A method and device for detecting and identifying alveolar bone defects in maxillary anterior teeth

Abstract

This invention discloses a method and apparatus for detecting and identifying alveolar bone defects in maxillary anterior teeth. The method includes: acquiring a target image set, which includes CBCT images of maxillary anterior teeth in both sagittal and coronal sections; performing data augmentation processing on the target image set according to a data augmentation strategy to obtain data augmentation results; the data augmentation processing includes image stitching and data amplification based on a self-learning data augmentation strategy; and performing feature processing on the data augmentation results according to a determined target network model to obtain feature processing results, which include first state information indicating normal alveolar bone or second state information indicating the presence of fenestration and / or cracking in the alveolar bone; wherein the feature processing operations include feature extraction, downsampling, feature fusion, loss calculation, and data prediction. Therefore, implementing this invention can improve the adaptability and accuracy of detecting and identifying small regions of alveolar bone on digital CBCT images of teeth.

Description

Technical Field

[0001] This invention relates to the field of deep learning and data recognition technology, and in particular to a method and device for detecting and recognizing alveolar bone defects in maxillary anterior teeth. Background Technology

[0002] Currently, deep learning methods for identifying and detecting types of dental abnormalities are mainly based on dental X-rays, using Convolutional Neural Networks (CNNs) as the mainstream network. With the development and improvement of CNNs, various CNN architectures have been developed for the identification and detection of dental abnormalities. Historical applications include: VGG-19 network, 7-layer CNN architecture, Region-CNN (R-CNN) series networks, Single Shot Multi Box Detector (SSD), and U-Net. In addition, researchers have used Faster R-CNN for tooth numbering to assist in the detection and identification of images of dental defects. However, improving the accuracy of dental abnormality image identification and detection remains a key technical challenge for current research and improvement. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method and device for detecting and identifying alveolar bone defects in maxillary anterior teeth, which can improve the adaptability and accuracy of detecting and identifying small areas of alveolar bone on digital CBCT images of teeth.

[0004] To address the aforementioned technical problems, the first aspect of this invention discloses a method for detecting and identifying alveolar bone defects in the maxillary anterior teeth, the method comprising:

[0005] A set of target images to be processed is obtained, which includes at least one CBCT image corresponding to the sagittal and coronal sections of the maxillary anterior teeth; the set of target images is labeled with preset data.

[0006] According to a preset data augmentation strategy, data augmentation processing is performed on the target image set to obtain the data augmentation result corresponding to the target image set; the data augmentation processing includes image stitching and data amplification based on a self-learning data augmentation strategy.

[0007] Based on the determined target network model, a preset feature processing operation is performed on the data augmentation result to obtain the feature processing result corresponding to the data augmentation result. The feature processing result includes alveolar bone state information, which includes first state information indicating that the alveolar bone is normal or second state information indicating that the alveolar bone has fenestration and / or cracking.

[0008] The feature processing operations include feature extraction, downsampling, feature fusion, loss calculation, and data prediction, each with its corresponding operations.

[0009] As an optional implementation, in the first aspect of the present invention, the step of performing data augmentation processing on the target image set according to a preset data augmentation strategy to obtain a data augmentation result corresponding to the target image set includes:

[0010] According to a preset data augmentation strategy, image stitching is performed on the target image set to obtain the image stitching result of the target image set. The image stitching includes random stitching or similar stitching based on image similarity. The image similarity between the two or more target images in the similar stitching is within a preset similarity threshold.

[0011] Based on the determined self-learning data augmentation strategy, an algorithm search is performed on multiple predefined image augmentation operations to obtain the target augmentation algorithm and its corresponding augmentation parameters that are suitable for the image stitching result.

[0012] Then, based on the target enhancement algorithm and its corresponding enhancement parameters, an enhancement operation is performed on the image stitching result to obtain the enhanced result of the image stitching result, which serves as the data enhancement result corresponding to the target image set.

[0013] As an optional implementation, in the first aspect of the present invention, the self-learning data augmentation strategy includes a search space and a search algorithm;

[0014] The search space includes a first preset number of sub-strategies, each sub-strategy includes a second preset number of target strategy operations, and each target strategy operation in a single sub-strategy has its operation order, while two adjacent target strategy operations have different operation types; each target strategy operation is used for image enhancement.

[0015] Each of the target strategy operations has its corresponding operation probability and operation intensity; and the operation probability corresponding to each target strategy operation is a third preset number of first discrete values, and the operation intensity is a fourth preset number of second discrete values; and all the first discrete values ​​follow a uniform distribution, and all the second discrete values ​​follow a uniform distribution.

[0016] As an optional implementation, in the first aspect of the present invention, the step of performing an algorithm search from a predefined plurality of image enhancement operations according to a determined self-learning data augmentation strategy to obtain a target enhancement algorithm and its corresponding enhancement parameters that are suitable for the image stitching result includes:

[0017] For each sub-strategy, according to the search algorithm, all the first discrete values ​​and all the second discrete values ​​corresponding to the sub-strategy are searched and paired with the image stitching result to obtain the pairing set corresponding to the sub-strategy. The pairing set corresponding to the sub-strategy includes multiple pairing groups. Each pairing group corresponds to a set of first discrete values ​​and second discrete values, and each pairing group has a pairing value with the image stitching result.

[0018] From all the pairing sets and all the pairing groups they include, determine the pairing group with the highest pairing value, and denote it as the target pairing group;

[0019] The sub-strategy corresponding to the target pairing group is determined as the target enhancement algorithm that is compatible with the image stitching result, and the first discrete value and the second discrete value corresponding to the target pairing group are determined as the enhancement parameters corresponding to the target enhancement algorithm.

[0020] As an optional implementation, in the first aspect of the present invention, the target network model is an optimized YOLO v8 model; the target network model includes an improved Backbone network, an improved Neck network, and a Head network.

[0021] The Backbone network includes a fifth preset number of CSPModule modules and a layer of SPPF modules, and these two types of modules are cascaded in the Backbone network.

[0022] The Neck network adopts an AFPN progressive feature pyramid structure;

[0023] The step of performing a preset feature processing operation on the data augmentation result based on the determined target network model to obtain the feature processing result corresponding to the data augmentation result includes:

[0024] Based on all the CSPModule modules and the SPPF module, a first processing operation is performed on the input data augmentation result to obtain a first processing result corresponding to the data augmentation result; the first processing operation includes feature extraction and downsampling;

[0025] The second processing operation is performed on the first processing result according to the Neck network to obtain the second processing result corresponding to the first processing result. The second processing operation includes at least three layers of progressive feature fusion operation; and the feature fusion operation corresponding to different layers corresponds to different spatial weights.

[0026] The third processing operation is performed on the second processing result according to the Head network to obtain the third processing result corresponding to the second processing result, which is used as the feature processing result corresponding to the data augmentation result.

[0027] The third processing operation includes at least three operations: loss calculation, loss weighting calculation, and backpropagation optimization.

[0028] As an optional implementation, in the first aspect of the present invention, the data augmentation result outputs a result corresponding to each layer of the CSPModule module after passing through it, which is denoted as the first sub-result; the data augmentation result outputs a result corresponding to the last layer of the CSPModule module and the SPPF module, which is denoted as the second sub-result; the first processing result includes all the first sub-results and the second sub-results.

[0029] The step of performing a second processing operation on the first processing result according to the Neck network to obtain a second processing result corresponding to the first processing result includes:

[0030] The second sub-result is selected as the first feature of the Neck network; the first and second adjacent first sub-results are selected as the second and third features of the Neck network, respectively.

[0031] The second feature and the third feature are combined with the first spatial weight and the second spatial weight assigned to them in the Neck network and input into the feature pyramid corresponding to the Neck network. Then, the first feature and its third spatial weight assigned to it in the Neck network are input into the feature pyramid to obtain the multi-scale features corresponding to the first feature, the second feature and the third feature, which are used as the second processing result.

[0032] As an optional implementation, in the first aspect of the present invention, the Head network consists of decoupled classification branches and regression branches, wherein the classification branch includes a VFL loss function and the regression branch includes a DFL loss function and a CIoU loss function.

[0033] The step of performing a third processing operation on the second processing result based on the Head network to obtain a third processing result corresponding to the second processing result includes:

[0034] The second processing result is sequentially input into the DFL loss function, the CIoU loss function, and the VFL loss function to calculate the first loss value, the second loss value, and the third loss value in sequence.

[0035] Multiply the first loss value, the second loss value, and the third loss value by their respective weighting values ​​and sum them to obtain the total network loss of the Head network.

[0036] The total loss of the network is minimized using the backpropagation algorithm, resulting in multiple minimization results.

[0037] Select the minimized result from all the minimized results where the total network loss is within the preset target loss threshold, and use it as the third result.

[0038] A second aspect of this invention discloses a device for detecting and identifying alveolar bone defects in the maxillary anterior teeth, the device comprising:

[0039] The acquisition module is used to acquire a set of target images to be processed, which includes at least one CBCT image corresponding to the sagittal and coronal sections of the maxillary anterior teeth; the set of target images has been labeled with preset data.

[0040] The data augmentation processing module is used to perform data augmentation processing on the target image set according to a preset data augmentation strategy to obtain the data augmentation result corresponding to the target image set; the data augmentation processing includes image stitching and data amplification based on a self-learning data augmentation strategy;

[0041] The feature processing module is used to perform a preset feature processing operation on the data augmentation result according to the determined target network model to obtain the feature processing result corresponding to the data augmentation result. The feature processing result includes alveolar bone state information, which includes first state information indicating that the alveolar bone is normal or second state information indicating that the alveolar bone has fenestration and / or cracking.

[0042] The feature processing operations include feature extraction, downsampling, feature fusion, loss calculation, and data prediction, each with its corresponding operations.

[0043] As an optional implementation, in the second aspect of the present invention, the data augmentation processing module performs data augmentation processing on the target image set according to a preset data augmentation strategy to obtain the data augmentation result corresponding to the target image set, specifically including:

[0044] According to a preset data augmentation strategy, image stitching is performed on the target image set to obtain the image stitching result of the target image set. The image stitching includes random stitching or similar stitching based on image similarity. The image similarity between the two or more target images in the similar stitching is within a preset similarity threshold.

[0045] Based on the determined self-learning data augmentation strategy, an algorithm search is performed on multiple predefined image augmentation operations to obtain the target augmentation algorithm and its corresponding augmentation parameters that are suitable for the image stitching result.

[0046] Then, based on the target enhancement algorithm and its corresponding enhancement parameters, an enhancement operation is performed on the image stitching result to obtain the enhanced result of the image stitching result, which serves as the data enhancement result corresponding to the target image set.

[0047] As an optional implementation, in a second aspect of the present invention, the self-learning data augmentation strategy includes a search space and a search algorithm;

[0048] The search space includes a first preset number of sub-strategies, each sub-strategy includes a second preset number of target strategy operations, and each target strategy operation in a single sub-strategy has its operation order, while two adjacent target strategy operations have different operation types; each target strategy operation is used for image enhancement.

[0049] Each of the target strategy operations has its corresponding operation probability and operation intensity; and the operation probability corresponding to each target strategy operation is a third preset number of first discrete values, and the operation intensity is a fourth preset number of second discrete values; and all the first discrete values ​​follow a uniform distribution, and all the second discrete values ​​follow a uniform distribution.

[0050] As an optional implementation, in the second aspect of the present invention, the data augmentation processing module searches for a target augmentation algorithm and its corresponding augmentation parameters that are suitable for the image stitching result based on a determined self-learning data augmentation strategy, specifically including:

[0051] For each sub-strategy, according to the search algorithm, all the first discrete values ​​and all the second discrete values ​​corresponding to the sub-strategy are searched and paired with the image stitching result to obtain the pairing set corresponding to the sub-strategy. The pairing set corresponding to the sub-strategy includes multiple pairing groups. Each pairing group corresponds to a set of first discrete values ​​and second discrete values, and each pairing group has a pairing value with the image stitching result.

[0052] From all the pairing sets and all the pairing groups they include, determine the pairing group with the highest pairing value, and denote it as the target pairing group;

[0053] The sub-strategy corresponding to the target pairing group is determined as the target enhancement algorithm that is compatible with the image stitching result, and the first discrete value and the second discrete value corresponding to the target pairing group are determined as the enhancement parameters corresponding to the target enhancement algorithm.

[0054] As an optional implementation, in a second aspect of the invention, the target network model is an optimized YOLO v8 model; the target network model includes an improved Backbone network, an improved Neck network, and a Head network.

[0055] The Backbone network includes a fifth preset number of CSPModule modules and a layer of SPPF modules, and these two types of modules are cascaded in the Backbone network.

[0056] The Neck network adopts an AFPN progressive feature pyramid structure;

[0057] The feature processing module includes:

[0058] The first processing submodule is configured to perform a first processing operation on the input data augmentation result based on all the CSPModule modules and the SPPF module, to obtain a first processing result corresponding to the data augmentation result; the first processing operation includes feature extraction and downsampling;

[0059] The second processing submodule is used to perform a second processing operation on the first processing result according to the Neck network to obtain a second processing result corresponding to the first processing result. The second processing operation includes at least three layers of progressive feature fusion operations; and the feature fusion operations corresponding to different layers correspond to different spatial weights.

[0060] The third processing submodule is used to perform a third processing operation on the second processing result according to the Head network to obtain the third processing result corresponding to the second processing result, which is used as the feature processing result corresponding to the data augmentation result.

[0061] The third processing operation includes at least three operations: loss calculation, loss weighting calculation, and backpropagation optimization.

[0062] As an optional implementation, in the second aspect of the present invention, the data augmentation result outputs a result corresponding to each layer of the CSPModule module after passing through it, which is denoted as the first sub-result; the data augmentation result outputs a result corresponding to the last layer of the CSPModule module and the SPPF module, which is denoted as the second sub-result; the first processing result includes all the first sub-results and the second sub-results.

[0063] The second processing submodule performs a second processing operation on the first processing result based on the Neck network to obtain the second processing result corresponding to the first processing result. Specifically, this includes:

[0064] The second sub-result is selected as the first feature of the Neck network; the first and second adjacent first sub-results are selected as the second and third features of the Neck network, respectively.

[0065] The second feature and the third feature are combined with the first spatial weight and the second spatial weight assigned to them in the Neck network and input into the feature pyramid corresponding to the Neck network. Then, the first feature and its third spatial weight assigned to it in the Neck network are input into the feature pyramid to obtain the multi-scale features corresponding to the first feature, the second feature and the third feature, which are used as the second processing result.

[0066] As an optional implementation, in a second aspect of the present invention, the Head network consists of decoupled classification branches and regression branches, wherein the classification branch includes a VFL loss function and the regression branch includes a DFL loss function and a CIoU loss function.

[0067] The third processing submodule performs a third processing operation on the second processing result based on the Head network to obtain the third processing result corresponding to the second processing result. Specifically, the method for obtaining the third processing result corresponding to the second processing result includes:

[0068] The second processing result is sequentially input into the DFL loss function, the CIoU loss function, and the VFL loss function to calculate the first loss value, the second loss value, and the third loss value in sequence.

[0069] Multiply the first loss value, the second loss value, and the third loss value by their respective weighting values ​​and sum them to obtain the total network loss of the Head network.

[0070] The total loss of the network is minimized using the backpropagation algorithm, resulting in multiple minimization results.

[0071] Select the minimized result from all the minimized results where the total network loss is within the preset target loss threshold, and use it as the third result.

[0072] A third aspect of this invention discloses another device for detecting and identifying alveolar bone defects in the maxillary anterior teeth, the device comprising:

[0073] Memory containing executable program code;

[0074] A processor coupled to the memory;

[0075] The processor calls the executable program code stored in the memory to execute the method for detecting and identifying alveolar bone defects in the maxillary anterior teeth disclosed in the first aspect of the present invention.

[0076] The fourth aspect of the present invention discloses a computer storage medium storing computer instructions, which, when invoked, are used to execute the method for detecting and identifying alveolar bone defects in maxillary anterior teeth disclosed in the first aspect of the present invention.

[0077] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0078] This invention provides a method for detecting and identifying alveolar bone defects in maxillary anterior teeth. The method includes: acquiring a set of target images to be processed, the set including at least one CBCT image corresponding to both sagittal and coronal sections of the maxillary anterior teeth; the set of target images undergoing preset data annotation; performing data augmentation processing on the set of target images according to a preset data augmentation strategy to obtain a data augmentation result corresponding to the set of target images; the data augmentation processing includes image stitching and data amplification based on a self-learning data augmentation strategy; performing preset feature processing operations on the data augmentation result according to a determined target network model to obtain a feature processing result corresponding to the data augmentation result, the feature processing result including alveolar bone state information, the state information including first state information indicating normal alveolar bone or second state information indicating the presence of fenestration and / or cracking in the alveolar bone; wherein the feature processing operations include feature extraction, downsampling, feature fusion, loss calculation, and data prediction operations. As can be seen, after acquiring the target image set, the present invention enriches and expands the dataset through a data augmentation module (which adds image stitching and self-learning augmentation strategies to the existing augmentation methods), which is beneficial for improving the robustness of the target network model when inputting the dataset into the model for training. Furthermore, feature processing is performed on the optimized target network model to make it adaptable to digital CBCT images of teeth with higher spatial resolution. At the same time, through improvements and settings to the feature extraction and fusion module, the optimized target network model improves the adaptability, detection accuracy, and recognition accuracy when processing the detection and recognition of small changes in areas of tooth bone loss. Attached Figure Description

[0079] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0080] Figure 1 This is a flowchart illustrating a method for detecting and identifying alveolar bone defects in the maxillary anterior teeth, as disclosed in an embodiment of the present invention.

[0081] Figure 2 This is a flowchart illustrating another method for detecting and identifying alveolar bone defects in the maxillary anterior teeth disclosed in an embodiment of the present invention.

[0082] Figure 3 This is a schematic diagram of the structure of a device for detecting and identifying alveolar bone defects in maxillary anterior teeth disclosed in an embodiment of the present invention;

[0083] Figure 4 This is a schematic diagram of another maxillary anterior alveolar bone defect detection and identification device disclosed in an embodiment of the present invention;

[0084] Figure 5 This is a schematic diagram of the structure of another maxillary anterior alveolar bone defect detection and identification device disclosed in an embodiment of the present invention;

[0085] Figure 6 This is a flowchart illustrating another method for detecting and identifying alveolar bone defects in the maxillary anterior teeth disclosed in an embodiment of the present invention.

[0086] Figure 7 This is a schematic diagram of the structure of a backbone network for feature extraction disclosed in an embodiment of the present invention;

[0087] Figure 8 This is a comparative structural diagram of a Neck network before and after improvement, as disclosed in an embodiment of the present invention. Detailed Implementation

[0088] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0089] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0090] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0091] This invention discloses a method and apparatus for detecting and identifying alveolar bone defects in the maxillary anterior teeth. After acquiring the target image set, a data augmentation module (which adds image stitching and self-learning augmentation strategies to existing augmentation methods) is used to enrich and expand the dataset, which is beneficial for improving the robustness of the target network model when inputting the dataset into the model for training. Furthermore, the optimized target network model undergoes feature processing to adapt it to digital CBCT images of teeth with higher spatial resolution. Simultaneously, improvements and settings to the feature extraction and fusion modules enhance the adaptability, detection accuracy, and recognition accuracy of the optimized target network model when handling the detection and identification of subtle changes in tooth bone defects. These are described in detail below.

[0092] Example 1

[0093] Please see Figure 1 as well as Figure 6 , Figure 1 This is a flowchart illustrating a method for detecting and identifying alveolar bone defects in the maxillary anterior teeth, as disclosed in an embodiment of the present invention. Figure 6 This is a flowchart illustrating another method for detecting and identifying alveolar bone defects in the maxillary anterior teeth disclosed in an embodiment of the present invention. Figure 1 and Figure 6 The described method for detecting and identifying alveolar bone defects in the maxillary anterior teeth can be applied to devices for detecting and identifying alveolar bone defects in the maxillary anterior teeth, and the embodiments of the present invention are not limited thereto. Figure 1 As shown, the method for detecting and identifying alveolar bone defects in the maxillary anterior teeth may include the following operations:

[0094] 101. Obtain the target image set to be processed. The target image set includes at least one CBCT image corresponding to the sagittal and coronal sections of the maxillary anterior teeth.

[0095] In this embodiment of the invention, the target image set is labeled with preset data.

[0096] In this embodiment of the invention, after initially acquiring CBCT images corresponding to the sagittal and coronal sections of the maxillary anterior teeth, the acquired images are standardly annotated to obtain standard format detection data, and the annotated detection data is used as the target image set.

[0097] In this embodiment of the invention, it should be noted that, in the target image set, the CBCT images of the maxillary anterior teeth sagittal plane and the CBCT images corresponding to the coronal dual sections can be stored in a mixed manner or stored separately according to their classification. This embodiment of the invention does not limit this.

[0098] 102. According to the preset data augmentation strategy, perform data augmentation processing on the target image set to obtain the data augmentation result corresponding to the target image set.

[0099] In this embodiment of the invention, data augmentation processing includes image stitching and data amplification based on a self-learning data augmentation strategy.

[0100] In this embodiment of the invention, before directly inputting the target image set into the target network model, the target image set is subjected to data augmentation processing through this improved data augmentation strategy, which is beneficial to the subsequent input of the target network model into the target network model and improves the robustness of the target network model.

[0101] 103. Perform preset feature processing operations on the data augmentation results based on the determined target network model to obtain the feature processing results corresponding to the data augmentation results.

[0102] In this embodiment of the invention, the feature processing result includes alveolar bone state information, which includes first state information indicating that the alveolar bone is normal or second state information indicating that the alveolar bone has fenestration and / or cracking.

[0103] In this embodiment of the invention, the feature processing operations include feature extraction, downsampling, feature fusion, loss calculation, and data prediction, each with its corresponding operations.

[0104] It is evident that implementation Figure 1The described method for detecting and identifying alveolar bone defects in maxillary anterior teeth, after acquiring the target image set, enriches and expands the dataset through a data augmentation module (adding image stitching and self-learning augmentation strategies to existing augmentation methods). This enhances the robustness of the target network model when the dataset is input into the target network model for training. Furthermore, the optimized target network model undergoes feature processing to adapt to higher spatial resolution digital CBCT images of teeth. Simultaneously, improvements and settings to the feature extraction and fusion modules improve the adaptability, detection accuracy, and recognition accuracy of the optimized target network model when handling the detection and identification of subtle changes in tooth bone defects.

[0105] In an optional embodiment, step 102, which involves performing data augmentation processing on the target image set according to a preset data augmentation strategy to obtain the data augmentation result corresponding to the target image set, specifically includes the following methods:

[0106] According to the preset data augmentation strategy, image stitching is performed on the target image set to obtain the image stitching result of the target image set. Image stitching includes random stitching or similar stitching based on image similarity. The image similarity between two or more target images that are similarly stitched is within the preset similarity threshold.

[0107] Based on the determined self-learning data augmentation strategy, algorithm search is performed on multiple predefined image augmentation operations to obtain the target augmentation algorithm and its corresponding augmentation parameters that are suitable for the image stitching result.

[0108] Then, based on the target enhancement algorithm and its corresponding enhancement parameters, enhancement operations are performed on the image stitching results to obtain the enhanced results of the image stitching results, which are used as the data enhancement results corresponding to the target image set.

[0109] As can be seen, in this optional embodiment, before inputting the target image set into the subsequent target network model, the dataset is expanded and enriched through an improved data augmentation strategy. This is beneficial for the subsequent training and application of the target network model based on the expanded and enriched dataset, thereby improving the robustness of the target network model and making it more suitable for actual application scenarios.

[0110] In this optional embodiment, the self-learning data augmentation strategy includes a search space and a search algorithm;

[0111] The search space includes a first preset number of sub-strategies, each sub-strategy includes a second preset number of target strategy operations, and each target strategy operation in a single sub-strategy has its operation order, while two adjacent target strategy operations have different operation types; each target strategy operation is used for image enhancement.

[0112] Each target strategy operation has its corresponding operation probability and operation intensity; and the operation probability corresponding to each target strategy operation is a third preset number of first discrete values, and the operation intensity is a fourth preset number of second discrete values; and all first discrete values ​​follow a uniform distribution, and all second discrete values ​​follow a uniform distribution.

[0113] In this optional embodiment, specifically, after comprehensively considering uniform distribution, computational complexity, and model performance, the search space can include 5 sub-strategies; each sub-strategy includes 2 target strategy operations, which can specifically be 2 simple image enhancement operations, and the 2 push-enhancement operations have a corresponding sequential application order.

[0114] The pre-set image enhancement operations include: Mosaic, SnapMix, CutMix, Mixup, and TranslateX / Y; two of these five image enhancement operations are randomly selected, and the application order of each enhancement operation is set before being added to a sub-strategy.

[0115] Furthermore, for each image enhancement operation in each sub-strategy, the image enhancement operation corresponds to a default operation intensity range. Before applying the image enhancement operation, it is discretized into 11 values ​​following a uniform distribution (corresponding to a third preset number of first discrete values). This uniform distribution discretization setting is beneficial for subsequent use of discrete search algorithms. Similarly, each image enhancement operation corresponds to a default operation probability, which is also discretized into 10 values ​​following a uniform distribution (corresponding to a fourth preset number of second discrete values).

[0116] It should be noted that the choice of 11 or 10 discrete values ​​represents a trade-off between average distribution, computational complexity, and model performance. This choice provides sufficient flexibility and diversity while limiting computational overhead and searching within a reasonable range. However, the specific number and range of discrete values ​​are not limited to 11 or 10 and can be adjusted according to specific task requirements and available computing resources.

[0117] Finally, in practical applications, each image augmentation operation is treated as a time hyperparameter. Through the 21 discrete values ​​corresponding to the image augmentation operation, reinforcement learning is used to continuously try and optimize, thereby finding the optimal data augmentation strategy.

[0118] Furthermore, the method described above, which involves performing algorithm searches from multiple predefined image enhancement operations based on the determined self-learning data augmentation strategy to obtain the target augmentation algorithm and its corresponding augmentation parameters suitable for the image stitching result, specifically includes:

[0119] For each sub-strategy, according to the search algorithm, all the first discrete values ​​and all the second discrete values ​​corresponding to the sub-strategy are searched and paired with the image stitching result to obtain the pairing set corresponding to the sub-strategy. The pairing set corresponding to the sub-strategy includes multiple pairing groups. Each pairing group corresponds to a set of first discrete values ​​and second discrete values, and each pairing group has its pairing value with the image stitching result.

[0120] From all the pairing sets and all the pairing groups they include, determine the pairing group with the highest pairing value, and denote it as the target pairing group;

[0121] The sub-strategy corresponding to the target pairing group is determined as the target enhancement algorithm that is compatible with the image stitching result, and the first discrete value and the second discrete value corresponding to the target pairing group are determined as the enhancement parameters corresponding to the target enhancement algorithm.

[0122] As can be seen, in this optional embodiment, the intelligent learning optimization of the data augmentation strategy is achieved by combining the search space (including sub-strategies and their uniformly discretized operation strengths and operation probabilities) with the search algorithm, thereby obtaining the optimal data augmentation strategy. This improves the accuracy and reliability of the determination of the data augmentation strategy, which is beneficial for improving the robustness of the target network model when the dataset augmented by the data augmentation strategy is input into the target network model for training.

[0123] Example 2

[0124] Please see Figure 2 , Figure 2 This is a flowchart illustrating another method for detecting and identifying alveolar bone defects in the maxillary anterior teeth disclosed in an embodiment of the present invention. Figure 2 The described method for detecting and identifying alveolar bone defects in the maxillary anterior teeth can be applied to devices for detecting and identifying alveolar bone defects in the maxillary anterior teeth, and the embodiments of the present invention are not limited thereto. Figure 2 As shown, the method for detecting and identifying alveolar bone defects in the maxillary anterior teeth may include the following operations:

[0125] 201. Obtain the target image set to be processed. The target image set includes at least one CBCT image corresponding to the sagittal and coronal sections of the maxillary anterior teeth.

[0126] 202. According to the preset data augmentation strategy, perform data augmentation processing on the target image set to obtain the data augmentation result corresponding to the target image set.

[0127] In this embodiment of the invention, the target network model is an optimized YOLO v8 model; the target network model includes an improved Backbone network, an improved Neck network, and a Head network;

[0128] The Backbone network includes a fifth preset number of CSPModule modules and a first layer of SPPF modules, and these two types of modules are cascaded in the Backbone network.

[0129] In this embodiment of the invention, specifically for the improved Backbone network, please refer to [link to relevant documentation]. Figure 7 , Figure 7 This is a schematic diagram of the structure of a backbone network for feature extraction disclosed in an embodiment of the present invention; as shown below. Figure 7 As shown, the backbone network can consist of a cascaded structure of four CSPModule layers and one SPPF layer. The four CSPModule layers are the basic building blocks of the backbone network, employing the CSPDarkNet-53 network structure. Their function is to perform feature extraction and downsampling to obtain higher-level feature representations. In practical applications, each CSPModule layer is specifically configured as follows: the kernel size of the first convolutional layer is set to 3, and the stride is 2.

[0130] The number of layers for the CSPModule can be increased or decreased depending on the actual application, but the number of layers should not be less than 3.

[0131] In this embodiment of the invention, the SPPF module is a module of the backbone network, consisting of three MaxPools connected in series. The purpose of this module is to introduce the concept of Spatial Pyramid Pooling to obtain feature representations at different scales (including shallow, medium, and deep features). In practical applications, the size of the convolutional kernel is set to 5*5.

[0132] In this embodiment of the invention, the improved Neck network adopts an AFPN progressive feature pyramid structure.

[0133] For further descriptions of steps 201-202 in this embodiment of the invention, please refer to the other specific descriptions of steps 101-102 in Embodiment 1. These descriptions will not be repeated in this embodiment of the invention.

[0134] 203. Based on all CSPModule modules and SPPF modules, perform the first processing operation on the input data augmentation result to obtain the first processing result corresponding to the data augmentation result; the first processing operation includes feature extraction and downsampling.

[0135] 204. Perform a second processing operation on the first processing result according to the Neck network to obtain the second processing result corresponding to the first processing result. The second processing operation includes at least three progressive feature fusion operations; and the feature fusion operations corresponding to different levels correspond to different spatial weights.

[0136] 205. Perform a third processing operation on the second processing result based on the Head network to obtain the third processing result corresponding to the second processing result, which is used as the feature processing result corresponding to the data augmentation result.

[0137] In this embodiment of the invention, the third processing operation includes at least three operations: loss calculation, loss weighting calculation, and backpropagation optimization.

[0138] It is evident that implementation Figure 2 The described method for detecting and identifying alveolar bone defects in the maxillary anterior teeth utilizes an improved YOLO v8 model (specifically including an improved Backbone network, Neck network, and Head network) that incorporates the concept of spatial pyramid pooling. The Backbone network captures visual information at different levels, transforming the input image into a more representative feature representation. This representation is then fed into the Neck fusion network for further processing and fusion, enhancing the expressive power and discriminative power of the features. Finally, in the prediction stage, the Head network is used for loss calculation, weighted processing, backpropagation optimization, and other techniques to improve prediction accuracy and robustness.

[0139] In an optional embodiment, such as Figure 7 As shown, for each layer of CSPModule that the data augmentation result passes through, the output result corresponding to that layer of CSPModule is output, which is denoted as the first sub-result; the output result corresponding to the last layer of CSPModule and SPPF module is denoted as the second sub-result; the first processing result includes all the first sub-results and the second sub-results.

[0140] Step 204 involves performing a second processing operation on the first processing result based on the Neck network to obtain the second processing result corresponding to the first processing result. The specific methods include:

[0141] The second sub-result is selected as the first feature of the Neck network; the first and second adjacent first sub-results are selected as the second and third features of the Neck network, respectively.

[0142] The second and third features are combined with their first and second spatial weights in the Neck network and input into the feature pyramid corresponding to the Neck network. Then, the first feature and its third spatial weight in the Neck network are input into the feature pyramid to obtain the multi-scale features corresponding to the first, second, and third features, which are used as the second processing result.

[0143] In this optional embodiment, specifically, please refer to... Figure 8 , Figure 8 This is a comparative structural diagram of a Neck network before and after improvement, as disclosed in an embodiment of the present invention. Figure 8 As shown, when applying this Neck feature fusion network, the last layer features of the last three CSPModule modules are extracted from all feature layers of the Backbone feature extraction backbone network, thus generating a set of features at different scales (shallow, mid-level, and deep features), denoted as {C3, C4, C5}. To perform deep feature fusion on features at different scales, the low-level and mid-level features {C3, C4} are first input into the feature pyramid network, and then the high-level feature C5 is added. After the feature fusion step, a set of multi-scale features is generated as {P3, P4, P5}.

[0144] In this optional embodiment, the architecture of AFPN is as follows: Figure 8 As shown, during the bottom-up feature extraction process of the Backbone network, AFPN progressively integrates shallow, mid-level, and deep features. Furthermore, in the multi-level feature fusion process, features from different levels are assigned different spatial weights using a proportional allocation method (shallow:mid:deep = 5:3:2) to enhance the importance of key levels and mitigate the influence of conflicting information from different objectives. The spatial weight allocation for each level can also be adjusted according to actual applications, and this embodiment of the invention is not limited to this.

[0145] As can be seen, in this optional embodiment, the improved Neck feature fusion network introduces a progressively hierarchical fusion structure and assigns different spatial weights to features at different levels. This helps to enhance the importance of key levels, reduce the influence of contradictory information from different targets, and improve the fusion accuracy and reliability of the improved Neck feature fusion network.

[0146] In another alternative embodiment, the Head network consists of decoupled classification and regression branches, with the classification branch including the VFL loss function and the regression branch including the DFL loss function and the CIoU loss function.

[0147] The head network employs a decoupled structure, extracting category and location features through two parallel branches (classification and regression branches), respectively. Each branch then uses a 1×1 convolutional layer to complete the classification and localization tasks. The classification branch calculates the VFL loss function, while the regression branch calculates the DFL and CIoU loss functions.

[0148] In this optional embodiment, step 205, which involves performing a third processing operation on the second processing result based on the Head network to obtain the third processing result corresponding to the second processing result, specifically includes:

[0149] The second processing result is input into the DFL loss function, CIoU loss function, and VFL loss function in sequence to calculate the first loss value, the second loss value, and the third loss value in sequence.

[0150] Multiply the first loss value, the second loss value, and the third loss value by their respective weighted values ​​and sum them up to obtain the total network loss of the Head network.

[0151] By using the backpropagation algorithm, the total network loss is minimized, resulting in multiple minimization outcomes.

[0152] Select the minimized result from all minimized results where the total network loss is within the preset target loss threshold, and use it as the third result.

[0153] In this optional embodiment, each loss function provides different constraints to guide the network's learning process. The DFL and CIoU loss functions constrain the regression branch, enabling it to accurately predict the position and shape of the target bounding box, while the VFL loss function constrains the classification branch, enabling it to accurately predict the target's category. Thus, the entire network can simultaneously optimize classification and regression tasks during training to improve prediction accuracy and robustness. The DFL loss models the location of the detection boxes as a general distribution, ensuring the network quickly focuses on the distribution of locations closest to the target. The specific calculation formula is shown in equation (1):

[0154] DFL(S i ,S i+1 )=-((y i+1 -y)log(S i )+(yy i (S) i+1 )) (1)

[0155] Among them, S i S i+1This represents the general distribution of the regression branch output, which can also be understood as the network's prediction of the target location. i and y i+1 This indicates that it corresponds to S i S i+1 The true value, that is, the true distribution of the target location.

[0156] The CIoU loss function is used for bounding box regression. It further considers the distance between bounding boxes and the compactness of overlapping portions based on IoU. Its calculation formula is shown in equation (2):

[0157]

[0158] Where IoU represents the intersection-union ratio. gt represents the predicted bounding box and the ground truth bounding box, respectively. represents the diagonal length of their overlapping portion. c represents the normalization factor. α represents the balance coefficient. αv represents the penalty term for non-overlapping bounding boxes.

[0159] The VFL loss function is used for loss calculation in classification tasks. It introduces variance to reduce the weight of easily classified samples, thereby increasing attention to difficult samples. Its calculation formula is shown in equation (3):

[0160]

[0161] Where p represents the model's predicted probability that a sample belongs to the positive class, q is the output of the classification branch, and represents the true label of the sample. α is the positive-negative sample balancing coefficient, used to adjust the weights between positive and negative samples, controlling the model's attention to samples of different classes. γ is the easy-difficulty sample balancing coefficient, used to adjust the weights between easy and difficult samples.

[0162] As can be seen, in this optional embodiment, the total loss of the target network model is composed of the weighted sum of the three loss functions mentioned above, and the backpropagation algorithm is used to minimize the loss function, thereby adjusting the network weights and parameters to make the prediction results as close as possible to the true target category and bounding box. Finally, through continuous iterative training, the network can gradually improve its prediction performance and identify areas of slight change and determine the state of the alveolar bone (normal, alveolar bone fenestration, and crack) during the prediction stage; that is, through the setting of this Head network, the final data prediction and recognition are processed, which helps to improve the detection accuracy and recognition accuracy of areas of slight change in the alveolar bone.

[0163] Example 3

[0164] Please see Figure 3 , Figure 3This is a schematic diagram of a device for detecting and identifying alveolar bone defects in the maxillary anterior teeth, as disclosed in an embodiment of the present invention. The device can be a terminal, equipment, system, or server for detecting and identifying alveolar bone defects in the maxillary anterior teeth. The server can be a local server, a remote server, or a cloud server (also known as a cloud-based server). When the server is not a cloud server, it can communicate with the cloud server; this embodiment of the present invention does not impose any limitations. Figure 3 As shown, the maxillary anterior alveolar bone defect detection and identification device may include an acquisition module 301, a data augmentation processing module 302, and a feature processing module 303, wherein:

[0165] The acquisition module 301 is used to acquire the target image set to be processed. The target image set includes at least one CBCT image corresponding to the sagittal and coronal sections of the maxillary anterior teeth. The target image set has been labeled with preset data.

[0166] The data augmentation processing module 302 is used to perform data augmentation processing on the target image set according to a preset data augmentation strategy, and obtain the data augmentation result corresponding to the target image set. The data augmentation processing includes image stitching and data amplification based on a self-learning data augmentation strategy.

[0167] Feature processing module 303 is used to perform preset feature processing operations on the data augmentation results according to the determined target network model to obtain the feature processing results corresponding to the data augmentation results. The feature processing results include alveolar bone state information, which includes first state information indicating that the alveolar bone is normal or second state information indicating that the alveolar bone has fenestration and / or cracking.

[0168] The feature processing operations include feature extraction, downsampling, feature fusion, loss calculation, and data prediction, each with its corresponding operations.

[0169] It is evident that implementation Figure 3 The described device for detecting and recognizing alveolar bone defects in maxillary anterior teeth, after acquiring the target image set, enriches and expands the dataset through a data augmentation module (adding image stitching and self-learning enhancement strategies to existing augmentation methods). This enhances the robustness of the target network model when the dataset is input into the target network model for training. Furthermore, the optimized target network model undergoes feature processing to adapt to higher spatial resolution digital CBCT images of teeth. Simultaneously, improvements and settings to the feature extraction and fusion module enhance the adaptability, detection accuracy, and recognition accuracy of the optimized target network model when detecting and recognizing subtle changes in areas of tooth bone loss.

[0170] In an optional embodiment, the data augmentation processing module 302 performs data augmentation processing on the target image set according to a preset data augmentation strategy, and the specific method for obtaining the data augmentation result corresponding to the target image set includes:

[0171] According to the preset data augmentation strategy, image stitching is performed on the target image set to obtain the image stitching result of the target image set. Image stitching includes random stitching or similar stitching based on image similarity. The image similarity between two or more target images that are similarly stitched is within the preset similarity threshold.

[0172] Based on the determined self-learning data augmentation strategy, algorithm search is performed on multiple predefined image augmentation operations to obtain the target augmentation algorithm and its corresponding augmentation parameters that are suitable for the image stitching result.

[0173] Then, based on the target enhancement algorithm and its corresponding enhancement parameters, enhancement operations are performed on the image stitching results to obtain the enhanced results of the image stitching results, which are used as the data enhancement results corresponding to the target image set.

[0174] As can be seen, in this optional embodiment, before inputting the target image set into the subsequent target network model, the dataset is expanded and enriched through an improved data augmentation strategy. This is beneficial for the subsequent training and application of the target network model based on the expanded and enriched dataset, thereby improving the robustness of the target network model and making it more suitable for actual application scenarios.

[0175] In another alternative embodiment, the self-learning data augmentation strategy includes a search space and a search algorithm;

[0176] The search space includes a first preset number of sub-strategies, each sub-strategy includes a second preset number of target strategy operations, and each target strategy operation in a single sub-strategy has its operation order, while two adjacent target strategy operations have different operation types; each target strategy operation is used for image enhancement.

[0177] Each target strategy operation has its corresponding operation probability and operation intensity; and the operation probability corresponding to each target strategy operation is a third preset number of first discrete values, and the operation intensity is a fourth preset number of second discrete values; and all first discrete values ​​follow a uniform distribution, and all second discrete values ​​follow a uniform distribution.

[0178] In this optional embodiment, the data augmentation processing module 302 further performs algorithm search from a predefined set of image augmentation operations based on the determined self-learning data augmentation strategy to obtain the target augmentation algorithm and its corresponding augmentation parameters that are suitable for the image stitching result. Specifically, this includes:

[0179] For each sub-strategy, according to the search algorithm, all the first discrete values ​​and all the second discrete values ​​corresponding to the sub-strategy are searched and paired with the image stitching result to obtain the pairing set corresponding to the sub-strategy. The pairing set corresponding to the sub-strategy includes multiple pairing groups. Each pairing group corresponds to a set of first discrete values ​​and second discrete values, and each pairing group has its pairing value with the image stitching result.

[0180] From all the pairing sets and all the pairing groups they include, determine the pairing group with the highest pairing value, and denote it as the target pairing group;

[0181] The sub-strategy corresponding to the target pairing group is determined as the target enhancement algorithm that is compatible with the image stitching result, and the first discrete value and the second discrete value corresponding to the target pairing group are determined as the enhancement parameters corresponding to the target enhancement algorithm.

[0182] As can be seen, in this optional embodiment, the intelligent learning optimization of the data augmentation strategy is achieved by combining the search space (including sub-strategies and their uniformly discretized operation strengths and operation probabilities) with the search algorithm, thereby obtaining the optimal data augmentation strategy. This improves the accuracy and reliability of the determination of the data augmentation strategy, which is beneficial for improving the robustness of the target network model when the dataset augmented by the data augmentation strategy is input into the target network model for training.

[0183] In another alternative embodiment, the target network model is an optimized YOLO v8 model; the target network model includes an improved Backbone network, an improved Neck network, and a Head network.

[0184] The Backbone network includes a fifth preset number of CSPModule modules and a first layer of SPPF modules, and these two types of modules are cascaded in the Backbone network.

[0185] The Neck network employs an AFPN progressive feature pyramid structure;

[0186] Feature processing module 303 includes:

[0187] The first processing submodule 3031 is used to perform a first processing operation on the input data augmentation result based on all CSPModule modules and SPPF modules to obtain the first processing result corresponding to the data augmentation result; the first processing operation includes feature extraction and downsampling;

[0188] The second processing submodule 3032 is used to perform a second processing operation on the first processing result according to the Neck network to obtain a second processing result corresponding to the first processing result. The second processing operation includes at least three layers of progressive feature fusion operations; and the feature fusion operations corresponding to different layers correspond to different spatial weights.

[0189] The third processing submodule 3033 is used to perform a third processing operation on the second processing result according to the Head network to obtain the third processing result corresponding to the second processing result, which is used as the feature processing result corresponding to the data augmentation result.

[0190] The third processing operation includes at least three operations: loss calculation, loss weighting calculation, and backpropagation optimization.

[0191] As can be seen, in this optional embodiment, the improved YOLO v8 model (specifically including the improved Backbone network, Neck network, and Head network) incorporates the idea of ​​spatial pyramid pooling. The Backbone network is used to capture visual information at different levels, thereby converting the input image into a more representative feature representation. This representation is then fed into the Neck fusion network for further processing and fusion, which helps to improve the expressive power and discriminative power of the features. Finally, in the prediction stage, the Head network is used to perform corresponding loss calculations, weighted processing, backpropagation optimization, etc., to improve prediction accuracy and robustness.

[0192] In another optional embodiment, each time the data augmentation result passes through a CSPModule, the output result corresponding to that CSPModule is output, which is denoted as the first sub-result; the output result of the data augmentation result after passing through the last CSPModule and the SPPF module is denoted as the second sub-result; the first processing result includes all the first sub-results and the second sub-results.

[0193] The second processing submodule 3032 performs a second processing operation on the first processing result based on the Neck network to obtain the second processing result corresponding to the first processing result. Specifically, the methods for obtaining the second processing result corresponding to the first processing result include:

[0194] The second sub-result is selected as the first feature of the Neck network; the first and second adjacent first sub-results are selected as the second and third features of the Neck network, respectively.

[0195] The second and third features are combined with their first and second spatial weights in the Neck network and input into the feature pyramid corresponding to the Neck network. Then, the first feature and its third spatial weight in the Neck network are input into the feature pyramid to obtain the multi-scale features corresponding to the first, second, and third features, which are used as the second processing result.

[0196] As can be seen, in this optional embodiment, the improved Neck feature fusion network introduces a progressively hierarchical fusion structure and assigns different spatial weights to features at different levels. This helps to enhance the importance of key levels, reduce the influence of contradictory information from different targets, and improve the fusion accuracy and reliability of the improved Neck feature fusion network.

[0197] In another alternative embodiment, the Head network consists of decoupled classification and regression branches, with the classification branch including the VFL loss function and the regression branch including the DFL loss function and the CIoU loss function.

[0198] The third processing submodule 3033 performs a third processing operation on the second processing result based on the Head network, and the specific methods for obtaining the third processing result corresponding to the second processing result include:

[0199] The second processing result is input into the DFL loss function, CIoU loss function, and VFL loss function in sequence to calculate the first loss value, the second loss value, and the third loss value in sequence.

[0200] Multiply the first loss value, the second loss value, and the third loss value by their respective weighted values ​​and sum them up to obtain the total network loss of the Head network.

[0201] By using the backpropagation algorithm, the total network loss is minimized, resulting in multiple minimization outcomes.

[0202] Select the minimized result from all minimized results where the total network loss is within the preset target loss threshold, and use it as the third result.

[0203] As can be seen, in this optional embodiment, the total loss of the target network model is composed of the weighted sum of the three loss functions mentioned above, and the backpropagation algorithm is used to minimize the loss function, thereby adjusting the network weights and parameters to make the prediction results as close as possible to the true target category and bounding box. Finally, through continuous iterative training, the network can gradually improve its prediction performance and identify areas of slight change and determine the state of the alveolar bone (normal, alveolar bone fenestration, and crack) during the prediction stage; that is, through the setting of this Head network, the final data prediction and recognition are processed, which helps to improve the detection accuracy and recognition accuracy of areas of slight change in the alveolar bone.

[0204] Example 4

[0205] Please see Figure 5 , Figure 5 This is a schematic diagram of the structure of another maxillary anterior tooth alveolar bone defect detection and identification device disclosed in an embodiment of the present invention. Figure 5As shown, the device for detecting and identifying alveolar bone defects in the maxillary anterior teeth may include:

[0206] Memory 401 storing executable program code;

[0207] Processor 402 coupled to memory 401;

[0208] The processor 402 calls the executable program code stored in the memory 401 to execute the steps in the method for detecting and identifying alveolar bone defects in maxillary anterior teeth as described in Embodiment 1 or Embodiment 2 of the present invention.

[0209] Example 5

[0210] This invention discloses a computer storage medium storing computer instructions. When these computer instructions are invoked, they are used to execute the steps in the method for detecting and identifying alveolar bone defects in the maxillary anterior teeth described in Embodiment 1 or Embodiment 2 of this invention.

[0211] Example 6

[0212] This invention discloses a computer program product, which includes a non-transitory computer storage medium storing a computer program, and the computer program is operable to cause a computer to perform the steps in the maxillary anterior alveolar bone defect detection and identification method described in Embodiment 1 or Embodiment 2.

[0213] The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0214] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0215] Finally, it should be noted that the method and apparatus for detecting and identifying alveolar bone defects in maxillary anterior teeth disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention and are only used to illustrate the technical solutions of the present invention, not to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting and identifying alveolar bone defects in the maxillary anterior teeth, characterized in that, The method includes: A set of target images to be processed is obtained, which includes at least one CBCT image corresponding to the sagittal and coronal sections of the maxillary anterior teeth; the set of target images is labeled with preset data. According to a preset data augmentation strategy, image stitching is performed on the target image set to obtain the image stitching result of the target image set. The image stitching includes random stitching or similar stitching based on image similarity. The image similarity between the two or more target images in the similar stitching is within a preset similarity threshold. The predetermined self-learning data augmentation strategy includes a search space and a search algorithm, and the search space includes a first preset number of sub-strategies; for each sub-strategy included in the search space, according to the search algorithm, all first discrete values ​​and all second discrete values ​​corresponding to the sub-strategy are searched and paired with the image stitching result to obtain a pairing set corresponding to the sub-strategy. The pairing set corresponding to the sub-strategy includes multiple pairing groups, each pairing group corresponds to a set of first discrete values ​​and second discrete values, and each pairing group has its pairing value with the image stitching result; From all the pairing sets and all the pairing groups they include, determine the pairing group with the highest pairing value, and denote it as the target pairing group; The sub-strategy corresponding to the target pairing group is determined as the target enhancement algorithm that is compatible with the image stitching result, and the first discrete value and the second discrete value corresponding to the target pairing group are determined as the enhancement parameters corresponding to the target enhancement algorithm; Then, based on the target enhancement algorithm and its corresponding enhancement parameters, an enhancement operation is performed on the image stitching result to obtain the enhanced result of the image stitching result, which is used as the data enhancement result corresponding to the target image set; Based on the determined target network model, a preset feature processing operation is performed on the data augmentation result to obtain the feature processing result corresponding to the data augmentation result. The feature processing result includes alveolar bone state information, which includes first state information indicating that the alveolar bone is normal or second state information indicating that the alveolar bone has fenestration and / or cracking. The feature processing operations include feature extraction, downsampling, feature fusion, loss calculation, and data prediction, each with its corresponding operations.

2. The method for detecting and identifying alveolar bone defects in the maxillary anterior teeth according to claim 1, characterized in that, Each sub-strategy includes a second preset number of target strategy operations, and each target strategy operation in a single sub-strategy has its operation order, while two adjacent target strategy operations have different operation types; each target strategy operation is used for image enhancement; Each of the target strategy operations has its corresponding operation probability and operation intensity; and the operation probability corresponding to each target strategy operation is a third preset number of first discrete values, and the operation intensity is a fourth preset number of second discrete values; and all the first discrete values ​​follow a uniform distribution, and all the second discrete values ​​follow a uniform distribution.

3. The method for detecting and identifying alveolar bone defects in the maxillary anterior teeth according to claim 1, characterized in that, The target network model is an optimized YOLO v8 model; the target network model includes an improved Backbone network, an improved Neck network, and a Head network; The Backbone network includes a fifth preset number of CSPModule modules and a layer of SPPF modules, and these two types of modules are cascaded in the Backbone network. The Neck network adopts an AFPN progressive feature pyramid structure; The step of performing a preset feature processing operation on the data augmentation result based on the determined target network model to obtain the feature processing result corresponding to the data augmentation result includes: Based on all the CSPModule modules and the SPPF module, a first processing operation is performed on the input data augmentation result to obtain a first processing result corresponding to the data augmentation result; the first processing operation includes feature extraction and downsampling; The second processing operation is performed on the first processing result according to the Neck network to obtain the second processing result corresponding to the first processing result. The second processing operation includes at least three layers of progressive feature fusion operation; and the feature fusion operation corresponding to different layers corresponds to different spatial weights. The third processing operation is performed on the second processing result according to the Head network to obtain the third processing result corresponding to the second processing result, which is used as the feature processing result corresponding to the data augmentation result. The third processing operation includes at least three operations: loss calculation, loss weighting calculation, and backpropagation optimization.

4. The method for detecting and identifying alveolar bone defects in the maxillary anterior teeth according to claim 3, characterized in that, Each time the data augmentation result passes through a CSPModule, it outputs the corresponding output result of that CSPModule, which is denoted as the first sub-result; the output result of the data augmentation result after passing through the last CSPModule and the SPPF module is denoted as the second sub-result; the first processing result includes all the first sub-results and the second sub-results. The step of performing a second processing operation on the first processing result according to the Neck network to obtain a second processing result corresponding to the first processing result includes: The second sub-result is selected as the first feature of the Neck network; The first and second adjacent first sub-results of the second sub-result are selected as the second and third features of the Neck network; The second feature and the third feature are combined with the first spatial weight and the second spatial weight assigned to them in the Neck network and input into the feature pyramid corresponding to the Neck network. Then, the first feature and its third spatial weight assigned to it in the Neck network are input into the feature pyramid to obtain the multi-scale features corresponding to the first feature, the second feature and the third feature, which are used as the second processing result.

5. The method for detecting and identifying alveolar bone defects in the maxillary anterior teeth according to claim 3 or 4, characterized in that, The Head network consists of a decoupled classification branch and a regression branch. The classification branch includes the VFL loss function, and the regression branch includes the DFL loss function and the CIoU loss function. The step of performing a third processing operation on the second processing result based on the Head network to obtain a third processing result corresponding to the second processing result includes: The second processing result is sequentially input into the DFL loss function, the CIoU loss function, and the VFL loss function to calculate the first loss value, the second loss value, and the third loss value in sequence. Multiply the first loss value, the second loss value, and the third loss value by their respective weighting values ​​and sum them to obtain the total network loss of the Head network. The total loss of the network is minimized using the backpropagation algorithm, resulting in multiple minimization results. Select the minimized result from all the minimized results where the total network loss is within the preset target loss threshold, and use it as the third result.

6. A device for detecting and identifying alveolar bone defects in maxillary anterior teeth, characterized in that, The device is used to perform the method for detecting and identifying alveolar bone defects in the maxillary anterior teeth as described in any one of claims 1-5, and the device comprises: The acquisition module is used to acquire a set of target images to be processed, which includes at least one CBCT image corresponding to the sagittal and coronal sections of the maxillary anterior teeth; the set of target images has been labeled with preset data. The data augmentation processing module is used to perform data augmentation processing on the target image set according to a preset data augmentation strategy to obtain the data augmentation result corresponding to the target image set; the data augmentation processing includes image stitching and data amplification based on a self-learning data augmentation strategy; The feature processing module is used to perform a preset feature processing operation on the data augmentation result according to the determined target network model to obtain the feature processing result corresponding to the data augmentation result. The feature processing result includes alveolar bone state information, which includes first state information indicating that the alveolar bone is normal or second state information indicating that the alveolar bone has fenestration and / or cracking. The feature processing operations include feature extraction, downsampling, feature fusion, loss calculation, and data prediction, each with its corresponding operations.

7. A device for detecting and identifying alveolar bone defects in the maxillary anterior teeth, characterized in that, The device includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the method for detecting and identifying alveolar bone defects in maxillary anterior teeth as described in any one of claims 1-5.

8. A computer storage medium, characterized in that, The computer storage medium stores computer instructions, which, when invoked, are used to execute the method for detecting and identifying alveolar bone defects in the maxillary anterior teeth as described in any one of claims 1-5.

Images