Three-stage tooth intelligent accurate segmentation method based on shape distance field

Abstract

The invention relates to a three-stage tooth intelligent accurate segmentation method based on a shape distance field, and the method comprises the following steps: S1, obtaining a tooth image, and carrying out the preprocessing; s2, extracting a region of interest, and determining a foreground region; s3, calculating the distance and direction from each foreground pixel in the foreground region to the centroid of the corresponding tooth to form a centroid offset vector; s4, generating a tooth centroid density map, and performing clustering identification on centroids of the teeth; s5, designing a segmentation mask loss function and a segmentation boundary loss function of the model; s6, designing a shape loss function based on a distance field; and the trained tooth segmentation model comprises a multi-label mask segmentation sub-network, a multi-label boundary segmentation sub-network and a shape sub-network based on a distance field, which are respectively used for executing a mask image generation task, a boundary detection task and a boundary fine description task, and outputting a final segmentation result. Compared with the prior art, the method has the advantages of improving the accuracy of tooth segmentation and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of single tooth segmentation, and in particular to a three-stage tooth intelligent and precise segmentation method based on shape distance field. Background Art

[0002] Dental 3D cone beam computed tomography (CBCT) is an important tool to assist dentists in tooth extraction, implantation or realignment. In clinical practice, dentists need to manually mark each tooth from the image slice by slice, which is laborious, time-consuming and highly dependent on the operator's experience. Therefore, an accurate and fully automatic method is needed to segment individual teeth from dental 3D cone beam computed tomography (CBCT) images.

[0003] However, due to the small size, close arrangement and complex structure of teeth, the boundaries between teeth and surrounding tissues are often unclear in existing segmentation results, and adjacent teeth are prone to adhesion. Even the most advanced learning-based methods often cannot handle such complex situations. This is mainly because such methods only use simple tooth representations (such as tooth centroids or bounding boxes), and therefore cannot capture the detailed shape changes of each tooth. In areas with low image contrast, the common boundaries where teeth contact during close occlusion and the interface between the root and the surrounding alveolar bone are even worse. Neither traditional methods nor learning-based networks can correctly segment teeth from background tissues in these areas. Summary of the invention

[0004] The purpose of the present invention is to provide a three-stage tooth intelligent and precise segmentation method based on shape distance field in order to improve the segmentation performance of tooth images.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A three-stage tooth intelligent and precise segmentation method based on shape distance field, the method comprising the following steps:

[0007] S1, acquiring a tooth image, and preprocessing the image to obtain a preprocessed image;

[0008] S2, binarizing the preprocessed image, extracting the region of interest, and determining the foreground areas of the maxillary and mandibular regions;

[0009] S3, calculating the distance and direction from each foreground pixel in the foreground area to the mass center of the corresponding tooth to form a mass center offset vector;

[0010] S4, combining the binarized preprocessed image and the centroid offset vector of S2, generating a tooth centroid density map by calculating the frequency of voxels pointed to by other voxels, locating the peak in the tooth centroid density map, and clustering to identify the centroid of the teeth;

[0011] S5. Design the segmentation mask loss function and segmentation boundary loss function of the model;

[0012] S6. Design a shape loss function based on distance field;

[0013] S7. A complete loss function is composed based on the shape loss function, the segmentation mask loss function and the segmentation boundary loss function. A trained tooth segmentation model is obtained based on the complete loss function. The single tooth image in the region of interest and the centroid of the identified single tooth are input into the tooth segmentation model. The tooth segmentation model includes a multi-label mask segmentation subnetwork, a multi-label boundary segmentation subnetwork and a shape subnetwork based on a distance field, which are respectively used to perform mask image generation tasks, boundary detection tasks and boundary fine description tasks. After the actual image is input into the tooth segmentation model, the multi-label mask segmentation subnetwork outputs a tooth category or region label image based on the actual image, the boundary detection task outputs a boundary detection result based on the tooth category or region label image, the boundary fine description task calculates the distance from each pixel in the image to its nearest boundary, and uses positive and negative signs to distinguish boundary areas, optimizes the boundary detection results, and outputs the final segmentation result as the output of the tooth segmentation model.

[0014] Furthermore, the centroid offset vector is:

[0015] p offset =p i -p cnt

[0016] Among them, p i is the position of the current foreground pixel, p cnt is the centroid position of the tooth to which the pixel belongs, p offset is the centroid offset vector.

[0017] Furthermore, the segmentation mask loss function is:

[0018] L seg =μ1L seg_dice +μ2L seg_cross

[0019] Among them, L seg_dice represents the Dice loss function between the segmentation mask and the corresponding true label; L seg_cross represents the cross entropy loss function between the segmentation mask and the corresponding true label; μ1, μ2 are the weights of the corresponding items respectively.

[0020] Furthermore, the segmentation boundary loss function is:

[0021]

[0022] Among them, L bd_dice represents the Dice loss function between the segmentation mask and the corresponding true label; L bd_cross represents the cross entropy loss function between the segmentation mask and the corresponding true label; μ3, μ4 are the weights of the corresponding items respectively.

[0023] Furthermore, the shape loss function is:

[0024]

[0025] Where N is the total number of pixels in the image, r i represents the predicted distance value, Γ r (y i ) represents the distance value calculated based on the true label.

[0026] Furthermore, the distance value calculated based on the true label is:

[0027]

[0028] in, and Represents the true value of segmentation y i Different pixels in is the Euclidean distance between them, S in and S out are the interior and exterior of an object, is the object boundary.

[0029] Furthermore, the complete loss function is:

[0030] L loss =λ1L seg +λ2L bdr +λ3L shape

[0031] Where L seg , L bdr , L shape They represent the loss functions for segmentation mask, boundary, and shape tasks respectively, and λ1, λ2, λ3 are the weights of each item.

[0032] Furthermore, the calculation of the frequency of voxels being pointed to by other voxels to generate the tooth centroid density map is specifically as follows:

[0033] Each voxel is traversed and the directional information of other voxels to it is recorded, so as to calculate the directional frequency of each voxel. The frequency reflects the centroid distribution characteristics of different teeth, embodies the dense and sparse areas of teeth, and generates a tooth centroid density map.

[0034] Furthermore, the binarization is specifically as follows: marking pixels in the foreground area as 1 and marking pixels in the background area as 0.

[0035] Furthermore, the preprocessing is specifically as follows:

[0036] The tooth image is converted into an appropriate data format, the minimum and maximum values ​​of the image are calculated and normalized to obtain the preprocessed image.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] The present invention sets up loss functions corresponding to the three tasks of mask image task, boundary detection task and boundary fine description task respectively. Through the gradually refined three-stage network segmentation method, the model can effectively deal with the complex morphology and close arrangement of teeth, and realize high-precision automatic segmentation and recognition of teeth. The multi-label mask segmentation subnetwork of the present invention explicitly models the mask semantic relationship between the target tooth and the adjacent teeth, taking into account the topological influence of the adjacent teeth, ensuring the accuracy and stability of the segmentation results when the complex network affects the tooth arrangement, and improving the accuracy of the target tooth mask segmentation; the multi-label boundary segmentation subnetwork proposes a new loss function, which explicitly models the boundary semantic relationship between the target tooth and the adjacent teeth, and improves the accuracy of the target tooth edge segmentation; the shape subnetwork based on the distance field accurately describes the edge information of adjacent teeth, and improves the accuracy of tooth edge segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a flow chart of the present invention;

[0040] Figure 2 Extraction flow chart for the region of interest of the present invention;

[0041] Figure 3 A schematic diagram of clustering-based centroid generation of the present invention;

[0042] Figure 4 A schematic diagram of a multi-label multi-task network of the present invention;

[0043] Figure 5 This is a schematic diagram of the experimental effect of the present invention in solving the problem of tooth adhesion. DETAILED DESCRIPTION

[0044] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0045] This case proposes a three-stage intelligent and precise tooth segmentation method based on shape distance field. The flowchart of the method is as follows: Figure 1 As shown, the method comprises the following steps:

[0046] S1, acquiring a tooth image, and preprocessing the image to obtain a preprocessed image;

[0047] S2, binarizing the preprocessed image, extracting the region of interest, and determining the foreground areas of the maxillary and mandibular regions;

[0048] S3, calculating the distance and direction from each foreground pixel in the foreground area to the mass center of the corresponding tooth to form a mass center offset vector;

[0049] S4, combining the binarized preprocessed image and the centroid offset vector of S2, generating a tooth centroid density map by calculating the frequency of voxels pointed to by other voxels, locating the peak in the tooth centroid density map, and identifying the tooth centroid;

[0050] S5. Design the segmentation mask loss function and segmentation boundary loss function of the model;

[0051] S6. Design a shape loss function based on distance field;

[0052] S7. A complete loss function is composed based on the shape loss function, the segmentation mask loss function and the segmentation boundary loss function. A trained tooth segmentation model is obtained based on the complete loss function. The single tooth image in the region of interest and the centroid of the identified single tooth are input into the tooth segmentation model. The tooth segmentation model includes a multi-label mask segmentation subnetwork, a multi-label boundary segmentation subnetwork and a shape subnetwork based on a distance field, which are respectively used to perform mask image generation tasks, boundary detection tasks and boundary fine description tasks. After the actual image is input into the tooth segmentation model, the multi-label mask segmentation subnetwork outputs a tooth category or region label image based on the actual image, the boundary detection task outputs a boundary detection result based on the tooth category or region label image, the boundary fine description task optimizes the boundary detection result, and outputs the final segmentation result as the output of the tooth segmentation model.

[0053] The flowchart of region of interest extraction is as follows Figure 2 As shown in Figure 2, the centroid generation diagram based on clustering is shown in Figure 2. Figure 3 As shown, the multi-label multi-task network diagram is as follows Figure 4 As shown in the figure, the experimental effect diagram of solving the problem of tooth adhesion is as follows Figure 5 shown.

[0054] When extracting the information of the region of interest containing the tooth organ, due to the small size of the tooth, the original image is preprocessed to extract the local area containing the tooth and reduce the interference of irrelevant background. The purpose of this stage is to limit the processing range and focus on the area where the tooth is located, thereby improving the accuracy and efficiency of subsequent segmentation.

[0055] When constructing the tooth centroid map, the centroid clustering method is used to distinguish the position information of each tooth in order to solve the problem of adhesion between teeth and adjacent structures. Because no matter what the shape, position and direction of the tooth is, the centroid point is a stable feature point inside the tooth shape, which is more reliable than the bounding box and more in line with the characteristics of small size and dense arrangement of teeth. The centroid can effectively identify the center position of each tooth, thereby creating an independent candidate area for each tooth and reducing the impact of adjacent tooth adhesion on the segmentation result.

[0056] In the multi-label mask segmentation subnetwork, the network uses a convolutional neural network structure to input a tooth image containing multiple labels and generate a corresponding multi-label mask image. The network contains multiple output channels, each corresponding to a specific label, which is used to predict the segmentation masks of different tooth structures or categories. By combining contextual information and local features, the network can accurately segment complex tooth structures, ensure accurate distinction between different label areas, and support joint segmentation of multiple tooth categories or regions.

[0057] In the multi-label boundary segmentation subnetwork, the network focuses on boundary detection in multi-label segmentation tasks through a convolutional neural network structure. This subnetwork contains multiple output channels, each of which is responsible for predicting the boundaries of different tooth structures or categories. By capturing edge features and local details in the image, the network can accurately identify the boundaries of different areas of the tooth and distinguish the boundary lines between various labels. The network uses multi-scale feature fusion technology to ensure that even small boundaries can be accurately segmented and distinguished under complex tooth morphology.

[0058] In the shape network based on the designed distance field, the signed distance field is introduced to describe the grid boundary in detail, thereby effectively solving the problem of distance field shape adhesion. The signed distance field can accurately depict the complex shape and boundary features of teeth by calculating the distance from each pixel in the image to its nearest boundary and using positive and negative signs to distinguish the boundary area. Especially in the case of multiple tooth boundaries and blurred boundaries, the signed distance field can provide geometric information of the structure, help the network effectively separate the shapes of neighboring teeth, avoid boundary adhesion, and thus improve the accuracy of the segmentation results.

[0059] Through this gradually refined segmentation method, the model can effectively deal with the complex morphology and close arrangement of teeth, and achieve high-precision automatic segmentation and recognition of teeth.

[0060] The present invention realizes automatic and precise tooth segmentation, accurately extracts the boundary and shape of each tooth from complex oral three-dimensional cone beam computed tomography (CBCT) images, and provides certain technical support for automatic tooth segmentation.

[0061] The present invention extracts information of a region of interest containing tooth organs based on the Mamba algorithm, and focuses on the region where the teeth are located by limiting the processing range, thereby improving the accuracy and efficiency of subsequent segmentation.

[0062] Clustering generates a tooth centroid map to obtain the guiding centroid. Due to the variable shape, position and orientation of teeth, the centroid point can more stably and accurately depict the feature points inside the tooth shape than the bounding box, which is more in line with the characteristics of small size and dense arrangement of teeth. By clustering the centroid, the center position of each tooth can be effectively identified, thereby creating an independent candidate area for each tooth.

[0063] The present invention designs a multi-label mask segmentation subnetwork. The network uses a convolutional neural network structure to input a tooth image containing multiple labels and generate a corresponding multi-label mask image. The network includes multiple output channels, each channel corresponds to a specific label, and is used to predict the segmentation masks of different tooth structures or categories. By combining contextual information and local features, the network can accurately segment complex tooth structures, ensure accurate distinction between different label areas, and support joint segmentation of multiple tooth categories or areas.

[0064] The present invention designs a multi-label boundary segmentation subnetwork. The network focuses on boundary detection in multi-label segmentation tasks through a convolutional neural network structure. The subnetwork contains multiple output channels, each of which is responsible for predicting the boundaries of different structures or categories of teeth. By capturing edge features and local details in the image, the network can accurately identify the boundaries of different areas of the teeth and distinguish the boundary lines between various labels. The network uses multi-scale feature fusion technology to ensure that even small boundaries can be accurately segmented and distinguished under complex tooth morphology.

[0065] The present invention designs a shape subnetwork based on a distance field. By introducing a signed distance field, the grid boundary is described in detail, thereby effectively solving the problem of distance field shape adhesion. The signed distance field can accurately depict the complex shape and boundary features of teeth by calculating the distance from each pixel in the image to its nearest boundary and using positive and negative signs to distinguish the boundary area. Especially in the case of multiple tooth boundaries and blurred boundaries, the signed distance field can provide geometric information of the structure, help the network effectively separate the shapes of neighboring teeth, avoid boundary adhesion, and thus improve the accuracy of the segmentation results.

[0066] The present invention has the following beneficial effects:

[0067] 1. The three-stage tooth segmentation framework of the present invention proposes a tooth segmentation framework based on signed distance field, which is divided into three parts: region of interest extraction, centroid positioning, and multi-label multi-task single tooth segmentation. It can effectively improve the accuracy and stability of tooth segmentation and provide strong support for subsequent dental analysis and treatment.

[0068] 2. The present invention extracts the region of interest containing the teeth, ensuring that the tooth features are clearly visible and reducing background interference in the image, so that the processing algorithm can be more focused on the teeth themselves, reducing the amount of calculation, improving the algorithm's operating efficiency, saving time and resources, and improving the accuracy of subsequent segmentation and analysis.

[0069] 3. The present invention uses a clustering algorithm to generate a guiding centroid to distinguish and locate teeth. The generated centroid indicates the target tooth area, creates an independent candidate area for each tooth, reduces the impact of adjacent tooth adhesion on the segmentation result, improves the accuracy of the segmentation result, and reduces the possibility of misjudgment.

[0070] 4. The present invention designs a multi-label mask segmentation subnetwork to explicitly model the mask semantic relationship between the target tooth and the adjacent teeth, taking into account the topological influence of the adjacent teeth, ensuring the accuracy and stability of the segmentation results when the tooth arrangement is affected by a complex network, and improving the accuracy of the target tooth mask segmentation.

[0071] 5. The present invention designs a multi-label boundary segmentation subnetwork and proposes a new loss function to explicitly model the boundary semantic relationship between the target tooth and the adjacent teeth, thereby improving the accuracy of the target tooth edge segmentation.

[0072] 6. The present invention designs a shape sub-network based on distance field to accurately describe the information of adjacent tooth edges and improve the accuracy of tooth edge segmentation.

[0073] The specific implementation of the steps is:

[0074] Step 1: Normalize the image.

[0075] In this example, the original image is read and converted into an appropriate data format (such as a NumPy array) to ensure that the value range of the image can be processed later. Next, the minimum and maximum values ​​of the image are calculated to determine the original value range. Then, the normalization formula is used to convert each pixel value of the image to a value between 0 and 1. The formula is:

[0076]

[0077] where x iIt is the value of a pixel in the current image, indicating the intensity or brightness of the pixel, usually ranging from 0 to 255; min It is the minimum value of all pixel values ​​in the image. It represents the darkest part (or the part with the lowest intensity) of the image. By subtracting the minimum value, the values ​​of all pixels can be "shifted" to zero. max It is the maximum value of all pixel values ​​in the image. It represents the brightest part (or the part with the highest intensity) of the image; im It is the normalized pixel value, whose range is adjusted to between 0 and 1. This value indicates the position of the pixel relative to the brightness distribution of the entire image.

[0078] The pixel values ​​of the images are scaled to the same range, and the contrast and brightness differences of the images are standardized, thereby eliminating the impact of brightness or contrast differences between different images. This allows the model to better capture features during training and inference, avoiding the problems of reduced learning efficiency and slow convergence speed due to different numerical ranges, and ensuring the stability of the model when processing images.

[0079] Step 2: Obtain the region of interest containing the teeth.

[0080] In this example, the region of interest is extracted. First, the image is binarized and converted into a binary image containing only two values, 0 and 1, where the pixels in the foreground area are marked as 1 and part of the background is marked as 0. This simplified representation not only reduces the complexity of the image, but also enhances the contrast between the target area and the background, allowing the segmentation algorithm to identify the target area more efficiently. Then, a mamba-based region of interest extraction is designed. By clipping the binary image of the target area and mapping it back to the original image, the foreground areas of the maxilla and mandible can be effectively located. These two steps lay the foundation for subsequent image analysis and segmentation.

[0081] Step 3: Design a clustering-based centroid generation network.

[0082] In this example, the guidance centroid is generated by a clustering algorithm. First, a centroid offset map is generated. In this step, the generated centroid offset map represents the three-dimensional vector of the centroid corresponding to each foreground pixel (such as a tooth). Specifically, the foreground pixels in the image are first determined, and the distance and direction from each foreground pixel to its corresponding tooth centroid are calculated. This can be achieved by the following formula:

[0083] p offset =p i -p cnt

[0084] Among them, p i is the position of the current foreground pixel, p cntis the centroid position of the tooth to which the pixel belongs, p offset is the centroid offset vector. The final centroid offset map not only provides the spatial information of the target point, but also reflects the relative position relationship between the teeth, providing basic data for subsequent analysis.

[0085] Combined with the previously generated binary image, the tooth centroid density map is generated by calculating the frequency of a certain voxel being pointed to by other voxels. In this process, each voxel is traversed, and the pointing information of other voxels to it is recorded, so as to calculate the pointing frequency of each voxel. This frequency can reflect the centroid distribution characteristics of different teeth, and then reveal the dense and sparse areas of teeth.

[0086] After the tooth centroid density map is generated, a fast search clustering method is used to locate the peaks in the density map to predict the tooth centroid. In this method, high-density areas in the density map are first identified to determine potential cluster centers. These cluster centers usually correspond to points with relatively high density values ​​and a large distance to the nearest point with a high density value. Ultimately, the centroid of the tooth is accurately identified and located through cluster analysis of the peaks, providing an accurate spatial reference for subsequent image analysis and processing.

[0087] Step 5: Design multi-label mask and boundary sub-network.

[0088] The Dice loss function is a commonly used metric for evaluating and optimizing image segmentation performance, and is applicable to the field of image processing. It is mainly used to measure the degree of overlap between sparse medical prediction results and true labels. The value of the Dice loss function cycles between 0 and 1, and a value greater than 1 indicates a degree of overlap.

[0089]

[0090] Where A is the predicted segmentation area and B is the true label area. |A∩B| represents the intersection of the predicted area and the true area, that is, the number of pixels that overlap between the two. |A| and |B| are the number of pixels in the predicted area and the true area, respectively.

[0091] Cross entropy loss is a loss function widely used in classification tasks. It is used to measure the difference between the probability distribution predicted by the model and the true label distribution. Usually, the cross entropy loss function is used in conjunction with the softmax activation function to generate a probability distribution for each category, making the model output more interpretable. In addition, cross entropy loss is highly sensitive to misclassification, especially when the model's predicted probability for a certain class is close to 0, the loss value will rise sharply. This feature helps the model converge faster and can effectively optimize the performance of the classification model. For the binary classification problem here, the cross entropy loss can be expressed as:

[0092]

[0093] Where M is the total number of samples; q j is the true label (0 or 1) of the jth sample; p j is the probability that the model predicts that the jth sample belongs to the positive class.

[0094] The segmentation mask loss function is responsible for optimizing the segmentation mask so that the predicted segmentation result is closer to the actual target area. It consists of two parts, specifically:

[0095] L seg =μ1L seg_dice +μ2L seg_cross

[0096] Where L seg_dice represents the Dice loss function between the segmentation mask and the corresponding true label; L seg_cross represents the cross entropy loss function between the segmentation mask and the corresponding true label; μ1, v2 are the weights of the corresponding items respectively.

[0097] The segmentation boundary loss function focuses on improving the boundary accuracy of the segmentation results, ensuring that the model can more accurately identify the edges of the target object. The specific form is:

[0098]

[0099] Where L bd_dice represents the Dice loss function between the segmentation mask and the corresponding true label; L bd-cross represents the cross entropy loss function between the segmentation mask and the corresponding true label; μ3, μ4 are the weights of the corresponding items respectively.

[0100] This loss function combines the advantages of Dice loss and cross entropy loss. It can not only enhance the overlapping area between the segmentation result and the true label through Dice loss, but also further optimize the accuracy of pixel-level classification through cross entropy loss. The introduction of weight coefficients makes this loss function flexible, which can better balance the model's attention to regional overlap (Dice loss) and classification accuracy (cross entropy loss) during training to improve the overall performance of segmentation. Through this design, the segmentation model can better balance segmentation accuracy and the ability to handle unbalanced data.

[0101] Step 6: Design a shape sub-network based on distance field.

[0102] In this example, the present invention designs a shape loss function based on the distance field, aiming to make the segmentation result generated by the distance field constraint model closer to the actual shape of the target object. The specific loss function is:

[0103]

[0104] Where N is the total number of pixels in the image, r i represents the predicted distance value, Γ r (y i ) represents the distance value calculated based on the true label, which is expressed as follows:

[0105]

[0106] in and Represents the true value of segmentation y i Different pixels in is the Euclidean distance between them. S in and S out are the interior and exterior of an object, is the object boundary. The loss function measures the difference between the predicted value and the true distance value through the mean square error (MSE). This shape loss can effectively constrain the segmentation boundary generated by the model to make it more accurately conform to the true contour of the target.

[0107] Step 7: Network solution.

[0108] In this example, the loss function of the multi-label multi-task learning network is an important design element. A weighted loss function is usually used to handle the weights of different tasks and labels to ensure that each task is reasonably optimized during the training process. The loss function can be calculated by summing the weights of the function through the losses of each task, and each task can have different weights.

[0109] The overall loss function of multi-label multi-task is designed as follows:

[0110] L loss =λ1L seg +λ2L bdr +λ3L shape (6.1)

[0111] Where L seg , L bdr , L shape Respectively represent the loss functions of segmentation mask, boundary and shape tasks, λ1, λ2, λ3, are the weights of each item. Through this weighted summation design, the present invention can simultaneously optimize the three key tasks of segmentation, boundary and shape in the multi-task, multi-label training process. The weight of each task can be flexibly adjusted according to actual needs, so as to achieve a balance between different tasks and avoid a task being over-optimized or neglected during the training process. Through the multi-label multi-task loss function design, the network can have stronger generalization ability and accuracy when facing complex scenes, meeting the needs of accurate segmentation, boundary distinction and shape constraints.

[0112] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A three-stage intelligent and precise tooth segmentation method based on shape distance field, characterized in that: The method comprises the following steps: S1, acquiring a tooth image, and preprocessing the image to obtain a preprocessed image; S2, binarizing the preprocessed image, extracting the region of interest, and determining the foreground areas of the maxillary and mandibular regions; S3, calculating the distance and direction from each foreground pixel in the foreground area to the mass center of the corresponding tooth to form a mass center offset vector; S4, combining the binarized preprocessed image and the centroid offset vector of S2, generating a tooth centroid density map by calculating the frequency of voxels pointed to by other voxels, locating the peak in the tooth centroid density map, and clustering to identify the centroid of the teeth; S5. Design the segmentation mask loss function and segmentation boundary loss function of the model; S6. Design a shape loss function based on distance field; S7. A complete loss function is composed based on the shape loss function, the segmentation mask loss function and the segmentation boundary loss function. A trained tooth segmentation model is obtained based on the complete loss function. The single tooth image in the region of interest and the centroid of the identified single tooth are input into the tooth segmentation model. The tooth segmentation model includes a multi-label mask segmentation subnetwork, a multi-label boundary segmentation subnetwork and a shape subnetwork based on a distance field, which are respectively used to perform mask image generation tasks, boundary detection tasks and boundary fine description tasks. After the actual image is input into the tooth segmentation model, the multi-label mask segmentation subnetwork outputs a tooth category or region label image based on the actual image, the boundary detection task outputs a boundary detection result based on the tooth category or region label image, the boundary fine description task calculates the distance from each pixel in the image to its nearest boundary, and uses positive and negative signs to distinguish boundary areas, optimizes the boundary detection results, and outputs the final segmentation result as the output of the tooth segmentation model.

2. The three-stage tooth intelligent and precise segmentation method based on shape distance field according to claim 1, characterized in that: The centroid offset vector is: p offset =p i -p cnt Among them, p i is the position of the current foreground pixel, p cnt is the centroid position of the tooth to which the pixel belongs, p offset is the centroid offset vector.

3. The three-stage tooth intelligent and precise segmentation method based on shape distance field according to claim 1, characterized in that: The segmentation mask loss function is: L seg =μ1L seg_dice +μ2L seg_cross Among them, L seg_dice represents the Dice loss function between the segmentation mask and the corresponding true label; L seg_cross represents the cross entropy loss function between the segmentation mask and the corresponding true label; μ1, μ2 are the weights of the corresponding items respectively.

4. The three-stage tooth intelligent and precise segmentation method based on shape distance field according to claim 1, characterized in that: The segmentation boundary loss function is: Among them, L bd _ dice represents the Dice loss function between the segmentation mask and the corresponding true label; L bd_cross represents the cross entropy loss function between the segmentation mask and the corresponding true label; μ3, μ4 are the weights of the corresponding items respectively.

5. The three-stage tooth intelligent and precise segmentation method based on shape distance field according to claim 1, characterized in that: The shape loss function is: Where N is the total number of pixels in the image, r i represents the predicted distance value, Γ r (y i ) represents the distance value calculated based on the true label.

6. The three-stage intelligent and precise tooth segmentation method based on shape distance field according to claim 5, characterized in that: The distance value calculated based on the true label is: in, and Represents the true value of segmentation y i Different pixels in is the Euclidean distance between them, S in and S out are the interior and exterior of an object, is the object boundary.

7. The three-stage tooth intelligent and precise segmentation method based on shape distance field according to claim 6, characterized in that: The complete loss function is: L loss =λ1L seg +λ2L bdr +λ3L shape Where L seg , L bdr , L shape They represent the loss functions for segmentation mask, boundary, and shape tasks respectively, and λ1, λ2, λ3 are the weights of each item.

8. The three-stage tooth intelligent and precise segmentation method based on shape distance field according to claim 1, characterized in that: The method of calculating the frequency of voxels being pointed to by other voxels and generating the tooth centroid density map is specifically as follows: Each voxel is traversed and the directional information of other voxels to it is recorded, so as to calculate the directional frequency of each voxel. The frequency reflects the centroid distribution characteristics of different teeth, embodies the dense and sparse areas of teeth, and generates a tooth centroid density map.

9. The three-stage tooth intelligent and precise segmentation method based on shape distance field according to claim 1, characterized in that: The binarization is specifically as follows: marking pixels in the foreground area as 1 and marking pixels in the background area as 0.

10. The three-stage tooth intelligent and precise segmentation method based on shape distance field according to claim 1, characterized in that: The pre-processing is specifically as follows: The tooth image is converted into an appropriate data format, the minimum and maximum values ​​of the image are calculated and normalized to obtain the preprocessed image.

Images