A tooth malformation classification method, device and medium based on each granularity

Abstract

The application relates to the technical field of image processing, and discloses a tooth deformity classification method and device based on different granularities and a medium, which comprises the following steps: acquiring multi-view intraoral images of a current sample, wherein the multi-view intraoral images at least include a front occlusion image, a left occlusion image and a right occlusion image; performing tooth region detection on the multi-view intraoral images so as to locate a front tooth key region used for analyzing the front tooth occlusion relationship; based on the front tooth key region, extracting local features reflecting the local occlusion relationship of upper and lower front teeth, and extracting global features of the multi-view intraoral images; performing coarse classification based on the features obtained by fusing the local features and the global features, and obtaining a coarse classification result containing at least one of a front tooth relationship abnormality class, a vertical abnormality class, a dental arch space abnormality class, a protrusion abnormality class and a normal or uncertain class; based on the coarse classification result, calling a sub-classifier corresponding to the coarse classification result to perform fine classification, and outputting a tooth deformity classification result.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method, device and medium for classifying dental malformations based on various granularities. Background Technology

[0002] Dental malformations refer to deformities of the teeth, jawbone, and craniofacial region caused by congenital or acquired factors during a child's growth and development. They not only affect facial aesthetics but can also lead to periodontal tissue damage, decreased chewing function, and speech disorders.

[0003] Currently, the classification of dental malocclusion largely relies on clinical examinations by professional dentists. Doctors typically combine intraoral visualization, model analysis, and two-dimensional photographs (such as frontal and lateral images) for assessment. However, this method has significant limitations. First, dental malocclusion classification is highly dependent on the dentist's subjective experience and professional level, making diagnostic consistency difficult to guarantee and resulting in low efficiency, making it unsuitable for large-scale population screening needs. Second, traditional two-dimensional photographs provide limited information and cannot comprehensively and accurately reflect the three-dimensional spatial relationships of teeth, especially for assessing subtle positional differences in the anterior-posterior and vertical directions of the upper and lower anterior teeth, which are crucial for evaluating anterior crossbite; their accuracy and reliability are insufficient. Summary of the Invention

[0004] This specification provides one or more embodiments of a method, device, and medium for classifying dental malformations based on various granularities, in order to solve the technical problems raised in the background art.

[0005] One or more embodiments of this specification employ the following technical solutions: This specification provides one or more embodiments of a method for classifying dental malformations based on various granularities, the method comprising: Acquire multi-view intraoral images of the current sample, wherein the multi-view intraoral images include at least a frontal occlusal image, a left occlusal image, and a right occlusal image; Tooth region detection is performed on the multi-view intraoral images to locate key anterior tooth regions for analyzing anterior occlusal relationships; Based on the aforementioned key anterior tooth region, local features reflecting the local occlusal relationship of the upper and lower anterior teeth are extracted, and global features of the multi-view intraoral images are also extracted. These local features include normalized anterior-posterior position difference parameters, coverage direction parameters, boundary contact ratio parameters, and anterior tooth contour angle parameters extracted based on the key anterior tooth region, wherein: The normalized position difference parameter is obtained based on the position of the lower anterior tooth reference point in the anterior-posterior direction, the position of the upper anterior tooth reference point in the anterior-posterior direction, and the average width of the upper central incisor or the reference width of the anterior tooth region. The normalized position difference parameter is used to characterize the degree of anterior displacement of the lower anterior teeth relative to the upper anterior teeth in the anterior-posterior direction. It is normalized by the average width of the upper central incisor or the reference width of the anterior tooth region to reduce the influence of differences in tooth size and imaging scale among different individuals on the results. When the normalized position difference parameter is greater than the first preset threshold, it indicates that the degree of anterior displacement of the lower anterior teeth is relatively large. The coverage direction parameter is obtained based on a normalized discriminant function constructed from the relative positional relationship, directional relationship, and coverage order of the upper anterior incisal edge line and the lower anterior incisal edge line. The coverage direction parameter is used to determine whether the relationship where the upper anterior teeth should cover the lower anterior teeth has reversed. When the coverage direction parameter satisfies the reverse coverage condition, it is marked as the anterior teeth reverse bite direction feature is established. The boundary contact ratio parameter is obtained based on the length of the local abnormal contact boundary of the anterior teeth and the total boundary length of the anterior teeth analysis area. The boundary contact ratio parameter is used to characterize the proportion of abnormal contact boundaries in the local area of ​​the anterior teeth and to reflect the degree of abnormality of the local contact mode of the upper and lower anterior teeth. The anterior tooth contour angle parameter is obtained based on the local contour direction angle of the upper anterior teeth and the local contour direction angle of the lower anterior teeth. The anterior tooth contour angle parameter is used to characterize the posture difference between the upper and lower anterior teeth in the local contour direction. The fusion weight of each view image is determined based on the quality score of each view image in the multi-view intraoral image, the visibility ratio of the key anterior tooth region, and the boundary clarity of the key anterior tooth region. Based on the fusion weights of the images from each viewpoint, the local features and the global features are fused together, and the determined features are coarsely classified to obtain a coarse classification result that includes at least one of the following: anterior tooth relationship abnormality, vertical abnormality, dental arch space abnormality, protrusion abnormality, and normal or uncertain class. During the classification process, if it is determined that the current sample meets the key discrimination rules for anterior crossbite, the output probability of the anterior crossbite category is increased. The key discrimination rule is determined based on whether at least two parameters contained in the local feature exceed the corresponding threshold. The output probability of the anterior crossbite category is determined based on the original probability of the anterior crossbite category and the rule satisfaction score. The rule satisfaction score is used to characterize the degree to which the current sample satisfies the key discrimination rule for anterior crossbite. The rule satisfaction score is determined based on whether each parameter in the local feature exceeds the corresponding preset threshold. The preset threshold corresponding to each parameter in the local feature is determined based on the preset conditions of each parameter in the local feature. Based on the coarse classification result, the sub-classifier corresponding to the coarse classification result is called to perform fine classification and output the dental malformation classification result.

[0006] This specification provides one or more embodiments of a dental malformation classification device based on various granularities, comprising: At least one processor and bus; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to: Acquire multi-view intraoral images of the current sample, wherein the multi-view intraoral images include at least a frontal occlusal image, a left occlusal image, and a right occlusal image; Tooth region detection is performed on the multi-view intraoral images to locate key anterior tooth regions for analyzing anterior occlusal relationships; Based on the aforementioned key anterior tooth region, local features reflecting the local occlusal relationship of the upper and lower anterior teeth are extracted, and global features of the multi-view intraoral images are also extracted. These local features include normalized anterior-posterior position difference parameters, coverage direction parameters, boundary contact ratio parameters, and anterior tooth contour angle parameters extracted based on the key anterior tooth region, wherein: The normalized position difference parameter is obtained based on the position of the lower anterior tooth reference point in the anterior-posterior direction, the position of the upper anterior tooth reference point in the anterior-posterior direction, and the average width of the upper central incisor or the reference width of the anterior tooth region. The normalized position difference parameter is used to characterize the degree of anterior displacement of the lower anterior teeth relative to the upper anterior teeth in the anterior-posterior direction. It is normalized by the average width of the upper central incisor or the reference width of the anterior tooth region to reduce the influence of differences in tooth size and imaging scale among different individuals on the results. When the normalized position difference parameter is greater than the first preset threshold, it indicates that the degree of anterior displacement of the lower anterior teeth is relatively large. The coverage direction parameter is obtained based on a normalized discriminant function constructed from the relative positional relationship, directional relationship, and coverage order of the upper anterior incisal edge line and the lower anterior incisal edge line. The coverage direction parameter is used to determine whether the relationship where the upper anterior teeth should cover the lower anterior teeth has reversed. When the coverage direction parameter satisfies the reverse coverage condition, it is marked as the anterior teeth reverse bite direction feature is established. The boundary contact ratio parameter is obtained based on the length of the local abnormal contact boundary of the anterior teeth and the total boundary length of the anterior teeth analysis area. The boundary contact ratio parameter is used to characterize the proportion of abnormal contact boundaries in the local area of ​​the anterior teeth and to reflect the degree of abnormality of the local contact mode of the upper and lower anterior teeth. The anterior tooth contour angle parameter is obtained based on the local contour direction angle of the upper anterior teeth and the local contour direction angle of the lower anterior teeth. The anterior tooth contour angle parameter is used to characterize the posture difference between the upper and lower anterior teeth in the local contour direction. The fusion weight of each view image is determined based on the quality score of each view image in the multi-view intraoral image, the visibility ratio of the key anterior tooth region, and the boundary clarity of the key anterior tooth region. Based on the fusion weights of the images from each viewpoint, the local features and the global features are fused together, and the determined features are coarsely classified to obtain a coarse classification result that includes at least one of the following: anterior tooth relationship abnormality, vertical abnormality, dental arch space abnormality, protrusion abnormality, and normal or uncertain class. During the classification process, if it is determined that the current sample meets the key discrimination rules for anterior crossbite, the output probability of the anterior crossbite category is increased. The key discrimination rule is determined based on whether at least two parameters contained in the local feature exceed the corresponding threshold. The output probability of the anterior crossbite category is determined based on the original probability of the anterior crossbite category and the rule satisfaction score. The rule satisfaction score is used to characterize the degree to which the current sample satisfies the key discrimination rule for anterior crossbite. The rule satisfaction score is determined based on whether each parameter in the local feature exceeds the corresponding preset threshold. The preset threshold corresponding to each parameter in the local feature is determined based on the preset conditions of each parameter in the local feature. Based on the coarse classification result, the sub-classifier corresponding to the coarse classification result is called to perform fine classification and output the dental malformation classification result.

[0007] This specification provides one or more embodiments of a non-volatile computer storage medium storing computer-executable instructions, which, when executed by a computer, can perform the following: Acquire multi-view intraoral images of the current sample, wherein the multi-view intraoral images include at least a frontal occlusal image, a left occlusal image, and a right occlusal image; Tooth region detection is performed on the multi-view intraoral images to locate key anterior tooth regions for analyzing anterior occlusal relationships; Based on the aforementioned key anterior tooth region, local features reflecting the local occlusal relationship of the upper and lower anterior teeth are extracted, and global features of the multi-view intraoral images are also extracted. These local features include normalized anterior-posterior position difference parameters, coverage direction parameters, boundary contact ratio parameters, and anterior tooth contour angle parameters extracted based on the key anterior tooth region, wherein: The normalized position difference parameter is obtained based on the position of the lower anterior tooth reference point in the anterior-posterior direction, the position of the upper anterior tooth reference point in the anterior-posterior direction, and the average width of the upper central incisor or the reference width of the anterior tooth region. The normalized position difference parameter is used to characterize the degree of anterior displacement of the lower anterior teeth relative to the upper anterior teeth in the anterior-posterior direction. It is normalized by the average width of the upper central incisor or the reference width of the anterior tooth region to reduce the influence of differences in tooth size and imaging scale among different individuals on the results. When the normalized position difference parameter is greater than the first preset threshold, it indicates that the degree of anterior displacement of the lower anterior teeth is relatively large. The coverage direction parameter is obtained based on a normalized discriminant function constructed from the relative positional relationship, directional relationship, and coverage order of the upper anterior incisal edge line and the lower anterior incisal edge line. The coverage direction parameter is used to determine whether the relationship where the upper anterior teeth should cover the lower anterior teeth has reversed. When the coverage direction parameter satisfies the reverse coverage condition, it is marked as the anterior teeth reverse bite direction feature is established. The boundary contact ratio parameter is obtained based on the length of the local abnormal contact boundary of the anterior teeth and the total boundary length of the anterior teeth analysis area. The boundary contact ratio parameter is used to characterize the proportion of abnormal contact boundaries in the local area of ​​the anterior teeth and to reflect the degree of abnormality of the local contact mode of the upper and lower anterior teeth. The anterior tooth contour angle parameter is obtained based on the local contour direction angle of the upper anterior teeth and the local contour direction angle of the lower anterior teeth. The anterior tooth contour angle parameter is used to characterize the posture difference between the upper and lower anterior teeth in the local contour direction. The fusion weight of each view image is determined based on the quality score of each view image in the multi-view intraoral image, the visibility ratio of the key anterior tooth region, and the boundary clarity of the key anterior tooth region. Based on the fusion weights of the images from each viewpoint, the local features and the global features are fused together, and the determined features are coarsely classified to obtain a coarse classification result that includes at least one of the following: anterior tooth relationship abnormality, vertical abnormality, dental arch space abnormality, protrusion abnormality, and normal or uncertain class. During the classification process, if it is determined that the current sample meets the key discrimination rules for anterior crossbite, the output probability of the anterior crossbite category is increased. The key discrimination rule is determined based on whether at least two parameters contained in the local feature exceed the corresponding threshold. The output probability of the anterior crossbite category is determined based on the original probability of the anterior crossbite category and the rule satisfaction score. The rule satisfaction score is used to characterize the degree to which the current sample satisfies the key discrimination rule for anterior crossbite. The rule satisfaction score is determined based on whether each parameter in the local feature exceeds the corresponding preset threshold. The preset threshold corresponding to each parameter in the local feature is determined based on the preset conditions of each parameter in the local feature. Based on the coarse classification result, the subclassifier corresponding to the coarse classification result is called to perform fine classification, and the dental malformation classification result is output.

[0008] The above-described at least one technical solution adopted in the embodiments of this specification can achieve the following beneficial effects: This application addresses the feature loss problem caused by blind spots in single-view imaging by fusing multi-view image information. It utilizes precise localization of key anterior tooth regions, enabling the extracted normalized positional differences and contour angles to quantify the subtle geometric relationships between teeth, thus overcoming the limitation of traditional global features in capturing local occlusal abnormalities. Furthermore, by introducing a hierarchical reasoning mechanism from coarse to fine classification, the coarse classification results are used to pre-divide the feature space, effectively constraining the retrieval range of the fine classifier. This allows the model to make targeted discriminations based on morphological differences of different abnormality types when dealing with complex coupled symptoms of arch morphology, protrusion, and vertical relationships, thereby avoiding misjudgments caused by decision boundary confusion in single-stage classification models and improving the certainty of complex malocclusion classification results.

[0009] At the local feature level, the positional difference parameter before and after normalization introduces the average width of the upper central incisors or the reference width of the anterior tooth region as a built-in scale benchmark, converting the absolute positional difference into a dimensionless relative quantity. This effectively eliminates scale interference caused by individual tooth size differences and changes in imaging distance, making the key indicator of the degree of mandibular anterior tooth movement comparable across different samples. The overbite direction parameter constructs a discriminant function that encodes the relative position, direction, and overbite order of the incisal edge, directly making a mathematical determination of the overbite direction. This clearly distinguishes between normal overbite and reverse overbite, avoiding misclassification of anterior reverse bite as deep overbite. The boundary contact ratio parameter measures the extent and continuity of abnormal contact by the proportion of the abnormal contact boundary length to the total boundary length of the analysis area. This suppresses single-point noise interference and focuses on true occlusal abnormalities with evidence of continuous boundaries. The anterior tooth region contour angle parameter provides additional evidence from the pose differences in the local contour directions of the upper and lower anterior teeth. It can identify samples where the positional difference is not yet obvious but the contour orientation has become abnormal, thus supplementing the discriminative dimension beyond position and overbite information. These four parameters provide independent evidence from the perspectives of position, direction, contact range, and posture, and together constitute a multi-dimensional quantitative representation of the local occlusal relationship of the anterior teeth.

[0010] In the feature fusion stage, this application introduces a dynamic fusion weighting mechanism based on multi-view image quality, the visible proportion of the key anterior tooth region, and boundary sharpness. This means that instead of simply stitching together or averaging the features from each viewpoint, a comprehensive reliability is calculated for each viewpoint. This reliability is obtained by weighting the image quality score, the visible proportion of the key anterior tooth region, and the sharpness of the key anterior tooth boundaries. The image quality score reflects whether the basic image quality (such as sharpness, brightness, and contrast) meets the analysis requirements; the visible proportion of the key anterior tooth region measures whether the core diagnostic area is fully presented and unobstructed from that viewpoint; and the boundary sharpness assesses the sharpness and recognizability of key geometric elements such as the incisal edge and contact boundary. These three factors together determine the reliability of the features extracted from that viewpoint. Subsequently, the reliability of each viewpoint is normalized into a fusion weight, and this weight is used to perform a weighted fusion of the local and global features extracted from each viewpoint.

[0011] Based on the comprehensive features obtained through the aforementioned dynamic weighted fusion, coarse classification can categorize samples into one of the following classes: anterior tooth relationship abnormalities, vertical abnormalities, dental arch spatial abnormalities, protrusion abnormalities, or normal / uncertain. Since the fused features have undergone reliability-weighted optimization, their signal-to-noise ratio and representativeness are stronger, resulting in clearer decision boundaries for coarse classification. This allows for more accurate assignment of samples to the correct major abnormality category, laying a solid foundation for subsequent fine classification. The entire technical solution, through precise quantification of local features, intelligent fusion of multi-perspective features, and hierarchical classification reasoning, is interconnected, collectively achieving a substantial improvement in the accuracy and robustness of dental malformation classification. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 A flowchart illustrating a method for classifying dental malformations based on various granularities, provided for one or more embodiments of this specification; Figure 2 This is a schematic diagram of a dental malformation classification device based on various granularities, provided for one or more embodiments of this specification. Detailed Implementation

[0013] This specification provides a method, device, and medium for classifying dental malformations based on various granularities.

[0014] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0015] Figure 1 This diagram illustrates a granular-based method for classifying dental malformations, provided for one or more embodiments of this specification. This process can be executed by a dental malformation classification system. Certain input parameters or intermediate results in the process can be manually adjusted to help improve accuracy.

[0016] The method flow steps of the embodiments in this specification are as follows: S101, acquire multi-view intraoral images of the current sample, wherein the multi-view intraoral images include at least a frontal occlusal image, a left occlusal image, and a right occlusal image.

[0017] In the embodiments described in this specification, the implementation of this step forms the basis for initiating the entire process. Specifically, a sequence of images of the user's oral cavity is acquired using the camera of a mobile terminal (such as a smartphone). At least three intraoral images from specific angles must be obtained: one frontal occlusal image, one left-side occlusal image, and one right-side occlusal image. To ensure the effectiveness of subsequent analysis, a standard viewfinder and shooting guidelines can be provided to the user during the acquisition phase. The input parameters that allow for manual adjustment in this step mainly include shooting condition parameters, such as the recommended shooting distance range, minimum image resolution requirements, and the minimum proportion of the dentition in the image. By adjusting these parameters, the user can be guided to capture images that better meet the analysis requirements, improving the quality of the input data from the source and laying a reliable data foundation for all subsequent steps.

[0018] S102, perform tooth region detection on the multi-view intraoral image to locate the key anterior tooth region for analyzing the anterior tooth occlusal relationship.

[0019] In the embodiments of this specification, the implementation of this step aims to accurately locate the anatomical structures related to the core diagnostic target from the original image acquired in S101. First, a pre-trained target detection model or image segmentation model (such as semantic segmentation or instance segmentation model) is used to process each multi-view intraoral image to separate the main tooth region from the background, obtaining accurate tooth region segmentation results. Then, based on this segmentation result, the anterior tooth region is further extracted from the entire dentition according to the geometric positional relationship of the tooth arrangement (such as the region at the front of the dental arch or near the midline). More refined localization includes: distinguishing and identifying the upper and lower anterior tooth regions within the anterior tooth region; simultaneously, for the left and right occlusal images, the regions extending posteriorly from the anterior teeth and the transition areas between the anterior and posterior teeth also need to be located. Finally, based on these located regions, key points or key lines such as the incisal edges of the upper and lower central incisors and the anterior midline are extracted for quantitative analysis. The output of this step (i.e., the precisely defined key anterior tooth regions and their geometric elements) is the direct basis for the next step of targeted feature extraction.

[0020] S103, based on the key anterior tooth region, extract local features reflecting the local occlusal relationship of the upper and lower anterior teeth, and extract global features of the multi-view intraoral image. The local features include normalized front-to-back position difference parameters, coverage direction parameters, boundary contact ratio parameters and anterior tooth region contour angle parameters extracted based on the key anterior tooth region.

[0021] The normalized position difference parameter is obtained based on the position of the lower anterior tooth reference point in the anterior-posterior direction, the position of the upper anterior tooth reference point in the anterior-posterior direction, and the average width of the upper central incisor or the reference width of the anterior tooth region. The normalized position difference parameter is used to characterize the degree of anterior displacement of the lower anterior teeth relative to the upper anterior teeth in the anterior-posterior direction. It is normalized by the average width of the upper central incisor or the reference width of the anterior tooth region to reduce the influence of differences in tooth size and imaging scale among different individuals on the results. When the normalized position difference parameter is greater than the first preset threshold, it indicates that the degree of anterior displacement of the lower anterior teeth is relatively large. The coverage direction parameter is obtained based on a normalized discriminant function constructed from the relative positional relationship, directional relationship, and coverage order of the upper anterior incisal edge line and the lower anterior incisal edge line. The coverage direction parameter is used to determine whether the relationship where the upper anterior teeth should cover the lower anterior teeth has reversed. When the coverage direction parameter satisfies the reverse coverage condition, it is marked as the anterior teeth reverse bite direction feature is established. The boundary contact ratio parameter is obtained based on the length of the local abnormal contact boundary of the anterior teeth and the total boundary length of the anterior teeth analysis area. The boundary contact ratio parameter is used to characterize the proportion of abnormal contact boundaries in the local area of ​​the anterior teeth and to reflect the degree of abnormality of the local contact mode of the upper and lower anterior teeth. The anterior tooth contour angle parameter is obtained based on the local contour direction angle of the upper anterior teeth and the local contour direction angle of the lower anterior teeth. The anterior tooth contour angle parameter is used to characterize the posture difference between the upper and lower anterior teeth in the local contour direction.

[0022] In the embodiments described in this specification, the implementation of this step constructs a core feature representation for classification, which directly depends on the output of S102. The implementation is divided into two parallel and complementary branches: Firstly, for local feature extraction, based on the key anterior tooth region located by S102 and the extracted key point lines, a set of parameters is specifically calculated to quantify the relative position and posture relationship between the upper and lower anterior teeth. These parameters include, but are not limited to: the normalized front-to-back position difference parameter used to measure the degree of forward movement of the lower anterior teeth relative to the upper anterior teeth; the overbite direction parameter used to determine whether the overbite relationship of the upper anterior teeth has reversed; the boundary contact ratio parameter used to reflect the proportion of abnormal contact; and the contour angle parameter used to characterize the difference in contour posture between the upper and lower anterior teeth.

[0023] Secondly, for global feature extraction, the complete multi-view intraoral images are input into a convolutional neural network backbone to extract global feature vectors that can characterize the overall arrangement of the entire dentition, the shape of the dental arch, and the occlusal profile.

[0024] This step allows for the adjustment of intermediate parameters, including reference widths and angular benchmarks involved in local feature calculations. By fine-tuning these parameters, the general morphological patterns of teeth in different age groups can be adapted.

[0025] S104. Determine the fusion weight of each view image based on the quality score of each view image in the multi-view intraoral image, the visibility ratio of the key anterior tooth region, and the boundary clarity of the key anterior tooth region.

[0026] In the embodiments of this specification, for each multi-view intraoral image (frontal occlusion image, left occlusion image, and right occlusion image), evaluation results in three aspects are obtained: The image quality score is derived from a preliminary image quality assessment. This assessment comprehensively considers factors such as image sharpness, brightness, contrast, occlusion ratio, closure integrity, image tilt angle, and viewing angle deviation, ultimately generating a comprehensive quality score for each image. This score reflects the quality of the image at that viewing angle in terms of basic image quality and shooting standardization.

[0027] Regarding the visibility ratio of key anterior tooth regions, based on anterior tooth region detection and localization, the visibility of key anterior tooth regions (i.e., the core regions used to analyze anterior tooth occlusion, such as the incisal edges of the maxillary and mandibular incisal edges, and the anterior midline) in the entire image from this perspective is calculated. If the key regions are obscured by the lips, tongue, or other tissues, or if some regions are not visible in the image due to the shooting angle, this ratio will decrease. A higher visibility ratio indicates that the perspective provides more complete information for analyzing anterior tooth occlusion.

[0028] For the boundary clarity of critical anterior tooth regions, also based on the results of anterior tooth region detection, the clarity of the boundaries (such as incisal edges, contact boundaries, and contour edges) of critical anterior tooth regions is evaluated. By analyzing the sharpness of these boundaries, the significance of gradient changes, and the presence of blurriness or jaggedness, it is determined whether the geometric features of the critical anterior tooth regions from this perspective can be accurately extracted. The clearer the boundaries, the more reliable the subsequent extracted local features.

[0029] After obtaining the three metrics mentioned above, the system calculates a comprehensive reliability score for each viewpoint. This reliability score is obtained by combining the image quality score, the visibility ratio of the critical anterior teeth region, and the sharpness of the critical anterior teeth region boundaries according to preset importance. For example, a viewpoint with good image quality but severely occluded critical regions will have a lower reliability score than a viewpoint with medium image quality but complete and sharp critical regions. Subsequently, the reliability scores of all viewpoints are normalized so that the sum of the fusion weights of each viewpoint is one. In this way, viewpoints with high reliability receive greater weights, while viewpoints with low reliability receive correspondingly smaller weights. If the reliability of a certain viewpoint is too low (e.g., the critical region is almost invisible), the weight of that viewpoint can be set to zero, meaning it will not participate in subsequent fusion, or a retake prompt can be triggered, requiring the user to retake an image from that viewpoint.

[0030] S105, based on the fusion weights of the images from each viewpoint, the local features and the global features are fused, and the determined features are coarsely classified to obtain a coarse classification result that includes at least one of the following: anterior tooth relationship abnormality, vertical abnormality, dental arch space abnormality, protrusion abnormality, and normal or uncertain class.

[0031] In the embodiments of this specification, for each viewpoint, two types of features have been extracted in the preceding steps. One type is local features reflecting the local occlusal relationship of the upper and lower anterior teeth (including normalized position difference parameters, overbite direction parameters, boundary contact ratio parameters, and anterior tooth contour angle parameters). The other type is global features reflecting the overall arrangement of the dental arch, the dental arch contour, the relationship between the upper and lower teeth, and the occlusal morphology. The current goal is to integrate these features from different viewpoints into a unified comprehensive feature.

[0032] For each viewpoint, its local and global features are multiplied by the corresponding fusion weights. Then, the weighted local features from all views are summed, and the weighted global features from all views are also summed. Finally, these two summed results are concatenated to form the final comprehensive feature used for classification. Since the fusion weights are normalized based on the reliability of each viewpoint, high-quality views with complete key regions and clear boundaries will dominate the comprehensive feature, while the contribution of low-quality views will be weakened.

[0033] After obtaining the comprehensive features, they are input into a pre-trained coarse classifier. The coarse classifier is a multi-classification model whose output space contains five categories: anterior tooth relation abnormalities, vertical abnormalities, dental arch spatial abnormalities, convexity abnormalities, and normal or uncertain categories. The coarse classifier calculates the probability of each category based on the input comprehensive features and then selects the category with the highest probability as the coarse classification result for the current sample. If the highest probability is lower than a preset threshold, the normal or uncertain category is directly output to avoid high-risk misclassification.

[0034] The coarse classification result will serve as the entry point for subsequent fine classification. The system will call the corresponding sub-classifier based on this result. For example, if the coarse classification result is an anterior tooth relationship abnormality, a sub-classifier that distinguishes between anterior crossbite and anterior deep overbite will be called; if it is a vertical abnormality, a sub-classifier that distinguishes between occlusion and deep overbite will be called, and so on. In this way, the hierarchical strategy of first coarse classification and then fine classification can effectively reduce the confusion problem when directly classifying multiple categories.

[0035] In the embodiments of this specification, the implementation of this step aims to perform preliminary sorting and high-level classification of complex features. Its input is a comprehensive feature representation composed of local and global features extracted and weighted through the aforementioned steps. This step utilizes a pre-trained coarse classifier (e.g., a classification model based on a convolutional neural network architecture) to analyze this comprehensive feature representation. The goal of the classifier is to perform a high-level discrimination, not to directly output seven specific malformation types, but to classify samples into a predefined, more general coarse category. These coarse categories are divided according to the core representation direction of the malformation, including at least: "Anterior Tooth Relationship Abnormality Class" focusing on the overbite and positional relationship of the upper and lower anterior teeth; "Vertical Abnormality Class" focusing on vertical occlusal problems; "Dental Arch Space Abnormality Class" reflecting dental arch density problems; and "Protrusion Abnormality Class" targeting the overall protrusion of the jawbone or teeth. In addition, a "Normal or Indeterminate Class" is set to accommodate samples without obvious abnormalities or with ambiguous features. The intermediate results or parameters that allow for manual intervention in this step mainly include the decision threshold of the coarse classification. By adjusting this threshold, the strictness with which a sample is classified into a certain abnormal category can be controlled. For example, increasing the threshold can reduce false positives, but may miss minor abnormalities, thus achieving a balance between screening sensitivity and specificity.

[0036] S106, During the classification process, if it is determined that the current sample meets the key discrimination rules for anterior crossbite, the output probability of the anterior crossbite category is increased.

[0037] The key discrimination rule is determined based on whether at least two parameters contained in the local feature exceed the corresponding threshold. The output probability of the anterior crossbite category is determined based on the original probability of the anterior crossbite category and the rule satisfaction score. The rule satisfaction score is used to characterize the degree to which the current sample satisfies the key discrimination rule for anterior crossbite. The rule satisfaction score is determined based on whether each parameter in the local feature exceeds the corresponding preset threshold. The preset threshold corresponding to each parameter in the local feature is determined based on the preset conditions of each parameter in the local feature.

[0038] In the embodiments described in this specification, rule triggering judgment is performed first, followed by probability correction execution.

[0039] During the rule-triggered judgment phase, the system examines a set of specific local feature parameters calculated for the current sample and used for classification. These parameters are indicators that directly reflect the occlusal relationship of the upper and lower anterior teeth, previously quantified and extracted from the image, such as the normalized anterior-posterior position difference parameter, coverage direction parameter, boundary contact ratio parameter, and anterior tooth contour angle parameter. The system compares the actual values ​​of these parameters with preset corresponding thresholds. The key discrimination rule is defined as: at least two (or more) of the above parameters simultaneously satisfy their abnormal conditions (e.g., the anterior-posterior position difference parameter is greater than its threshold, and the coverage direction parameter shows reverse coverage). Only when the system determines that the current sample meets this composite rule will the subsequent probability enhancement action be triggered. This rule design allows for manual intervention and adjustment; experts can change the strictness of rule triggering by adjusting the judgment thresholds of each parameter and the minimum number of parameters required to constitute the rule, thereby finely controlling the balance between screening sensitivity and specificity.

[0040] During the probability correction execution phase, once a rule is triggered, the system does not change the probabilities of other malocclusion categories. Instead, it specifically corrects the output probability of the anterior crossbite category upwards. The specific increase is not a fixed value but is achieved through a computational model. This model first calculates a "rule satisfaction score" between 0 and 1 based on the satisfaction of each sub-rule. This score reflects the overall strength of the sample's conformity to the morphological features of anterior crossbite. Then, this score is multiplied by a configurable "enhancement coefficient," and the product is used as the correction amount, which is added to the original output probability of the "anterior crossbite" category to produce the final increased probability. In this step, the "enhancement coefficient" is an important parameter that allows for manual intervention. By adjusting the size of this coefficient, the correction strength of rule-based logical evidence to the initial data-driven probability results can be controlled. This ensures that the final result both respects the patterns learned by the model from big data and strengthens the hard evidence defined by key domain knowledge, significantly improving the stability and clinical credibility of the system's screening results for the key category of anterior crossbite.

[0041] S107, Based on the coarse classification result, call the sub-classifier corresponding to the coarse classification result to perform fine classification, and output the dental malformation classification result.

[0042] In the embodiments described in this specification, the system intelligently routes the data to corresponding specialized sub-classifiers for processing based on the coarse classification results. This is not a general classifier handling all cases, but rather a set of expert sub-classifiers, each with its own focus.

[0043] Specifically, if the coarse classification result is "anterior tooth relationship abnormality," a sub-classifier specifically distinguishing between "anterior crossbite" and "anterior deep overbite" is invoked. This sub-classifier is designed to prioritize parameters such as overbite direction and anteroposterior position extracted from local features. If it is "vertical abnormality," a sub-classifier specifically distinguishing between "open bite" and "deep overbite" is invoked. If it is "arch space abnormality," a sub-classifier specifically distinguishing between "dental crowding" and "dental gaps" is invoked. If it is "protrusion abnormality," a sub-classifier specifically used to identify "bimaxillary protrusion" is invoked. Each sub-classifier, within its focused subdivision, uses a more suitable combination of input features (e.g., the sub-classifier for anterior tooth relationship abnormalities assigns higher weights to local features) for final discrimination. This step ultimately outputs one of seven dental malformation classification results (dental crowding, dental gaps, bimaxillary protrusion, open bite, deep overbite, anterior deep overbite, and anterior crossbite). The parts of this step that allow for manual intervention include the discrimination thresholds within each subclassifier and fine-tuning the weights of different features. By optimizing these parameters through expert experience, the discrimination accuracy and reliability of specific malformation subclasses, especially the key category of "anterior crossbite," can be further improved.

[0044] It should be noted that in the automatic classification of dental malformations, directly performing a single, detailed classification of multiple specific malformations is highly complex and prone to affecting accuracy due to feature confusion between different types. The two-stage classification strategy adopted in this invention effectively solves this problem. First, the method uses fused features for coarse classification, categorizing samples into limited broad classes such as "anterior tooth relationship abnormalities" or "vertical abnormalities." This step constructs a high-level, generalized discriminative framework. Subsequently, instead of using a single complex model to handle all cases, the system intelligently calls sub-classifiers specifically optimized for that broad category of malformation to complete the fine classification based on the coarse classification results. This design allows each sub-classifier to focus only on learning and discriminating the subtle differences between several specific malformations within a specific direction, without being forced to establish complex global decision boundaries between all malformation types. Therefore, the decision path of the entire classification process is clearer and more focused, reducing the risk of misclassification caused by a large, comprehensive single model under the interference of different types of features, thereby improving the overall accuracy and reliability of the final output classification results.

[0045] It should be noted that this application solves the problem of feature loss caused by blind spots in single-view imaging by fusing multi-view image information. This application utilizes the precise localization of key areas of the anterior teeth, so that the extracted normalized position difference and contour angle and other parameters can quantify the subtle geometric relationship between teeth, thereby making up for the deficiency of traditional global features in capturing local malocclusion abnormalities. Furthermore, by introducing a hierarchical reasoning mechanism from coarse classification to fine classification, the feature space is pre-divided using the coarse classification results, which effectively constrains the retrieval range of the fine classifier. This allows the model to make targeted discrimination based on the morphological differences of different abnormality types when dealing with complex coupled symptoms of dental arch morphology, protrusion and vertical relationship, thereby avoiding misjudgment caused by decision boundary confusion in single-stage classification models and improving the certainty of complex malocclusion classification results.

[0046] At the local feature level, the positional difference parameter before and after normalization introduces the average width of the upper central incisors or the reference width of the anterior tooth region as a built-in scale benchmark, converting the absolute positional difference into a dimensionless relative quantity. This effectively eliminates scale interference caused by individual tooth size differences and changes in imaging distance, making the key indicator of the degree of mandibular anterior tooth movement comparable across different samples. The overbite direction parameter constructs a discriminant function that encodes the relative position, direction, and overbite order of the incisal edge, directly making a mathematical determination of the overbite direction. This clearly distinguishes between normal overbite and reverse overbite, avoiding misclassification of anterior reverse bite as deep overbite. The boundary contact ratio parameter measures the extent and continuity of abnormal contact by the proportion of the abnormal contact boundary length to the total boundary length of the analysis area. This suppresses single-point noise interference and focuses on true occlusal abnormalities with evidence of continuous boundaries. The anterior tooth region contour angle parameter provides additional evidence from the pose differences in the local contour directions of the upper and lower anterior teeth. It can identify samples where the positional difference is not yet obvious but the contour orientation has become abnormal, thus supplementing the discriminative dimension beyond position and overbite information. These four parameters provide independent evidence from the perspectives of position, direction, contact range, and posture, and together constitute a multi-dimensional quantitative representation of the local occlusal relationship of the anterior teeth.

[0047] In the feature fusion stage, this application introduces a dynamic fusion weighting mechanism based on multi-view image quality, the visible proportion of the key anterior tooth region, and boundary sharpness. This means that instead of simply stitching together or averaging the features from each viewpoint, a comprehensive reliability is calculated for each viewpoint. This reliability is obtained by weighting the image quality score, the visible proportion of the key anterior tooth region, and the sharpness of the key anterior tooth boundaries. The image quality score reflects whether the basic image quality (such as sharpness, brightness, and contrast) meets the analysis requirements; the visible proportion of the key anterior tooth region measures whether the core diagnostic area is fully presented and unobstructed from that viewpoint; and the boundary sharpness assesses the sharpness and recognizability of key geometric elements such as the incisal edge and contact boundary. These three factors together determine the reliability of the features extracted from that viewpoint. Subsequently, the reliability of each viewpoint is normalized into a fusion weight, and this weight is used to perform a weighted fusion of the local and global features extracted from each viewpoint.

[0048] Based on the comprehensive features obtained through the aforementioned dynamic weighted fusion, coarse classification can categorize samples into one of the following classes: anterior tooth relationship abnormalities, vertical abnormalities, dental arch spatial abnormalities, protrusion abnormalities, or normal / uncertain. Since the fused features have undergone reliability-weighted optimization, their signal-to-noise ratio and representativeness are stronger, resulting in clearer decision boundaries for coarse classification. This allows for more accurate assignment of samples to the correct major abnormality category, laying a solid foundation for subsequent fine classification. The entire technical solution, through precise quantification of local features, intelligent fusion of multi-perspective features, and hierarchical classification reasoning, is interconnected, collectively achieving a substantial improvement in the accuracy and robustness of dental malformation classification.

[0049] Furthermore, after acquiring the multi-view intraoral image of the current sample, the method flow steps of this embodiment are as follows: S201, perform image quality assessment on the multi-view intraoral images to obtain assessment results. The image quality assessment includes the assessment of at least one of the following indicators: sharpness, brightness, contrast, visible proportion of tooth area, occlusion proportion, occlusal closure integrity, image tilt angle, and degree of viewpoint deviation.

[0050] In the embodiments of this specification, this step is a preliminary quality control measure that must be performed after acquiring the original image and before any in-depth analysis. Its core is a systematic, multi-indicator quantitative evaluation of each multi-view intraoral image (frontal, left, and right occlusal images) acquired in step S101. The evaluation is not a single-dimensional judgment, but a comprehensive examination of a set of predefined indicators directly related to the success or failure of subsequent analysis. These indicators include basic image attributes such as sharpness, brightness, and contrast; professional content integrity indicators directly related to dental malformation analysis, such as the visible proportion of tooth areas, the occlusion proportion, and the integrity of occlusal closure; in addition, it covers shooting standardization indicators, such as image tilt angle and the degree of viewpoint deviation. By calculating these indicators, the system generates a comprehensive evaluation result for each image. The input parameters that allow manual intervention in this step are mainly the specific thresholds or weights corresponding to each evaluation indicator, such as the minimum acceptable value for sharpness and the minimum visible proportion of the key anterior tooth area. By adjusting these parameters, it is possible to flexibly adapt to different application scenario standards or common shooting conditions of different populations, thereby customizing the definition of what constitutes a "qualified" input image.

[0051] S202, in response to the evaluation result not meeting the preset requirements, a retake prompt is provided.

[0052] In the embodiments described in this specification, the implementation of this step is a logical extension and action response of the S201 quality assessment. It executes decisions based on the assessment results generated by the S201 step. The system compares the assessment results with preset quality requirements (i.e., the adjustable threshold standards in S201). Once it is determined that the assessment results of any one or more viewpoint images do not meet the requirements, the system will not allow images with quality defects to enter subsequent core analysis processes such as tooth region detection and feature extraction. Instead, it will immediately interrupt the current analysis chain and initiate explicit interactive feedback to the user (such as the image photographer), namely, a "retake prompt". The implementation of this mechanism ensures that only images that pass the quality screening can enter the subsequent complex calculation process, thereby eliminating misjudgments or analysis failures caused by quality problems such as blurriness, missing key areas, and angle deviations from the source, greatly improving the practicality and reliability of the entire method in non-professional shooting environments. The part of this step that allows manual intervention and adjustment is mainly the strictness of triggering the retake prompt, such as whether it is triggered when any indicator fails to meet the standard, or only when the comprehensive score is below a certain threshold. This can balance the relationship between user experience and result accuracy.

[0053] It should be noted that, addressing the common problem of blurry, overly dark, overly bright, incomplete tooth areas, or improper angles in oral imaging under non-professional conditions, leading to unreliable or ineffective subsequent analysis, this invention introduces a systematic image quality assessment process immediately after image acquisition. This process does not rely on a single judgment but comprehensively considers basic image quality such as sharpness, brightness, and contrast, as well as professional indicators highly relevant to subsequent analysis, such as the visible proportion of tooth areas, occlusion, complete occlusion, image tilt, and viewing angle deviation. This comprehensive and targeted assessment system accurately identifies low-quality or invalid images that do not meet analysis requirements. If the assessment result is unsatisfactory, the system immediately provides the user with a clear prompt to retake the image. This mechanism fundamentally prioritizes quality control, ensuring that only images of acceptable quality enter the subsequent complex feature extraction and classification process. This significantly improves the robustness and practicality of the entire method in real-world applications, enabling even non-professional users to capture images suitable for effective analysis under guidance, thus guaranteeing the initial reliability of screening results.

[0054] It should be noted that in the automatic classification of dental malformations, directly performing a single, detailed classification of multiple specific malformations is highly complex and prone to affecting accuracy due to feature confusion between different types. The two-stage classification strategy adopted in this invention effectively solves this problem. First, the method uses fused features for coarse classification, categorizing samples into limited broad classes such as "anterior tooth relationship abnormalities" or "vertical abnormalities." This step constructs a high-level, generalized discriminative framework. Subsequently, instead of using a single complex model to handle all cases, the system intelligently calls sub-classifiers specifically optimized for that broad category of malformation to complete the fine classification based on the coarse classification results. This design allows each sub-classifier to focus only on learning and discriminating the subtle differences between several specific malformations within a specific direction, without being forced to establish complex global decision boundaries between all malformation types. Therefore, the decision path of the entire classification process is clearer and more focused, reducing the risk of misclassification caused by a large, comprehensive single model under the interference of different types of features, thereby improving the overall accuracy and reliability of the final output classification results.

[0055] Furthermore, the output probability of the anterior crossbite category satisfies ,in, The output probability is the anterior crossbite category. The original probability for the anterior crossbite category. Let λ be the rule satisfaction score, and λ be the enhancement coefficient. The rule satisfaction score is defined as follows: ;in, The parameter representing the position difference before and after normalization Whether it exceeds the first preset threshold, if it meets the threshold, then set to 1, otherwise set to 0; Indicates the coverage direction parameter Whether the reverse cover condition is met; if it is, set to 1, otherwise set to 0. Indicates boundary contact ratio parameter Whether it exceeds the second preset threshold, if it meets the threshold, then set to 1, otherwise set to 0; Indicates the parameters of the anterior tooth contour angle. Whether it exceeds the third preset threshold, if it meets the threshold, then set to 1, otherwise set to 0; This indicates whether at least two perspectives show a consistent anterior teeth reversal relationship; if so, the value is 1, otherwise it is 0. to These are the weighting coefficients for the corresponding rules.

[0056] It should be noted that in the process of focusing on the identification of anterior crossbite, if only a simple "yes" or "no" hard rule is used to cover the model's original judgment, the decision may be too rigid and unable to effectively handle complex or atypical samples that meet some features but whose overall evidence strength is at a critical state, thus potentially leading to new misjudgments. The probability correction mechanism designed in this invention provides a more refined and adaptive solution. This mechanism does not directly give a binary conclusion, but transforms the degree of rule satisfaction based on multiple local features and multi-perspective consistency into a continuous "rule satisfaction score" through an explicit mathematical formula. This score, through an adjustable enhancement coefficient, gently corrects the probability value of the model's original output in a linear superposition manner. The core advantage of this method lies in its smoothness and interpretability of decision-making: it allows for a dynamic trade-off and fusion between the model's own probability prediction and logical judgment based on explicit rules. For samples with highly significant rule features, a higher rule score significantly increases the final probability, ensuring high-confidence detection. For samples that only partially meet the rules or exhibit atypical features, the correction provided by the rule score is moderate, achieving a balance between model prediction and rule implication in the final result, avoiding the limitations of a single judgment source. This design ensures that the final determination of anterior crossbite fully respects the complex patterns learned by the data-driven model from a large number of samples, while firmly incorporating key domain knowledge logic. This improves the reliability of the criterion while maintaining the system's flexibility and robustness in handling multiple sample cases.

[0057] Furthermore, in the process of detecting tooth regions in the multi-view intraoral images to locate key anterior tooth regions for analyzing anterior occlusal relationships, the method flow steps of this embodiment are as follows: S301, using an object detection model fused with a semantic segmentation model or an instance segmentation model, the main tooth region is detected and segmented in the multi-view intraoral image to obtain the segmentation result.

[0058] In the embodiments described in this specification, the implementation plan of this step is the starting point and foundation of the entire localization process, aiming to initially separate the target object from the original image. Its implementation involves using a pre-trained computer vision model (such as a deep learning-based object detection, semantic segmentation, or instance segmentation model) to process the input multi-view intraoral images (front, left, and right sides). The model's task is to identify each tooth instance in the image or at least separate all tooth regions from background regions such as gums, lips, and tongue, outputting the segmentation result of the main tooth region for each image (i.e., a mask image indicating whether each pixel belongs to a tooth). The aspects of this step that allow for manual intervention mainly involve model selection and fine-tuning. For example, a model pre-trained on a larger oral dataset can be selected, and the model can be fine-tuned using specific data based on the differences in tooth morphology among specific populations (such as children and adults) to improve its segmentation accuracy and robustness in the current application scenario.

[0059] S302, extract the anterior tooth region based on geometric positional relationships from the segmentation results.

[0060] In the embodiments of this specification, the implementation of this step is to perform initial region focusing based on the overall tooth region obtained in S301. This is based on the fixed anatomical positional relationship of the teeth arranged in the dental arch. In specific implementation, the system analyzes the overall spatial distribution of all tooth regions in the segmentation results (e.g., by calculating the center point or contour of all tooth regions, fitting a rough dental arch curve). The anterior tooth region (usually referring to the central incisors, lateral incisors, and canines) is geometrically located at the very front of the dental arch (the region closest to the center of the lower part of the image). Based on such preset geometric rules (e.g., selecting all tooth segmentation regions located in the first third of the dental arch curve), the system automatically delineates the anterior tooth region from the entire tooth body segmentation results. In this step, the geometric rules used to define the "anterior tooth region" (such as specific position thresholds and proportions) are parameters that can be manually adjusted according to clinical definitions or actual application needs.

[0061] S303, in the anterior tooth region, the upper anterior tooth region and the lower anterior tooth region are identified.

[0062] In the embodiments of this specification, the implementation of this step is a further refinement of S302 and is crucial for achieving the separation of the maxillary and mandibular anterior teeth. Based on the overall mask of the anterior tooth region obtained in S302, the system needs to distinguish which pixels belong to the maxillary anterior teeth and which belong to the mandibular anterior teeth. A typical implementation method is to use the occlusal line or the horizontal midline of the image as a reference. Since the maxillary and mandibular anterior teeth are usually roughly divided along the horizontal direction in a frontal occlusal image, the system can analyze the pixel distribution of the anterior tooth region in the vertical direction, find a natural dividing position (e.g., find the gap with the lowest density through horizontal projection), identify the upper region as the maxillary anterior tooth region, and the lower region as the mandibular anterior tooth region. The accuracy of this identification step is crucial for the subsequent calculation of the relative positional features of the maxillary and mandibular anterior teeth, and its segmentation logic (such as the method for determining the reference line) allows for manual calibration and adjustment.

[0063] S304, in the left occlusal image and the right occlusal image, locate the anterior tooth extension area and the anterior-posterior tooth transition area.

[0064] In the embodiments described in this specification, the implementation of this step is a specialized processing of lateral view images, designed to more accurately define the lateral boundaries of the anterior teeth from different angles. In the segmentation results of the left and right occlusal images, the system not only identifies the anterior teeth but also needs to locate two specific regions: the "anterior tooth extension area" and the "anterior-posterior tooth transition area." The "anterior tooth extension area" typically refers to a range extending laterally and posterolaterally from the most prominent point of the anterior tooth, used to define the lateral contour of the anterior teeth. The "anterior-posterior tooth transition area" refers to the area near the junction of the anterior and posterior teeth (such as premolars). This can be achieved by delineating the area along the dental arch curve, combined with anatomical landmarks of the anterior teeth (such as the cusps of canines), on the tooth segmentation results of the lateral image. The localization results of this step help to more accurately define the range of the anterior teeth used for analysis in a three-dimensional spatial concept. The anatomical landmarks and range definition rules used for localization can be adjusted based on clinical experience.

[0065] S305, based on the upper anterior tooth region, the lower anterior tooth region, the anterior tooth extension region and the anterior-posterior tooth transition region, extract at least one key point or key line for characterizing the anterior tooth occlusal relationship, the key point or key line includes: the incisal edge point of the upper central incisor, the incisal edge point of the lower central incisor, the local contour center point of the upper and lower anterior teeth, the anterior tooth midline, the incisal edge connection line and the local anterior segment contour line of the dental arch.

[0066] In the embodiments of this specification, the implementation of this step is the final goal of the preceding four steps, aiming to extract calculable geometric elements. The system will comprehensively utilize the precise spatial information obtained in steps S302, S303, and S304, such as the upper anterior tooth region, lower anterior tooth region, anterior tooth extension region, and anterior-posterior tooth transition region, and apply image processing and geometric calculation algorithms to extract a series of points or lines with clear anatomical significance. For example, the most prominent point on the lower edge of the upper anterior tooth region is found as the incisal edge point of the upper central incisor, and the corresponding point on the upper edge of the lower anterior tooth region is found as the incisal edge point of the lower central incisor; the center points of the upper and lower anterior tooth regions are calculated to obtain the local contour center points of the upper and lower anterior teeth; the anterior tooth midline is determined by connecting specific points of the left and right central incisors; an incisal edge line is formed by connecting multiple incisal edge points; and a curve is fitted on the anterior tooth extension region to obtain the local anterior segment contour line of the dental arch. These key points and lines are the original geometric basis for subsequent calculation of all local feature parameters (such as anterior-posterior position difference, contour angle, etc.). In this step, the specific extraction algorithm for each key point or key line (such as contour point detection and curve fitting methods) can be selected and its parameters optimized according to the requirements of accuracy and stability. This is an important step that allows for manual intervention to optimize the quality of feature extraction.

[0067] It should be noted that when analyzing anterior occlusion from multi-view intraoral images, if only a rough rectangular region containing all anterior teeth is located or a simple overall segmentation is performed, the resulting region will still contain a large amount of background or other tooth information unrelated to the occlusion analysis, and will not be able to accurately distinguish the anatomical structures of the upper and lower anterior teeth. This directly leads to the subsequent extraction of mixed and impure local features, thereby reducing the accuracy of parameter calculations used to determine anterior crossbite and other malocclusions. The localization method described in this invention solves this problem through a coarse-to-fine, multi-step collaborative process. It first uses a mature detection or segmentation model to obtain a precise main tooth region, eliminating most of the interference from non-tooth tissues. Subsequently, its core lies not in stopping there, but in further separating the upper and lower anterior tooth regions from the anterior tooth region based on geometric relationships, which creates the premise for separately quantifying the independent and relative positions of the upper and lower anterior teeth. Specifically, the method also requires locating the "anterior tooth extension zone" and "transition zone" in lateral view images. This essentially provides a more precise definition of the three-dimensional spatial extent of the anterior tooth region from different observation angles. Ultimately, based on these finely segmented and located sub-regions, the method extracts geometric elements with clear anatomical significance, such as incisal edge points, contour center points, and midlines. These key points and lines form the stable and reproducible mathematical foundation upon which all subsequent quantitative local features (such as anterior-posterior position difference, coverage direction, and contour angle) depend. Therefore, this entire localization process, through progressive refinement and multi-view cross-definition of the anterior tooth analysis area, ensures that the source data for subsequent feature extraction—the key geometric elements characterizing occlusal relationships—can be acquired with high precision and consistency, thus laying a reliable foundation for the accurate judgment of the entire classification system.

[0068] Furthermore, in the process of extracting the global features of the multi-view intraoral image, the method flow steps of this embodiment are as follows: S401, the multi-view intraoral images are input into the feature extraction network to obtain global features representing the overall arrangement of the dental arch, the outline of the dental arch, the relationship between the upper and lower teeth and the occlusal morphology, wherein the feature extraction network adopts a convolutional neural network backbone structure.

[0069] In the embodiments described in this specification, this step is a crucial process for obtaining a comprehensive morphological representation of the dentition. It begins with multi-view intraoral images (frontal, left, and right occlusal images) acquired in previous steps and which have passed quality assessment. These images will serve as the direct input for this step.

[0070] In practice, the system initiates an independent feature extraction process for each viewpoint image. Each image first undergoes necessary standardization preprocessing, such as scaling to the fixed resolution required by the network and normalizing pixel values ​​to adapt to the network's input format. Subsequently, each preprocessed image is input into a shared or structurally identical feature extraction network. The core architecture of this network employs a convolutional neural network backbone pre-trained on large-scale image data. As the image passes through this deep network, multi-layer convolutions, pooling, and other operations abstract and transform image information step by step, ultimately forming a high-dimensional feature vector at a specific layer at the end of the network (usually before the last fully connected layer or global pooling layer). This feature vector is not specific to any particular tooth but integrates the contextual information of all teeth and their background in the entire image, effectively representing the overall arrangement of the dentition, the approximate outline of the dental arch, the relative relationship between the upper and lower dentitions, and the overall occlusal morphology from that viewpoint. Therefore, for a set of three viewpoint images, the output of this step is three independent global feature vectors, each carrying macroscopic information about the entire dentition from different viewing angles.

[0071] This step, which allows for manual intervention, primarily involves the selection and fine-tuning of the feature extraction network itself. For example, convolutional neural network backbones of different depths or complexities (such as residual networks or their lightweight variants) can be chosen to balance feature representation power and computational efficiency. Furthermore, depending on computational resource constraints or feature granularity requirements in practical applications, it can be decided whether to use network weights further fine-tuned on a specific oral image dataset to ensure that the extracted global features better meet the task requirements of dental malformation analysis. By adjusting these network-related parameters, the quality and relevance of the global features can be optimized, thereby affecting the performance of subsequent fusion and classification steps. The global features extracted in this step, along with local features extracted from the fine anterior tooth region, will be fed into the subsequent feature fusion step to achieve a comprehensive judgment of macroscopic morphology and microscopic occlusal relationships.

[0072] It is important to note that in the automatic classification of dental malformations, the identification of many types of malformations (such as crowded teeth and abnormal dental arch morphology) highly depends on a comprehensive assessment of the overall arrangement of the entire dentition, the outline of the dental arch, and the macroscopic relationship between the upper and lower teeth, rather than just the local details of the anterior teeth region. If the classification model relies solely on local features extracted from images that focus on anterior occlusion, it may fail to fully perceive and utilize this global morphological information, leading to insensitivity or inaccuracy in the identification of the aforementioned types of malformations. This invention incorporates a global feature extraction step in parallel throughout the technical process. Its core implementation leverages the powerful hierarchical feature learning capabilities of the convolutional neural network backbone structure. This network performs end-to-end depth processing on each multi-view intraoral image, automatically learning and extracting high-dimensional feature vectors that characterize the "overall arrangement of the dentition, the outline of the dental arch, the relationship between the upper and lower teeth, and the occlusal morphology." These global features encode the overall spatial configuration and morphological context of the dentition at a macroscopic scale, effectively complementing the local features extracted from the fine anterior teeth region that reflect microscopic positional relationships. By combining global and local features in subsequent classification, the model's decisions can be based simultaneously on direct evidence of local occlusal relationships and macroscopic clues of overall dentition morphology. This combined local and global feature representation significantly enhances the classification system's ability to comprehensively distinguish various types of dental malformations, particularly improving the accuracy and robustness of identification for malformation categories that rely more on overall morphological assessment (compared to categories that rely solely on local anterior tooth relationships), thus achieving more comprehensive and reliable automated screening.

[0073] It should be noted that the embodiments of this specification also provide a method for classifying dental malformations based on multi-view intraoral images. This method uses intraoral images of the frontal occlusion, left occlusion, and right occlusion as the main inputs. Within the overall framework of seven types of dental malformations, it focuses on screening and classifying anterior crossbite. Through image quality assessment, tooth region detection, localization of key anterior tooth regions, extraction of local occlusal relationship features, multi-view reliability weighted fusion, and classification determination, it achieves the screening and classification of seven types of dental malformations.

[0074] (1) Image acquisition and input construction The user acquires at least three intraoral images via a mobile terminal, including: a frontal occlusal intraoral image, a left occlusal intraoral image, and a right occlusal intraoral image. In some implementations, an optional maxillary occlusal image and / or mandibular occlusal image may also be acquired to enhance the auxiliary judgment of categories such as crowding and interdental spaces.

[0075] In some implementations, the user may be required to meet at least one of the following shooting conditions during image acquisition: the shooting distance is 8cm to 20cm; the image resolution is not less than 720×720 pixels, preferably not less than 1024×1024 pixels; the dental arch area occupies not less than 40% of the image area; the upper and lower teeth are in a naturally closed or nearly naturally closed state in the biting state; and the shooting angle deviation is not greater than a preset angle threshold.

[0076] Before acquisition, the system displays a standard viewfinder and shooting guidelines to the user to reduce problems such as tooth misalignment, partial missing teeth, and excessive tilt in the image.

[0077] (2) Image quality assessment and retake prompts To prevent low-quality images from directly entering the classification model, this invention performs a quality assessment on each input image before classification. The quality assessment includes at least one or more of the following metrics: sharpness, brightness, contrast, visible proportion of the tooth region, visible proportion of the key anterior tooth region, occlusion proportion, occlusion integrity, image tilt angle, and degree of viewpoint deviation.

[0078] Sharpness is characterized by variance based on the Laplacian operator; brightness is characterized by the mean gray level; contrast is characterized by the standard deviation of gray levels; the visible proportion of the tooth region and the visible proportion of the key anterior tooth region are calculated from the tooth region detection results. A comprehensive quality score is calculated for each image. It can be expressed by the formula: ; in, Indicates the clarity score. Indicates the brightness / contrast score. This indicates that the score is visible in the key area. This indicates a penalty for occlusion. The score indicates the completeness of occlusal closure.

[0079] When the overall quality score of any key viewpoint image is lower than a preset threshold, the system does not output a classification result, but instead outputs a retake prompt. To enhance the reliability of the screening of anterior crossbite, in one embodiment, if the proportion of visible anterior teeth in the frontal bite image is lower than a preset value, a retake prompt is triggered first, even if the quality scores of the other images meet the requirements.

[0080] (3) Image preprocessing and standardization Images that pass the quality assessment are preprocessed, including image size unification, brightness normalization, contrast enhancement, noise reduction, image orientation correction, and color space normalization.

[0081] Image preprocessing includes the following operations: First, the original image is scaled to a first preset size for quality analysis, which is 512×512; then, the image entering the classification model is scaled to a second preset size, which is 224×224, 320×320, or 384×384; local contrast enhancement is performed using CLAHE; noise reduction is performed using median filtering or Gaussian filtering, with the kernel size preferably being 3×3 or 5×5; the image is rotated and corrected according to the dental midline or the main direction of the dental arch, so that the angle between the main direction of the dental arch and the standard direction does not exceed 5°.

[0082] In some implementations, image preprocessing may also include cropping non-target regions such as the lips and corners of the mouth to reduce interference from irrelevant regions in subsequent analysis.

[0083] (4) Tooth area detection and localization of key anterior teeth areas The preprocessed image is subjected to tooth region detection to obtain the main tooth region; further, the upper anterior tooth region and the lower anterior tooth region are identified within the main tooth region.

[0084] Tooth region detection is achieved using object detection models, semantic segmentation models, or instance segmentation models. First, the main tooth regions are segmented from the entire intraoral image. Then, the anterior tooth regions are extracted based on the segmentation results and geometric positional relationships.

[0085] The localization of key anterior tooth regions includes: locating the anterior tooth extension area and the transition area between the anterior and posterior teeth in the left and right occlusal images; locating the maxillary and mandibular anterior tooth regions in the anterior tooth region images obtained in the previous step; and extracting the incisal edges, labial contour lines, and contact boundaries of the maxillary and mandibular anterior teeth. Further key points or lines are extracted, including: the incisal edge point of the maxillary central incisor, the incisal edge point of the mandibular central incisor, the local contour center points of the maxillary and mandibular anterior teeth, the anterior tooth midline, the incisal edge connection line, and the local anterior arch contour line.

[0086] (5) Feature extraction 1. Global Feature Extraction The frontal occlusion image, left occlusion image, and right occlusion image are respectively input into the feature extraction network to obtain global feature vectors representing the overall arrangement of the dental arch, the dental arch contour, the relationship between the upper and lower teeth, and the occlusal morphology.

[0087] The feature extraction network employs a convolutional neural network backbone, such as a residual network or a lightweight variant thereof.

[0088] 2. Local Feature Extraction For targeted screening of anterior crossbite, this invention extracts local occlusal relationship features based on key anterior tooth regions. These key anterior tooth regions include the upper and lower anterior tooth regions in the frontal occlusal diagram, as well as the anterior tooth extension area and the anterior-posterior tooth transition area in the left and right occlusal diagrams. Based on these key anterior tooth regions, local features reflecting the local occlusal relationship of the upper and lower anterior teeth can be extracted.

[0089] The local features include at least one or more of the following: parameters used to construct the position difference before and after normalization. Anterior and posterior position features of the reference points for the upper and lower anterior teeth and the reference width characteristics of the anterior tooth region Used to construct coverage direction parameters Features of the incisal edge of the upper anterior teeth and the incisal edge features of the lower anterior teeth Used to construct boundary contact ratio parameters Anterior tooth local abnormal contact boundary length characteristics and the total boundary length characteristics of the anterior tooth analysis region ; Used to construct the contour angle parameters of the anterior tooth region Local contour direction angle features of the upper anterior teeth and the local contour direction angle features of the lower anterior teeth In addition, the key point distance features of the anterior tooth region, the contour curvature features of the anterior tooth region, and the texture and edge intensity features of the anterior tooth region.

[0090] In one embodiment, to improve the consistency and comparability of local relationship feature representations among different samples, four normalized local relationship parameters are further constructed based on the aforementioned local features to quantify the anterior-posterior positional relationship, overlay direction relationship, local contact relationship, and contour pose relationship of the upper and lower anterior teeth: First, the position difference parameter before and after normalization. ,in This indicates the position of the lower anterior teeth reference point in the anteroposterior direction. This indicates the position of the upper anterior teeth reference point in the anteroposterior direction. This indicates the average width of the upper central incisors or the reference width of the anterior region. The meaning of this parameter is: to characterize the degree of anterior displacement of the lower anterior teeth relative to the upper anterior teeth in the anteroposterior direction, and to be expressed as a reference width. Normalization was performed to reduce the impact of differences in tooth size and imaging scale among individuals on the results. When the value exceeds the preset threshold, it indicates that the lower anterior teeth have moved forward to a greater extent.

[0091] Second, the coverage direction parameter This is used to characterize whether the overbite direction of the upper and lower anterior teeth has reversed. Among them, Indicates the incisal edge line of the upper anterior teeth. Indicates the incisal edge line of the lower anterior teeth. This represents a normalized discriminant function constructed based on the relative positional relationship, directional relationship, and overbite sequence between the incisal edges of the upper and lower anterior teeth. The meaning of this parameter is to characterize whether the overbite direction between the upper and lower anterior teeth is normal, especially to determine whether the relationship where the upper anterior teeth should cover the lower anterior teeth has reversed. When the overbite direction parameter satisfies the reverse overbite condition, it is marked as a valid anterior crossbite direction feature.

[0092] Third, boundary contact ratio parameters ,in Indicates the length of the boundary of a local abnormal contact in the anterior teeth. This parameter represents the total boundary length of the anterior tooth analysis area. Its meaning is to characterize the proportion of abnormal contact boundaries within the local anterior tooth region, reflecting the degree of abnormality in the local contact patterns of the upper and lower anterior teeth.

[0093] Fourth, the parameters of the anterior tooth contour angle. ,in Indicates the direction angle of the local contour of the upper anterior teeth. This indicates the direction angle of the local contour of the lower anterior teeth. If... If the angle exceeds a preset threshold, it indicates an abnormality in the local relationship between the upper and lower anterior teeth. This parameter represents the difference in posture between the upper and lower anterior teeth in the local contour direction, that is, the size of the angle between the local shapes of the upper and lower anterior teeth.

[0094] By adding local relational feature modeling, this invention enables clearer distinction between anterior crossbite and other categories such as anterior deep overbite and deep overjet.

[0095] (6) Key rules for identifying anterior crossbite To enhance the recognition stability of anterior crossbite in the seven classification categories, this invention can set a key discrimination rule layer for anterior crossbite in addition to the classification model.

[0096] In a preferred embodiment, the probability of outputting the anterior crossbite category is increased when the anterior crossbite discrimination rule meets at least two of the following conditions: Exceeding the preset threshold; Satisfy reverse coverage; Exceeding the preset threshold; Exceeding the preset threshold; at least two perspectives show consistent anterior tooth reversal.

[0097] In one embodiment, if the original predicted probability of anterior crossbite is Then, when the key discrimination rule is met, the execution probability is increased: Where R is the rule satisfaction score and λ is the enhancement coefficient. The rule satisfaction score R is used to characterize the degree to which the current sample meets the key discrimination rule for anterior crossbite, and is defined as: ; in, Indicates parameters Whether the preset threshold is exceeded; if it is met, the value is 1; otherwise, the value is 0. Indicates parameters Whether the reverse cover condition is met; if it is, set to 1, otherwise set to 0. Indicates parameters Whether the preset threshold is exceeded; if it is met, the value is 1; otherwise, the value is 0. Indicates parameters Whether the preset threshold is exceeded; if it is met, the value is 1; otherwise, the value is 0. This indicates whether at least two perspectives show a consistent anterior teeth reversal relationship; if so, the value is 1, otherwise it is 0. to These are the weight coefficients for the corresponding rules. Preferably, when all rules have equal weight: ; at this time: ; The value of R ranges from 0 to 1. The larger the value of R, the higher the degree to which the current sample meets the key discrimination rule for anterior crossbite.

[0098] (7) Multi-perspective reliability weighted fusion After obtaining the global and local features of the images from each viewpoint, this invention determines the fusion weights of different viewpoints based on the image quality score, the visible proportion of the key anterior teeth region, and the boundary clarity of each viewpoint, rather than directly performing equal-weighted stitching.

[0099] In a preferred embodiment, the weight of the i-th viewpoint It can be calculated using the following formula: ; in This represents the quality score of the i-th viewpoint. This indicates that the score is visible in the key area. The score represents the clarity of the boundary. , , , where is the weighting coefficient. The critical boundary sharpness score of the anterior teeth can be calculated based on the edge response intensity, gradient magnitude, boundary continuity, or sharpness evaluation results of the anterior incisal line, local contact boundary, or local contour boundary of the anterior teeth.

[0100] The fusion method can employ weighted summation, weighted concatenation, or attention-weighted fusion. Preferably, local features are fused first, and then combined with global features before being fed into the classifier, to highlight the contribution of anterior tooth local analysis in key categories.

[0101] (8) Seven-class classification model and hierarchical classification strategy To highlight the identification of anterior crossbite while retaining the classification range of the seven types of dental malformations, this invention preferably employs a hierarchical classification strategy. This hierarchical classification strategy includes a coarse classification stage and a fine classification stage. The coarse classification stage initially divides the input samples into larger abnormal categories, while the fine classification stage further outputs the final category results within the corresponding coarse classification branches. By adopting a "coarse classification first, then fine classification" approach, category confusion during direct classification of the seven categories can be reduced, improving the discrimination stability of key categories, especially anterior crossbite.

[0102] In one implementation, the input to the classification stage is a comprehensive feature obtained by weighted fusion of multi-view reliability. This comprehensive feature is obtained by fusing global and local features extracted from the frontal occlusion image, left occlusion image, and right occlusion image, respectively. The global features characterize the overall dental arch arrangement, arch contour, upper and lower tooth relationship, and occlusal morphology, while the local features primarily characterize the local relationships of the anterior teeth, including the positional difference parameters before and after normalization. Coverage direction parameters Boundary contact ratio parameters Anterior tooth contour angle parameters In addition, auxiliary features such as key point distance, contour curvature, texture and edge intensity.

[0103] The first stage is coarse classification.

[0104] The coarse classifier is used to roughly classify samples into categories such as anterior tooth relationship abnormalities, vertical abnormalities, dental arch space abnormalities, protrusion abnormalities, and normal or indeterminate categories. Among them, the anterior tooth relationship abnormality category includes anterior crossbite and anterior deep overbite; the vertical abnormality category includes open bite and deep overbite; the dental arch space abnormality category includes dentition crowding and dentition gaps; and the protrusion abnormality category includes bimaxillary protrusion.

[0105] In a preferred embodiment, the coarse classifier can be implemented using a convolutional neural network classifier, such as ResNet, MobileNet, EfficientNet, or lightweight variants thereof, and outputs the probability of each coarse classification category through fully connected layers and Softmax. The coarse classification stage mainly uses the overall morphological features after multi-view fusion for preliminary discrimination. Among them, the anterior tooth relationship abnormality category focuses more on the local relationship features of the anterior teeth, the vertical abnormality category focuses more on the vertical overlap and separation features of the upper and lower anterior teeth, the dental arch space abnormality category focuses more on the density and spacing features of the tooth arrangement, and the protrusion abnormality category focuses more on the overall protrusion degree of the anterior teeth and the lateral contour forward movement features. The system takes the coarse classification result corresponding to the highest probability to proceed to the next stage; if the highest probability is lower than a preset threshold, a normal or uncertain prompt is output.

[0106] The second stage involves detailed classification.

[0107] Based on the coarse classification results, the corresponding sub-classifiers are called to output the final seven-class results. Each sub-classifier can be implemented using a convolutional neural network classifier, a multilayer perceptron classifier, or a two-branch classifier that uses a combination of local and global features as input.

[0108] When the anterior tooth relation abnormality class is selected, the anterior tooth relation abnormality subclassifier is invoked to perform a binary classification between anterior crossbite and anterior deep overbite. This subclassifier is preferentially used. , , , Including local features of the anterior teeth, to enhance the ability to discriminate local occlusal relationships of the anterior teeth; When a vertical anomaly is selected, a vertical anomaly subclassifier is invoked to perform binary classification between open bite and deep overbite. This subclassifier primarily uses features such as anterior vertical overbite, open bite distance, and local contact height distribution. When a dental arch spatial anomaly is selected, a dental arch spatial anomaly subclassifier is invoked to perform binary classification between dentition crowding and interdental spaces. This subclassifier primarily uses features such as dentition density, tooth overlap, interdental space width, and dental arch continuity. When the protrusion abnormality class is selected, the protrusion abnormality subclassifier is invoked to confirm and distinguish bimaxillary protrusion. This subclassifier focuses on features such as the overall degree of protrusion of the upper and lower anterior teeth, the direction of labial inclination, and the trend of lateral contour forward movement.

[0109] In a preferred embodiment, the first-stage coarse classifier and each sub-classifier in the second stage can be implemented using a structure of "shared feature extraction backbone network + independent classification head" to reduce redundant computation and facilitate lightweight deployment. For the branch of abnormal anterior tooth relationship, it is preferable to increase the weight of local anterior tooth relationship features in the input of the sub-classifier; when the fine classification result is anterior crossbite, the output probability can be further modified by combining the key discrimination rule of anterior crossbite to improve the screening stability of this key category.

[0110] Fine-classify and output the probability values ​​of the seven categories. These correspond to dentition crowding, dentition gaps, bimaxillary protrusion, open bite, deep overbite, deep overjet, and anterior crossbite, respectively. The system selects the category with the highest probability as the candidate category; if the candidate category is anterior crossbite, the final confidence level is adjusted by combining key discrimination rules; if the confidence level of the coarse or fine classification is insufficient, the system outputs an uncertain result or a further inspection suggestion. Through the above hierarchical classification strategy, the system can maintain coverage of the seven types of malformations while focusing on using more targeted local features in key category branches, thereby improving the classification accuracy and robustness under ordinary intraoral image conditions.

[0111] (9) Classification output and confidence control The system output includes: seven categories of malocclusion classification results, corresponding probability values ​​for each category, classification confidence level, key screening prompts for anterior crossbite, and prompts for re-examination or further examination.

[0112] (10) Training dataset construction and label processing To support the training of a seven-class classification model, this invention can construct a multi-view intraoral image dataset. The dataset construction process includes: acquiring the patient's frontal occlusal image, left occlusal image, and right occlusal image; acquiring the corresponding medical record text or label information; associating images with labels based on patient number, shooting time, and medical records; standardizing and mapping the labels to a seven-class classification system; and manually verifying some samples.

[0113] The training samples are divided into training, validation, and test sets at 70%, 15%, and 15% respectively. To enhance model robustness, the following data augmentation methods can be used: random rotation; random brightness perturbation; random contrast perturbation; random cropping; random noise addition; and slight translation or affine transformation.

[0114] When the number of samples of anterior crossbite is relatively small, the learning of key categories can be enhanced by class reweighting, hard sample mining, or oversampling.

[0115] (11) Model training parameters The training parameters for the classification model can be set as follows: the optimizer uses Adam or AdamW; the initial learning rate is 1.e-4 ~5 e-4 Batch size: 8–32; Training epochs: 50–200; Weight decay: 1. e-5 ~1 e-3 The number of rounds to be stopped early is 8 to 20.

[0116] The loss function can be a weighted cross-entropy loss. In practice, to strengthen the anterior crossbite category, a higher category weight can be assigned to this category.

[0117] (12) Performance evaluation indicators In this invention, the model evaluation metrics include at least: accuracy, precision, recall, F1 score, confusion matrix, Top-K accuracy, etc. For details, please refer to Table 1 for model evaluation metrics.

[0118] Table 1 Model Evaluation Indicators

[0119] For the key category of anterior crossbite, it is preferable to separately calculate the anterior crossbite accuracy rate, anterior crossbite recall rate, anterior crossbite F1 value, and confusion rate between anterior deep overbite and deep overjet, in order to verify the improvement effect of the present invention in the key category.

[0120] Figure 2 A dental malformation classification device based on various granularities, provided for one or more embodiments of this specification, includes: At least one processor and bus; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to: Acquire multi-view intraoral images of the current sample, wherein the multi-view intraoral images include at least a frontal occlusal image, a left occlusal image, and a right occlusal image; Tooth region detection is performed on the multi-view intraoral images to locate key anterior tooth regions for analyzing anterior occlusal relationships; Based on the aforementioned key anterior tooth region, local features reflecting the local occlusal relationship of the upper and lower anterior teeth are extracted, and global features of the multi-view intraoral images are also extracted. These local features include normalized anterior-posterior position difference parameters, coverage direction parameters, boundary contact ratio parameters, and anterior tooth contour angle parameters extracted based on the key anterior tooth region, wherein: The normalized position difference parameter is obtained based on the position of the lower anterior tooth reference point in the anterior-posterior direction, the position of the upper anterior tooth reference point in the anterior-posterior direction, and the average width of the upper central incisor or the reference width of the anterior tooth region. The normalized position difference parameter is used to characterize the degree of anterior displacement of the lower anterior teeth relative to the upper anterior teeth in the anterior-posterior direction. It is normalized by the average width of the upper central incisor or the reference width of the anterior tooth region to reduce the influence of differences in tooth size and imaging scale among different individuals on the results. When the normalized position difference parameter is greater than the first preset threshold, it indicates that the degree of anterior displacement of the lower anterior teeth is relatively large. The coverage direction parameter is obtained based on a normalized discriminant function constructed from the relative positional relationship, directional relationship, and coverage order of the upper anterior incisal edge line and the lower anterior incisal edge line. The coverage direction parameter is used to determine whether the relationship where the upper anterior teeth should cover the lower anterior teeth has reversed. When the coverage direction parameter satisfies the reverse coverage condition, it is marked as the anterior teeth reverse bite direction feature is established. The boundary contact ratio parameter is obtained based on the length of the local abnormal contact boundary of the anterior teeth and the total boundary length of the anterior teeth analysis area. The boundary contact ratio parameter is used to characterize the proportion of abnormal contact boundaries in the local area of ​​the anterior teeth and to reflect the degree of abnormality of the local contact mode of the upper and lower anterior teeth. The anterior tooth contour angle parameter is obtained based on the local contour direction angle of the upper anterior teeth and the local contour direction angle of the lower anterior teeth. The anterior tooth contour angle parameter is used to characterize the posture difference between the upper and lower anterior teeth in the local contour direction. The fusion weight of each view image is determined based on the quality score of each view image in the multi-view intraoral image, the visibility ratio of the key anterior tooth region, and the boundary clarity of the key anterior tooth region. Based on the fusion weights of the images from each viewpoint, the local features and the global features are fused together, and the determined features are coarsely classified to obtain a coarse classification result that includes at least one of the following: anterior tooth relationship abnormality, vertical abnormality, dental arch space abnormality, protrusion abnormality, and normal or uncertain class. During the classification process, if it is determined that the current sample meets the key discrimination rules for anterior crossbite, the output probability of the anterior crossbite category is increased. The key discrimination rule is determined based on whether at least two parameters contained in the local feature exceed the corresponding threshold. The output probability of the anterior crossbite category is determined based on the original probability of the anterior crossbite category and the rule satisfaction score. The rule satisfaction score is used to characterize the degree to which the current sample satisfies the key discrimination rule for anterior crossbite. The rule satisfaction score is determined based on whether each parameter in the local feature exceeds the corresponding preset threshold. The preset threshold corresponding to each parameter in the local feature is determined based on the preset conditions of each parameter in the local feature. Based on the coarse classification result, the sub-classifier corresponding to the coarse classification result is called to perform fine classification and output the dental malformation classification result.

[0121] This specification provides one or more embodiments of a non-volatile computer storage medium storing computer-executable instructions, which, when executed by a computer, can perform the following: Acquire multi-view intraoral images of the current sample, wherein the multi-view intraoral images include at least a frontal occlusal image, a left occlusal image, and a right occlusal image; Tooth region detection is performed on the multi-view intraoral images to locate key anterior tooth regions for analyzing anterior occlusal relationships; Based on the aforementioned key anterior tooth region, local features reflecting the local occlusal relationship of the upper and lower anterior teeth are extracted, and global features of the multi-view intraoral images are also extracted. These local features include normalized anterior-posterior position difference parameters, coverage direction parameters, boundary contact ratio parameters, and anterior tooth contour angle parameters extracted based on the key anterior tooth region, wherein: The normalized position difference parameter is obtained based on the position of the lower anterior tooth reference point in the anterior-posterior direction, the position of the upper anterior tooth reference point in the anterior-posterior direction, and the average width of the upper central incisor or the reference width of the anterior tooth region. The normalized position difference parameter is used to characterize the degree of anterior displacement of the lower anterior teeth relative to the upper anterior teeth in the anterior-posterior direction. It is normalized by the average width of the upper central incisor or the reference width of the anterior tooth region to reduce the influence of differences in tooth size and imaging scale among different individuals on the results. When the normalized position difference parameter is greater than the first preset threshold, it indicates that the degree of anterior displacement of the lower anterior teeth is relatively large. The coverage direction parameter is obtained based on a normalized discriminant function constructed from the relative positional relationship, directional relationship, and coverage order of the upper anterior incisal edge line and the lower anterior incisal edge line. The coverage direction parameter is used to determine whether the relationship where the upper anterior teeth should cover the lower anterior teeth has reversed. When the coverage direction parameter satisfies the reverse coverage condition, it is marked as the anterior teeth reverse bite direction feature is established. The boundary contact ratio parameter is obtained based on the length of the local abnormal contact boundary of the anterior teeth and the total boundary length of the anterior teeth analysis area. The boundary contact ratio parameter is used to characterize the proportion of abnormal contact boundaries in the local area of ​​the anterior teeth and to reflect the degree of abnormality of the local contact mode of the upper and lower anterior teeth. The anterior tooth contour angle parameter is obtained based on the local contour direction angle of the upper anterior teeth and the local contour direction angle of the lower anterior teeth. The anterior tooth contour angle parameter is used to characterize the posture difference between the upper and lower anterior teeth in the local contour direction. The fusion weight of each view image is determined based on the quality score of each view image in the multi-view intraoral image, the visibility ratio of the key anterior tooth region, and the boundary clarity of the key anterior tooth region. Based on the fusion weights of the images from each viewpoint, the local features and the global features are fused together, and the determined features are coarsely classified to obtain a coarse classification result that includes at least one of the following: anterior tooth relationship abnormality, vertical abnormality, dental arch space abnormality, protrusion abnormality, and normal or uncertain class. During the classification process, if it is determined that the current sample meets the key discrimination rules for anterior crossbite, the output probability of the anterior crossbite category is increased. The key discrimination rule is determined based on whether at least two parameters contained in the local feature exceed the corresponding threshold. The output probability of the anterior crossbite category is determined based on the original probability of the anterior crossbite category and the rule satisfaction score. The rule satisfaction score is used to characterize the degree to which the current sample satisfies the key discrimination rule for anterior crossbite. The rule satisfaction score is determined based on whether each parameter in the local feature exceeds the corresponding preset threshold. The preset threshold corresponding to each parameter in the local feature is determined based on the preset conditions of each parameter in the local feature. Based on the coarse classification result, the subclassifier corresponding to the coarse classification result is called to perform fine classification, and the dental malformation classification result is output.

[0122] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0123] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0124] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0125] In the embodiments provided in this application, it should be understood that the disclosed apparatus / network devices and methods can be implemented in other ways. For example, the apparatus / network device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0126] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0127] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The aforementioned units can be implemented in hardware or software.

[0128] If the integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0129] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 this application, and should all be included within the protection scope of this application.

Claims

1. A method for classifying dental malformations based on various granularities, characterized in that, include: Acquire multi-view intraoral images of the current sample, wherein the multi-view intraoral images include at least a frontal occlusal image, a left occlusal image, and a right occlusal image; Tooth region detection is performed on the multi-view intraoral images to locate key anterior tooth regions for analyzing anterior occlusal relationships; Based on the aforementioned key anterior tooth region, local features reflecting the local occlusal relationship of the upper and lower anterior teeth are extracted, and global features of the multi-view intraoral images are also extracted. These local features include normalized anterior-posterior position difference parameters, coverage direction parameters, boundary contact ratio parameters, and anterior tooth contour angle parameters extracted based on the key anterior tooth region, wherein: The normalized position difference parameter is obtained based on the position of the lower anterior tooth reference point in the anterior-posterior direction, the position of the upper anterior tooth reference point in the anterior-posterior direction, and the average width of the upper central incisor or the reference width of the anterior tooth region. The normalized position difference parameter is used to characterize the degree of anterior displacement of the lower anterior teeth relative to the upper anterior teeth in the anterior-posterior direction. It is normalized by the average width of the upper central incisor or the reference width of the anterior tooth region to reduce the influence of differences in tooth size and imaging scale among different individuals on the results. When the normalized position difference parameter is greater than the first preset threshold, it indicates that the degree of anterior displacement of the lower anterior teeth is relatively large. The coverage direction parameter is obtained based on a normalized discriminant function constructed from the relative positional relationship, directional relationship, and coverage order of the upper anterior incisal edge line and the lower anterior incisal edge line. The coverage direction parameter is used to determine whether the relationship where the upper anterior teeth should cover the lower anterior teeth has reversed. When the coverage direction parameter satisfies the reverse coverage condition, it is marked as the anterior teeth reverse bite direction feature is established. The boundary contact ratio parameter is obtained based on the length of the local abnormal contact boundary of the anterior teeth and the total boundary length of the anterior teeth analysis area. The boundary contact ratio parameter is used to characterize the proportion of abnormal contact boundaries in the local area of ​​the anterior teeth and to reflect the degree of abnormality of the local contact mode of the upper and lower anterior teeth. The anterior tooth contour angle parameter is obtained based on the local contour direction angle of the upper anterior teeth and the local contour direction angle of the lower anterior teeth. The anterior tooth contour angle parameter is used to characterize the posture difference between the upper and lower anterior teeth in the local contour direction. The fusion weight of each view image is determined based on the quality score of each view image in the multi-view intraoral image, the visibility ratio of the key anterior tooth region, and the boundary clarity of the key anterior tooth region. Based on the fusion weights of the images from each viewpoint, the local features and the global features are fused together, and the determined features are coarsely classified to obtain a coarse classification result that includes at least one of the following: anterior tooth relationship abnormality, vertical abnormality, dental arch space abnormality, protrusion abnormality, and normal or uncertain class. During the classification process, if it is determined that the current sample meets the key discrimination rules for anterior crossbite, the output probability of the anterior crossbite category is increased. The key discrimination rule is determined based on whether at least two parameters contained in the local feature exceed the corresponding threshold. The output probability of the anterior crossbite category is determined based on the original probability of the anterior crossbite category and the rule satisfaction score. The rule satisfaction score is used to characterize the degree to which the current sample satisfies the key discrimination rule for anterior crossbite. The rule satisfaction score is determined based on whether each parameter in the local feature exceeds the corresponding preset threshold. The preset threshold corresponding to each parameter in the local feature is determined based on the preset conditions of each parameter in the local feature. Based on the coarse classification result, the subclassifier corresponding to the coarse classification result is called to perform fine classification, and the dental malformation classification result is output.

2. The method according to claim 1, characterized in that, After acquiring the multi-view intraoral image of the current sample, the method further includes: The image quality of the multi-view intraoral images is evaluated to obtain evaluation results. The image quality evaluation includes the evaluation of at least one of the following indicators: sharpness, brightness, contrast, visible proportion of tooth area, occlusion proportion, occlusal closure integrity, image tilt angle, and degree of view deviation. If the evaluation result does not meet the preset requirements, a retake prompt will be sent.

3. The method according to claim 1, characterized in that, The output probability of the anterior crossbite category satisfies ,in, The output probability is the anterior crossbite category. The original probability for the anterior crossbite category. The rule satisfaction score is defined as follows: λ is the enhancement coefficient. ； in, This indicates whether the position difference parameter before and after normalization exceeds the first preset threshold. If it does, the value is 1; otherwise, it is 0. This indicates whether the coverage direction parameter satisfies the reverse coverage condition; if it does, it is set to 1, otherwise it is set to 0. This indicates whether the boundary contact ratio parameter exceeds the second preset threshold. If it does, the value is 1; otherwise, the value is 0. This indicates whether the anterior tooth region contour angle parameter exceeds the third preset threshold. If it meets the threshold, the value is 1; otherwise, the value is 0. This indicates whether at least two perspectives show a consistent reverse relationship between the anterior teeth; if so, the value is 1, otherwise it is 0. to These are the weighting coefficients for the corresponding rules.

4. The method according to claim 1, characterized in that, The step of detecting tooth regions in the multi-view intraoral images to locate key anterior tooth regions for analyzing anterior occlusal relationships includes: The tooth main region is detected and segmented in the multi-view intraoral image by using an object detection model fused with a semantic segmentation model or an instance segmentation model to obtain the segmentation result. The anterior tooth region is extracted from the segmentation results based on geometric positional relationships; Within the anterior tooth region, the upper anterior tooth region and the lower anterior tooth region are identified; In the left occlusal image and the right occlusal image, locate the anterior tooth extension area and the anterior-posterior tooth transition area; Based on the upper anterior tooth region, the lower anterior tooth region, the anterior tooth extension region, and the anterior-posterior tooth transition region, at least one key point or key line is extracted to characterize the anterior tooth occlusal relationship. The key point or key line includes: the incisal edge point of the upper central incisor, the incisal edge point of the lower central incisor, the local contour center point of the upper and lower anterior teeth, the anterior tooth midline, the incisal edge connection line, and the local anterior dental arch contour line.

5. The method according to claim 1, characterized in that, The extraction of global features from the multi-view intraoral images includes: The multi-view intraoral images are input into a feature extraction network to obtain global features that characterize the overall arrangement of the dental arch, the dental arch contour, the relationship between the upper and lower teeth, and the occlusal morphology. The feature extraction network adopts a convolutional neural network backbone structure.

6. A dental malformation classification device based on various granularities, characterized in that, include: At least one processor and bus; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to implement the granularity-based dental malformation classification method as described in any one of claims 1-5.

7. A non-volatile computer storage medium, characterized in that, The device stores computer-executable instructions, which, when executed by a computer, enable the tooth malformation classification method based on various granularities as described in any one of claims 1-5.

Images