CBCT image tooth segmentation method based on deep learning

Abstract

The invention provides a CBCT image tooth segmentation method based on deep learning, and relates to the technical field of oral cavity image segmentation, and the method comprises the steps: extracting a two-dimensional slice from a CBCT image as an input image; carrying out normalization, artifact suppression and data enhancement on the image to improve the preprocessing quality; constructing a segmentation network structure fusing self-calibration multi-scale attention and a mixed attention mechanism, wherein the segmentation network structure is used for extracting features and enhancing context information; training a model based on the tooth image data set, and performing parameter learning by adopting a joint loss function and an optimizer; and finally, tooth segmentation is realized through the trained model, and three-dimensional reconstruction is carried out. The method has the advantages of high segmentation precision, high processing speed, light structure and the like, and is suitable for automatic identification and diagnosis assistance of teeth in stomatology.

Description

Technical Field

[0001] This invention relates to the field of oral image segmentation technology, and in particular to a deep learning-based CBCT image tooth segmentation method. Background Technology

[0002] With the rapid development of digital medical imaging technology, cone-beam computed tomography (CBCT) has been widely used in the field of dentistry. CBCT can generate high-resolution three-dimensional images, clearly presenting the structure of teeth and maxillofacial bones, providing important imaging evidence for the diagnosis and treatment of dental diseases. Compared with traditional two-dimensional X-ray imaging, CBCT provides more comprehensive and accurate three-dimensional structural information, and is widely used in clinical scenarios such as dental implants, orthodontics, and root canal treatment, helping doctors to achieve more accurate assessments and treatment plans.

[0003] However, existing tooth segmentation methods still suffer from low accuracy, boundary adhesion, and incomplete segmentation when dealing with complex tooth structures, dense arrangements, blurred boundaries, and severe artifact interference in CBCT images. Traditional methods are sensitive to image quality and have poor generalization ability; while deep learning methods such as convolutional neural networks have strong feature extraction capabilities, they are still limited in modeling long-range dependencies and preserving spatial details; at the same time, although some Transformer-based models improve global modeling capabilities, their high computational complexity and poor real-time performance make it difficult to meet the dual requirements of segmentation accuracy and efficiency in clinical scenarios. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide a deep learning-based tooth segmentation method for CBCT images, which can solve several key problems existing in the tooth image segmentation methods of the prior art when processing CBCT images, including the difficulty of traditional convolutional neural networks in modeling long-range dependencies, blurred segmentation boundaries, weak recognition ability of small structures, and the high computational complexity of existing Transformer methods, which are difficult to meet the needs of real-time clinical processing.

[0005] To address the aforementioned technical problems, this invention provides a deep learning-based CBCT image tooth segmentation method, specifically comprising the following steps: S1: Extract two-dimensional slice images from CBCT images to construct a dental image dataset; S2: Perform preprocessing on the two-dimensional slice images in the tooth image dataset, including normalization, artifact suppression, and data augmentation, to improve image quality; S3: Construct a two-dimensional tooth segmentation model that integrates self-calibrating multi-scale attention and hybrid attention mechanisms to extract features and enhance contextual information; S4: The model is trained based on the preprocessed tooth image dataset, and the parameters are learned using a joint loss function and optimizer. S5: Use the trained model to segment teeth in the two-dimensional CBCT slice image to be processed, and perform three-dimensional reconstruction to obtain a complete three-dimensional segmented image of the dental arch.

[0006] Furthermore, step S1 specifically includes the following steps: S11: Slice the original 3D CBCT image according to the preset axial direction; S12: Select slices containing the dental arch region to construct a dental image dataset.

[0007] Furthermore, step S2 specifically includes the following steps: S21: By linear normalization, the pixel gray values ​​in the two-dimensional slice image are uniformly mapped to a fixed range of 0 to 1; S22: A comprehensive algorithm based on image reconstruction, filtering, and local enhancement is used to detect and suppress metal artifact regions in the image; S23: Perform various enhancement operations on the image, including random rotation, horizontal flipping, center cropping, and affine transformation, to construct a more diverse training sample set.

[0008] Furthermore, S3 specifically includes: S31: ResNet34 is used as the encoder backbone network, and a skip connection structure is introduced in its residual module to effectively alleviate gradient vanishing and improve the ability to express deep features; combined with the U-Net architecture, high and low layer feature fusion is achieved through the encoder-decoder structure; S32: Integrates the SCMA module to enhance the perception of complex structures. The SCMA module is used to model the scale inconsistency between structures of different sizes in dental images and enhances the model’s ability to resolve complex regions through a self-calibration mechanism. S33: Integrating the HAS module to coordinate local details and global context. The HAS module is used to combine local attention with global attention to achieve collaborative modeling of local edges and global semantic information, thereby improving tooth contour and contextual understanding. S34: Add multiple cross-layer connections between the encoder and decoder to directly pass shallow high-resolution features to the decoder, thereby improving the ability to reconstruct tooth boundary structures.

[0009] Furthermore, step S4 specifically includes: The preprocessed two-dimensional slice image samples were divided into training set, validation set and test set in a 6:2:2 ratio; The SCMA-HAS-UNet algorithm was used to train two-dimensional slices of teeth. The training process included setting the number of images processed per batch, the random number seed s, and the number of training rounds. The model used the AdamW optimizer with a learning rate of lr and a weight decay coefficient of p. We employ a weighted combination of the Dice loss function and the binary cross-entropy loss function for joint supervision to simultaneously optimize the overlap of segmented regions and pixel-level classification accuracy. When the model's performance on the validation set reaches a stable level, stop training and save the best model.

[0010] Furthermore, S5 specifically includes: The trained model preweights are input into the 2D CBCT slice image to be processed, and the corresponding tooth region segmentation mask is output. The segmentation mask results in the continuous slices are voxel aligned and stacked for reconstruction to obtain 3D tooth segmentation data consistent with the original CBCT image space. The 3D tooth regions are fused and stitched in spatial order to output a 3D segmentation image containing the complete dental arch structure.

[0011] To achieve the above objectives, the present invention also provides a deep learning-based CBCT image tooth segmentation device, including one or more processors, for implementing the above-described deep learning-based CBCT image tooth segmentation method.

[0012] To achieve the above objectives, the present invention also provides an electronic device, including a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-described deep learning-based CBCT image tooth segmentation method.

[0013] To achieve the above objectives, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described deep learning-based CBCT image tooth segmentation method.

[0014] The beneficial effects of this invention are: This invention proposes a deep learning-based tooth segmentation method for CBCT images, which integrates a self-calibrating multi-scale attention mechanism and a local-global hybrid attention mechanism, effectively improving the segmentation accuracy of tooth structures in complex CBCT images. The constructed SCMA-HAS-UNet network not only possesses strong multi-scale feature extraction capabilities but also captures contextual semantics while preserving detailed information, achieving robust adaptation to various complex clinical scenarios such as tightly packed dentition, blurred boundaries, and morphological variations. This method is applicable to dental CBCT images from different sources, exhibiting strong versatility and promotional value, and can be widely applied in oral healthcare scenarios such as clinical auxiliary diagnosis, digital orthodontic planning, and implant surgery navigation. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall process of the CBCT image tooth segmentation method based on deep learning provided by the present invention. Figure 2 This is a flowchart of the steps in the CBCT image preprocessing process of the present invention; Figure 3 This is a schematic diagram of the SCMA-HAS-UNet two-dimensional tooth image segmentation network structure constructed in this invention; Figure 4 This is a schematic diagram of the structure of the device of the present invention; Figure 5 This is a schematic diagram of an electronic device according to the present invention. Detailed Implementation

[0016] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions 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, not all embodiments. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0017] The following description, in conjunction with the accompanying drawings, details the deep learning-based CBCT image tooth segmentation method provided by the present invention through specific embodiments and application scenarios.

[0018] Reference manual attached Figure 1 The tooth segmentation method based on deep learning in CBCT images in this embodiment specifically includes the following steps: S1: Extract two-dimensional slice images from CBCT images to construct a dental image dataset; This invention uses an original three-dimensional CBCT image dataset provided by a hospital for training, testing, and evaluation; Informed consent from the hospital was obtained before using the dataset, allowing its three-dimensional CBCT images to be used for research purposes; S2: Perform preprocessing on the two-dimensional slice images in the tooth image dataset, including normalization, artifact suppression, and data augmentation, to improve image quality; S3: Construct a two-dimensional tooth segmentation model that integrates self-calibrating multi-scale attention and hybrid attention mechanisms to extract features and enhance contextual information; S4: The model is trained based on the preprocessed tooth image dataset, and the parameters are learned using a joint loss function and optimizer. S5: Use the trained model to segment teeth in the two-dimensional CBCT slice image to be processed, and perform three-dimensional reconstruction to obtain a complete three-dimensional segmented image of the dental arch.

[0019] In this preferred embodiment, step S1 involves extracting two-dimensional slice images from the input three-dimensional CBCT volume image for model training and prediction, specifically including the following steps.

[0020] S11: Slice the original 3D CBCT image according to a preset axial direction. Preferably, the slice spacing is consistent with the voxel size of the CBCT image, and the spatial coordinates of the image are kept consistent during the slicing process.

[0021] S12: After slicing, select valid slice ranges containing the dental arch region from the generated 2D image. This selection process is based on the image histogram distribution, preferably retaining continuous slices containing dental structures and removing invalid image regions in the upper and lower layers that contain only air or non-target structures, thereby constructing a high-quality input image set.

[0022] In this preferred embodiment, combined with Figure 2 In step S2, the two-dimensional slice image extracted from the CBCT image undergoes image preprocessing to improve the input data quality, reduce noise interference, and enhance the model's robustness. Specifically, this includes the following steps: S21: Map the pixel grayscale values ​​in the two-dimensional slice image to a fixed range [0, 1]. This step is accomplished through linear normalization, which standardizes the original pixel values ​​of the image according to their maximum and minimum values ​​to unify the image brightness distribution.

[0023] S22: A comprehensive algorithm based on image reconstruction, filtering, and local enhancement is used to detect and suppress metallic artifact regions in the image. Preferably, edge detection and frequency domain analysis methods can be combined to identify artifact feature regions, and median filtering algorithm can be used to locally reconstruct and enhance artifact blocks.

[0024] S23: Perform various data augmentation operations on the normalized training images to expand sample diversity and improve the model's generalization ability. Augmentation operations include: random rotation (±10°), horizontal flipping, center cropping (cropping the original image to 536×536), and affine transformations (such as translation, scaling, and shearing).

[0025] In this preferred embodiment, combined with Figure 3 In step S3, a deep image segmentation network structure integrating multi-scale attention and local-global context awareness is constructed to efficiently extract tooth region features from CBCT images. Specifically, this includes: S31: ResNet34 is used as the encoder backbone network, and skip connections are introduced into its residual modules to effectively alleviate gradient vanishing and improve the ability to represent deep features. Combined with the U-Net architecture, high- and low-level features are fused through an encoder-decoder structure.

[0026] S32: Integrates the SCMA module to enhance the perception of complex structures. The SCMA (Self-Calibrated Multi-scale Attention) module is used to model the scale inconsistency between different sized structures (such as crowns and roots) in dental images and enhances the model's ability to distinguish complex regions (such as overlaps and fractures) through a self-calibration mechanism.

[0027] Specifically, its core includes two processes: multi-scale feature extraction and scale fusion. Input feature map , Indicates a channel. Indicates altitude, The width is represented by three convolutional kernels of different scales. After processing, we get: ; in, This represents the feature map after processing by the i-th convolutional kernel. This indicates a convolution operation performed using kernel Ki.

[0028] Channel-level fusion of features at various scales: ; in, This represents the feature map after multi-scale feature fusion. Channel-level splicing operation, This represents the feature map after processing by the first convolutional kernel. This represents the feature map after processing by the second convolutional kernel. This represents the feature map after processing by the third convolutional kernel.

[0029] Weight vectors are generated through a self-calibrating attention mechanism. ,right Adaptive weighting is applied to each channel: in, This represents the feature map after being weighted by the self-calibrating attention mechanism.

[0030] This module improves the model's sensitivity to features in scenarios such as metal artifacts and blurred edges.

[0031] S33: The HAS module is integrated to coordinate local details and global context. The HAS (Hybrid Attention Synergy) module is used to combine local attention (Local Self-Attention) and global attention (Global ContextAttention) to achieve collaborative modeling of local edges and global semantic information, thereby improving the ability to understand tooth contours and context.

[0032] Specifically, the HAS structure includes two branches: Local attention branch: in, Represents the query matrix. Represents the key matrix. Represents a value matrix, This represents the learnable weights of the query matrix. The learnable weights represent the key matrix. The learnable weights represent the values ​​in the matrix. This represents the output of the local attention branch. Represents the scaling factor, input features , It is a learnable relative positional encoding. Indicates the batch size of the input features. Indicates the number of attention heads. This indicates the spatial dimensions of the feature map.

[0033] Global context branch: in, This indicates a downsampling operation. This represents the output of the global context branch. Indicates the number of attention heads. The weights of the i-th attention head are represented by the input features. .

[0034] Ultimately, through learnable parameters To merge: in, This represents the final output of the hybrid attention mechanism. This indicates the weight balance of attention fusion, which has the ability to be dynamically adjusted.

[0035] S34: To avoid the loss of spatial localization information in deep networks, multiple skip connections are added between the encoder and decoder to directly pass shallow high-resolution features to the decoder, improving the reconstruction capability of tooth boundary structures. Furthermore, a Scconv (Spatial and Channel Reconstruction Convolution) convolutional module is integrated to reduce spatial and channel redundancy between features in the convolutional neural network, thereby compressing the CNN model and improving its performance.

[0036] The SCMA-HAS-UNet network structure constructed in this embodiment has both multi-scale modeling capabilities and local-global collaborative perception capabilities. It has significant advantages in handling various CBCT scenarios such as complex tooth structures, blurred boundaries, and metal artifact interference, and can provide a strong structural foundation for subsequent model training.

[0037] In this preferred embodiment, in step S4, the constructed SCMA-HAS-UNet model is trained based on the tooth image dataset. The training process of the two-dimensional tooth segmentation model specifically includes the following steps: The preprocessed two-dimensional slice image samples were divided into training set, validation set and test set in a ratio of 6:2:2.

[0038] In this embodiment, the preprocessed two-dimensional slice image samples are specifically divided into a training set of 3000 images, a validation set of 1000 images, and a test set of 1000 images.

[0039] The SCMA-HAS-UNet algorithm was used to train two-dimensional slices of teeth. Specifically, during the model training phase, the batch size, random number seed s, and training epochs were set. The AdamW optimizer was used, with a learning rate of lr and a weight decay coefficient of p.

[0040] In this embodiment, specifically, 8 images are processed per batch (Batch size is 8), the random seed is set to 42, the number of training rounds is set to 40, the learning rate is set to 0.001, and the weight decay coefficient is set to 1e-8.

[0041] We employ a weighted combination of the Dice loss function and the Binary Cross Entropy (BCE) loss function for joint supervision to simultaneously optimize the overlap of segmented regions and pixel-level classification accuracy.

[0042] In this embodiment, specifically, the Dice loss function Used to measure the overlap between the predicted region and the true label, it is robust to small targets and class imbalance; BCE loss function Point-by-point supervision of pixel-level predictions helps improve the ability to learn boundary details. Joint loss function. The expression is as follows: ; in, These are the weighting coefficients for the two sub-losses, which are set to be equal by default. .

[0043] When the model's performance on the validation set reaches a stable level, stop training and save the best model.

[0044] In this preferred embodiment, in step S5, after loading the pre-trained weights onto the trained SCMA-HAS-UNet model, the corresponding two-dimensional CBCT slice image to be processed is input, and the corresponding tooth region segmentation mask image is output. All slice mask results are voxel aligned and stacked according to their spatial order in the original CBCT volumetric image to obtain three-dimensional segmentation volume data consistent with the original image space. Finally, the tooth regions are stitched and fused to output a complete three-dimensional segmentation image of the dentition, which can be used for subsequent clinical visualization, analysis, or assisted treatment.

[0045] Corresponding to the aforementioned embodiments of the deep learning-based CBCT image tooth segmentation method, the present invention also provides embodiments of a deep learning-based CBCT image tooth segmentation device.

[0046] See Figure 4The deep learning-based CBCT image tooth segmentation device provided in this embodiment of the invention includes one or more processors for implementing the deep learning-based CBCT image tooth segmentation method in the above embodiment.

[0047] The embodiments of the deep learning-based CBCT image tooth segmentation device of the present invention can be applied to any device with data processing capabilities, such as a computer. The device embodiments can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 4 The diagram shown is a hardware structure diagram of any data processing-capable device, including the deep learning-based CBCT image tooth segmentation device of this invention. (Except for...) Figure 4 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0048] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0049] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0050] Corresponding to the aforementioned embodiments of the deep learning-based CBCT image tooth segmentation method, this application also provides an electronic device, including: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the deep learning-based CBCT image tooth segmentation method as described above. Figure 5The diagram shown is a hardware structure diagram of any device with data processing capabilities for the deep learning-based CBCT image tooth segmentation method provided in this application embodiment, except... Figure 5 In addition to the processor, memory, DMA controller, disk, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0051] Corresponding to the aforementioned embodiments of the deep learning-based CBCT image tooth segmentation method, this embodiment of the invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the deep learning-based CBCT image tooth segmentation method described in the above embodiments.

[0052] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0053] Finally, it should be noted that the specific embodiments described above are only for illustrating the technical solutions of the present invention, and not for limiting it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still modify or make equivalent substitutions to the technical features without departing from the spirit and substance of the present invention, and such modifications or substitutions should all be covered within the protection scope of the present invention.

Claims

1. A deep learning-based CBCT image tooth segmentation method, characterized in that, include: S1: Extract two-dimensional slice images from CBCT images to construct a dental image dataset; S2: Perform preprocessing on the two-dimensional slice images in the tooth image dataset, including normalization, artifact suppression, and data augmentation, to improve image quality; S3: Construct a two-dimensional tooth segmentation model that integrates self-calibrating multi-scale attention and hybrid attention mechanisms to extract features and enhance contextual information; S4: The model is trained based on the preprocessed tooth image dataset, and the parameters are learned using a joint loss function and optimizer. S5: Use the trained model to segment teeth in the two-dimensional CBCT slice image to be processed, and perform three-dimensional reconstruction to obtain a complete three-dimensional segmented image of the dental arch.

2. The deep learning-based CBCT image tooth segmentation method according to claim 1, characterized in that: Step S1 specifically includes the following steps: S11: Slice the original 3D CBCT image according to the preset axial direction; S12: Select slices containing the dental arch region to construct a dental image dataset.

3. The deep learning-based CBCT image tooth segmentation method according to claim 1, characterized in that: S2 specifically includes the following steps: S21: By linear normalization, the pixel gray values ​​in the two-dimensional slice image are uniformly mapped to a fixed range of 0 to 1; S22: A comprehensive algorithm based on image reconstruction, filtering, and local enhancement is used to detect and suppress metal artifact regions in the image; S23: Perform various enhancement operations on the image, including random rotation, horizontal flipping, center cropping, and affine transformation, to construct a more diverse training sample set.

4. The deep learning-based CBCT image tooth segmentation method according to claim 1, characterized in that: S3 specifically includes: S31: ResNet34 is used as the encoder backbone network, and a skip connection structure is introduced in its residual module to effectively alleviate gradient vanishing and improve the ability to express deep features; combined with the U-Net architecture, high and low layer feature fusion is achieved through the encoder-decoder structure; S32: Integrates the SCMA module to enhance the perception of complex structures. The SCMA module is used to model the scale inconsistency between structures of different sizes in dental images and enhances the model’s ability to resolve complex regions through a self-calibration mechanism. S33: Integrating the HAS module to coordinate local details and global context. The HAS module is used to combine local attention with global attention to achieve collaborative modeling of local edges and global semantic information, thereby improving tooth contour and contextual understanding. S34: Add multiple cross-layer connections between the encoder and decoder to directly pass shallow high-resolution features to the decoder, thereby improving the ability to reconstruct tooth boundary structures.

5. The deep learning-based CBCT image tooth segmentation method according to claim 1, characterized in that: Step S4 specifically includes: The preprocessed two-dimensional slice image samples were divided into training set, validation set and test set in a 6:2:2 ratio; The SCMA-HAS-UNet algorithm was used to train two-dimensional slices of teeth. The training process included setting the number of images processed per batch, the random number seed s, and the number of training rounds. The model used the AdamW optimizer with a learning rate of lr and a weight decay coefficient of p. We employ a weighted combination of the Dice loss function and the binary cross-entropy loss function for joint supervision to simultaneously optimize the overlap of segmented regions and pixel-level classification accuracy. When the model's performance on the validation set reaches a stable level, stop training and save the best model.

6. The deep learning-based CBCT image tooth segmentation method according to claim 1, characterized in that: S5 specifically includes: The trained model preweights are input into the 2D CBCT slice image to be processed, and the corresponding tooth region segmentation mask is output. The segmentation mask results in the continuous slices are voxel aligned and stacked for reconstruction to obtain 3D tooth segmentation data consistent with the original CBCT image space. The 3D tooth regions are fused and stitched in spatial order to output a 3D segmentation image containing the complete dental arch structure.

7. A deep learning-based CBCT image tooth segmentation device, characterized in that, It includes one or more processors for implementing the deep learning-based CBCT image tooth segmentation method according to any one of claims 1-6.

8. An electronic device comprising a memory and a processor, characterized in that, The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the deep learning-based CBCT image tooth segmentation method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the deep learning-based CBCT image tooth segmentation method as described in any one of claims 1-6.

Images