Tooth segmentation method and device

Abstract

The invention discloses a tooth segmentation method and device, and the method comprises the steps: inputting a tooth three-dimensional image into a tooth segmentation model, and obtaining a first tooth prediction frame and a first tooth segmentation result of a first tooth; determining first tooth segmentation results with segmentation conflicts according to the first tooth segmentation results of the plurality of first teeth; performing image classification identification on the first to-be-segmented image, and determining an error prediction frame in the at least one first tooth prediction frame; deleting the first tooth segmentation result corresponding to the error prediction frame from each first tooth segmentation result to obtain a second tooth segmentation result; shielding the second tooth segmentation result in the tooth three-dimensional image to obtain a first tooth shielding image; inputting the first tooth mask image into a tooth segmentation model to obtain a second tooth prediction frame of the second tooth and a third tooth segmentation result; and determining a tooth segmentation result of the tooth three-dimensional image based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth.

Description

Technical Field

[0001] This application relates to the field of image recognition technology, and in particular to a tooth segmentation method and apparatus. Background Technology

[0002] Cone beam computer tomography (CBCT) images of the oral and maxillofacial region are the primary basis for oral diagnosis, providing comprehensive oral information. 3D tooth reconstruction based on CBCT images has a wide range of applications.

[0003] Current tooth segmentation methods first determine the target region where the tooth is located, and then segment the tooth within that target region. This method performs well on CBCT images with relatively standard tooth growth patterns. However, it has lower detection rates and accuracy for abnormal teeth, such as impacted teeth with random tooth axis orientation, small and crowded deciduous teeth, or teeth with highly overlapping detection frames.

[0004] Therefore, there is an urgent need for a solution to improve the accuracy of detecting abnormal tooth types. Summary of the Invention

[0005] This application provides a tooth segmentation method and apparatus to improve the accuracy of detecting abnormal tooth types.

[0006] In a first aspect, this application provides a method for tooth segmentation, the method comprising:

[0007] A 3D image of a tooth is input into a tooth segmentation model to obtain at least one first tooth prediction bounding box and at least one first tooth segmentation result for at least one first tooth. Based on the first tooth segmentation results for multiple first teeth, first tooth segmentation results with segmentation conflicts are identified. The segmentation conflict refers to the overlap range between two tooth segmentation results being greater than a set threshold. Image classification and recognition are performed on the first image to be segmented to identify at least one incorrect prediction bounding box in the first tooth prediction bounding box. The first image to be segmented is the image to be segmented corresponding to the first tooth prediction bounding box of the first tooth segmentation result with segmentation conflicts. The first tooth segmentation result corresponding to the incorrect prediction bounding box is deleted from each first tooth segmentation result to obtain a second tooth segmentation result. The second tooth segmentation result in the 3D image of the tooth is masked to obtain a first tooth masking image. The first tooth masking image is input into the tooth segmentation model to obtain a second tooth prediction bounding box for the second tooth and a third tooth segmentation result for the second tooth. Based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth, the tooth segmentation result of the 3D image of the tooth is determined.

[0008] In the above technical solution, the accuracy of the first predicted bounding box corresponding to the first tooth segmentation result with segmentation conflict is first determined, and the first tooth segmentation result corresponding to the incorrect predicted bounding box is deleted. Then, the segmented teeth are masked to obtain the first tooth mask image, which eliminates the interference of the segmented teeth and constructs an environment where the targets are not crowded or overlapping. The first tooth mask image is input into the tooth segmentation model to detect teeth that may be missed, which can improve the accuracy of detecting abnormal types of teeth.

[0009] In one possible implementation, after obtaining the second tooth prediction box of the second tooth and the third tooth segmentation result of the second tooth, the method further includes: determining whether there is a segmentation conflict between the third tooth segmentation result of each second tooth and the second tooth segmentation result of the first tooth; and determining the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth, including: determining the tooth segmentation result of the three-dimensional image of the tooth according to the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth where there is no segmentation conflict.

[0010] In the above technical solution, since the third segmentation result of the second tooth serves as a supplement to the second segmentation result of the first tooth, when there is a segmentation conflict between the third segmentation result of the second tooth and the second segmentation result of the first tooth, the third segmentation result of the second tooth is deleted. The second segmentation result of the first tooth and the third segmentation result of the second tooth where there is no segmentation conflict are determined as the tooth segmentation result of the three-dimensional image of the tooth.

[0011] In one possible implementation, after obtaining the second tooth prediction box of the second tooth and the third tooth segmentation result of the second tooth, the method further includes: determining whether there is a segmentation conflict between the third tooth segmentation results of each second tooth; if there is a segmentation conflict, determining whether the second tooth prediction box corresponding to the third tooth segmentation result with the segmentation conflict is an incorrect prediction box; and determining the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth, including: determining the tooth segmentation result of the three-dimensional image of the tooth according to the second tooth segmentation result of the first tooth and the third tooth segmentation result corresponding to the non-incorrect prediction box in the second tooth prediction box.

[0012] In the above technical solution, when there is a segmentation conflict between the segmentation results of the second tooth and the third tooth, it is determined whether the predicted box of the second tooth corresponding to the segmentation result of the third tooth with the segmentation conflict is an incorrect predicted box. If the predicted box of the second tooth is determined to be an incorrect predicted box, the segmentation result of the third tooth corresponding to the predicted box of the second tooth is deleted to further improve the accuracy of the supplementary segmentation results.

[0013] In one possible implementation, the following method is used to determine whether a tooth prediction box is an incorrect prediction box: the image of the tooth to be segmented within the range of the tooth prediction box, and the corresponding tooth segmentation result are input into a first classification model to determine whether the tooth prediction box is an incorrect prediction box; wherein, the positive samples used to train the first classification model are correct tooth prediction boxes and corresponding tooth segmentation results; the negative samples used to train the first classification model are historically incorrect prediction boxes and corresponding tooth segmentation results obtained based on the tooth segmentation model.

[0014] In the above technical solution, when two tooth segmentation results conflict, there must be a misidentified prediction box in the two corresponding prediction boxes. By identifying the misidentified prediction box and removing the corresponding tooth segmentation result through the first classification model, the accuracy of tooth segmentation can be improved.

[0015] In one possible implementation, after obtaining the second tooth prediction box and the third tooth segmentation result of the second tooth, the method further includes: inputting a first tooth masking image within the range of any second tooth prediction box and the second tooth prediction box into a second classification model to determine whether the second tooth prediction box is a valid prediction box; and determining the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth, including: determining the tooth segmentation result of the three-dimensional image of the tooth according to the second tooth segmentation result of the first tooth and the third tooth segmentation result corresponding to the valid prediction box in the second tooth prediction box.

[0016] In the above technical solution, if it is determined that the second tooth prediction box is not a valid prediction box, the third tooth segmentation result corresponding to the second tooth prediction box is deleted to further improve the accuracy of the supplementary segmentation result.

[0017] In one possible implementation, negative samples for training the second classification model are obtained by masking all teeth in the sample image to obtain a first masked sample image; and scaling and / or rotating and / or translating the tooth prediction box corresponding to each masked tooth in the first masked sample image to obtain negative samples for the second classification model.

[0018] In the above technical solution, scaling and / or rotating and / or translating the predicted tooth bounding boxes corresponding to each masked tooth in the first masked sample image can enrich the shape of the predicted tooth bounding boxes of the negative samples. Furthermore, training the second classification model with the processed sample images can improve the accuracy of the second classification model.

[0019] In one possible implementation, positive samples for training the second classification model are obtained by: randomly selecting a portion of teeth from the sample image and masking them to obtain a second masked sample image; masking another portion of teeth to obtain a third masked sample image; and enlarging the predicted bounding boxes of the teeth corresponding to the unmasked teeth in the second masked sample image and the third masked sample image to obtain positive samples for the second classification model.

[0020] In the above technical solution, a portion of the teeth in the sample image is randomly selected for masking to obtain a second masked sample image, and another portion of the teeth is masked to obtain a third masked sample image. This allows the presence or absence of masking around the teeth in the sample image to be random. Enlarging the predicted bounding boxes of the unmasked teeth in the second and third masked sample images enriches the shape of the predicted bounding boxes for positive samples. Furthermore, training the second classification model using these processed sample images improves the accuracy of the second classification model.

[0021] In one possible implementation, after obtaining the second tooth prediction box of the second tooth and the third tooth segmentation result of the second tooth, the method further includes: determining whether the second tooth prediction box is within the effective region of the segmented tooth; and determining the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth, including: determining the tooth segmentation result of the three-dimensional image of the tooth according to the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth located within the effective region.

[0022] In the above technical solution, the coordinates of the second predicted box are used to determine whether the second predicted box is within the effective area of ​​the segmented teeth. If the second predicted box is not within the effective area of ​​the segmented teeth, it is considered to be an incorrectly identified predicted box, and the tooth segmentation result corresponding to the second predicted box is deleted to further improve the accuracy of the supplementary segmentation results.

[0023] In one possible implementation, inputting a first tooth masking image into a tooth segmentation model to obtain a second tooth prediction bounding box and a third tooth segmentation result for the second tooth includes: performing at least one spatial transformation on the first tooth masking image to obtain at least one second tooth masking image with a different principal axis orientation from the first tooth masking image; obtaining a third tooth prediction bounding box for the second tooth by passing the first tooth masking image through the 3D detection model of the tooth segmentation model; obtaining a fourth tooth prediction bounding box for the second tooth by passing at least one second tooth masking image through the 3D detection model of the tooth segmentation model; obtaining a second tooth prediction bounding box for the second tooth by deduplicating the third and fourth tooth prediction bounding boxes; and obtaining a third tooth segmentation result for the second tooth by passing the second tooth prediction bounding box for the second tooth through the 3D segmentation model of the tooth segmentation model.

[0024] In the above technical solution, since the main axis orientation of most teeth in the training data of the tooth segmentation model is along the Z-axis, the tooth segmentation model is prone to missing teeth with random main axis orientations, such as impacted teeth, during detection. In this embodiment, the first tooth mask image is subjected to at least one spatial transformation to obtain at least one second tooth mask image with a different main axis orientation from the first tooth mask image. Then, the first tooth mask image is used to obtain a third predicted bounding box for the second tooth through the 3D detection model in the tooth segmentation model; the second tooth mask image is then input into the 3D detection model in the tooth segmentation model to obtain a fourth predicted bounding box for the second tooth. The fourth predicted bounding box for the second tooth may overlap with the third predicted bounding box, so deduplication processing is required to obtain the second predicted bounding box for the second tooth. Finally, the second predicted bounding box for the second tooth is input into the 3D segmentation model of the tooth segmentation model to obtain the third tooth segmentation result for the second tooth, thus detecting and segmenting teeth whose main axis orientation has an angle of deviation from the Z-axis.

[0025] Secondly, embodiments of this application provide a tooth-splitting device, the device comprising:

[0026] The detection module is used to input a three-dimensional image of a tooth into a tooth segmentation model to obtain at least one first tooth prediction box and at least one first tooth segmentation result of at least one first tooth.

[0027] The detection module is further configured to determine, based on multiple first tooth segmentation results, a first tooth segmentation result with segmentation conflict; the segmentation conflict refers to the overlap between two tooth segmentation results being greater than a set threshold; perform image classification and recognition on the first image to be segmented to determine at least one incorrect prediction box within the first tooth prediction box; the first image to be segmented is the image to be segmented corresponding to the first tooth prediction box of the first tooth segmentation result with segmentation conflict; delete the first tooth segmentation result corresponding to the incorrect prediction box from each first tooth segmentation result to obtain the second tooth segmentation result;

[0028] The detection module is also used to mask the second tooth segmentation result in the three-dimensional image of the tooth to obtain a first tooth masking image; and to input the first tooth masking image into the tooth segmentation model to obtain the second tooth prediction box of the second tooth and the third tooth segmentation result of the second tooth.

[0029] The processing module is used to determine the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth.

[0030] In one possible design, after obtaining the second tooth prediction box of the second tooth and the third tooth segmentation result of the second tooth, the detection module is further used to determine whether there is a segmentation conflict between the third tooth segmentation result of each second tooth and the second tooth segmentation result of the first tooth; when the processing module determines the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth, it is specifically used to determine the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth where there is no segmentation conflict.

[0031] In one possible design, after obtaining the second tooth prediction box of the second tooth and the third tooth segmentation result of the second tooth, the detection module is further used to determine whether there is a segmentation conflict between the third tooth segmentation results of each second tooth; if there is a segmentation conflict, it is determined whether the second tooth prediction box corresponding to the third tooth segmentation result with the segmentation conflict is an incorrect prediction box; when the processing module determines the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth, it is specifically used to determine the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result corresponding to the non-incorrect prediction box in the second tooth prediction box.

[0032] In one possible design, the detection module determines whether a tooth prediction box is an incorrect prediction box by: inputting the image of the tooth to be segmented within the range of the tooth prediction box and the corresponding tooth segmentation result into a first classification model, and determining whether the tooth prediction box is an incorrect prediction box; wherein, the positive samples used to train the first classification model are correct tooth prediction boxes and corresponding tooth segmentation results; the negative samples used to train the first classification model are historical incorrect prediction boxes and corresponding tooth segmentation results obtained based on the tooth segmentation model.

[0033] In one possible design, after the detection module obtains the second tooth prediction box and the third tooth segmentation result of the second tooth, the processing module is used to input the first tooth masking image within the range of any second tooth prediction box and the second tooth prediction box into the second classification model to determine whether the second tooth prediction box is a valid prediction box; when the processing module determines the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth, it is specifically used to determine the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result corresponding to the valid prediction box in the second tooth prediction box.

[0034] In one possible design, the device further includes a training module for obtaining negative samples for training the second classification model by: masking all teeth in the sample image to obtain a first masked sample image; scaling and / or rotating and / or translating the tooth prediction box corresponding to each masked tooth in the first masked sample image as a negative sample for the second classification model.

[0035] In one possible design, the training module is further configured to obtain positive samples for training the second classification model by: randomly selecting a portion of teeth from the sample image to mask them to obtain a second masked sample image, and masking another portion of teeth to obtain a third masked sample image; and enlarging the predicted bounding boxes of the teeth corresponding to the unmasked teeth in the second masked sample image and the third masked sample image to use as positive samples for the second classification model.

[0036] In one possible design, after obtaining the second tooth prediction bounding box of the second tooth and the third tooth segmentation result of the second tooth, the detection module is further used to determine whether the second tooth prediction bounding box is within the effective area of ​​the segmented tooth; when the processing module determines the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth, it is specifically used to determine the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth located within the effective area.

[0037] The detection module, when inputting a first tooth masking image into a tooth segmentation model to obtain a second tooth prediction bounding box and a third tooth segmentation result for the second tooth, specifically performs at least one spatial transformation on the first tooth masking image to obtain at least one second tooth masking image with a different principal axis orientation from the first tooth masking image; obtains a third tooth prediction bounding box for the second tooth by passing the first tooth masking image through the three-dimensional detection model of the tooth segmentation model, and obtains at least one fourth tooth prediction bounding box for the second tooth by passing at least one second tooth masking image through the three-dimensional detection model of the tooth segmentation model; obtains a second tooth prediction bounding box for the second tooth by deduplicating the third and fourth tooth prediction bounding boxes; and obtains a third tooth segmentation result for the second tooth by passing the second tooth prediction bounding box for the second tooth through the three-dimensional segmentation model of the tooth segmentation model.

[0038] Thirdly, embodiments of this application also provide a computing device, including:

[0039] Memory, used to store program instructions;

[0040] A processor is configured to invoke program instructions stored in the memory and execute the method described in either the first or second aspect according to the obtained program instructions.

[0041] Fourthly, embodiments of this application also provide a computer-readable storage medium storing computer-readable instructions that, when read and executed by a computer, cause the method described in any possible design of the first or second aspect to be implemented. Attached Figure Description

[0042] Figure 1 A schematic flowchart illustrating a tooth segmentation method provided in an embodiment of this application;

[0043] Figure 2 A schematic diagram of a three-dimensional image of a tooth with segmentation conflict provided for an embodiment of this application;

[0044] Figure 3 This is a schematic diagram illustrating the detection of a tooth mask image, as provided in an embodiment of this application.

[0045] Figure 4 A schematic diagram illustrating the process of detecting a tooth mask image provided in an embodiment of this application;

[0046] Figure 5 This is a schematic diagram illustrating spatial transformation of a teeth mask image, provided as an embodiment of this application.

[0047] Figure 6 A schematic diagram of sample images for a second classification model provided in an embodiment of this application;

[0048] Figure 7 A schematic diagram illustrating the determination of the x-coordinate range of a tooth prediction frame, provided in an embodiment of this application;

[0049] Figure 8 A schematic diagram illustrating the determination of the y-coordinate range of a tooth prediction frame, provided in an embodiment of this application;

[0050] Figure 9 This is a schematic diagram of a tooth-splitting device provided in an embodiment of this application;

[0051] Figure 10 This is a schematic diagram of the structure of a computing device provided in an embodiment of this application. Detailed Implementation

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

[0053] In the embodiments of this application, "multiple" refers to two or more. Terms such as "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or order.

[0054] Figure 1 This is a flowchart illustrating a tooth segmentation method provided in an embodiment of this application, as shown below. Figure 1 As shown, the method includes:

[0055] Step 101: Input the 3D image of the tooth into the tooth segmentation model to obtain at least one first tooth prediction box and at least one first tooth segmentation result of at least one first tooth.

[0056] In this embodiment, the three-dimensional image of the tooth can be a cone beam computer tomography (CBCT) image of the oral and maxillofacial region. CBCT images are the primary basis for oral diagnosis and can provide sufficient oral information. The area enclosed by the tooth prediction box is the target region that may be a tooth detected in the three-dimensional image of the tooth, and the tooth segmentation result is the tooth image segmented from the three-dimensional image of the tooth within the tooth prediction box.

[0057] The tooth segmentation model includes a 3D detection model and a 3D segmentation model. First, the 3D image of the tooth is input into the 3D detection model to obtain at least one predicted bounding box for at least one first tooth. Then, the 3D image of the tooth within the predicted bounding box is input into the 3D segmentation model to obtain at least one segmentation result for at least one first tooth.

[0058] The 3D detection model is used to detect potential target regions in a 3D image of teeth. The 3D detection model can be a deep learning-based 3D detection model such as RetinaNet3D. During training, the ATSS sample matching strategy is used. The input 3D image of the teeth is (128, 128, 128). The NMS non-maximum suppression threshold can be set to 0.6 to release more prediction results.

[0059] A 3D segmentation model is used to segment the 3D image within the predicted tooth bounding box output by the 3D detection model, obtaining a segmented tooth mask image. The 3D segmentation model can be a deep learning-based 3D segmentation model such as UNet, which uses a Transformer as the encoder. The size of the 3D tooth image within the predicted tooth bounding box is adjusted to (96, 96, 96) and then input into the 3D tooth model to obtain the mask image, where a mask of 1 represents a tooth image and a mask of 0 represents a non-tooth image. The mask image is then subjected to a series of processing steps, including morphological erosion, preservation of major connected regions, morphological dilation, resizing to the original size, and Gaussian smoothing, to obtain the final tooth segmentation result.

[0060] Step 102: Based on the segmentation results of multiple first teeth, determine the segmentation results of the first teeth that have segmentation conflicts.

[0061] In this context, a segmentation conflict refers to a situation where the overlap between two tooth segmentation results exceeds a set threshold. After obtaining multiple first tooth segmentation results, it is determined whether the overlap between two first tooth segmentation results exceeds the set threshold. If it does, a segmentation conflict is identified between the two first tooth segmentation results.

[0062] In one possible implementation, the existence of a segmentation conflict can be determined by the volume of the overlapping region of two first tooth segmentation results. For example, assuming two first tooth segmentation results m1 and m2, and an overlapping region m, the volume (m) of the overlapping region is checked to see if it exceeds a first preset threshold. If volume (m) exceeds the first preset threshold, a segmentation conflict is determined to exist between the two first tooth segmentation results. The first preset threshold can be set based on practical experience, and this application does not specifically limit it.

[0063] In another possible implementation, the existence of a segmentation conflict can be determined by the ratio of the volume of the overlapping region of the two first tooth segmentation results to the total volume of the first tooth segmentation results. Specifically, assuming two first tooth segmentation results m1 and m2, and an overlapping region m, the ratio of the volume of the overlapping region to the volume of either of the two first tooth segmentation results is determined to be greater than a second preset threshold. This is calculated as ios1 = volume(m) / volume(m1) and ios2 = volume(m) / volume(m2), where ios represents the ratio of the volume of the overlapping region to the total volume of the tooth segmentation results. If at least one of ios1 and ios2 is greater than the second preset threshold, a segmentation conflict is determined to exist between the two first tooth segmentation results. The second preset threshold can be set based on practical experience; for example, setting the second preset threshold to 0.1. If at least one of ios1 and ios2 is greater than 0.1, a segmentation conflict is determined to exist between the two first tooth segmentation results.

[0064] Step 103: Perform image classification and recognition on the first image to be segmented to identify at least one incorrect prediction box in the first tooth prediction box.

[0065] The first image to be segmented is the image corresponding to the predicted bounding box of the first tooth in the segmentation results of the first tooth where segmentation conflicts exist. (Reference) Figure 2 , Figure 2-2 If tooth segmentation results A and B conflict, then the predicted bounding boxes for the conflicting tooth segmentation results A and B are ( ). Figure 2-1 There must be incorrect prediction boxes in a) and b), and similarly, Figure 2-2 Tooth segmentation results C and D in the image also have segmentation conflicts. Therefore, the predicted bounding boxes for the teeth in segmentation results C and D that have segmentation conflicts are (…). Figure 2-1 There must be erroneous bounding boxes in c) and d). Therefore, after determining that there is a segmentation conflict in the first tooth segmentation result, it is necessary to determine whether the first tooth bounding box corresponding to the segmentation result with the segmentation conflict is an incorrectly identified bounding box.

[0066] When determining whether a tooth prediction box is an incorrect prediction box, the image of the tooth to be segmented within the range of the tooth prediction box and the corresponding tooth segmentation result can be input into the first classification model to determine whether the tooth prediction box is an incorrect prediction box.

[0067] In this model, the positive samples used to train the first classification model consist of correct tooth prediction boxes and their corresponding tooth segmentation results; the negative samples used to train the first classification model consist of historically incorrect prediction boxes obtained based on the tooth segmentation model and their corresponding tooth segmentation results. The ratio of positive to negative samples is maintained at approximately 1:1.

[0068] Step 104: Delete the first tooth segmentation results corresponding to the erroneous prediction boxes from each first tooth segmentation result to obtain the second tooth segmentation results.

[0069] Step 105: Mask the segmentation result of the second tooth in the 3D image of the tooth to obtain the masked image of the first tooth.

[0070] Step 106: Input the first tooth mask image into the tooth segmentation model to obtain the second tooth prediction box of the second tooth and the third tooth segmentation result of the second tooth.

[0071] The second tooth segmentation result is the identified tooth image in the 3D tooth image, for reference. Figure 3 , Figure 3 -a masks the identified teeth in the 3D tooth image before inputting it into the tooth segmentation model to obtain the second tooth's predicted bounding box and the third tooth's segmentation result. (Refer to...) Figure 3 -b, detects teeth that were not detected previously. Since the influence of teeth already identified in the 3D image is masked, this step can detect teeth that may have been missed in step 101, such as deciduous teeth and impacted teeth, as well as teeth that were mistakenly deleted in step 104.

[0072] It should be noted that if all teeth in the 3D image have been accurately identified before step 106, then after inputting the first tooth mask image into the segmentation model tooth segmentation model in step 106, there will be no output of tooth prediction boxes and tooth segmentation results.

[0073] In one possible implementation, refer to Figure 4 The process involves inputting the masked image of the first tooth into the tooth segmentation model to obtain the predicted bounding box of the second tooth and the segmentation result of the third tooth of the second tooth. Specifically, this includes the following steps:

[0074] Step 401: After performing at least one spatial transformation on the first tooth masking image, obtain at least one second tooth masking image with a different main axis orientation from the first tooth masking image.

[0075] Because the training data for the tooth segmentation model contains teeth whose principal axis orientation is along the Z-axis, the model may miss teeth with random principal axis orientations, such as impacted teeth. This embodiment of the application performs at least one spatial transformation on the first tooth mask image to obtain at least one second tooth mask image with a different principal axis orientation. The second tooth mask image is then input into the tooth segmentation model, enabling the model to detect and segment teeth whose principal axis orientation is offset from the Z-axis.

[0076] For example, the first tooth mask image can be subjected to at least one spatial transformation in the following manner:

[0077] Example 1, for reference Figure 5 ,Will Figure 5-1 Swapping the x-axis and z-axis coordinates of the first tooth mask image data-m(x, y, z) shown, and rotating the tooth mask image 90 degrees around the y-axis, yields... Figure 5-2 The second tooth mask image shown is `data-m.transpose(z, y, x)`. By swapping the y-axis and z-axis coordinates of the first tooth mask image, and rotating the tooth mask image 90 degrees around the x-axis, we obtain... Figure 5-3 The second tooth mask image shown is data-m.transpose(x, z, y). The sign of the transposed coordinates can be determined based on whether the rotation is clockwise or counterclockwise.

[0078] Example 2: The first tooth masking image is rotated by at least one predetermined angle to obtain at least one second tooth masking image. The at least one predetermined angle is one of several common tilt angles of impacted teeth obtained through large-scale statistical analysis based on the actual growth of the impacted teeth.

[0079] Step 402: Obtain the third tooth prediction box of the second tooth by passing the first tooth mask image through the three-dimensional detection model of the tooth segmentation model, and obtain the fourth tooth prediction box of the second tooth by passing at least one second tooth mask image through the three-dimensional detection model of the tooth segmentation model.

[0080] Continuing with Example 1 above, the first tooth masking image data-m(x, y, z) is input into the 3D detection model of the tooth segmentation model to obtain the third tooth prediction box (x, y, z) for the second tooth; the second tooth masking image m.transpose(z, y, x) is input into the 3D detection model of the tooth segmentation model to obtain the fourth tooth prediction box (z, y, x); the second tooth masking image m.transpose(x, z, y) is input into the 3D detection model of the tooth segmentation model to obtain the fourth tooth prediction box (x, z, y) for the second tooth.

[0081] Step 403: After deduplicating the prediction boxes of the third and fourth teeth, the prediction box of the second tooth is obtained.

[0082] Since the above steps involve detecting the same tooth mask image from different angles, there may be duplicate tooth prediction boxes between the fourth and third tooth prediction boxes for the second tooth. Therefore, it is necessary to remove duplicates from the fourth and third tooth prediction boxes to obtain the second tooth prediction box for the second tooth.

[0083] Continuing with Example 1 above, after deduplicating the predicted boxes for the third tooth (x, y, z), the fourth tooth (z, y, x), and the second tooth (x, z, y), we obtain the predicted box for the second tooth.

[0084] Step 404: Obtain the third tooth segmentation result of the second tooth by using the three-dimensional segmentation model of the tooth segmentation model to obtain the second tooth's second tooth prediction box.

[0085] Step 107: Based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth, determine the tooth segmentation result of the three-dimensional image of the tooth.

[0086] The segmentation result of the third tooth of the second tooth is used as a supplement to the segmentation result of the second tooth of the first tooth to obtain the tooth segmentation result of the three-dimensional image of the tooth.

[0087] In the above technical solution, the accuracy of the first predicted bounding box corresponding to the first tooth segmentation result with segmentation conflict is first determined, and the first tooth segmentation result corresponding to the incorrect predicted bounding box is deleted. Then, the segmented teeth are masked to obtain the first tooth mask image, which eliminates the interference of the segmented teeth and constructs an environment where the targets are not crowded or overlapping. The first tooth mask image is input into the tooth segmentation model to detect teeth that may be missed, which can improve the accuracy of detecting abnormal types of teeth.

[0088] In one possible implementation, after obtaining the second tooth prediction box and the third tooth segmentation result of the second tooth, the second tooth prediction box and the third tooth segmentation result can be further processed by at least one of the following methods to obtain a more accurate tooth segmentation result of the three-dimensional image.

[0089] Method 1

[0090] Determine whether there is a segmentation conflict between the third tooth segmentation result of each second tooth and the second tooth segmentation result of the first tooth.

[0091] Specifically, the method for determining whether a segmentation conflict exists is the same as step 102, and will not be elaborated here. Since the segmentation result of the third tooth of the second tooth serves as a supplement to the segmentation result of the second tooth of the first tooth, when there is a segmentation conflict between the segmentation result of the third tooth of the second tooth and the segmentation result of the second tooth of the first tooth, the segmentation result of the third tooth of the second tooth is deleted. The segmentation result of the second tooth of the first tooth and the segmentation result of the third tooth of the second tooth, which do not have a segmentation conflict, are determined as the tooth segmentation result of the 3D image of the tooth.

[0092] Method 2

[0093] Determine whether there is a segmentation conflict between the segmentation results of the third teeth of each second tooth. If there is a segmentation conflict, determine whether the predicted box of the second tooth corresponding to the segmentation result of the third tooth with the segmentation conflict is an incorrect predicted box.

[0094] Specifically, the method for determining whether segmentation conflicts exist is the same as step 102, and will not be elaborated here. When determining whether the second tooth prediction box corresponding to the third tooth segmentation result with segmentation conflict is an incorrect prediction box, the image of the tooth to be segmented within the range of the second tooth prediction box and the corresponding third tooth segmentation result can be input into the first classification model to determine whether the second tooth prediction box is an incorrect prediction box. If the second tooth prediction box is determined to be an incorrect prediction box, the third tooth segmentation result corresponding to the second tooth prediction box is deleted to further improve the accuracy of the supplementary segmentation results. The second tooth segmentation result of the first tooth and the third tooth segmentation result corresponding to the non-incorrect prediction box in the second tooth prediction box are determined as the tooth segmentation result of the 3D tooth image.

[0095] Method 3

[0096] The masked image of the first tooth within any second tooth prediction bounding box and its range is input into the second classification model to determine whether the second tooth prediction bounding box is a valid prediction bounding box. If the second tooth prediction bounding box is not a valid prediction bounding box, the third tooth segmentation result corresponding to that second tooth prediction bounding box is deleted to further improve the accuracy of the supplementary segmentation results. The second tooth segmentation result of the first tooth and the third tooth segmentation result corresponding to the valid prediction bounding box in the second tooth prediction bounding box are determined as the tooth segmentation result of the 3D tooth image.

[0097] The second classification model is used to determine whether the second tooth prediction box is a valid prediction box. In one possible implementation, the negative samples for training the second classification model can be obtained by masking all the teeth in the sample image to obtain a first masked sample image, and scaling and / or rotating and / or translating the tooth prediction box corresponding to each masked tooth in the first masked sample image as a negative sample for the second classification model.

[0098] Specifically, refer to Figure 6 , Figure 6 -(a) Mask all teeth in the sample image to obtain the first masked sample image data-m0. For the tooth prediction box corresponding to each masked tooth, perform proportional scaling and / or rotation and / or translation while keeping the center unchanged. For example, randomly scale the tooth prediction box with a ratio of r = random(-0.05, 0.1), rotate it 90° around the X-axis with a probability of 0.25, rotate it 90° around the Y-axis with a probability of 0.25, rotate it 90° around the Z-axis with a probability of 0.25, and do not rotate it with a probability of 0.25. Then, while keeping the tooth prediction box from exceeding the tooth boundary, randomly translate it to a position. Use the scaled and / or rotated and / or translated tooth prediction boxes and the masked sample images within the tooth prediction boxes as negative samples for the second classification model.

[0099] In one possible implementation, positive samples for training the second classification model can be obtained by randomly selecting a portion of the teeth in the sample image to obtain a second masked sample image, masking another portion of the teeth to obtain a third masked sample image, and enlarging the tooth prediction boxes corresponding to the unmasked teeth in the second and third masked sample images as positive samples for the second classification model.

[0100] Specifically, refer to Figure 6 -(b) and Figure 6 -(c), taking the random selection of half of the teeth for masking as an example, half of the teeth are randomly selected from the sample image to mask to obtain the second masked sample image data-m1, and the other half of the teeth are masked to obtain the third masked sample image data-m2. The tooth prediction boxes corresponding to the unmasked teeth in the second masked sample image data-m1 and the third masked sample image data-m2 are enlarged. For example, the tooth prediction boxes are randomly enlarged at a ratio of r = random(0,0.1). The enlarged tooth prediction boxes and the masked sample images in the tooth prediction boxes are used as positive samples of the second classification model.

[0101] Method 4

[0102] Determine whether the predicted bounding box for the second tooth is within the effective region of the segmented tooth. If the predicted bounding box is not within the effective region, it is considered an incorrectly identified bounding box, and the corresponding tooth segmentation result is deleted to further improve the accuracy of the supplementary segmentation results. The segmentation results of the second tooth for the first tooth and the third tooth segmentation results for the second tooth within the effective region are determined as the tooth segmentation results of the 3D tooth image.

[0103] For example, it can be determined whether the second tooth prediction box is within the effective area of ​​the segmented tooth by comparing the coordinates of the second prediction box with the coordinates of the segmented tooth. Specifically, it can be determined whether the x-coordinate of the second tooth prediction box is within the x-coordinate range of the segmented tooth, and whether the y-coordinate of the second tooth prediction box is within the y-coordinate range of the segmented tooth.

[0104] For example, refer to Figure 7 When determining whether the x-coordinate of the second tooth prediction box is within the x-coordinate range of the segmented teeth, for each second tooth prediction box, first determine the leftmost tooth prediction box (box-L) and the rightmost tooth prediction box (box-R) that are in the same dental row (upper or lower dentition) as the second tooth prediction box. Then, the prediction box that is closest to the second tooth prediction box between the leftmost and rightmost tooth prediction boxes (box-L and box-R) is taken as the prediction box to be compared (box-compare), and it is determined whether the x-coordinate of the second tooth prediction box is less than the x-coordinate of the prediction box to be compared (box-compare).

[0105] For example, refer to Figure 8 To determine whether the y-coordinate of the second tooth prediction frame is within the range of the y-coordinates of the segmented teeth, first determine the projection S1 of the segmented teeth on the XY plane and the projection S2 of the second tooth prediction frame on the XY plane. Then, draw a straight line L parallel to the Y-axis through the center of the projection S2 of the second tooth prediction frame on the XY plane. Determine the two intersection points of this line L with the projection S1 of the segmented teeth on the XY plane. The y-coordinate of the intersection point with the smallest y-coordinate is denoted as y-min, and the y-coordinate of the intersection point with the largest y-coordinate is denoted as y-max. Determine whether the y-coordinate of the second tooth prediction frame is within the range of the y-coordinates of the two intersection points (from y-min to y-max).

[0106] It should be noted that the above four methods can be used individually or in combination, and this application does not specify the order in which they are used in combination.

[0107] Based on the same technological concept Figure 9 This is a schematic diagram of a tooth-splitting device provided in an embodiment of this application, as shown below. Figure 9 As shown, the device 900 includes:

[0108] The detection module 901 is used to input a three-dimensional image of a tooth into a tooth segmentation model to obtain at least one first tooth prediction box and at least one first tooth segmentation result of at least one first tooth.

[0109] The detection module 901 is further configured to determine, based on multiple first tooth segmentation results, a first tooth segmentation result with segmentation conflict; the segmentation conflict refers to the overlap between two tooth segmentation results being greater than a set threshold; perform image classification and recognition on the first image to be segmented to determine at least one incorrect prediction box in the first tooth prediction box; the first image to be segmented is the image to be segmented corresponding to the first tooth prediction box of the first tooth segmentation result with segmentation conflict; delete the first tooth segmentation result corresponding to the incorrect prediction box from each first tooth segmentation result to obtain a second tooth segmentation result;

[0110] The detection module 901 is further configured to mask the second tooth segmentation result in the three-dimensional image of the tooth to obtain a first tooth masking image; and input the first tooth masking image into the tooth segmentation model to obtain the second tooth prediction box of the second tooth and the third tooth segmentation result of the second tooth.

[0111] The processing module 902 is used to determine the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth.

[0112] In one possible design, after obtaining the second tooth prediction box of the second tooth and the third tooth segmentation result of the second tooth, the detection module 901 is further used to determine whether there is a segmentation conflict between the third tooth segmentation result of each second tooth and the second tooth segmentation result of the first tooth; when determining the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth, the processing module 902 is specifically used to determine the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth of the second tooth where there is no segmentation conflict.

[0113] In one possible design, after obtaining the second tooth prediction box of the second tooth and the third tooth segmentation result of the second tooth, the detection module 901 is further used to determine whether there is a segmentation conflict between the third tooth segmentation results of each second tooth; if there is a segmentation conflict, it is determined whether the second tooth prediction box corresponding to the third tooth segmentation result with the segmentation conflict is an incorrect prediction box; the processing module 902, when determining the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth, is specifically used to determine the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result corresponding to the non-incorrect prediction box in the second tooth prediction box.

[0114] In one possible design, the detection module 901 determines whether a tooth prediction box is an incorrect prediction box by: inputting the image of the tooth to be segmented within the range of the tooth prediction box and the corresponding tooth segmentation result into a first classification model, and determining whether the tooth prediction box is an incorrect prediction box; wherein, the positive samples for training the first classification model are correct tooth prediction boxes and corresponding tooth segmentation results; the negative samples for training the first classification model are historical incorrect prediction boxes and corresponding tooth segmentation results obtained based on the tooth segmentation model.

[0115] In one possible design, after the detection module 901 obtains the second tooth prediction box and the third tooth segmentation result of the second tooth, the processing module 902 is used to input the first tooth masking image within the range of any second tooth prediction box and the second tooth prediction box into the second classification model to determine whether the second tooth prediction box is a valid prediction box; when the processing module 902 determines the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth, it is specifically used to determine the tooth segmentation result of the three-dimensional image of the tooth according to the second tooth segmentation result of the first tooth and the third tooth segmentation result corresponding to the valid prediction box in the second tooth prediction box.

[0116] In one possible design, the device further includes a training module 903 for obtaining negative samples for training the second classification model by: masking all teeth in the sample image to obtain a first masked sample image; scaling and / or rotating and / or translating the tooth prediction box corresponding to each masked tooth in the first masked sample image as a negative sample for the second classification model.

[0117] In one possible design, the training module 903 is further configured to obtain positive samples for training the second classification model by: randomly selecting a portion of teeth from the sample image to mask them to obtain a second masked sample image, and masking another portion of teeth to obtain a third masked sample image; and enlarging the predicted bounding boxes of the teeth corresponding to the unmasked teeth in the second masked sample image and the third masked sample image to use as positive samples for the second classification model.

[0118] In one possible design, after obtaining the second tooth prediction bounding box of the second tooth and the third tooth segmentation result of the second tooth, the detection module 901 is further used to determine whether the second tooth prediction bounding box is within the effective area of ​​the segmented tooth; when determining the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth, the processing module 902 is specifically used to determine the tooth segmentation result of the three-dimensional image of the tooth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth located within the effective area.

[0119] The detection module 901, when inputting a first tooth masking image into a tooth segmentation model to obtain a second tooth prediction box and a third tooth segmentation result for the second tooth, specifically performs at least one spatial transformation on the first tooth masking image to obtain at least one second tooth masking image with a different principal axis orientation from the first tooth masking image; obtains a third tooth prediction box for the second tooth by passing the first tooth masking image through the three-dimensional detection model of the tooth segmentation model, and obtains a fourth tooth prediction box for the second tooth by passing at least one second tooth masking image through the three-dimensional detection model of the tooth segmentation model; obtains a second tooth prediction box for the second tooth by deduplicating the third and fourth tooth prediction boxes; and obtains a third tooth segmentation result for the second tooth by passing the second tooth prediction box for the second tooth through the three-dimensional segmentation model of the tooth segmentation model.

[0120] Based on the same technical concept, embodiments of this application provide a computing device, such as... Figure 10 As shown, it includes at least one processor 1001 and a memory 1002 connected to at least one processor. In this embodiment, the specific connection medium between the processor 1001 and the memory 1002 is not limited. Figure 10 Taking the connection between processor 1001 and memory 1002 via a bus as an example. The bus can be divided into address bus, data bus, control bus, etc.

[0121] In this embodiment of the application, the memory 1002 stores instructions that can be executed by at least one processor 1001. By executing the instructions stored in the memory 1002, at least one processor 1001 can perform the above-described tooth segmentation method.

[0122] The processor 1001 is the control center of the computing device. It can connect to various parts of the computer device through various interfaces and lines, and perform resource settings by running or executing instructions stored in the memory 1002 and calling data stored in the memory 1002.

[0123] Optionally, the processor 1001 may include one or more processing units. The processor 1001 may integrate an application processor and a modem processor, wherein the application processor primarily handles the operating system, user interface, and applications, while the modem processor primarily handles wireless communication. It is understood that the modem processor may not be integrated into the processor 1001. In some embodiments, the processor 1001 and the memory 1002 may be implemented on the same chip; in some embodiments, they may be implemented separately on independent chips.

[0124] The processor 1001 can be a general-purpose processor, such as a central processing unit (CPU), digital signal processor, application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0125] Memory 1002, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. Memory 1002 may include at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type memory, random access memory (RAM), static random access memory (SRAM), programmable read-only memory (PROM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), magnetic memory, magnetic disk, optical disk, etc. Memory 1002 can be any other medium capable of carrying or storing desired program code in the form of instructions or data structures that can be accessed by a computer, but is not limited thereto. In the embodiments of this application, memory 1002 can also be a circuit or any other device capable of implementing storage functions for storing program instructions and / or data.

[0126] Based on the same technical concept, embodiments of this application also provide a computer-readable storage medium storing a computer-executable program for causing a computer to perform any of the tooth segmentation methods listed above.

[0127] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0128] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0129] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0130] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0131] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0132] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for tooth segmentation, characterized in that, The method includes: The three-dimensional image of the tooth is input into the tooth segmentation model to obtain at least one first tooth prediction box and at least one first tooth segmentation result of at least one first tooth; Based on multiple first tooth segmentation results, a first tooth segmentation result with segmentation conflict is identified; the segmentation conflict refers to the overlap between two tooth segmentation results being greater than a set threshold; image classification and recognition are performed on the first image to be segmented to identify at least one incorrect prediction box in the first tooth prediction box; the first image to be segmented is the image to be segmented corresponding to the first tooth prediction box of the first tooth segmentation result with segmentation conflict; the first tooth segmentation result corresponding to the incorrect prediction box is deleted from each first tooth segmentation result to obtain the second tooth segmentation result. The second tooth segmentation result in the three-dimensional image of the tooth is masked to obtain the first tooth masking image; the first tooth masking image is input into the tooth segmentation model to obtain the second tooth prediction box of the second tooth and the third tooth segmentation result of the second tooth. Based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth, the tooth segmentation result of the three-dimensional image of the tooth is determined.

2. The method according to claim 1, characterized in that, After obtaining the second tooth prediction box of the second tooth and at least one third tooth segmentation result of the second tooth, the process also includes: Determine whether there is a segmentation conflict between the third tooth segmentation result of each second tooth and the second tooth segmentation result of the first tooth; Based on the segmentation results of the second tooth from the first tooth and the third tooth from the second tooth, the tooth segmentation results of the three-dimensional image of the tooth are determined, including: Based on the segmentation results of the second tooth of the first tooth and the segmentation results of the third tooth of the second tooth where there is no segmentation conflict, the tooth segmentation results of the three-dimensional image of the tooth are determined.

3. The method according to claim 1, characterized in that, After obtaining the second tooth prediction box and the third tooth segmentation results of the second tooth, the following steps are also included: Determine whether there are any segmentation conflicts between the segmentation results of the second and third teeth; If there is a segmentation conflict, determine whether the second tooth prediction box corresponding to the segmentation result of the third tooth with the segmentation conflict is an incorrect prediction box. Based on the segmentation results of the second tooth from the first tooth and the third tooth from the second tooth, the tooth segmentation results of the three-dimensional image of the tooth are determined, including: Based on the second tooth segmentation result of the first tooth and the third tooth segmentation result corresponding to the non-erroneous prediction box in the second tooth prediction box, the tooth segmentation result of the three-dimensional image of the tooth is determined.

4. The method according to claim 1, characterized in that, The following method can be used to determine whether a tooth prediction bounding box is an incorrect prediction bounding box: The image of the tooth to be segmented within the predicted tooth bounding box and the corresponding tooth segmentation result are input into the first classification model to determine whether the predicted tooth bounding box is an incorrect predicted box. The positive samples used to train the first classification model are correct tooth prediction boxes and corresponding tooth segmentation results; the negative samples used to train the first classification model are historically incorrect prediction boxes and corresponding tooth segmentation results obtained based on the tooth segmentation model.

5. The method according to claim 1, characterized in that, After obtaining the second tooth prediction box and the third tooth segmentation results of the second tooth, the following steps are also included: Input the first tooth mask image within any second tooth prediction box and the range of the second tooth prediction box into the second classification model to determine whether the second tooth prediction box is a valid prediction box; Based on the segmentation results of the second tooth from the first tooth and the third tooth from the second tooth, the tooth segmentation results of the three-dimensional image of the tooth are determined, including: Based on the second tooth segmentation result of the first tooth and the third tooth segmentation result corresponding to the effective prediction box in the second tooth prediction box, the tooth segmentation result of the three-dimensional image of the tooth is determined.

6. The method according to claim 5, characterized in that, The negative samples used to train the second classification model are obtained in the following way: The first masked sample image is obtained by masking all the teeth in the sample image. The predicted bounding boxes of each masked tooth in the first masked sample image are scaled and / or rotated and / or translated, and used as negative samples for the second classification model.

7. The method according to claim 5, characterized in that, Positive samples for training the second classification model are obtained in the following way: A second masked sample image is obtained by randomly selecting a portion of the teeth from the sample image and masking another portion of the teeth; a third masked sample image is obtained by masking another portion of the teeth. The predicted bounding boxes of unmasked teeth in the second and third masked sample images are enlarged and used as positive samples for the second classification model.

8. The method according to claim 1, characterized in that, After obtaining the second tooth prediction box and the third tooth segmentation results of the second tooth, the following steps are also included: Determine whether the second tooth prediction box is within the effective area of ​​the segmented tooth; Based on the segmentation results of the second tooth from the first tooth and the third tooth from the second tooth, the tooth segmentation results of the three-dimensional image of the tooth are determined, including: Based on the segmentation results of the second tooth of the first tooth and the segmentation results of the third tooth of the second tooth located within the effective area, the tooth segmentation results of the three-dimensional image of the tooth are determined.

9. The method according to any one of claims 1 to 8, characterized in that, The first tooth mask image is input into the tooth segmentation model to obtain the second tooth prediction box and the third tooth segmentation result, including: After performing at least one spatial transformation on the first tooth masking image, at least one second tooth masking image with a different main axis orientation from the first tooth masking image is obtained; The first tooth masking image is used to obtain the third tooth prediction box of the second tooth through the three-dimensional detection model of the tooth segmentation model, and at least one second tooth masking image is used to obtain the fourth tooth prediction box of the second tooth through the three-dimensional detection model of the tooth segmentation model. After deduplication of the third and fourth tooth prediction frames, the second tooth prediction frame is obtained for the second tooth. The second tooth prediction box of the second tooth is used to obtain the third tooth segmentation result of the second tooth through the three-dimensional segmentation model of the tooth segmentation model.

10. A tooth-splitting device, characterized in that, include: The detection module is used to input a three-dimensional image of a tooth into a tooth segmentation model to obtain at least one first tooth prediction box and at least one first tooth segmentation result of at least one first tooth. The detection module is further configured to determine, based on the first tooth segmentation results of multiple first teeth, a first tooth segmentation result with segmentation conflict; the segmentation conflict refers to the overlap between two tooth segmentation results being greater than a set threshold. Perform image classification and recognition on the first image to be segmented to identify at least one incorrectly predicted box within the first tooth prediction box; The first image to be segmented is the image corresponding to the first tooth prediction box of the first tooth segmentation result with segmentation conflict; the first tooth segmentation result corresponding to the incorrect prediction box is deleted from each first tooth segmentation result to obtain the second tooth segmentation result; The detection module is also used to mask the second tooth segmentation result in the three-dimensional image of the tooth to obtain a first tooth mask image; The first tooth mask image is input into the tooth segmentation model to obtain the second tooth prediction box of the second tooth and the third tooth segmentation result of the second tooth. The processing module is used to determine the tooth segmentation result of the three-dimensional image of the teeth based on the second tooth segmentation result of the first tooth and the third tooth segmentation result of the second tooth.

11. A computing device, characterized in that, include: Memory, used to store program instructions; A processor is configured to invoke program instructions stored in the memory and execute the method as described in any one of claims 1 to 9 according to the obtained program instructions.

12. A computer-readable storage medium, characterized in that, Includes computer-readable instructions that, when read and executed by a computer, cause the method as described in any one of claims 1 to 9 to be implemented.

Images