Classification method for children's supernumerary teeth based on double cooperation mechanism

Abstract

The application discloses a kind of children's supernumerary teeth classification methods based on double cooperation mechanism, to children's dental curve tomogram image as global image, from which the ROI feature map of the multiple supernumerary teeth position of children is positioned as local image;Global image and local image are used respectively to classify supernumerary teeth, to obtain first classification result and second classification result;Using the fusion strategy based on class label probability, the first classification result and second classification result are fused, to obtain final classification result.The application designs "positioning-classification" cooperation mechanism, by accurately positioning ROI area, provide high-quality target area for classification module;And combined with "global-local" cooperation double branch mechanism, global branch captures overall structure of image, local branch focuses on ROI detail features, using probability fusion obtains final classification result, alleviates the deviation of sample imbalance and the influence on classification, further improves recognition effect, and enhances model generalization ability.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and relates to medical image classification, and in particular to a method for classifying supernumerary teeth in children based on a dual-cooperative mechanism. Background Technology

[0002] Supernumerary teeth are a type of dental developmental abnormality, characterized by extra teeth beyond the normal number of teeth in the dentition, such as... Figure 1 Impacted teeth are commonly seen in pediatric patients. Early screening and intervention are of great significance in preventing complications such as impacted teeth, reducing treatment complexity, and improving treatment success rates, and are crucial for the healthy development of children's teeth.

[0003] Currently, the diagnosis of supernumerary teeth mainly relies on the manual interpretation of imaging examinations (including panoramic radiographs, CBCT, and periapical radiographs). Due to the high radiation dose of CBCT, it is generally not the first choice and is only considered when routine low-dose radiographic examinations cannot provide sufficient clinical diagnostic information. Periapical radiographs, as a two-dimensional imaging tool, have limitations in identifying supernumerary teeth, such as missing spatial information, which can easily lead to missed or misdiagnosed cases. Furthermore, in complex cases, it is difficult to accurately assess the spatial position of the teeth and their relationship with adjacent structures, which may adversely affect the formulation of subsequent treatment plans. In contrast, panoramic radiographs, due to their ease of operation, relatively low radiation dose, and ability to provide a panoramic view of the jawbone and dentition, have become a commonly used primary screening method in clinical practice, especially suitable for early screening of supernumerary teeth in children and large-scale epidemiological studies.

[0004] With the rapid development of deep learning technology, it has shown significant advantages in the field of medical image analysis, enabling accurate detection and classification of oral lesions such as dental caries and periodontal disease, effectively reducing misdiagnosis and missed diagnosis rates, alleviating the burden on doctors, and optimizing the allocation of medical resources. However, research on the identification of supernumerary teeth in children started relatively late. Common problems include low recall rates and unconvincing test data. Existing studies mostly involve the direct application of models, failing to adequately adapt to the characteristics of supernumerary teeth in children for targeted exploration, making it difficult to accurately locate regions of interest (ROIs), and resulting in insufficient feature extraction capabilities, leading to difficulties in identifying malformed supernumerary teeth (such as...). Figure 2 The model performs poorly in recognizing morphological variations (as shown in the image), with low recall rates, and its overall classification performance needs improvement. Furthermore, existing public datasets lack specific annotations for children with supernumerary teeth, exhibiting issues such as ambiguous age labeling, large age ranges, and insufficient sample size. Private datasets are also insufficient and do not accurately reflect reality, failing to guarantee the model's generalization ability. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, the present invention aims to provide a method for classifying supernumerary teeth in children based on a dual-cooperative mechanism, so as to solve the problems of identification and classification difficulties caused by the small proportion of supernumerary tooth area in curved tomographic films, strong background interference, and complex onset location.

[0006] To achieve the above object, the technical scheme adopted by the present application is:

[0007] A child multiple supernumerary tooth classification method based on a double cooperation mechanism, comprising the following steps:

[0008] Step 1, taking a child tooth curve surface tomogram image as a global image, positioning a child multiple supernumerary tooth multiple position from the global image to obtain an ROI feature map as a local image;

[0009] Step 2, respectively using the global image and the local image for supernumerary tooth classification to obtain a first classification result and a second classification result;

[0010] Step 3, adopting a fusion strategy based on a category label probability to fuse the first classification result and the second classification result to obtain a final classification result.

[0011] Compared with the prior art, the present application has the following beneficial effects:

[0012] To solve the above problems, the method proposed in the present application contains the following three key innovations and contributions:

[0013] (1) Inspired by the experience and knowledge of clinical artificial observation of curve surface tomogram for diagnosing multiple supernumerary teeth, a "positioning-classification" and "global-local" double cooperation mechanism classification method for multiple supernumerary teeth is proposed. In view of the problem that the multiple supernumerary tooth region accounts for a small proportion in the curve surface tomogram and the background interference is strong, the "positioning-classification" cooperation mechanism is designed, and a high-quality target region is provided for the classification module through accurate positioning of the ROI region; at the same time, the MPDIoU loss function is introduced to realize the alignment of the pixel-level prediction box and the real box, accurately cover the target features and suppress irrelevant background noise; in view of the complex multiple supernumerary tooth disease position, the "global-local" cooperation double branch mechanism is combined, the global branch captures the overall structure of the image, the local branch focuses on the detailed features of the ROI, and the final classification result is obtained by using probability fusion, which alleviates the deviation of the sample imbalance to the classification and further improves the recognition effect and enhances the model generalization ability.

[0014] (2) In view of the problem that the multiple supernumerary tooth feature representation is insufficient and the adaptability is poor due to the complex background of the curve surface tomogram, a morphological fusion space self-adaptive attention module (MFSA) is introduced in the classification module, the feature weight is adaptively adjusted, the redundant information is suppressed and the key features of the multiple supernumerary teeth are strengthened, the tooth feature representation ability is effectively improved, and the missed detection problem is solved;

[0015] (3) In view of the problems of complex and nonlinear enhancement of feature distribution of normal teeth and supernumerary teeth caused by unbalanced data set categories, and insufficient fitting capacity of traditional full connection layer, an Efficient-KAN is used to replace the traditional full connection layer to solve the bottleneck of nonlinear fitting capacity in the mapping from features to diagnostic categories, better capture the relationship between features, and effectively handle more complex nonlinear relationships. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a schematic diagram of a supernumerary tooth.

[0017] Figure 2 is a schematic diagram of a supernumerary tooth with deformity.

[0018] Figure 3 is a framework diagram of the dual collaboration mechanism classification module of the present application.

[0019] Figure 4 is a "positioning-classification" collaboration mechanism architecture diagram.

[0020] Figure 5 is a recall rate trend chart.

[0021] Figure 6 is a contrast case of identifying supernumerary teeth (left lower for other methods, right lower for the method of the present application).

[0022] Figure 7 is a heat map of different models. DETAILED DESCRIPTION

[0023] The embodiments of the present application will be described in detail below with reference to the accompanying drawings and examples.

[0024] Supernumerary teeth are common dental developmental abnormalities in children. If not timely intervention, it may cause abnormal eruption of permanent teeth, crowded dentition, jaw deformity and adjacent tooth caries, etc. Current diagnosis relies on the experience of doctors, which has strong subjectivity, high misdiagnosis rate and other challenges, and a certain proportion of cases are accidentally found by non-target image examination.

[0025] As a basic examination of children's dentistry, curved-surface fault pieces combined with computer assistance can reduce the subjective influence of doctors and improve the detection rate. However, the existing models lack pertinence in feature extraction, making it difficult to meet the actual clinical needs. Combined with the judgment process of doctors in clinical diagnosis of multiple teeth, the present application proposes a clinical knowledge inspired multiple teeth classification method for children based on a double collaboration mechanism, which effectively improves the recall rate by constructing two collaboration mechanisms of "positioning-classification" and "global-local".

[0026] The double collaboration mechanism of the multiple teeth classification method for children proposed by the present application is shown in Figure 3 The main steps of the method can be described as:

[0027] Step 1: Taking the curved-surface fault piece image of children's teeth as a global image, the ROI feature map is located as a local image;

[0028] Step 2: Classify multiple teeth using the global image and the local image respectively to obtain the first classification result and the second classification result;

[0029] Step 3: Using a fusion strategy based on class label probability to fuse the first classification result and the second classification result to obtain the final classification result.

[0030] The principle of the present application, wherein the "positioning-classification" collaboration mechanism is to solve the problem that the existing model cannot accurately focus on the ROI region in the identification of multiple teeth for children, as Figure 4 The mechanism adopts the idea of "positioning first, then classification", which first uses the positioning module to accurately extract the ROI region and takes it as the input of the subsequent classification model, so as to focus on the key region and extract and enhance the features, and finally output the classification result with better discrimination.

[0031] The "positioning-classification" collaborative mechanism is based on the design concept of decoupling-information focusing, aiming to break through the performance bottleneck of traditional classification models due to the difficulty in accurately capturing the position correlation of multiple supernumerary teeth and focusing on the differences in position characteristics. First, through task decoupling, the process of classifying children's multiple supernumerary teeth is divided into two relatively independent stages: the positioning module completes the coarse-grained target positioning, and the classification module focuses on the fine-grained lesion discrimination, thereby avoiding the mutual interference of positioning and classification tasks in the feature sharing process, making the detection more focused on the identification of the target area, and the classification focusing on the fine judgment of the local structure. Second, through information focusing, the high-confidence ROI screening mechanism in the positioning stage is used to extract the target area containing certain background from the complex background of the curved tomographic film (such as normal teeth, bone structure), effectively suppressing the interference of irrelevant information on the classification model. Finally, through the calculation optimization strategy, the classification module only processes a small number of ROI regions output by the detection. This mechanism structurally fits the clinical thinking of "positioning first, then diagnosis" in oral medical imaging, while achieving the professional decoupling of the model structure and the collaborative optimization of the calculation efficiency.

[0032] In the ROI region positioning stage, due to the large size of the curved tomographic film, the small size of the supernumerary teeth region, and the complex background, the target recognition is difficult. Therefore, in step 1 of the present application, YOLOv8 is used to adapt to the small target detection task of children's multiple supernumerary teeth. YOLOv8 uses multi-scale feature fusion and dynamic receptive field mechanism to significantly improve the detection ability of small and edge blurred targets. It achieves a good balance between detection accuracy, efficiency and resource consumption, providing a stable and reliable foundation for the subsequent classification task.

[0033] Although YOLOv8 has strong basic positioning ability, its default use of traditional loss function still has certain limitations when dealing with complex targets such as children's multiple supernumerary teeth. Common target detection loss functions such as GIoU and DioU mainly focus on the overlapping area between the predicted box and the real box in the calculation process, and it is difficult to fully describe the geometric relationship between the boundary boxes, especially when the target shape is irregular or the size is small, which can easily lead to a decrease in positioning accuracy. In children's curved tomographic film, multiple supernumerary teeth often have irregular structures such as cone-shaped, curved or deformed, and traditional IoU-based loss functions do not fully consider factors such as center point offset and aspect ratio difference, resulting in deviations in the detection results. Therefore, the present application introduces MPDIoU (Multi-Perspective Distance-IoU) as an alternative loss function, which effectively improves the positioning accuracy of the model on complex structure targets by integrating the spatial relationship and scale difference between the boundary boxes from multiple angles.

[0034] The MPDIoU loss function realizes multi-direction distance perception by calculating the distance of the predicted frame and the real frame in the horizontal, vertical and diagonal directions, and enhances the adaptability to irregular shapes; a length-width ratio penalty term is introduced to ensure that the proportion of the predicted frame is closer to the real multiple teeth, and to realize the pixel-level alignment of the predicted frame and the real frame; and the gradient of the non-overlapping area is also smoothed to avoid training shock. The formula is:

[0035]

[0036] Wherein, the IoU is the ratio of the intersection and union between the predicted bounding box and the real bounding box. d1 is the Euclidean distance between the top left corner point of the predicted bounding box and the top left corner point of the real bounding box. d2 is the Euclidean distance between the right bottom corner point of the predicted bounding box and the right bottom corner point of the real bounding box. w and h are the width and height of the input image respectively.

[0037] Accordingly, in step 1 of the present application, the YOLOv8 network based on the MPDIoU loss function is used as a positioning module to detect the ROI feature map from the global image.

[0038] In the classification stage, the classification module of the present application selects ResNeXt as the backbone network, and the input of the classification module is the global image or the local image, that is, the global image and the local image are respectively input to perform multiple supernumerary tooth classification, and the first classification result and the second classification result are obtained.

[0039] ResNeXt can efficiently capture the subtle differences between multiple supernumerary teeth and normal teeth in shape, texture and other aspects while improving feature expression ability through grouped convolution mechanism. Especially in the small sample scene, it still has strong feature extraction stability and is suitable for classification tasks such as children's multiple supernumerary teeth which need to be finely distinguished. Its main structure can be embodied in Figure 4 , which includes multiple residual blocks.

[0040] However, ResNeXt has difficulty focusing on the target area when processing large-format panoramic images. In the classification task of children's multiple supernumerary teeth, the complex tooth morphology and the high complexity of the background structure in the curved surface tomogram bring significant challenges to feature and accurate classification. The ability to capture small structures and edge details is limited, which limits its performance in practical applications.

[0041] To this end, the application designs a double attention mechanism that fuses spatial priori and tooth morphology features, namely a morphology-fused spatial adaptive attention module (MFSA), and specifically arranges the MFSA in part or all of the residual blocks of ResNeXt to enhance the attention ability of the model to the multiple supernumerary teeth region. The MFSA module can be specifically arranged after the last convolution block of the residual block.

[0042] Reference Figure 4 The MFSA module of the application mainly performs the following steps:

[0043] First, the input tooth feature map F is obtained as follows: C×H×W Average pooling and maximum pooling are performed along the channel axis (C axis) to obtain two two-dimensional feature maps and The two two-dimensional feature maps can capture the spatial information of the curved surface tomogram, and the two two-dimensional feature maps are spliced along the channel axis to form a new feature map

[0044]

[0045] The tooth feature map F is the output of the last convolution block of the residual block, as shown in Figure 4 The residual block has three convolution blocks, and the MFSA module is arranged at the output position of the last 1x1 convolution block. The output of the 1x1 convolution block is the tooth feature map F, and the output of the MFSA module is element-wise summed with the input of the residual block where the MFSA module is located, as the final output of the residual block.

[0046] Secondly, the is applied to a convolution kernel to generate a spatial attention map M S .

[0047] Specifically, the spliced feature map is applied to a 7x7 convolution kernel to generate a spatial attention map M S ∈R 1×H×W , which is represented as follows:

[0048]

[0049] Where σ is the Sigmoid activation function, which is used to normalize the value of the attention map to the range of [0, 1].

[0050] Then, the obtained spatial attention map M S is subjected to global average pooling (GAP) to compress its spatial dimension to a single feature vector F gap ∈R C×1×1This process generates a feature vector with dimension C x 1 x 1 by averaging the feature values of each channel along the spatial dimension (i.e., H x W). Global average pooling effectively preserves the global information of the channel dimension while eliminating the differences in spatial positions. Next, a linear embedding function is applied to F gap to enhance the feature representation, resulting in an enhanced feature map F r .

[0051] Specifically, F gap is mapped to F r using a 1 x 1 convolution, which reduces the number of channels of intermediate features and reduces the number of parameters from C x C to C x (C / r), where C is the number of channels of the tooth feature map F, and r is a predefined compression ratio. This design significantly reduces the computational burden while maintaining the ability of feature representation. The process of obtaining F gap and F r can be described as follows:

[0052] F gap = GAP(F) e R C×1×1

[0053] F r = Conv 1×1 (F gap ) e R C×1×1

[0054] Finally, based on the adjacency matrix A, the enhanced feature map F r is subjected to a dynamic graph convolution operation to obtain the feature weight F out , which is mapped back to F gap . The resulting result is concatenated with F gao to obtain the recalibrated feature map f refined , which is the output of the MFSA module.

[0055] The core of the classification module of the present application is to adjust the weight between the feature vertices through dynamic graph convolution operation. This process is based on the adjacency matrix A = A0 x A1 + A2, which consists of three parts:

[0056] 1) Unit matrix A0: represents the weight of each feature vertex itself, ensuring that each feature vertex has the same weight in the initial state.

[0057] 2) Self-attention mechanism A1: normalize F r to a probability distribution through the softmax function, and then diagonalize the result to form A1, which reflects the importance of each feature vertex.

[0058] 3) Learnable adjacency matrix A2: represents the relationship between feature vertices, dynamically learned through backpropagation, which can capture the complex relationship between feature vertices.

[0059] Based on the adjacency matrix A, the dynamic graph convolution operation is realized by the following formula:

[0060]

[0061] where W is the weight of the 1D convolution layer, is the normalized adjacency matrix, and the normalization operation is realized by the degree matrix D, which ensures that the weight sum of each feature vertex is 1:

[0062]

[0063] where each element D ii of the degree matrix D is defined as ∑ j A ij + ∈, ∈ is a very small value, which is used to avoid division by zero. A ij is the element in the i-th row and j-th column of the adjacency matrix A, which represents the association strength between two channels (feature vertices).

[0064] The feature weight F out generated by the graph convolution operation is then mapped back to F gap as the weight of channel attention. This process is realized by a linear embedding function: F out is passed through a 1x1 convolution layer, then a sigmoid function is applied to normalize it to the range [0, 1], and finally element-wise multiplication is performed with F gao .

[0065] The feature map F refined after recalibration by the present application can more effectively suppress redundant information and strengthen key features.

[0066] Integrating the above steps, the complete MFSA can be represented as:

[0067]

[0068] In the classification research of children's supernumerary teeth, due to the significant difference in morphological structure, the extracted tooth features have high complexity and diversity, and the feature distribution is complex and nonlinear due to the unbalanced categories. When the traditional fully connected layer corresponds these complex features to specific diagnostic categories, it is difficult to capture the nonlinear relationship between features due to the limitation of linear transformation characteristics, thereby affecting the classification performance. To solve this problem, the Efficient-KAN is introduced to replace the traditional fully connected layer in the present application, further improving the nonlinear fitting ability of ResNeXt to complex input features, thereby enhancing the overall classification performance. In the present application, this is specifically manifested as the ability to enhance the expression and discrimination of ResNeXt to the complex features of supernumerary teeth, more effectively model the nonlinear relationship between features, and thus improve the accuracy and robustness of classification. The specific implementation steps are shown in the following algorithm.

[0069] Step 001: Construct spline node grid: According to the input range grid range =[a,b] and interval Construct node: grid=[a-kh,...,b+kh](k=spline_order);

[0070] Step 002: Initialize linear weight base_weight, initialize weight using Kaiming method:

[0071] Step 003: Initialize spline weight spline weight ; Get spline coefficient:

[0072] Step 004: Initialize spline scaling factor, enable spline_scaler, and initialize scaling coefficient:

[0073] Step 005: Apply activation function to input: x'=σ(x);

[0074] Step 006: Map the activated input through linear weight: y_base=x'·W_base T ;

[0075] Step 007: Calculate B-spline basis function: B_i(x)=B-spline basis;

[0076] Step 008: Calculate spline interpolation output, expand spline basis function, and multiply by spline weight: y_spline=flatten(B)·W_spline T ;

[0077] Step 009: Add linear output and spline output to get total output: y = y_base + y_spline;

[0078] Step 010: Update spline grid to dynamically adjust spline nodes according to data distribution: grid = ε grid + (1 - ε) grid_adaptive; uniform

[0079] Step 011: Spline weight refitting, using current input and old output to refit spline weight: W_spline = curve2coeff(x, y_spline);

[0080] Step 012: Calculate regularization term, including L1 and information entropy constraints: L_reg = λ1 ∑ |W| + λ2 H(p), p = |W| / ∑ |W|;

[0081] Step 013: Stack multiple KANLinear layers in series to form a complete network:

[0082] The original KAN needs to expand the input into a high-dimensional tensor to realize the non-linear expression of the activation function, although it has certain modeling ability, but it faces the problems of large memory overhead and low computational efficiency, and the fitting ability of complex distribution characteristics is limited. To solve this bottleneck, Efficient-KAN introduces B-spline function as a fixed basis function of linear combination, which significantly improves the model's expression ability for non-linear relationships, while avoiding redundant high-dimensional expansion operations, and realizes efficient forward and backward propagation. Specifically, the B-spline function in Efficient-KAN is calculated as follows:

[0083]

[0084] where B k (x) is the B-spline basis function, c k is the corresponding coefficient, and K is the number of basis functions. Compared with the complex tensor construction of the original KAN, this method describes the non-linear mapping relationship between input features and output at a lower computational cost, especially suitable for highly nonlinear structures existing in children's mixed dentition stage multiple supernumerary teeth images, such as tooth ectopic growth, overlapping arrangement, morphological variation and boundary ambiguity, etc.

[0085] In terms of sparse regularization, the L1 regularization on the original input sample is replaced by the commonly used weight L1 regularization in neural networks. The regularization loss of the original KAN is:

[0086]

[0087] ​where f(x i ) is the network output, and λ is the regularization coefficient. The regularization loss of Efficient-KAN is:

[0088]

[0089] where ω j is the network weight. This strategy not only reduces the computational overhead, but also effectively suppresses non-target region interference terms in child supernumerary tooth detection, highlighting a small number of lesions with clinical value, thereby improving the discriminability and interpretability of the detection results.

[0090] In addition, for the learnable scale of the activation function in the original implementation, Efficient-KAN, the enable_standalone_scale_spline option (enabled by default) is proposed to control the learnable scale in the activation function, which is specifically represented as:

[0091] enable_spline_weight = spline_weight x spline_scaler

[0092] where spline_weight is the initial weight corresponding to the B-spline basis function, and spline_scaler is a learnable parameter used to adjust the scale of the B-spline weight, which can further enhance the model's adaptability to feature scale differences. In the case of significant gray scale changes or size differences in the supernumerary tooth target area, the model can still maintain stable and effective nonlinear modeling capabilities. The final effective weight enable_spline_weight is obtained by multiplying the initial weight by the learnable scale factor spline_scaler.

[0093] In the clinical diagnosis of child supernumerary teeth, doctors need to focus on observing the number of supernumerary teeth (single or multiple), the arrangement (whether to occupy the space of normal dentition), and the growth direction (vertical, horizontal, or inverted), and make judgments through the linkage analysis of global dentition structure and local lesion characteristics. The growth and development characteristics of child dentition lead to diverse tooth morphology, and the image is often disturbed by complex backgrounds such as permanent tooth germs and unerupted teeth, making classification models prone to misjudgment and missed detection, especially in fine-grained lesion feature recognition.

[0094] To solve the problem, the application proposes a "global-local" cooperation mechanism: the global feature focuses on the overall structure of the child's dentition, such as the symmetry of the dental arch during the transition period of deciduous teeth and permanent teeth, the change of the interproximal space, etc., which can assist in judging the number and arrangement of supernumerary teeth on the overall dentition, avoiding misjudgment caused by local structure interference in the macro level, which is consistent with the doctor's idea of grasping the overall abnormality through panoramic film first; the local feature deeply analyzes the micro details of the supernumerary teeth, accurately identifies typical forms such as conical tooth crown and curved tooth root, and captures the subtle differences in growth direction, effectively identifying hidden embedded supernumerary teeth, making up for the short board of the global model in the judgment of small lesions details, and finally realizing more stable and comprehensive classification decision through the probability fusion strategy of the label layer.

[0095] In the supernumerary tooth classification task, due to the high individual differences in tooth number, arrangement and growth direction, and the complex image background, the performance of different base classifiers on different samples often differs significantly, resulting in blurred decision boundaries between classes. In this context, the fusion strategy based on class label probability (i.e. soft voting) is not only a common result integration method, but also a preferred solution highly consistent with the task characteristics. Compared with hard voting which only relies on majority voting, soft voting can fully utilize the confidence information of the model by weighting and averaging the class probabilities output by each base classifier, and selecting the class with the highest probability as the final prediction result, achieving more accurate and stable discrimination. This strategy effectively balances the information between global structure and local details, alleviates the fluctuations in model performance caused by scene differences, and significantly improves the generalization ability and prediction accuracy of the integrated model, outperforming other fusion methods.

[0096] For each class, calculate the probability average of the first classification result and the second classification result and the healthy probability average

[0097]

[0098] Then compare and , and select the class with higher probability as the final classification result:

[0099] If , then abnormal;

[0100] If , then normal;

[0101] In the verification experiment of the present application, the data set (CSTD) is derived from the Department of Pediatric Dentistry of the Oral Hospital of Xi'an Jiaotong University, and the curved tomographic images are taken by using the cone beam computed tomography equipment (Kavo Shanghai Dental Medical Instruments Co., Ltd., Model: OP300-1). The data is the curved tomographic image of children aged 5 to 8 years old. This age group is in the mixed dentition stage, that is, the key period of coexistence and gradual replacement of deciduous teeth and permanent teeth. The characteristics of teeth at this stage are relatively complex, and it is the best period for early detection of supernumerary teeth. The data collection period spans from 2019 to 2023, covering 406 cases of supernumerary teeth and 2498 cases of normal controls.

[0102] In this study, the experimental environment is based on PyTorch 2.3.0 and Python 3.12 (running on Ubuntu 22.04 operating system), and the hardware configuration includes RTX 3080 GPU (10 GB video memory) and 12-core Intel(R) Xeon(R) Platinum 8255C CPU, with CUDA 12.1 acceleration framework. During the model training process, the CrossEntropyLoss function, which is a standard loss function widely used in classification tasks, was used. In addition, the Adam optimizer was used, and the learning rate was set to 0.0001 to achieve efficient and stable gradient optimization.

[0103] The Kappa coefficient is an index for evaluating the consistency of a classification model, considering the random consistency factor that may exist in the classification results. Its calculation formula is as follows:

[0104]

[0105] where p o is the sum of the number of correctly classified samples in each class divided by the total number of samples, that is, the overall classification accuracy. Assuming that the true sample size of each class is a1, a2, …, a c , and the predicted sample size of each class is b1, b2, …, b c , the total sample size is n, then

[0106]

[0107] The confusion matrix (Confusion Matrix) is a core tool for evaluating the performance of a classification model, which directly presents the mapping relationship between the true class and the predicted class in the form of a matrix. The matrix elements represent the matching relationship between the actual class and the predicted class, and its performance evaluation is mainly based on accuracy (Accuracy), precision (Precision), recall (Recall), and F1 score (F1-Score) and other indicators. Specifically:

[0108] (1) Accuracy reflects the overall classification ability of the model, defined as the proportion of correctly predicted samples to the total samples;

[0109] (2) Precision measures the reliability of positive prediction results, that is, the proportion of actual true positives among samples predicted as supernumerary teeth;

[0110] (3) Recall evaluates the detection ability of the model for target lesions, calculating the proportion of correctly identified samples among actual supernumerary teeth samples;

[0111] (4) F1 score is the harmonic mean of precision and recall, effectively balancing the performance trade-off between the two, suitable for scenarios with uneven data distribution.

[0112] Through comprehensive analysis of multi-dimensional indicators, the diagnostic performance of the model in supernumerary tooth detection tasks can be systematically evaluated. The performance measurement value is calculated using the following formula:

[0113] Confusion Matrix Predicted Positive Predicted Negative ActualPositive TP FN ActualNegative FP TN

[0114]

[0115] Where TP is true positive, TN is true negative, FP is false positive, and FN is false negative count of case detection.

[0116] To verify the influence of each component of the network on the identification performance of children's supernumerary teeth, the following ablation experiments are designed to verify the effectiveness of each module and loss function modification one by one through the control variable method. Based on the original ResNeXt classification model, each component is gradually introduced, and the variables and corresponding experimental groups (such as E1 represents experimental group 1) are as shown in Table 1.

[0117] Table 1 Ablation experiments

[0118]

[0119] Table 2 Experimental results

[0120] Experimental Group Accuracy Precision Recall F1 score E1 0.9268 0.7846 0.6456 0.7082 E2 0.9459 0.8636 0.7215 0.7856 E3 0.9356 0.7763 0.7468 0.7613 E4 0.9547 0.8955 0.7595 0.8214 E5 0.9721 0.8889 0.9114 0.9000 E6 0.9808 0.9250 0.9367 0.9308 E7 0.9851 0.9383 0.9620 0.9500

[0121] The experimental results are shown in Table 2. Through systematic verification of the proposed dual collaboration mechanism and its key modules, it can be seen that the performance indicators continue to optimize in the process of gradually integrating each component, showing a good performance improvement trend. Through the comparison of the performance of the network in the experimental groups, it can be seen that the introduction of each component has a certain effect on the performance of the network, and the dual collaboration mechanism has a significant effect on the performance of the network. Figure 5The recall rate trend chart shows that, after introducing the "positioning-classification" collaborative mechanism, the recall rate jumps from 0.7595 to 0.9114 by comparing E4 and E5, with an increase of 15.19 percentage points, verifying the effectiveness of the mechanism in enhancing key feature representation and reducing missed detection. Under the synergistic effect of the double collaborative mechanism, the recall rate is significantly improved by comparing E6 and E7, and the F1 score is also significantly improved.

[0122] Notably, due to the improvement of the full connection layer structure, the model tends to prioritize the recall rate of positive classes during training, which is clinically valuable for reducing supernumerary tooth misdiagnosis, resulting in a slight decrease in precision. The results show that the proposed method has achieved substantial breakthroughs in ROI precise positioning and key feature extraction, and each module in the mechanism plays a key role in supernumerary tooth identification, synergistically promoting significant improvement in overall classification performance.

[0123] To comprehensively verify the advantages of the proposed supernumerary tooth classification method for children in terms of accuracy, recall rate, and overall model performance, the invention designs a comparative experiment, selecting current mainstream models and representative methods from related literature as references. The experimental results are shown in Table 3, where the methods are sorted by recall rate from low to high for intuitive comparison of misdiagnosis control differences.

[0124] Table 3 Comparative experiment

[0125]

[0126]

[0127] From the performance differences of each model in the table, the performance of different network structures in the supernumerary tooth classification task of children is closely related to their feature extraction ability and adaptability to the task scenario. Lightweight networks such as SqueezeNet and AlexNet have limited feature extraction ability due to their small number of network layers and small parameter size, making it difficult to capture subtle features of supernumerary teeth, resulting in low accuracy and precision.

[0128] Classic deep networks such as EfficientNetV2 and Inception v3, although they have improved feature extraction ability by deepening the network or optimizing the convolution method, they are not well adapted to supernumerary teeth, which have high morphological variability. Some models can cover more samples (high recall rate), but their ability to distinguish similar features is weak, resulting in low precision and overall performance at a medium level.

[0129] High-performance backbone networks such as ConvNeXt, ResNeXt, and ResNet can extract more rich hierarchical features and model the complex morphology of multiple supernumerary teeth and the relationship with surrounding tissues more accurately, thus breaking through 0.84 in accuracy. Among them, ResNeXt enhances feature diversity through grouped convolution to improve precision, and the residual structure of ResNet focuses more on stable feature transmission, so the recall rate is better.

[0130] In comparative analysis, for the multiple supernumerary teeth samples that some existing networks cannot accurately identify or completely miss, the method proposed in the application shows better identification performance and can successfully identify them. For example, Figure 6 .

[0131] Finally, the method improves the identification ability of the deformed multiple supernumerary teeth. For example, as shown in Figure 2 , the method can correctly judge, and further combined with the heat map as shown in Figure 7 , it can be visually seen that the optimization of this identification performance is due to the effective capture of key features by the model, and the accuracy of feature extraction is intuitively verified in the visualization result. This method not only improves the accurate identification ability of deformed multiple supernumerary teeth, but also reduces the missed detection of variant samples, and finally realizes the significant leading of accuracy 0.9851 and recall rate 0.962, fully verifying its adaptability to the classification task of children's multiple supernumerary teeth.

[0132] In summary, the application proposes a feature-adaptive dual collaboration mechanism for children's multiple supernumerary teeth classification method to address the complex image background, difficulty in recognizing local feature differences, inability to focus on ROI area, and clinical auxiliary diagnosis needs in children's multiple supernumerary teeth classification. This method optimizes feature extraction and constructs a "location-classification" and "global-local" dual collaboration strategy to effectively improve the auxiliary diagnosis performance. Experimental results show that the model performs excellently in accuracy (98.51%), precision (93.83%), and recall rate of 96.2%, significantly overcoming the bottleneck of traditional methods that are difficult to focus on the children's multiple supernumerary teeth area (ROI), and has strong clinical application potential.

[0133] From the perspective of clinical value, the kappa coefficient reached 0.94, showing a high degree of consistency with expert diagnosis, effectively reducing the subjective differences in manual diagnosis, and significantly improving the standardization level of diagnosis in primary medical institutions. Based on this, the precise identification ability of the method for supernumerary teeth not only assists in making more accurate diagnostic decisions in clinical practice, but also lays a solid foundation for subsequent multi-classification tasks such as supernumerary teeth typing. Future research will focus on in-depth exploration of multi-classification tasks, including the identification of single and multiple supernumerary teeth, as well as the recognition of different growth directions such as orthosteric and inverted. This will promote the development of the model to a higher dimension of "precise typing diagnosis", providing more comprehensive and detailed intelligent auxiliary support for clinical decision-making.

Claims

1. A method for classifying children's multiple supernumerary teeth based on a double cooperation mechanism, characterized in that, Comprising the following steps: Step 1, taking the child tooth curved surface tomogram image as a global image, positioning the ROI feature map as a local image from the multiple supernumerary tooth multiple position of the child; Step 2, respectively using the global image and the local image for supernumerary tooth classification to obtain the first classification result and the second classification result, the method is as follows: Taking the global image or the local image as the input of the classification module to perform the supernumerary tooth classification, the classification module adopts ResNeXt as the backbone network, and introduces a morphological fusion spatial adaptive attention module in part or all of the residual blocks to enhance the attention ability to the supernumerary tooth region; wherein the morphological fusion spatial adaptive attention module performs the following steps: First, the input tooth feature map Average pooling and max pooling are performed along the channel axis to obtain two two-dimensional feature maps and A new feature map is obtained by splicing along the channel axis The tooth feature map is the output of the last convolution block of the residual block; Second, to Apply a convolution kernel to generate a spatial attention map ; Then, the global average pooling is performed to compress the spatial dimensions into a single feature vector The feature mapping is performed by a linear embedding function to enhance the feature representation, obtaining an enhanced feature map ;​​ Finally, based on the adjacency matrix , the enhanced feature map is subjected to a dynamic graph convolution operation to obtain feature weights , which are mapped back to , and the resulting results are spliced with to obtain a recalibrated feature map , which is output as the output of the morphological fusion space adaptive attention module; Step 3, adopting a fusion strategy based on class label probability to fuse the first classification result and the second classification result to obtain the final classification result.

2. The method for classifying children's multiple supernumerary teeth based on a dual cooperation mechanism according to claim 1, characterized in that, In the step 1, the local image is detected from the global image based on the YOLOv8 network adopting the MPDIoU loss function.

3. The method for classifying children's multiple supernumerary teeth based on a dual cooperation mechanism according to claim 1, characterized in that, To A 7x7 convolution kernel is applied to generate the spatial attention map , which is represented as follows: wherein, is an activation function for normalizing the values of the attention map to the range [0, 1].

4. The method for classifying children's multiple supernumerary teeth based on a dual cooperation mechanism according to claim 1, characterized in that, By convolution pair feature mapping to obtain , the number of parameters is reduced from to , wherein is the number of channels of the tooth feature map is a predefined compression ratio.​ 5. The method for classifying children's multiple supernumerary teeth based on a dual cooperation mechanism according to claim 1, characterized in that, The adjacency matrix ,in It is an identity matrix, representing the weight of each feature vertex, ensuring that each feature vertex has the same weight in the initial state; As a self-attention mechanism, through The function will Normalize to a probability distribution, then diagonalize the result to form a diagonal matrix, i.e. This reflects the importance of each feature vertex; It is a learnable adjacency matrix that represents the relationships between feature vertices. Through backpropagation and dynamic learning, it can capture the complex relationships between feature vertices.

6. The method for classifying children's multiple supernumerary teeth based on a dual cooperation mechanism according to claim 1, characterized in that, The dynamic graph convolution operation is represented as: wherein, is a weight of the one-dimensional convolution layer, is a normalized adjacency matrix, the normalization operation being implemented by the degree matrix ensuring that the sum of the weights for each feature vertex is 1: where each element of the degree matrix is defined as , is a very small value to avoid division by zero, is the element in the th row and i th column of the adjacency matrix j , representing the strength of association between two channels.​ 7. The method for classifying children's multiple supernumerary teeth based on a dual cooperation mechanism according to claim 1, characterized in that, mapping the feature weights back through the linear embedding function , as follows: ​ Will By one Convolutional layer, then apply The function is normalized to The range, and finally with Element-wise multiplication.

8. The method according to any one of claims 1 to 7, wherein the method is based on a dual cooperation mechanism. The backbone network adopts Efficient-KAN as a full connection layer, and the enable_standalone_scale_spline option is enabled by default, which is used to control the learnable scale in the activation function, and finally takes effect on the weight Specifically represented as: wherein, is the initial weight corresponding to the spline basis function, is a learnable parameter to adjust the scale of the spline weight.

9. The method according to any one of claims 1 to 7, wherein the method is based on a dual cooperation mechanism, characterized in that, The fusion strategy based on the category label probability calculates an average of the disease probability of the first classification result and the second classification result The fusion strategy based on the category label probability calculates an average of the health probability of the first classification result and the second classification result Then, the and are compared, and the category with higher probability is selected as the final classification result.

Images