Tooth image processing method, device, equipment, storage medium and program product

Abstract

The application provides a tooth image processing method, device, equipment, storage medium and program product, and relates to the technical field of image processing, wherein the method comprises: acquiring a tooth magnetic resonance imaging to be segmented; and performing segmentation processing on the tooth magnetic resonance imaging by using a segmentation model to obtain a segmentation image comprising a periodontal ligament and a pulp cavity region; wherein the segmentation model is used for sequentially performing feature extraction and feature fusion on the tooth magnetic resonance imaging, and performing segmentation processing on the tooth magnetic resonance imaging based on global-local fusion features obtained through feature fusion; the segmentation model is trained by using a preset loss function, the preset loss function is determined based on a dice loss and a topology loss, and the topology loss is determined based on a pixel-level loss, a predicted mask, an actual mask and key pixels. The application can improve the accuracy of tooth image segmentation results and realize fine segmentation of the periodontal ligament and the pulp cavity region in the tooth magnetic resonance imaging.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method, apparatus, device, storage medium, and program product for processing dental images. Background Technology

[0002] Teeth are primarily composed of enamel, dentin, pulp, and periodontal tissues. The pulp, a loose connective tissue located within the pulp cavity, plays a crucial role in maintaining normal tooth function, providing sensory, nutritional, and immune protection functions. Surrounded by inflexible dentin, nerves and blood vessels can only enter through the narrow apical foramen, lacking effective collateral blood circulation and thus limiting its self-repair capabilities. Under external stimuli such as infection and trauma, pulpitis or even liquefactive necrosis can develop, causing significant pain for patients and increasing the social medical burden. Untreated pulpitis can worsen and ultimately lead to tooth loss. Periodontal tissues refer to the functional system surrounding the teeth, composed of the gingiva, alveolar bone, periodontal ligament, and cementum. The gingiva and periodontal ligament are key components of the periodontal tissues. The pathological process of periodontal disease usually begins with dental plaque. Initially, it manifests as inflammatory reactions such as gingival redness, swelling, and bleeding. As the disease progresses, it can develop into periodontitis, characterized by loss of periodontal fiber attachment and alveolar bone resorption, ultimately leading to tooth loss. The prevalence of periodontitis in adults is as high as 80%-90%, making it the leading cause of tooth loss in adults.

[0003] These dental diseases severely impact people's oral health and quality of life. Therefore, accurate assessment of the condition of the internal and surrounding soft tissues (including pulp and periodontal tissues) is crucial for accurate clinical diagnosis, rational treatment planning, and treatment implementation. Currently, the assessment of pulp and periodontal tissues relies primarily on clinical examination. However, the accuracy and reliability of traditional assessment methods can be significantly affected when patients cannot accurately describe their symptoms or when examination is difficult due to specific anatomical structures. Therefore, obtaining high-quality oral imaging data and achieving high-precision segmentation and visualization of anatomical structures is of significant practical importance for improving clinical diagnosis and treatment.

[0004] Currently, clinical dental imaging primarily relies on cone-beam computed tomography (CBCT) images. However, CBCT images cannot show the specific distribution of blood vessels and nerves in the pulp chamber, provide insufficient imaging of bone microstructure, and often fail to visualize the periodontal ligament structure. Existing segmentation models are also not well-suited for segmenting the periodontal ligament and pulp chamber, resulting in generally poor accuracy. Summary of the Invention

[0005] This invention provides a method, apparatus, device, storage medium, and program product for processing dental images, in order to solve the defect of poor accuracy in dental image segmentation results in the prior art.

[0006] This invention provides a method for processing dental images, comprising:

[0007] Obtain magnetic resonance images of the teeth to be segmented;

[0008] The dental magnetic resonance imaging was segmented using a segmentation model to obtain a segmented image including the periodontal ligament and pulp chamber regions.

[0009] The segmentation model is used to sequentially extract and fuse features from the dental magnetic resonance imaging, and to segment the dental magnetic resonance imaging based on the global-local fusion features obtained by feature fusion.

[0010] The segmentation model is trained using a preset loss function, which is determined based on dice loss and topological loss. The topological loss is determined based on pixel-level loss, predicted mask, real mask, and key pixels.

[0011] According to a tooth image processing method provided by the present invention, the method involves segmenting the dental magnetic resonance imaging using a segmentation model to obtain a segmented image including the periodontal ligament and pulp chamber regions, comprising:

[0012] The local feature extraction layer of the segmentation model is used to extract local features from the dental magnetic resonance imaging to obtain local features;

[0013] The global feature extraction layer of the segmentation model is used to extract global features from the dental magnetic resonance imaging to obtain global features;

[0014] The feature fusion layer of the segmentation model is used to fuse the local features and the global features to obtain global-local fused features.

[0015] The image segmentation layer of the segmentation model is used to segment the dental magnetic resonance imaging based on the global-local fusion features to obtain a segmented image including the periodontal ligament and pulp chamber regions.

[0016] According to a tooth image processing method provided by the present invention, the preset loss function is as follows:

[0017] L = L dice + λL topo ;

[0018] in, L For the total loss,L dice Loss due to dice, λ The weighting factor for the topological loss. L topo The topology loss is as follows:

[0019] ;

[0020] in, L pixel For pixel-level loss, To predict the mask, For the real mask, For Hadama accumulation, V These are key pixels, as shown below:

[0021] ;

[0022] in, As background mask, For dilating the background mask, For teeth masking, The expanded tooth mask is obtained by expanding the background mask.

[0023] According to a tooth image processing method provided by the present invention, the tooth image processing method further includes:

[0024] Spatial analysis was performed on the magnetic resonance imaging of the teeth to obtain three-dimensional spatial information;

[0025] Based on the three-dimensional spatial information, the dental magnetic resonance imaging is divided into voxels to obtain the vertex values ​​of each voxel after voxel division.

[0026] Based on the vertex values ​​of each voxel and the preset isosurface threshold, isosurface points are determined;

[0027] Based on the segmented image including the periodontal ligament and pulp chamber region, the boundary information of the periodontal ligament and pulp chamber is extracted, and the boundary points are determined based on the boundary information;

[0028] Based on the isopleth points and the boundary points, triangular patches are generated, and a three-dimensional surface model is constructed based on the generated triangular patches.

[0029] According to a tooth image processing method provided by the present invention, after constructing a three-dimensional surface model based on generated triangular patches, the method further includes:

[0030] The three-dimensional surface model is optimized to obtain an optimized three-dimensional surface model;

[0031] The optimization process includes one or more of mesh optimization, smoothing, and feature mapping.

[0032] According to a method for processing dental images provided by the present invention, the step of acquiring magnetic resonance images of teeth to be segmented includes:

[0033] Obtain the raw magnetic resonance image;

[0034] The target detection model was used to detect the regions of each tooth in the original magnetic resonance imaging.

[0035] The original magnetic resonance image is cropped based on the regions detected for each tooth to obtain a segmented magnetic resonance image of the tooth.

[0036] The present invention also provides a dental image processing device, comprising the following modules:

[0037] The acquisition module is used to acquire magnetic resonance images of the teeth to be segmented.

[0038] The segmentation module is used to segment the dental magnetic resonance imaging using a segmentation model to obtain segmented images including the periodontal ligament and pulp chamber regions.

[0039] The segmentation model is used to sequentially extract and fuse features from the dental magnetic resonance imaging, and to segment the dental magnetic resonance imaging based on the global-local fusion features obtained by feature fusion.

[0040] The segmentation model is trained using a preset loss function, which is determined based on dice loss and topological loss. The topological loss is determined based on pixel-level loss, predicted mask, real mask, and key pixels.

[0041] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the dental image processing method as described above.

[0042] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the tooth image processing method as described above.

[0043] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the tooth image processing method as described above.

[0044] This invention provides a method, apparatus, device, storage medium, and program product for processing dental images. It acquires magnetic resonance imaging (MRI) images of teeth to be segmented, and then uses a segmentation model to segment these images, obtaining segmented images including the periodontal ligament and pulp chamber regions. This invention utilizes MRI, which offers superior imaging performance. The segmentation model extracts and fuses global and local features from the dental MRI images. Based on the fused global-local features, the dental MRI images are segmented. Compared to extracting single local or global features, this invention improves the accuracy of dental image segmentation. Furthermore, the segmentation model is trained using a pre-defined loss function. This pre-defined loss function introduces topological loss on top of dice loss. Topological loss is determined based on pixel-level loss, the predicted mask, the ground truth mask, and key pixels. For predicted masks that differ from corresponding key pixels in the ground truth mask, pixel-level penalties are applied. By introducing this topological loss, the model is better encouraged to optimize the segmentation results, thereby further improving the segmentation performance. Through the above methods, the present invention can improve the accuracy of tooth image segmentation results and achieve fine segmentation of the periodontal ligament and pulp cavity regions in dental magnetic resonance imaging. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0046] Figure 1 This is a schematic flowchart of the tooth image processing method provided by the present invention.

[0047] Figure 2 This is a model architecture diagram of the segmentation model provided by the present invention.

[0048] Figure 3 This is one of the comparison diagrams showing the segmentation effect of the periodontal ligament and pulp chamber region provided by the present invention.

[0049] Figure 4 This is the second comparison diagram of the segmentation effect of the periodontal ligament and pulp chamber region provided by the present invention.

[0050] Figure 5 This is a schematic diagram of the dental image processing device provided by the present invention.

[0051] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0053] Currently, clinical dental imaging relies primarily on CBCT images. However, CBCT images cannot clearly show the specific distribution of blood vessels and nerves within the pulp chamber, provide insufficient imaging of bone microstructure, and often fail to visualize the periodontal ligament structure. Existing segmentation models are also inadequate for segmenting the periodontal ligament and pulp chamber, resulting in generally poor accuracy.

[0054] To address the aforementioned problems, this invention proposes a method, apparatus, device, storage medium, and program product for processing dental images. The following details are related to these methods. Figures 1-6 Describe it.

[0055] Figure 1 This is a flowchart illustrating the tooth image processing method provided by the present invention, as shown below. Figure 1 As shown, the method includes:

[0056] Step S110: Obtain magnetic resonance imaging of the tooth to be segmented.

[0057] Currently, the main diagnostic methods for oral imaging are X-rays and CBCT images. Compared to X-rays, oral CBCT images can reflect the tissue condition from a three-dimensional perspective, detecting lesions that cannot be seen by the projection angle of oral X-rays, or those that are more subtle. However, CBCT images also have significant limitations: in terms of tissue imaging, CBCT cannot show the specific distribution of blood vessels and nerves in the pulp cavity, and the imaging of bone microstructure is not clear enough. CBCT can only roughly see cortical bone and cancellous bone, and cannot see the details of soft tissue and trabecular bone in the cancellous bone; furthermore, the periodontal ligament structure is usually not visible; and in terms of biosafety, the ionizing radiation of CBCT remains a concern.

[0058] Therefore, in this embodiment of the invention, magnetic resonance imaging (MRI) images with higher resolution and tissue contrast are used, which have significantly better imaging effects on dental pulp and periodontal soft tissues than CBCT images.

[0059] As one implementation method, dental MRI is high-field MRI, with a magnetic field strength ≥5T corresponding to a high field. It should be understood that compared to low-field MRI, high-field MRI offers higher resolution and has unique advantages in imaging the structure and function of the dental pulp and periodontal soft tissues, helping to explore subtle structural changes in diseases. However, since high-field MRI is often accompanied by more pronounced motion artifacts, overcoming artifacts and noise interference by breaking through existing multi-channel imaging techniques with coils is a technical bottleneck that high-field MRI needs to overcome for clinical translation. This invention, through model improvement and the introduction of topological loss into the loss function, effectively solves the above problems, achieving refined segmentation of the periodontal ligament and pulp cavity regions in dental MRI images.

[0060] Furthermore, considering that MRI images are in DICOM (Digital Imaging and Communications in Medicine) format, the DICOM format can be converted to PNG (Portable Network Graphics) format, and the grayscale values ​​can be copied on each channel, thus converting a single-channel DICOM image into a three-channel RGB image.

[0061] In one implementation, the dental magnetic resonance image to be segmented is obtained by cropping the dental region of the original magnetic resonance image, wherein the original magnetic resonance image can be an oral magnetic resonance image or a head magnetic resonance image.

[0062] Step S120: The dental magnetic resonance imaging is segmented using a segmentation model to obtain a segmented image including the periodontal ligament and pulp chamber regions.

[0063] The segmentation model is used to sequentially extract and fuse features from the dental magnetic resonance imaging (MRI) images, and to segment the dental MRI images based on the global-local fusion features obtained by feature fusion. The segmentation model is trained using a preset loss function, which is determined based on dice loss and topological loss. The topological loss is determined based on pixel-level loss, predicted mask, real mask, and key pixels.

[0064] As one implementation method, the segmentation model is trained based on magnetic resonance imaging of sample teeth and the corresponding image segmentation region annotation results.

[0065] In one implementation, the segmentation model includes a local feature extraction layer, a global feature extraction layer, a feature fusion layer, and an image segmentation layer. The local feature extraction layer extracts local features from the input image; the global feature extraction layer extracts global features from the input image; the feature fusion layer converts the extracted features into the final output, specifically fusing local and global features to obtain global-local fused features; and the image segmentation layer performs segmentation processing on the dental MRI images to be segmented.

[0066] As one implementation method, the preset loss function can be a weighted sum of dice loss and topology loss.

[0067] As one implementation method, the dice loss can be determined based on the predicted mask and the true mask, using the following formula:

[0068] .

[0069] in, L dice Loss due to dice, To predict the mask, This is the real mask.

[0070] As one implementation method, the topology loss is as follows:

[0071] .

[0072] in, L topo For topological loss, L pixel For pixel-level loss, To predict the mask, For the real mask, For Hadama accumulation, V For key pixels. Pixel-level loss can be mean squared error (MSE) loss or cross-entropy (CE) loss.

[0073] The tooth image processing method provided by this invention acquires a magnetic resonance imaging (MRI) image of a tooth to be segmented, and then uses a segmentation model to segment the MRI image, obtaining a segmented image including the periodontal ligament and pulp cavity regions. This invention utilizes MRI, which offers superior imaging performance. The segmentation model extracts and fuses global and local features from the MRI image, and segments the image based on the fused global-local features. Compared to extracting single local or global features, this invention improves the accuracy of tooth image segmentation. Simultaneously, the segmentation model is trained using a preset loss function, which introduces topological loss in addition to dice loss. The topological loss is determined based on pixel-level loss, predicted mask, ground truth mask, and key pixels. For predicted masks that differ from corresponding key pixels in the ground truth mask, a pixel-level penalty is applied. By introducing this topological loss, the model is better encouraged to optimize the segmentation results, thereby further improving the segmentation performance. Through these methods, this invention improves the accuracy of tooth image segmentation, achieving refined segmentation of the periodontal ligament and pulp cavity regions in dental MRI images.

[0074] In one embodiment, step S120 includes steps S121, S122, S123 and S124.

[0075] Step S121: Local features are extracted from the dental magnetic resonance imaging using the local feature extraction layer of the segmentation model to obtain local features.

[0076] Step S122: Use the global feature extraction layer of the segmentation model to extract global features from the dental magnetic resonance imaging to obtain global features.

[0077] Step S123: Use the feature fusion layer of the segmentation model to perform feature fusion on the local features and the global features to obtain global-local fused features.

[0078] Step S124: The image segmentation layer of the segmentation model is used to segment the dental magnetic resonance imaging based on the global-local fusion features to obtain a segmented image including the periodontal ligament and pulp chamber regions.

[0079] Considering that the pure Vision Transformer (VT) model, which uses a multi-head self-attention mechanism to model global information and capture long-range dependencies between different locations, performs poorly, we aim to improve its performance when training from scratch on small datasets. To accelerate VT's convergence and significantly improve its performance, we employ a "convolution + transformer" approach for the dental MRI images to be segmented in the segmentation model. We introduce a lightweight Convolutional Neural Network (CNN) trained on the same dataset used by the VT model. By mimicking the features of a pre-trained CNN, we provide local guidance to VT to help it learn local information. The segmentation model architecture is as follows: Figure 2 As shown.

[0080] like Figure 2 As shown, the segmentation model includes a local feature extraction layer, a global feature extraction layer, a feature fusion layer, and an image segmentation layer. The local feature extraction layer extracts local features from the input image and can be composed of multiple convolutional neural network layers. The global feature extraction layer may include an image patching module and multiple Transformer Blocks. The image patching module segments the input image into multiple small patches; the Transformer Block is the core component of the Transformer architecture, consisting of two main sub-components: a multi-head attention mechanism and a feedforward neural network, used to extract global features from the input image. The feature fusion layer converts the features extracted by the model into the final output, specifically fusing local and global features to obtain global-local fused features, and can be a multilayer perceptron (MLP head). The image segmentation layer performs segmentation processing on the dental MRI images to be segmented.

[0081] In this embodiment, the dental magnetic resonance imaging (MRI) image is input to the convolutional neural network layer of the segmentation model, where the convolutional neural network layer extracts local features. Before inputting the MRI image into the convolutional neural network, the dental MRI image can be downsampled, which can reduce computational complexity and enhance the robustness of the model while retaining key information.

[0082] While extracting local features, the dental magnetic resonance imaging can be input into the Transformer module of the segmentation model to extract global features from the dental magnetic resonance imaging.

[0083] Then, the extracted local and global features are input into the feature fusion layer of the segmentation model. The feature fusion layer fuses the local and global features to obtain global-local fused features.

[0084] Finally, the image segmentation layer of the segmentation model is used to segment the dental magnetic resonance imaging based on global-local fusion features to obtain segmented images including the periodontal ligament and pulp chamber regions.

[0085] In this embodiment, a lightweight convolutional neural network is introduced into the VT model. This allows for feature extraction from the dental MRI images to be segmented using a "CNN+transformer" approach. Local features extracted by the CNN provide local guidance to the VT model, aiding in learning local information. Simultaneously, global features extracted by the transformer are used to fuse the local and global features, resulting in a global-local fusion feature for segmentation. Compared to using only CNN or transformer for feature extraction, this embodiment improves the accuracy of dental image segmentation results.

[0086] In one embodiment, the preset loss function is as follows:

[0087] L=L dice +λL topo .

[0088] in, L For the total loss, L dice Loss due to dice, λ The weighting factor for the topological loss. L topo This is the topological loss.

[0089] The dice losses are as follows:

[0090] .

[0091] in, To predict the mask, This is the real mask.

[0092] The topology loss is as follows:

[0093] .

[0094] in, L pixel For pixel-level loss, To predict the mask, For the real mask, For Hadama accumulation,V These are key pixels, as shown below;

[0095] .

[0096] in, As background mask, For dilating the background mask, For teeth masking, The expanded tooth mask is obtained by expanding the background mask.

[0097] In this embodiment, considering the segmentation task of the periodontal ligament and pulp cavity region, the periodontal ligament tightly surrounds the tooth root, and the pulp is located in the center of the tooth, maintaining a specific spatial relationship with the surrounding tissues. Analysis shows that the challenge in the segmentation task is to preserve its topological structure and shape; that is, the periodontal ligament wraps around the tooth and exhibits a characteristic ring-shaped appearance. By appropriately applying specific constraints for the region of interest and utilizing prior knowledge of human tooth anatomy, the efficiency and accuracy of segmentation can be significantly improved. Based on the above analysis, to better preserve the ring-shaped shape of the blood vessel wall, we employ a topological loss function. Based on the characteristic that the periodontal ligament separates the pulp from the surrounding tissues, we check whether the positional relationship between the periodontal ligament and pulp in the segmentation result conforms to the actual anatomical structure, applying an additional penalty factor to the voxels at their boundary. In this way, even if the shape of the blood vessel wall is not usually circular, a complete ring-shaped blood vessel wall structure can be generated.

[0098] The specific calculation process for topology loss is as follows:

[0099] First obtain the background mask (denoted as ). ), tooth mask (denoted as ), predictive mask (denoted as ) and the real mask (denoted as Then, the background mask is dilated to obtain the dilated background mask (denoted as ); The tooth mask is then expanded to obtain an expanded tooth mask (denoted as ). ).

[0100] During the expansion process, a template similar to a rhombus-4-adjacent kernel can be used. The rhombus-4-adjacent kernel uses... k This indicates that the dilated background mask and dilated tooth mask are calculated using this kernel. Specifically:

[0101] .

[0102] .

[0103] in, This represents the convolution operation. As background mask, For dilating the background mask, For teeth masking, For expanding tooth masks.

[0104] Then, based on the background mask, dilated background mask, tooth mask, and dilated tooth mask, the key pixels (denoted as ) are determined. V The specific formula is as follows:

[0105] .

[0106] in, It represents the Hadamardi (or Hadama) stack.

[0107] Next, for pixels that differ from their corresponding key pixels in the ground truth, thus violating the expected topological relationship, a pixel-level penalty needs to be applied. Specifically, this applies to fragmented segmentation results that do not exhibit a circular structure. Correspondingly, the topological loss is determined based on pixel-level loss, predicted mask, ground truth mask, and key pixels, as shown below:

[0108] .

[0109] in, L pixel For pixel-level loss, To predict the mask, For the real mask, V For key pixels.

[0110] Pixel-level loss can be either mean squared error (MSE) loss or cross-entropy (CE) loss. In this embodiment, the default configuration of cross-entropy loss is used as the pixel-level loss.

[0111] Finally, based on pixel-level loss and topological loss, the preset loss function is determined, as shown in the following formula:

[0112] L=L dice +λL topo .

[0113] By introducing the aforementioned topological loss into the preset loss function, the model can be better encouraged to optimize the segmentation results, thereby further improving the segmentation performance of the segmentation model.

[0114] In one embodiment, the segmentation model is trained based on magnetic resonance imaging of sample teeth and the corresponding image segmentation region annotation results.

[0115] As one implementation method, the sample tooth magnetic resonance imaging is human high-field oral MRI.

[0116] As another implementation method, the magnetic resonance imaging of the sample teeth includes human high-field oral MRI and beagle high-field oral MRI. It should be understood that, considering the limited availability of human high-field oral MRI data, animal high-field oral MRI can be selected as the sample data. The present invention chooses the beagle as the animal model because its teeth and periodontal tissues share many similarities with humans in terms of anatomical morphology, tissue structure, and physiological characteristics. On the one hand, as an animal model of periodontitis, the periodontal ligament collagen fibers in the permanent teeth of beagles are neatly arranged, the average width of the periodontal ligament at the alveolar ridge crest is 0.2 mm, and the trabeculae in the alveolar bone are thick and dense. Beagles are susceptible to periodontitis, exhibiting gingival redness, bleeding, and inflammatory cell infiltration in gingivitis, and gingival recession, attachment loss, and alveolar bone resorption in periodontitis. Compared with rat and mouse models, the pathological changes are closer to those in humans. On the other hand, as a model of pulpitis, the dental arch morphology of a beagle is similar to that of a human, with a distinction between anterior and posterior teeth. Its teeth and pulp chamber volume are larger than those of mice, making pulpotomy easier, and the cost is significantly lower than that of primates. Therefore, beagles are ideal animals for studying and observing pulpitis, periapical periodontitis, and periodontitis. Furthermore, each beagle has 6 anterior teeth and 16 premolars, with relatively wide interdental spaces, minimizing the impact of adjacent teeth. This allows for a reduction in the number of dogs required; therefore, beagles were chosen for this study to construct a model of pulpitis and periodontitis.

[0117] As another implementation method, the sample tooth resonance imaging includes human high-field oral MRI, beagle high-field oral MRI, human low-field oral MRI, beagle low-field oral MRI, and oral CBCT. This is because the amount of high-field MRI data is significantly less than that of low-field MRI and CBCT data. Therefore, pre-training can be carried out using CBCT and low-field MRI data first. Then, the pre-trained model can be transferred to the high-field MRI data task. During the transfer process, some structural parts of the model that need to be fine-tuned can be made. Finally, the fine-tuned model can be trained based on the high-field MRI data.

[0118] For example, a 5T high-field oral MRI dataset of humans and a 7T high-field oral MRI dataset of beagles were used as sample dental MRI images. All DICOM images were converted to PNG format, and grayscale values ​​were copied on each channel. Single-channel DICOM images were converted to three-channel RGB images, and the dataset was hierarchically divided into training, validation, and test sets in a 6:2:2 ratio.

[0119] For example, pre-training was performed using 20 CBCT data points, 20 human 1.5T low-field MRI data points, and 24 beagle dog 3T low-field MRI data points. At this stage, the model initially established some feature extraction and representation capabilities. After pre-training, the trained model parameters were transferred to tasks involving high-field MRI data. This transfer process typically requires fine-tuning some aspects of the model's structure. During fine-tuning, the model further optimized its parameters based on the characteristics of high-field MRI data.

[0120] Furthermore, data augmentation processing is performed on the sample tooth MRI images. Data augmentation methods include, but are not limited to, random image resizing, random cropping, and random horizontal flipping. Through data augmentation, the sample tooth MRI images can be further amplified, generating new data from the same data distribution but different from the original images, thereby increasing the diversity of training data and the generalization ability of the model.

[0121] As one implementation method, the image segmentation region can be labeled manually using 3D Slicer software to create a segmentation mask and perform one-hot encoding to facilitate subsequent training of the segmentation model.

[0122] In one embodiment, the evaluation metrics for model training can employ three metrics: DSC (Dice Similarity Coefficient), ASD (Average Surface Distance), and HD (Hausdorff Distance) to assess the performance of the segmentation model. This is because in high-field oral MRI, the periodontal ligament and pulp typically exhibit low contrast and blurring between surrounding tissues, and the edges and shapes of the periodontal ligament and pulp chamber vary significantly between different teeth, posing a major challenge to segmentation. Therefore, this invention focuses on the edge performance of image segmentation, hence the selection of the aforementioned three metrics.

[0123] DSC quantifies the degree of similarity between the automatically segmented region (AS) and the labeled region (GT) by calculating the ratio of their intersection to their union. DSC is the most commonly used metric for evaluating neural network segmentation performance in medical image segmentation, representing the similarity between the segmented result and the labeled region, and effectively reflecting the internal filling quality of the segmentation. The DSC range is 0 to 1, where 1 represents a segmentation result completely identical to manual segmentation; therefore, a higher DSC indicates a higher degree of similarity between the segmented result and the labeled region. The formula for calculating DSC is as follows:

[0124] .

[0125] ASD and HD are distance-related parameters used to evaluate the segmentation effect of medical images, and they can reflect the segmentation effect of image edges well.

[0126] Where ASD represents all points on the edge of the segmented image (B AS ) and the edge points of the labeled area (B) GT The average distance (ASD) can reliably reflect the overall segmentation result; a smaller ASD indicates a better segmentation effect from the neural network. The formula for calculating ASD is as follows:

[0127] .

[0128] HD represents the maximum distance between the nearest edge point (AS) of the segmented image and the nearest edge point of the labeled region (GT). Compared to ASD, HD is more sensitive to outliers; a smaller HD indicates better segmentation performance from the neural network. The formula for calculating HD is as follows:

[0129] .

[0130] Based on any of the above embodiments, the tooth image processing method further includes:

[0131] Step S130: Perform spatial analysis on the magnetic resonance imaging of the teeth to obtain three-dimensional spatial information.

[0132] Step S140: Based on the three-dimensional spatial information, the dental magnetic resonance imaging is divided into voxels to obtain the vertex values ​​of each voxel after voxel division.

[0133] Step S150: Determine the isopleth points based on the vertex values ​​of each voxel and the preset isopleth threshold.

[0134] Step S160: Based on the segmented image including the periodontal ligament and pulp cavity region, extract the boundary information of the periodontal ligament and pulp cavity, and determine the boundary points based on the boundary information.

[0135] Step S170: Generate triangular patches based on the isopleth points and the boundary points, and construct a three-dimensional surface model based on the generated triangular patches.

[0136] Considering the limitations of high-field oral MRI, three-dimensional reconstruction of the periodontal ligament and pulp cavity tissues has not yet been achieved. Therefore, this embodiment utilizes an improved algorithm to realize three-dimensional reconstruction of dental MRI.

[0137] It should be noted that the execution order of steps S130-S150 is after step S110, and the execution order of step S160 is after step S120.

[0138] Specifically, the process begins with spatial analysis of the dental MRI images to obtain three-dimensional spatial information. This information represents one or more aspects of the tooth's geometry, size, volume, and color in the three-dimensional space, specifically depth values ​​and point cloud data. Next, based on this three-dimensional spatial information, the dental MRI images are finely segmented into voxels, and the vertex values ​​of each voxel are obtained. Then, the vertex values ​​of each voxel are compared with a preset isosurface threshold. The intersection of the isosurface and the voxel is determined based on the comparison results; the intersection points are the isopoints. After obtaining segmented images of the periodontal ligament and pulp chamber regions based on the segmentation model, the boundary information of these regions can be extracted, and boundary points are determined based on this boundary information. Finally, triangular patches are generated based on the isopoints and boundary points. These triangular patches are then tightly stitched together to gradually construct a three-dimensional surface model of the periodontal ligament and pulp chamber.

[0139] In one embodiment, during the construction process, an adaptive mesh generation technique is used to dynamically adjust the mesh density according to the complexity of the structure, thereby effectively controlling the amount of data in the model while ensuring the accuracy of the three-dimensional surface model.

[0140] In one embodiment, a curvature-based smoothing algorithm is introduced during the connection of isopleths and boundary points to ensure a natural transition between patches and avoid jagged or discontinuous surfaces.

[0141] In one embodiment, morphological operations are combined to refine and optimize the boundaries, further improving the accuracy and integrity of the three-dimensional surface model.

[0142] In this embodiment, based on the use of the Marching Cubes (MC) algorithm to determine contour points, the boundary points of the periodontal ligament and pulp cavity regions are also determined based on the segmented images including these regions. Combining the contour points and the boundary points, triangular patches are generated, and a three-dimensional surface model is constructed based on these patches. By introducing the method of determining the boundary points of the periodontal ligament and pulp cavity regions based on the segmentation results, high-precision three-dimensional reconstruction of the periodontal ligament and pulp cavity can be achieved.

[0143] Based on the above embodiments, after step S170, the tooth image processing method further includes:

[0144] The three-dimensional surface model is optimized to obtain an optimized three-dimensional surface model; wherein the optimization process includes one or more of mesh optimization, smoothing, and feature mapping.

[0145] After the initial completion of the 3D surface model, further optimization can be performed to achieve a more realistic 3D surface model. Optimization processes include, but are not limited to: mesh optimization, smoothing, and feature mapping.

[0146] Mesh optimization can employ energy function-based mesh simplification algorithms. By minimizing the model's energy function, the number of triangle faces can be significantly reduced without affecting the model's topology and key features, thereby reducing the complexity of the 3D surface model and improving rendering efficiency.

[0147] Smoothing can be achieved by using a bilateral filtering algorithm to smooth the model surface. While preserving the model's edge features, it removes surface noise and minor fluctuations, making the model more consistent with the real morphology of the periodontal ligament and pulp cavity.

[0148] Feature mapping involves analyzing the histological features of the periodontal ligament and pulp cavity, including but not limited to color, transparency, and texture features. Based on the analysis results, the actual color, transparency, and texture features are then simulated on a 3D surface model to enhance the model's realism and visualization.

[0149] Through the above optimization processes, highly realistic three-dimensional reconstruction of the periodontal ligament and pulp cavity can be achieved.

[0150] Furthermore, after obtaining the optimized 3D surface model, parameters can be measured on the optimized 3D surface model. The measurement indicators include, but are not limited to: the minimum and maximum thickness of the periodontal ligament, its inner surface area, outer surface area and volume, as well as the volume and grayscale value of the pulp cavity. Then, the measurement data is displayed on the 3D surface model so that medical personnel can intuitively obtain the tooth measurement data, facilitating diagnosis.

[0151] Based on any of the above embodiments, in this tooth image processing method, step S110 includes steps S111, S112 and S113.

[0152] Step S111: Obtain the raw magnetic resonance image.

[0153] Here, the raw magnetic resonance imaging (MRI) refers to the acquired MRI image. This raw MRI can be a raw human oral MRI or a raw human head MRI.

[0154] Step S112: Use the target detection model to detect the regions of each tooth in the original magnetic resonance imaging.

[0155] Specifically, the object detection model can use the YOLO (You Only Look Once, a deep learning-based object detection algorithm) model. If the original MRI is a raw human oral MRI, the object detection model can be used to directly locate the region of a single tooth to obtain the region of each tooth in the original oral MRI. If the original MRI is a raw human head MRI, the oral cavity can be located first, and then the region of a single tooth can be located to obtain the region of each tooth in the original oral MRI.

[0156] Step S113: Based on the detected regions of each tooth, the original magnetic resonance image is cropped to obtain the tooth magnetic resonance image to be segmented.

[0157] The original oral magnetic resonance imaging (MRI) images are cropped based on the regions of each detected tooth to obtain the tooth MRI images for segmentation, i.e., the tooth MRI images to be segmented.

[0158] In this embodiment, the region of each tooth is determined by the target detection model, and the original magnetic resonance image is cropped to remove irrelevant information, thereby improving the accuracy of subsequent segmentation.

[0159] Furthermore, in order to verify the effectiveness of the segmentation model of the present invention, we conducted relevant comparative tests.

[0160] Comparative Example 1 uses the conventional U-Net (U-shaped network) model, which is a deep learning model for semantic segmentation; Comparative Example 2 uses the VT model, which, compared to the model in this embodiment, does not introduce a lightweight convolutional neural network; Comparative Example 3 has the same model architecture as the model in this embodiment, but the loss function used during training is dice loss, that is, the loss function does not introduce topological loss.

[0161] The test data selected were human 5T high-field oral MRI and beagle dog 7T high-field oral MRI. The corresponding model evaluation results are shown in Tables 1 and 2 below. The segmentation results of the periodontal ligament and pulp chamber regions are shown below. Figure 3 and Figure 4 .

[0162] Table 1. Comparison of the model of this invention with other segmentation models on human 5T high-field oral MRI.

[0163]

[0164] Table 2. Comparison of the model of this invention with other segmentation models on 7T high-field oral MRI of beagles.

[0165]

[0166] Analysis of the data in Tables 1 and 2 shows that introducing the topology loss term significantly contributes to the performance improvement of the model framework proposed in this embodiment. When the topology loss is removed, the DSC of human 5T high-field oral MRI and beagle dog 7T high-field oral MRI decreases, while ASD and HD increase, indicating that the topology loss is important for accurate segmentation of the periodontal ligament and pulp chamber regions. Furthermore, compared with the existing U-Net model and VT model, the segmentation effect of this embodiment is clearly better.

[0167] Figure 3 This image shows the segmentation results of the periodontal ligament and pulp chamber regions output by human 5T high-field oral MRI under different models. Figure 3 The groups a, b, and c shown are the first, third, and seventh teeth of humans, respectively. Figure 4 This shows the segmentation results of the periodontal ligament and pulp chamber regions output by 7T high-field oral MRI in beagle dogs under different models. Figure 4 The groups a, b, and c shown are the teeth of different normal beagle dogs, while group d is the teeth of a diseased beagle dog.

[0168] Depend on Figure 3 and Figure 4 It can be seen that the segmentation model of this invention has better segmentation results. Even in challenging areas such as the pulp chamber and periodontal ligament with abnormal shapes or poor image quality, the method of this invention can produce reliable and accurate segmentation results. Furthermore, from... Figure 3 and Figure 4 It can also be seen that the introduction of topological functions effectively alleviates the problem of discontinuous segmentation of the periodontal ligament ring structure, which is conducive to improving the accuracy of the segmentation model.

[0169] The above experimental results highlight the robustness and consistency of the tooth image processing method proposed in the embodiments of the present invention, and demonstrate its great potential in improving downstream analysis of pulpitis and periodontitis.

[0170] The dental image processing apparatus provided by the present invention will be described below. The dental image processing apparatus described below can be referred to in correspondence with the dental image processing method described above.

[0171] Figure 5 This is a schematic diagram of the dental image processing device provided by the present invention, as shown below. Figure 5 As shown, the device includes an acquisition module 510 and a segmentation module 520; wherein:

[0172] The acquisition module 510 is used to acquire magnetic resonance images of the teeth to be segmented.

[0173] The segmentation module 520 is used to segment the dental magnetic resonance imaging using a segmentation model to obtain segmented images including the periodontal ligament and pulp chamber regions.

[0174] The segmentation model is used to sequentially extract and fuse features from the dental magnetic resonance imaging (MRI) images, and to segment the dental MRI images based on the global-local fusion features obtained by feature fusion. The segmentation model is trained using a preset loss function, which is determined based on dice loss and topological loss. The topological loss is determined based on pixel-level loss, predicted mask, real mask, and key pixels.

[0175] The dental image processing apparatus provided by this invention acquires a dental magnetic resonance imaging (MRI) image of a tooth to be segmented, and then uses a segmentation model to segment the MRI image, obtaining a segmented image including the periodontal ligament and pulp cavity regions. This invention utilizes MRI, which offers superior imaging performance. The segmentation model extracts and fuses global and local features from the dental MRI image, and segments the image based on the fused global-local features. Compared to extracting single local or global features, this invention improves the accuracy of dental image segmentation. Simultaneously, the segmentation model is trained using a preset loss function, which introduces topological loss in addition to dice loss. The topological loss is determined based on pixel-level loss, predicted mask, ground truth mask, and key pixels. For predicted masks that differ from corresponding key pixels in the ground truth mask, a pixel-level penalty is applied. By introducing this topological loss, the model is better encouraged to optimize the segmentation results, thereby further improving the segmentation performance. Through these methods, this invention improves the accuracy of dental image segmentation, achieving refined segmentation of the periodontal ligament and pulp cavity regions in dental MRI images.

[0176] It should be noted that the dental image processing apparatus provided in this embodiment of the invention can implement all the method steps implemented in the above-described dental image processing method embodiment and can achieve the same technical effect. Therefore, the parts and beneficial effects that are the same as those in the method embodiment will not be described in detail here.

[0177] Figure 6 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 6 As shown, the electronic device may include: a processor 610, a communications interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communications interface 620, and the memory 630 communicate with each other via the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute a tooth image processing method, which includes:

[0178] Obtain magnetic resonance images of the teeth to be segmented;

[0179] The dental magnetic resonance imaging was segmented using a segmentation model to obtain images of the periodontal ligament and pulp chamber regions;

[0180] The segmentation model is used to extract and fuse features from the dental magnetic resonance imaging, and to segment the dental magnetic resonance imaging to be segmented based on the obtained global-local fusion features.

[0181] The segmentation model is trained using a preset loss function, which is determined based on dice loss and topological loss. The topological loss is determined based on pixel-level loss, predicted mask, real mask, and key pixels.

[0182] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0183] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is able to execute the tooth image processing method provided by the above methods, the method comprising:

[0184] Obtain magnetic resonance images of the teeth to be segmented;

[0185] The dental magnetic resonance imaging was segmented using a segmentation model to obtain images of the periodontal ligament and pulp chamber regions;

[0186] The segmentation model is used to extract and fuse features from the dental magnetic resonance imaging, and to segment the dental magnetic resonance imaging to be segmented based on the obtained global-local fusion features.

[0187] The segmentation model is trained using a preset loss function, which is determined based on dice loss and topological loss. The topological loss is determined based on pixel-level loss, predicted mask, real mask, and key pixels.

[0188] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the tooth image processing method provided by the methods described above, the method comprising:

[0189] Obtain magnetic resonance images of the teeth to be segmented;

[0190] The dental magnetic resonance imaging was segmented using a segmentation model to obtain images of the periodontal ligament and pulp chamber regions;

[0191] The segmentation model is used to extract and fuse features from the dental magnetic resonance imaging, and to segment the dental magnetic resonance imaging to be segmented based on the obtained global-local fusion features.

[0192] The segmentation model is trained using a preset loss function, which is determined based on dice loss and topological loss. The topological loss is determined based on pixel-level loss, predicted mask, real mask, and key pixels.

[0193] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. 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 this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0194] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0195] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for processing dental images, characterized in that, include: Obtain magnetic resonance imaging of the tooth to be segmented; wherein, the magnetic resonance imaging of the tooth is a high-field magnetic resonance imaging with a magnetic field strength ≥5T; The dental magnetic resonance imaging was segmented using a segmentation model to obtain a segmented image including the periodontal ligament and pulp chamber regions. The segmentation model is used to sequentially extract and fuse features from the dental magnetic resonance imaging (MRI) images, and to segment the dental MRI images based on the global-local fusion features obtained by feature fusion. The segmentation model is a model constructed using CNN branches and Transformer branches. The CNN branch is used to extract local feature information from the dental MRI images, and the Transformer branch is used to extract global feature information from the dental MRI images, guided by the local feature information extracted by the CNN branch. The segmentation model is trained using a preset loss function, which is determined based on dice loss and topological loss. The topological loss is determined based on pixel-level loss, predicted mask, ground truth mask, and key pixels. The key pixels are shown below: ； Where V represents the key pixel. As background mask, For dilating the background mask, For teeth masking, The expanded tooth mask is obtained by expanding the background mask. The dental image processing method further includes: Spatial analysis was performed on the magnetic resonance imaging of the teeth to obtain three-dimensional spatial information; Based on the three-dimensional spatial information, the dental magnetic resonance imaging is divided into voxels to obtain the vertex values ​​of each voxel after voxel division. Based on the vertex values ​​of each voxel and the preset isosurface threshold, isosurface points are determined; Based on the segmented image including the periodontal ligament and pulp chamber region, the boundary information of the periodontal ligament and pulp chamber is extracted, and the boundary points are determined based on the boundary information; Using a curvature-based smoothing algorithm, triangular patches are generated based on the isopleths and boundary points, and a three-dimensional surface model is constructed based on the generated triangular patches.

2. The tooth image processing method according to claim 1, characterized in that, The segmentation model is used to segment the dental magnetic resonance imaging to obtain a segmented image including the periodontal ligament and pulp chamber regions, including: The local feature extraction layer of the segmentation model is used to extract local features from the dental magnetic resonance imaging to obtain local features; The global feature extraction layer of the segmentation model is used to extract global features from the dental magnetic resonance imaging to obtain global features; The feature fusion layer of the segmentation model is used to fuse the local features and the global features to obtain global-local fused features. The image segmentation layer of the segmentation model is used to segment the dental magnetic resonance imaging based on the global-local fusion features to obtain a segmented image including the periodontal ligament and pulp chamber regions.

3. The tooth image processing method according to claim 1, characterized in that, The preset loss function is as follows: L=L dice +λL topo ; Where L is the total loss, L dice Let L be the dice loss, λ be the weighting factor for the topology loss, and L be the weighting factor for the topology loss. topo The topology loss is as follows: ； Among them, L pixel For pixel-level loss, To predict the mask, For the real mask, For Hadama accumulation.

4. The tooth image processing method according to claim 1, characterized in that, After constructing the three-dimensional surface model based on the generated triangular patches, the process also includes: The three-dimensional surface model is optimized to obtain an optimized three-dimensional surface model; The optimization process includes one or more of mesh optimization, smoothing, and feature mapping.

5. The tooth image processing method according to any one of claims 1 to 3, characterized in that, The acquisition of the magnetic resonance imaging of the tooth to be segmented includes: Obtain the raw magnetic resonance image; The target detection model was used to detect the regions of each tooth in the original magnetic resonance imaging. The original magnetic resonance image is cropped based on the regions detected for each tooth to obtain a segmented magnetic resonance image of the tooth.

6. A dental image processing device, characterized in that, include: An acquisition module is used to acquire magnetic resonance imaging of the tooth to be segmented; wherein the magnetic resonance imaging of the tooth is a high-field magnetic resonance imaging with a magnetic field strength ≥5T; The segmentation module is used to segment the dental magnetic resonance imaging using a segmentation model to obtain segmented images including the periodontal ligament and pulp chamber regions. The segmentation model is used to sequentially extract and fuse features from the dental magnetic resonance imaging (MRI) images, and to segment the dental MRI images based on the global-local fusion features obtained by feature fusion. The segmentation model is a model constructed using CNN branches and Transformer branches. The CNN branch is used to extract local feature information from the dental MRI images, and the Transformer branch is used to extract global feature information from the dental MRI images, guided by the local feature information extracted by the CNN branch. The segmentation model is trained using a preset loss function, which is determined based on dice loss and topological loss. The topological loss is determined based on pixel-level loss, predicted mask, ground truth mask, and key pixels. The key pixels are shown below: ； Where V represents the key pixel. As background mask, For dilating the background mask, For teeth masking, The expanded tooth mask is obtained by expanding the background mask. The dental image processing device further includes a module for: Spatial analysis was performed on the magnetic resonance imaging of the teeth to obtain three-dimensional spatial information; Based on the three-dimensional spatial information, the dental magnetic resonance imaging is divided into voxels to obtain the vertex values ​​of each voxel after voxel division. Based on the vertex values ​​of each voxel and the preset isosurface threshold, isosurface points are determined; Based on the segmented image including the periodontal ligament and pulp chamber region, the boundary information of the periodontal ligament and pulp chamber is extracted, and the boundary points are determined based on the boundary information; Using a curvature-based smoothing algorithm, triangular patches are generated based on the isopleths and boundary points, and a three-dimensional surface model is constructed based on the generated triangular patches.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the tooth image processing method as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the tooth image processing method as described in any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the tooth image processing method as described in any one of claims 1 to 5.

Images