a learnable parameter

[0128] Then, the gating mechanism is used to dynamically adjust the channel attention strength and calculate the gating factor:

[0129] g CA =σ(W g1 X CA +b g1 )

[0130] Among them, W g1 、b g1 is a learnable parameter, σ is the Softmax activation function, and the result is compressed into the interval [0,1].

[0131] The interleaved spatial attention method captures the dependencies between spatial positions and also calculates the gating factor gSA:

[0132] g SA =σ(Wg2 X SA +b g2 )

[0133] Among them, X SA is the spatial dimension feature output obtained by the spatial attention method, W g2 、b g2 is another set of gated learnable parameters.

[0134] The final feature interaction method outputs features by multiplying the interleaved channel attention output and the spatial attention output by the corresponding gating factor g CA and g SA , and summing over the channel and spatial dimensions to obtain:

[0135] X ICSA =g CA ·X CA +g SA ·X SA

[0136] in, The final output feature map of the feature interaction method integrates the channel and spatial attention information, making the feature expression more accurate and robust.

[0137] S4: Decoding stage: The encoder features are restored to spatial resolution step by step, ultimately achieving fine segmentation of the tooth image. The decoding stage includes step-by-step upsampling, skip connections, and final pixel-level classification. The specific process and formulas are described as follows:

[0138] The decoding stage uses three consecutive interpolation upsampling layers to gradually restore the spatial resolution of the feature map. Each upsampling layer performs the following operations in sequence: Let the input feature map of the lth upsampling layer be U l , then perform interpolation upsampling operation to obtain a feature map U with higher spatial resolution l+1 , the specific formula is:

[0139]

[0140] Among them, the Interpolation function uses trilinear interpolation, that is, the feature map is interpolated in three dimensions: depth, height, and width, so that the feature size is enlarged to gradually restore the resolution of the original image.

[0141] To preserve the details of the tooth anatomy extracted during the encoding phase and maintain semantic consistency of the features, after upsampling each interpolation, it is fused with the corresponding encoder features through a multi-level skip connection. The specific process is as follows:

[0142] Assume that the feature after interpolation upsampling of the lth decoding layer is The corresponding encoder level l feature is The fusion method of the two is splicing in the feature channel dimension:

[0143]

[0144] Then the fused features Further fusion and optimization through a convolution operation:

[0145]

[0146] are the learnable weights and biases of the l+1th layer decoder convolution;

[0147] The above process is repeated three times to gradually restore the feature map from low resolution to the original input resolution.

[0148] After the feature recovery process is completed, the final feature map with the same size as the input CBCT image is obtained. Then, the convolution is used to map the pixel-level segmentation prediction results, and the Sigmoid function is used:

[0149]

[0150] P h,w,d is the probability of belonging to the target category at position (h,w,d);

[0151] The pixel-level mask generated by the above process is the final tooth segmentation result of the CBCT image.

[0152] In order to verify the technical effect of the method for tooth segmentation in CBCT images of the present invention, the following test is provided:

[0153] 1) Dataset preprocessing, as follows:

[0154] The experimental data was derived from CBCT dental images, including dental scans from different individuals. All images were normalized before being fed into the network to ensure consistent pixel value distribution. Furthermore, to improve the model's generalization capabilities, data augmentation was performed during the training phase, including random rotation, flipping, contrast adjustment, and artifact simulation. Because CBCT images are 3D structures, 3D patch sampling was used for data slicing to accommodate deep learning network input and improve training efficiency. Experiments were conducted using a dataset of 98 CBCT dental images, 78 of which served as training and 20 as testing. A series of network training and testing procedures were designed.

[0155] (2) The network training process is as follows:

[0156] The present invention adopts the PyTorch deep learning framework, such as Figure 1 As shown in the figure, the network structure mainly includes an encoder, a feature fusion and interaction module, and a decoder. The specific training process is as follows:

[0157] The normalized CBCT image is input and sent to the network structure encoder to extract features. The encoder contains 3 downsampling layers. Each downsampling layer performs the following operations: 3D convolution (Conv3D) extracts local features and uses the ReLU activation function for nonlinear transformation.

[0158] Feature extraction: Dense connections enhance the transferability of low-level and high-level features, and recursive filtering is used to build a cross-layer global context memory pool to improve semantic representation. Max pooling reduces the resolution of feature maps to extract higher-level features. During the feature fusion stage, dynamic weight allocation is implemented to optimize the fusion of features at different scales. Furthermore, feature interaction methods are introduced, leveraging cross-dimensional feature collaboration to enhance the model's feature representation capabilities and provide clearer tooth boundary information.

[0159] Decoding: Three upsampling layers are used to gradually restore spatial resolution, combined with skip connections to facilitate cross-layer information transfer. Skip connections allow for the fusion of low-level detail features with high-level semantic features, ensuring the integrity of the tooth's anatomical structure and mitigating the impact of information loss during upsampling. The final output is calculated using a sigmoid activation function to generate a pixel-level segmentation probability map for the tooth.

[0160] Loss Function and Optimization Strategy: Dice Loss (cross entropy loss) was used to optimize tooth segmentation and improve prediction of small objects. The Adam optimizer (initial learning rate 0.0006, weight decay 1e-4) was used, combined with ReduceLROnPlateau learning rate scheduling, to improve model training stability and convergence speed. Batch normalization was used during training to stabilize model training and improve generalization.

[0161] Training details: The batch size is 1, the number of training rounds is 20, and mixed precision (AMP) training is used to improve computational efficiency.

[0162] Testing Methods and Evaluation Metrics: After training, the model was evaluated using 20 test CBCT images. The specific testing steps were as follows: The test images were fed into the trained network model and inference was performed. The model automatically predicted tooth regions and generated pixel-level segmentation results.

[0163] The following indicators are used to evaluate segmentation accuracy:

[0164] Dice Similarity Coefficient (DSC) evaluates the similarity between the segmentation result and the true label.

[0165] IoU (Intersection over Union) measures the intersection over union ratio of segmented areas.

[0166] HD (Hausdorff Distance), calculates the spatial consistency of the segmentation boundary.

[0167] Result analysis: The test results show that the network structure of the present invention has achieved excellent performance in the CBCT tooth segmentation task. The visualization results show that the present invention can accurately segment the tooth area in the CBCT image and maintain a high level of detail retention at the boundary between the tooth root and the crown.

[0168] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

[0169] Although the present invention has been described above with reference to embodiments, various modifications may be made thereto and equivalent components may be substituted without departing from the scope of the present invention. In particular, as long as there are no structural conflicts, the various features of the embodiments disclosed herein may be combined with each other in any manner, and the omission of an exhaustive description of such combinations in this specification is solely for the sake of space and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for tooth segmentation in CBCT images, characterized in that: Here are the steps: S1. Preprocessing: The CBCT image dataset is standardized and data enhanced in sequence; S2, encoding: The CBCT image preprocessed in S1 is input into the encoder part of the network structure to complete the step-by-step extraction of image features and obtain multi-scale features {X1, X2, X3}; S3, feature fusion and interaction: further fuse the multi-scale feature maps generated by each downsampling layer of the encoder network structure to obtain richer and more effective feature expression; S4. Decoding: The encoder features are restored to a certain spatial resolution level by level, ultimately achieving fine segmentation of the tooth image.

2. The method for tooth segmentation of a CBCT image according to claim 1, characterized in that: Step S1 also includes the following steps: using a 3D Patch extraction method to process large-scale CBCT data into blocks.

3. The method for tooth segmentation of CBCT images according to claim 1, characterized in that: In step S2, the encoding includes three consecutive downsampling layers, each of which performs the following operations in sequence: First, the three-dimensional convolution operation is used to extract local features to improve the feature expression ability of the network. The input feature of the first layer is set to X l-1 , then the convolution operation formula is defined as: in, is the output feature of the lth layer, Conv3D represents 3×3 convolution, and They represent the weight matrix and bias term of the convolution kernel of the lth layer respectively, and σ is a nonlinear activation function used to enhance the nonlinear feature learning ability of the model; Secondly, the features output by the convolution operation are input into the feature extraction method. The specific implementation of the feature extraction method is as follows: Let the output of the kth sublayer in the feature extraction method be Its mathematical expression is: In the above formula, [] represents the feature concatenation operation in the channel dimension. is the learning parameter of the feature extraction method. Mamba is a state space model, which is defined as follows: Mamba(X)=BN(X+f SSM (LN(X))); Among them, X is the input feature, BN is the layer normalization function, f SSM Represents the calculation of the state space model. The specific formula is as follows: f SSM (X)=Conv1D(X;W in )*h SSM (X); Among them, W in is the convolution kernel weight, * represents element-by-element multiplication, and the recursive filter function h SSM The specific expression is: h t =A·h t-1 +B·X t ,Y t =C·h t +D·X t ; Among them, A, B, C, and D are the state space model parameters of the feature extraction method, and h t is the hidden state, h t-1 is the hidden state of the previous time step, which retains the past information, X t and Y t They are input and output features respectively; Finally, the maximum pooling operation is used to reduce the spatial resolution of the feature map to extract higher-level and more abstract semantic features. The pooling operation formula is: in, Represents the output value at position (i, j, k) in the lth layer, Represents the value of position (p,q,r) in the original feature map, Ω i,j,k Represents the local area window used for pooling operation in the feature map. The above process is repeated three times in sequence, and each execution produces a higher level of semantic features, thus forming a multi-scale feature set output by the encoder: {X1, X2, X3}.

4. The method for tooth segmentation of a CBCT image according to claim 1, characterized in that: In step S3, the specific steps of feature fusion and interaction are as follows: First, perform global average pooling (GAP) on each level feature map to obtain the global feature vector. For the i-th level feature map X i , the calculation formula is: in: Among them, z i Represents the feature map calculated by the i-th layer, H i ×W i ×D i Represent the height, width and depth of the feature map of the i-th layer respectively, Represents the eigenvalue of the feature map of layer i at channel c, located at coordinates (h, w, d). Subsequently, the global eigenvectors of all levels are concatenated into a comprehensive eigenvector Z: Among them, C1+C2+C3 is the total number of channels after splicing, z 1 、z 2 、z 3 They are features from different layers, and then the comprehensive feature vector is mapped to the weight value W through a 1×1 convolution; Where W=[W 1, W 2, W3], respectively represent different weights of different feature maps, W fuse is the convolution weight matrix, b fuse is the bias term, both of which are learnable parameters; to ensure the validity of the weight value, W is normalized using the Softmax function to obtain the normalized weight value α i : Among them, W i are different weights of different feature maps. The fused feature map is obtained by weighted summation of the feature maps of each level in the scale dimension and their corresponding weights: Then, the feature interaction method is used to further enhance the interaction ability between fusion features. First, the fusion feature map output by the feature fusion method is transformed into Flatten into 2D features: Next, channel attention and spatial attention calculations are performed separately. The specific process is as follows: Through the learnable projection matrix W Q ,W K ,W V Input features X flat Mapped to a matrix of query Q, key K and value V: Q=W Q X flat ,K=W K X flat ,V=W V X flat ; Then, we calculate the scaled dot product attention: Among them, d k is the feature dimension of each attention head, QK T Indicates the matching degree of each query with all keys. Softmax is used to normalize the attention weight. After the output of all attention heads is spliced ​​in the channel dimension, the interleaved channel attention output X is obtained through linear mapping. CA ; X CA =W O [head1;head2;…;head n ],W O is a learnable parameter; Among them, head n Represents the number of attention heads. Then, the gating mechanism is used to dynamically adjust the channel attention strength and calculate the gating factor: g CA =σ(W g1 X CA +b g1 ); Among them, W g1 、b g1 is a learnable parameter, X CA is the channel dimension feature, σ is the Softmax activation function, which compresses the result into the interval [0,1]; The interleaved spatial attention method captures the dependencies between spatial positions and also calculates the gating factor gSA: g SA =σ(W g2 X SA +b g2 ); Among them, X SA is the spatial dimension feature output obtained by the spatial attention method, W g2 、b g2 is another set of gated learnable parameters; The final feature interaction method outputs features by multiplying the interleaved channel attention output and the spatial attention output by the corresponding gating factor g CA and g SA , and summing over the channel and spatial dimensions to obtain: X IA =g CA ·X CA +g SA ·X SA ; in, The final output feature map of the feature interaction method integrates the channel and spatial attention information, making the feature expression more accurate and robust.

5. The method for tooth segmentation of CBCT images according to claim 1, characterized in that: In step S4, decoding includes step-by-step upsampling, skip connections, and final pixel-level classification. The specific process is as follows: The decoding uses three consecutive interpolation upsampling layers to gradually restore the spatial resolution of the feature map. Each upsampling layer performs the following operations in sequence: Set the input feature map of the lth upsampling layer to U l , then perform interpolation upsampling operation to obtain a feature map U with higher spatial resolution l+1 , the specific formula is: Among them, the Interpolation function uses trilinear interpolation, that is, the feature map is interpolated in three dimensions: depth, height, and width, so that the feature size is enlarged to gradually restore the resolution of the original image. To preserve the details of the tooth anatomy extracted during the encoding phase and maintain semantic consistency of the features, after upsampling each interpolation, it is fused with the corresponding encoder features through a multi-level skip connection. The specific process is as follows: Assume that the feature after interpolation upsampling of the lth decoding layer is The corresponding encoder level l feature is The fusion method of the two is splicing in the feature channel dimension: Then the fused features Further fusion and optimization through a convolution operation: are the learnable weights and biases of the l+1th layer decoder convolution; The above process is repeated three times to gradually restore the feature map from low resolution to the original input resolution; After the feature recovery process is completed, the final feature map with the same size as the input CBCT image is obtained. Then, the convolution is used to map the pixel-level segmentation prediction results, and the Sigmoid function is used: P h,w,d is the probability of belonging to the target category at position (h,w,d); The pixel-level mask generated by the above process is the final tooth segmentation result of the CBCT image.

Images