Dental site recognition and tooth segmentation method and system based on deep learning

Abstract

The application belongs to the technical field of dental image processing, and provides a tooth position recognition and tooth segmentation method and system based on deep learning. The method comprises the following steps: extracting global and local features of a natural light oral cavity image to obtain feature maps with different sizes and scales; fusing the feature maps of all scales to generate a feature pyramid and determine all candidate regions of interest; sequentially filtering and normalizing the candidate regions of interest; processing the normalized regions of interest through a pre-trained classification model and an instance segmentation process respectively to obtain preliminary tooth position recognition results and preliminary instance segmentation results of each tooth; performing non-maximum suppression processing on the preliminary instance segmentation results of each tooth to ensure that each tooth has and only has one tooth; removing repeated tooth positions in the preliminary tooth position recognition results according to the tooth centroids after the non-maximum suppression processing; and performing centroid sorting on the deduplicated tooth positions to obtain final tooth position recognition results and tooth segmentation results.

Description

Technical Field

[0001] This invention belongs to the field of dental image processing technology, and particularly relates to a method and system for tooth position recognition and tooth segmentation based on deep learning. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In practice, a complete panoramic X-ray of the oral cavity is required before tooth positioning can be identified. This not only increases the user's expenses but also poses a slight risk to their health. Most existing technologies that use artificial intelligence to process oral images simply use pre-trained neural network models to infer tooth positions. However, the inventors discovered that oral images taken under natural light with poor angles or unclear images cannot be accurately segmented and identified simply through network model inference. Summary of the Invention

[0004] To address the technical problems mentioned above, this invention provides a tooth position recognition and tooth segmentation method and system based on deep learning, which can efficiently and accurately identify the position of teeth in natural light oral images and complete tooth segmentation, and further correct the tooth recognition and segmentation results.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] The first aspect of the present invention provides a method for tooth position recognition and tooth segmentation based on deep learning.

[0007] A deep learning-based method for tooth position recognition and tooth segmentation, comprising:

[0008] Global and local features of natural light oral images are extracted to obtain feature maps of different scales;

[0009] By fusing feature maps at all scales, a feature pyramid is generated, and all candidate regions of interest are identified.

[0010] The candidate regions of interest are filtered and normalized sequentially.

[0011] The normalized regions of interest are processed by a pre-trained classification model and instance segmentation to obtain preliminary tooth position identification results and preliminary instance segmentation results for each tooth.

[0012] Non-maximum suppression is applied to the initial segmentation results of each tooth to ensure that there is exactly one tooth for each instance.

[0013] Based on the tooth centroid after nonmaximum suppression processing, duplicate tooth positions in the preliminary tooth position identification results are removed;

[0014] The deduplicated tooth positions are sorted by centroid to obtain the final tooth position recognition and tooth segmentation results.

[0015] As one implementation method, the process of obtaining preliminary tooth position identification results for each tooth is as follows:

[0016] The normalized region of interest is processed by a pre-trained classification model to obtain the category of the natural light oral cavity image; wherein, the category of the natural light oral cavity image includes maxillary image, mandibular image and normal tooth occlusion image;

[0017] Preliminary tooth position identification results were determined based on the category of natural light oral images.

[0018] As one implementation method, the process of filtering and normalizing the candidate regions of interest sequentially is as follows:

[0019] Binary classification and regression are performed on all candidate regions of interest to determine the regions of interest to be retained;

[0020] Based on the foreground score, the retained regions of interest are sorted, and then the non-maximum suppression method is used to filter out duplicate regions of interest. Finally, the size of the remaining regions of interest is normalized.

[0021] As one implementation method, the centroid ordering of the deduplicated posterior teeth is determined by the category of the natural light oral cavity image; wherein, the category of the natural light oral cavity image includes maxillary image, mandibular image and normal tooth occlusion image.

[0022] As one implementation, if the type of natural light oral image is a maxillary image, the centroid coordinates of each tooth are arranged in a downward-opening parabola to determine the position of each tooth in the maxillary image.

[0023] As one implementation, if the category of the natural light oral image is a mandibular image, then the centroid coordinates of each tooth are arranged in the order of an upward-opening parabola to determine the position of each tooth in the mandibular image.

[0024] As one implementation method, if the category of the natural light oral image is a normal tooth occlusion image, then the centroid ordinate of each tooth is used to determine whether each tooth belongs to the maxilla or mandible, and the horizontal arrangement order of the maxilla and mandible teeth is determined by the centroid abscissa of each tooth in ascending order.

[0025] As one implementation method, the centroid coordinates of each tooth are used to correct tooth position identification misalignment.

[0026] A second aspect of the present invention provides a tooth position recognition and tooth segmentation system based on deep learning.

[0027] A deep learning-based tooth position recognition and tooth segmentation system, comprising:

[0028] The feature extraction module is used to extract global and local features from natural light oral cavity images, resulting in feature maps of different sizes.

[0029] The candidate region of interest generation module is used to fuse feature maps of all scales, generate a feature pyramid, and determine all candidate regions of interest.

[0030] The Region of Interest (ROI) filtering and normalization module is used to filter and normalize candidate ROIs sequentially.

[0031] The preliminary classification and segmentation result acquisition module is used to process the normalized region of interest through a pre-trained classification model and instance segmentation to obtain the preliminary tooth position identification result and the preliminary instance segmentation result for each tooth.

[0032] The non-maximum suppression processing module is used to perform non-maximum suppression processing on the preliminary results of each tooth instance segmentation to ensure that there is exactly one tooth for each tooth.

[0033] The duplicate tooth position removal module is used to remove duplicate tooth positions in the preliminary tooth position identification results based on the tooth centroid after non-maximum suppression processing.

[0034] The centroid sorting module is used to sort the centroids of the deduplicated tooth positions to obtain the final tooth position recognition result and tooth segmentation result.

[0035] As one implementation method, the process of obtaining the preliminary tooth position identification result for each tooth in the preliminary classification and segmentation result acquisition module is as follows:

[0036] The normalized region of interest is processed by a pre-trained classification model to obtain the category of the natural light oral cavity image; wherein, the category of the natural light oral cavity image includes maxillary image, mandibular image and normal tooth occlusion image;

[0037] Preliminary tooth position identification results were determined based on the category of natural light oral images.

[0038] As one implementation method, in the region of interest filtering and normalization module, the process of sequentially filtering and normalizing the candidate regions of interest is as follows:

[0039] Binary classification and regression are performed on all candidate regions of interest to determine the regions of interest to be retained;

[0040] Based on the foreground score, the retained regions of interest are sorted, and then the non-maximum suppression method is used to filter out duplicate regions of interest. Finally, the size of the remaining regions of interest is normalized.

[0041] As one implementation method, the centroid ordering of the deduplicated posterior teeth is determined by the category of the natural light oral cavity image; wherein, the category of the natural light oral cavity image includes maxillary image, mandibular image and normal tooth occlusion image.

[0042] In one implementation, if the type of natural light oral image is a maxillary image, the centroid coordinates of each tooth are arranged in a downward-opening parabola to determine the position of each tooth in the maxillary image.

[0043] In one implementation, if the type of natural light oral image is a mandibular image, the centroid coordinates of each tooth are arranged in the order of an upward-opening parabola to determine the position of each tooth in the mandibular image.

[0044] As one implementation method, if the category of the natural light oral image is a normal tooth occlusion image, then the centroid ordinate of each tooth is used to determine whether each tooth belongs to the maxilla or mandible, and the horizontal arrangement order of the maxilla and mandible teeth is determined by the centroid abscissa of each tooth in ascending order.

[0045] As one implementation method, the centroid coordinates of each tooth are used to correct tooth position identification misalignment.

[0046] A third aspect of the present invention provides a computer-readable storage medium.

[0047] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the deep learning-based tooth position recognition and tooth segmentation method described above.

[0048] A fourth aspect of the present invention provides an electronic device.

[0049] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the deep learning-based tooth position recognition and tooth segmentation method described above.

[0050] Compared with the prior art, the beneficial effects of the present invention are:

[0051] This invention processes natural light oral images, extracts global and local features, and performs category recognition. The preliminary results of instance segmentation and tooth position recognition for each tooth are then processed with non-maximum suppression and tooth centroid position determination. Duplicate tooth positions are removed from the preliminary tooth position recognition results, and centroid sorting is performed to obtain the final tooth position recognition and tooth segmentation results. This solves the problem of inaccurate tooth position recognition and segmentation due to poor shooting angles and unclear images in user-captured oral images, thus improving the accuracy of tooth position recognition and segmentation.

[0052] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0053] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0054] Figure 1 This is a flowchart of a deep learning-based tooth position recognition and tooth segmentation method according to an embodiment of the present invention;

[0055] Figure 2 This is a post-processing flowchart of an embodiment of the present invention;

[0056] Figure 3 This is the final result of tooth position recognition and tooth segmentation of a normal occlusion image according to an embodiment of the present invention;

[0057] Figure 4 This is the final result of tooth position recognition and tooth segmentation of the maxillary image in this embodiment of the invention;

[0058] Figure 5 This is the final result of tooth position recognition and tooth segmentation of the mandibular image in this embodiment of the invention;

[0059] Figure 6 This is a preliminary result of maxillary intraoral tooth position identification and tooth segmentation in an embodiment of the present invention;

[0060] Figure 7 This is the result of correcting the incorrect tooth position after post-processing in an embodiment of the present invention;

[0061] Figure 8 These are preliminary results of mandibular image tooth position recognition and tooth segmentation according to an embodiment of the present invention;

[0062] Figure 9 This is the result of correcting the incorrect tooth position after post-processing in an embodiment of the present invention;

[0063] Figure 10These are preliminary results of tooth position recognition and tooth segmentation in normal occlusion images according to embodiments of the present invention;

[0064] Figure 11 This is the result of correcting the incorrect tooth position after post-processing in an embodiment of the present invention. Detailed Implementation

[0065] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0066] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0067] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0068] Example 1

[0069] This embodiment provides a deep learning-based method for tooth position recognition and tooth segmentation, which specifically includes the following steps:

[0070] Step 1: Extract global and local features from natural light oral cavity images to obtain feature maps of different scales.

[0071] according to Figure 1 It can be seen that before extracting the global and local features of the natural light oral cavity image in step 1, there are also data acquisition and data processing steps.

[0072] The data acquisition step involves acquiring images of the user's oral cavity under natural light.

[0073] The data preprocessing steps are as follows: the acquired natural light oral cavity images are subjected to data enhancement operations such as horizontal flipping, normalization, standardization, random adjustment of contrast and brightness, and histogram equalization.

[0074] This embodiment uses a backbone network to extract global and local features from natural light oral cavity images. Specifically, the Swin Transformer is chosen instead of convolutional networks like ResNet or VGG in Mask R-CNN. In the backbone network, features of intraoral images are extracted through a series of convolutional operations to obtain corresponding feature maps. The Swin Transformer is chosen because it innovatively introduces a sliding window mechanism, restricting attention computation to a single window. This not only introduces the locality of convolutional operations but also significantly reduces computational cost, allowing the model to learn information across windows. Simultaneously, downsampling layers gradually increase the receptive field, enabling the attention mechanism to attend to global features. Using the Swin Transformer as the backbone network saves computation while still focusing on both global and local information features. After feature extraction by the backbone network, feature maps of different scales are obtained.

[0075] Step 2: Fuse feature maps at all scales to generate a feature pyramid and identify all candidate regions of interest.

[0076] In step 2, a Feature Pyramid Network (FPN) is used to better fuse the feature maps obtained from the backbone network to generate a feature pyramid. Most network models directly use the feature maps from the last layer. Although the feature maps from the last layer, after a series of convolutions, have better semantic information, their resolution and location feature information are relatively low, making it easy to overlook small objects in the original in-aperture image. Therefore, the role of the FPN is to fuse feature maps of different resolutions, thereby obtaining features from each stage and reducing missed detections.

[0077] Step 3: Filter and normalize the candidate regions of interest in sequence.

[0078] The process of filtering and normalizing the candidate regions of interest sequentially is as follows:

[0079] Binary classification and regression are performed on all candidate regions of interest to determine the regions of interest to be retained;

[0080] Based on the foreground score, the retained regions of interest are sorted, and then the non-maximum suppression method is used to filter out duplicate regions of interest. Finally, the size of the remaining regions of interest is normalized.

[0081] Specifically, a series of ROIs (Regions of Interest) are generated on the pixels of the fused feature map obtained by the FPN network, thereby obtaining multiple candidate ROIs. The role of the RPN (Region Proposal Network) is to perform binary classification (foreground or background) and regression on the candidate ROIs, filtering out a portion of the candidate ROIs.

[0082] The Proposal Layer is used to filter the ROIs generated by RPN, and they are sorted according to the foreground score. Then, NMS (Non-Maximum Suppression) is used to filter out duplicate boxes.

[0083] The RoI Align layer makes a slight modification to RoI pooling. This is because RoIPooling performs two rounding operations during pooling, which leads to significant deviations when reconstructing the feature map from the original intraoral image. RoI Align, however, abandons this direct rounding method and instead uses bilinear interpolation to obtain pixel values ​​at four fixed coordinates. This makes the discontinuous operations in RoI Pooling more continuous, reducing the error in reconstructing the intraoral image. The main function of RoI Align is to unify the ROIs obtained from the Proposal Layer to a uniform size for subsequent multi-classification and regression operations.

[0084] Step 4: The normalized regions of interest are processed by the pre-trained classification model and instance segmentation to obtain the preliminary tooth position identification results and preliminary instance segmentation results for each tooth.

[0085] In the specific implementation process, the process of obtaining the preliminary tooth position identification results for each tooth is as follows:

[0086] The normalized region of interest is processed by a pre-trained classification model to obtain the category of the natural light oral cavity image; wherein, the category of the natural light oral cavity image includes maxillary image, mandibular image and normal tooth occlusion image;

[0087] Preliminary tooth position identification results were determined based on the category of natural light oral images.

[0088] Specifically, FCN (Fully Convolution Network) is used to generate a mask (instance segmentation) for each ROI.

[0089] For each natural light oral cavity image input into this invention, a pre-trained classification model is used to determine whether the input image belongs to the maxilla, mandible, or normal occlusion.

[0090] Among them, steps 1-4 belong to Figure 1The algorithm processing steps in the process.

[0091] After algorithmic processing, the teeth are divided into the following 53 categories: 32 permanent teeth (11-18, 21-28, 31-38, 41-48), 20 deciduous teeth (51-55, 61-65, 71-75, 81-85), and supernumerary teeth.

[0092] Steps 5-7 below belong to Figure 1 Post-processing steps, such as Figure 2 As shown.

[0093] Step 5: Perform non-maximum suppression processing on the preliminary segmentation results of each tooth instance to ensure that there is exactly one tooth for each instance.

[0094] The reason for choosing to use the mask of each tooth for NMS in this embodiment is that when using the detection frame of the tooth for NMS, there may be a situation where the IOU between the detection frames is relatively high, but the IOU between the actual masks is relatively low. This may filter out an extra tooth. However, using the mask directly for NMS can avoid this situation, ensuring that each tooth has exactly one mask.

[0095] Step 6: Based on the tooth centroid after non-maximum suppression processing, remove duplicate tooth positions from the preliminary tooth position identification results.

[0096] After NMS is completed, it can be confirmed that each tooth has one and only one appearance. However, different teeth may be labeled as the same tooth position. This problem can be solved by the following strategy: the centroid of the teeth with duplicate positions can be calculated, and the x and y coordinate values ​​of the centroid can be compared to make corrections until the prediction results no longer show duplicate tooth positions. Here, the origin of the coordinate system is the top left corner of each image. Starting from the top left corner of each image (0, 0), the x-coordinate gradually increases to the left, and the y-coordinate gradually increases downwards.

[0097] Step 7: Perform centroid sorting on the deduplicated tooth positions to obtain the final tooth position recognition and tooth segmentation results. The final tooth position recognition and tooth segmentation results for the maxillary image, mandibular image, and normal tooth occlusion image are shown below. Figure 3 , Figure 4 and Figure 5 As shown.

[0098] The centroid ordering of the deduplicated posterior teeth is determined by the category of the natural light oral image; wherein, the category of the natural light oral image includes maxillary image, mandibular image and normal tooth occlusion image.

[0099] If the category of the natural light oral image is a maxillary image, then the centroid coordinates of each tooth are arranged in a downward-opening parabola to determine the position of each tooth in the maxillary image.

[0100] If misalignment occurs in the tooth position recognition of the maxillary image, it can be corrected using the centroid coordinates of each tooth. Preliminary results of maxillary image tooth position recognition and tooth segmentation are as follows: Figure 6 As shown, the result after post-processing to correct the incorrect tooth position is as follows: Figure 7 As shown.

[0101] If the category of the natural light oral image is a mandibular image, then the centroid coordinates of each tooth are arranged in the order of an upward-opening parabola, thereby determining the position of each tooth in the mandibular image.

[0102] If misalignment occurs in the tooth position recognition of the mandibular image, it can be corrected using the centroid coordinates of each tooth. Preliminary results of mandibular image tooth position recognition and tooth segmentation are as follows: Figure 8 As shown, the result after post-processing to correct the incorrect tooth position is as follows: Figure 9 As shown.

[0103] If the category of the natural light oral image is a normal tooth occlusion image, then the vertical coordinate of the centroid of each tooth determines whether each tooth belongs to the maxilla or mandible. Then, the horizontal arrangement order of the maxilla and mandible teeth is determined by the ascending order of the horizontal coordinate of the centroid of each tooth.

[0104] If misalignment occurs in tooth positioning recognition of a normal occlusion image, it can be corrected using the centroid coordinates of each tooth. Preliminary results of tooth positioning recognition and tooth segmentation in a normal occlusion image are as follows: Figure 10 As shown, the result after post-processing to correct the incorrect tooth position is as follows: Figure 11 As shown.

[0105] Example 2

[0106] This embodiment provides a deep learning-based tooth position recognition and tooth segmentation system, which includes:

[0107] (1) Feature extraction module, which is used to extract global and local features of natural light oral cavity images to obtain feature maps of different sizes;

[0108] (2) Candidate region of interest generation module, which is used to fuse feature maps of all scales, generate feature pyramid, and determine all candidate regions of interest;

[0109] (3) Region of Interest Filtering and Normalization Module, which is used to filter and normalize the candidate regions of interest in sequence;

[0110] In the region of interest (ROI) filtering and normalization module, the process of sequentially filtering and normalizing candidate ROIs is as follows:

[0111] Binary classification and regression are performed on all candidate regions of interest to determine the regions of interest to be retained;

[0112] Based on the foreground score, the retained regions of interest are sorted, and then the non-maximum suppression method is used to filter out duplicate regions of interest. Finally, the size of the remaining regions of interest is normalized.

[0113] (4) Preliminary classification and segmentation result acquisition module, which is used to process the normalized region of interest through a pre-trained classification model and instance segmentation to obtain the preliminary tooth position identification result and the preliminary instance segmentation result for each tooth.

[0114] Specifically, in the preliminary classification and segmentation result acquisition module, the process of obtaining the preliminary tooth position identification result for each tooth is as follows:

[0115] The normalized region of interest is processed by a pre-trained classification model to obtain the category of the natural light oral cavity image; wherein, the category of the natural light oral cavity image includes maxillary image, mandibular image and normal tooth occlusion image;

[0116] Preliminary tooth position identification results were determined based on the category of natural light oral images.

[0117] (5) Non-maximum suppression processing module, which is used to perform non-maximum suppression processing on the preliminary results of instance segmentation of each tooth to ensure that there is one and only one tooth for each tooth;

[0118] (6) Duplicate tooth position removal module, which is used to remove duplicate tooth positions in the preliminary tooth position identification results based on the tooth centroid after non-maximum suppression processing;

[0119] (7) Centroid sorting module, which is used to sort the centroids of the deduplicated tooth positions to obtain the final tooth position recognition result and tooth segmentation result.

[0120] The centroid ordering of the deduplicated posterior teeth is determined by the category of the natural light oral cavity image; wherein the category of the natural light oral cavity image includes maxillary image, mandibular image and normal tooth occlusion image.

[0121] If the category of the natural light oral image is a maxillary image, then the centroid coordinates of each tooth are arranged in a downward-opening parabola to determine the position of each tooth in the maxillary image.

[0122] If the category of the natural light oral image is a mandibular image, then the centroid coordinates of each tooth are arranged in the order of an upward-opening parabola, thereby determining the position of each tooth in the mandibular image.

[0123] If the category of the natural light oral image is a normal tooth occlusion image, then the vertical coordinate of the centroid of each tooth determines whether each tooth belongs to the maxilla or mandible. Then, the horizontal arrangement order of the maxilla and mandible teeth is determined by the ascending order of the horizontal coordinate of the centroid of each tooth.

[0124] In practice, the centroid coordinates of each tooth are used to correct misalignment of tooth position identification.

[0125] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation processes are the same, so they will not be repeated here.

[0126] Example 3

[0127] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the deep learning-based tooth position recognition and tooth segmentation method described above.

[0128] Example 4

[0129] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the deep learning-based tooth position recognition and tooth segmentation method described above.

[0130] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as 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.

[0131] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for tooth position recognition and tooth segmentation based on deep learning, characterized in that, include: Global and local features of natural light oral images are extracted to obtain feature maps of different scales; The swin_transformer is used as the backbone to extract global and local features from natural light oral images. The swin_transformer introduces a sliding window mechanism to restrict attention calculation to one window, introduces the locality of convolution operation, saves computation, and allows the model to learn information across windows. At the same time, through downsampling layers, the receptive field is gradually increased, so that the attention mechanism can pay attention to global features. After the features are extracted by the backbone, feature maps of different sizes are obtained. By fusing feature maps at all scales, a feature pyramid is generated, and all candidate regions of interest are identified. The candidate regions of interest are filtered and normalized sequentially. The normalized regions of interest are processed by a pre-trained classification model and instance segmentation to obtain preliminary tooth position identification results and preliminary instance segmentation results for each tooth. Non-maximum suppression is applied to the initial segmentation results of each tooth to ensure that there is exactly one tooth for each instance. The non-maximum suppression process uses the instance segmentation mask of each tooth to perform non-maximum suppression, avoiding the problem of filtering out an extra tooth when using the detection boxes of teeth for non-maximum suppression because the IOU between detection boxes is high but the IOU between actual masks is low. Based on the tooth centroid after nonmaximum suppression processing, duplicate tooth positions in the preliminary tooth position identification results are removed; The deduplicated tooth positions are sorted by centroid to obtain the final tooth position recognition and tooth segmentation results; If the category of the natural light oral image is a maxillary image, then the centroid coordinates of each tooth are arranged in a downward-opening parabola, which determines the position of each tooth in the maxillary image. or If the category of the natural light oral image is a mandibular image, then the centroid coordinates of each tooth are arranged in the order of an upward-opening parabola, thereby determining the position of each tooth in the mandibular image. or If the category of the natural light oral image is a normal tooth occlusion image, then the centroid ordinate of each tooth is used to determine whether each tooth belongs to the maxilla or mandible, and the horizontal arrangement order of the maxilla and mandible teeth is determined by the order of the centroid abscissa of each tooth from small to large. The centroid coordinates of each tooth are used to correct tooth misalignment.

2. The tooth position recognition and tooth segmentation method based on deep learning as described in claim 1, characterized in that, The process of obtaining preliminary tooth position identification results for each tooth is as follows: The normalized region of interest is processed by a pre-trained classification model to obtain the category of the natural light oral cavity image; wherein, the category of the natural light oral cavity image includes maxillary image, mandibular image and normal tooth occlusion image; Preliminary tooth position identification results were determined based on the category of natural light oral images.

3. The tooth position recognition and tooth segmentation method based on deep learning as described in claim 1, characterized in that, The process of filtering and normalizing the candidate regions of interest sequentially is as follows: Binary classification and regression are performed on all candidate regions of interest to determine the regions of interest to be retained; Based on the foreground score, the retained regions of interest are sorted, and then the non-maximum suppression method is used to filter out duplicate regions of interest. Finally, the size of the remaining regions of interest is normalized.

4. The tooth position recognition and tooth segmentation method based on deep learning as described in claim 1, characterized in that, The centroid ordering of the deduplicated posterior teeth is determined by the category of the natural light oral image; wherein, the category of the natural light oral image includes maxillary image, mandibular image and normal tooth occlusion image.

5. A deep learning-based tooth position recognition and tooth segmentation system, employing the deep learning-based tooth position recognition and tooth segmentation method as described in any one of claims 1-4, characterized in that, include: The feature extraction module is used to extract global and local features from natural light oral cavity images, resulting in feature maps of different sizes. The candidate region of interest generation module is used to fuse feature maps of all scales, generate a feature pyramid, and determine all candidate regions of interest. The Region of Interest (ROI) filtering and normalization module is used to filter and normalize candidate ROIs sequentially. The preliminary classification and segmentation result acquisition module is used to process the normalized region of interest through a pre-trained classification model and instance segmentation to obtain the preliminary tooth position identification result and the preliminary instance segmentation result for each tooth. The non-maximum suppression processing module is used to perform non-maximum suppression processing on the preliminary results of each tooth instance segmentation to ensure that there is exactly one tooth for each tooth. The duplicate tooth position removal module is used to remove duplicate tooth positions in the preliminary tooth position identification results based on the tooth centroid after non-maximum suppression processing. The centroid sorting module is used to sort the centroids of the deduplicated tooth positions to obtain the final tooth position recognition result and tooth segmentation result.

6. The deep learning-based tooth position recognition and tooth segmentation system as described in claim 5, characterized in that, In the preliminary classification and segmentation result acquisition module, the process of obtaining the preliminary tooth position identification result for each tooth is as follows: The normalized region of interest is processed by a pre-trained classification model to obtain the category of the natural light oral cavity image; wherein, the category of the natural light oral cavity image includes maxillary image, mandibular image and normal tooth occlusion image; Preliminary tooth position identification results are determined based on the category of natural light oral images; or In the region of interest (ROI) filtering and normalization module, the process of filtering and normalizing the candidate ROIs sequentially is as follows: Binary classification and regression are performed on all candidate regions of interest to determine the regions of interest to be retained; Based on the foreground score, the retained regions of interest are sorted, and then the non-maximum suppression method is used to filter out duplicate regions of interest. The size of the remaining regions of interest is then normalized. or The centroid ordering of deduplicated posterior teeth is determined by the category of natural light oral images; or If the category of the natural light oral image is a maxillary image, then the centroid coordinates of each tooth are arranged in a downward-opening parabola, which determines the position of each tooth in the maxillary image. or If the category of the natural light oral image is a mandibular image, then the centroid coordinates of each tooth are arranged in the order of an upward-opening parabola, thereby determining the position of each tooth in the mandibular image. or If the category of the natural light oral image is a normal tooth occlusion image, then the centroid ordinate of each tooth is used to determine whether each tooth belongs to the maxilla or mandible, and the horizontal arrangement order of the maxilla and mandible teeth is determined by the order of the centroid abscissa of each tooth from small to large. or The centroid coordinates of each tooth are used to correct tooth misalignment.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the deep learning-based tooth position recognition and tooth segmentation method as described in any one of claims 1-4.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the deep learning-based tooth position recognition and tooth segmentation method as described in any one of claims 1-4.

Images