Children multi-generation tooth classification method based on double-cooperation mechanism

Abstract

The invention discloses a child multi-generation tooth classification method based on a double-cooperation mechanism, and the method comprises the steps: taking a child tooth curved surface sectional image as a global image, positioning a child multi-generation tooth multi-generation position from the global image, and obtaining an ROI feature map as a local image; performing multi-generation tooth classification by using the global image and the local image to obtain a first classification result and a second classification result; and fusing the first classification result and the second classification result by adopting a fusion strategy based on category label probability to obtain a final classification result. According to the invention, a positioning-classification cooperation mechanism is designed, and a high-quality target area is provided for a classification module through precise positioning of an ROI area; and a global-local cooperative double-branch mechanism is combined, a global branch captures an image overall structure, a local branch focuses ROI detail features, a final classification result is obtained through probability fusion, the influence of deviation generated by sample imbalance on classification is relieved, the recognition effect is further improved, and the model generalization ability is enhanced.

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 objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for classifying supernumerary teeth in children based on a dual-cooperative mechanism includes the following steps:

[0008] Step 1: Using the panoramic tomographic images of children's teeth as global images, locate the locations of multiple supernumerary teeth in the images to obtain ROI feature maps as local images;

[0009] Step 2: Use the global image and local image to classify supernumerary teeth respectively, and obtain the first classification result and the second classification result;

[0010] Step 3: Using a fusion strategy based on category label probabilities, the first classification result and the second classification result are fused to obtain the final classification result.

[0011] Compared with the prior art, the beneficial effects of the present invention are:

[0012] To address the above problems, the method proposed in this invention includes the following three key innovations and contributions:

[0013] (1) Inspired by the clinical experience of diagnosing supernumerary teeth through manual observation of curved tomographic images, a classification method with a dual collaborative mechanism of "localization-classification" and "global-local" is proposed for supernumerary teeth. To address the problem of the small proportion of supernumerary tooth areas and strong background interference in curved tomographic images, a "localization-classification" collaborative mechanism is designed. By accurately locating the ROI region, a high-quality target region is provided for the classification module. At the same time, the MPDIoU loss function is introduced to achieve pixel-level alignment between the predicted bounding box and the ground truth bounding box, accurately covering the target features and suppressing irrelevant background noise. For the complex location of supernumerary teeth, the dual branch mechanism of "global-local" collaboration is combined. The global branch captures the overall structure of the image, while the local branch focuses on the detailed features of the ROI. The final classification result is obtained by using probability fusion, which alleviates the influence of the bias caused by sample imbalance on the classification, further improves the recognition effect, and enhances the generalization ability of the model.

[0014] (2) To address the problem of insufficient representation and poor adaptability of supernumerary teeth features due to the complex background of curved tomographic sections, a morphological fusion spatial adaptive attention module (MFSA) is introduced into the classification module. By adaptively adjusting the feature weights, redundant information is suppressed and key features of supernumerary teeth are strengthened, which effectively improves the tooth feature representation capability and solves the problem of missed detection.

[0015] (3) To address the problem that the imbalance of the dataset class leads to the complex feature distribution and nonlinear enhancement of normal teeth and supernumerary teeth, and the insufficient fitting ability of traditional fully connected layers, Efficient-KAN is used to replace the traditional fully connected layer. This is used to solve the bottleneck of nonlinear fitting ability in the feature-to-diagnostic category mapping, better capture the relationship between features, and effectively handle more complex nonlinear relationships. Attached Figure Description

[0016] Figure 1 This is a diagram of supernumerary teeth.

[0017] Figure 2 This is a diagram illustrating supernumerary teeth with abnormalities.

[0018] Figure 3 This is a framework diagram of the dual-cooperation mechanism classification module of the present invention.

[0019] Figure 4 This is a diagram of the "location-classification" collaborative mechanism architecture of the present invention.

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

[0021] Figure 6 This is a comparative case of identifying supernumerary teeth (the bottom left shows other methods, and the bottom right shows the method of this invention).

[0022] Figure 7 These are heatmaps of different models. Detailed Implementation

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

[0024] Supernumerary teeth are a common dental developmental abnormality in children. If not addressed promptly, they can lead to problems such as abnormal eruption of permanent teeth, crowded teeth, jaw deformities, and decay of adjacent teeth. Current diagnosis relies heavily on physician experience, presenting challenges such as strong subjectivity and a high rate of missed diagnoses. A certain percentage of cases are discovered incidentally during non-targeted imaging examinations.

[0025] Curved radiographs, as a fundamental examination in pediatric dentistry, can reduce the subjective influence of doctors and improve detection rates when combined with computer assistance. However, existing models lack specificity in feature extraction, making it difficult to meet actual clinical needs. Based on the clinical diagnostic process of doctors regarding supernumerary teeth, this invention proposes a clinically inspired classification method for pediatric supernumerary teeth based on a dual-collaboration mechanism. By constructing two collaborative mechanisms—"localization-classification" and "global-local"—it effectively improves recall. Drawing on the doctor's "localization first, judgment later" approach, a morphological fusion spatial adaptive attention module is designed with ResNeXt as the backbone network to highlight key regional features, and Efficient-KAN is introduced as a fully connected layer to enhance the nonlinear representation of complex features, achieving accurate classification of supernumerary teeth. Addressing the doctor's habit of combining overall image with local details in diagnosis, the "global-local" mechanism independently classifies the entire image and the ROI region separately, integrating the results of the two branches by fusing category probability information. Experiments show that this method performs excellently in the task of classifying supernumerary teeth in children, with an accuracy of 0.99, precision of 0.94, recall of 0.96, F1 score of 0.95, and Kappa coefficient of 0.94. Compared with existing studies, its accuracy and recall are higher, providing stronger technical support for clinical practice.

[0026] The dual-cooperative mechanism for classifying supernumerary teeth in children proposed in this invention is as follows: Figure 3 As shown, this method employs a dual-cooperation mechanism design of "location-classification" collaboration and "global-local" collaboration. Its main steps can be described as follows:

[0027] Step 1: Using the panoramic tomographic image of a child's teeth as the global image, locate the ROI feature map from it as the local image;

[0028] Step 2: Use the global image and local image to classify supernumerary teeth respectively, and obtain the first classification result and the second classification result;

[0029] Step 3: Using a fusion strategy based on category label probabilities, the first classification result and the second classification result are fused to obtain the final classification result.

[0030] The principle of this invention, in which the "localization-classification" collaborative mechanism is designed to address the problem of existing models failing to accurately focus on the Region of Interest (ROI) in the identification of supernumerary teeth in children, such as... Figure 4 This mechanism adopts the idea of ​​"localization first, classification later". First, the localization module accurately extracts the ROI region and uses it as the input of the subsequent classification model. This allows for feature extraction and enhancement based on focusing on key regions, ultimately outputting a more discriminative classification result.

[0031] This "localization-classification" collaborative mechanism, based on the design concept of "decoupling-information focusing," aims to overcome the performance bottleneck of traditional classification models, which struggle to accurately capture the locational correlations of supernumerary teeth and focus on differences in locational features. First, through task decoupling, the classification process for pediatric supernumerary teeth is divided into two relatively independent stages: the localization module performs coarse-grained target localization, while the classification module focuses on fine-grained lesion discrimination. This avoids mutual interference between localization and classification tasks during feature sharing, allowing detection to focus more on target region identification, and classification to focus on the fine-grained judgment of local structures. Second, through information focusing, a high-confidence ROI screening mechanism in the localization stage is used to extract target regions containing certain background information from complex backgrounds (such as normal teeth and bone structures) in curved tomographic images, effectively suppressing the interference of irrelevant information on the classification model. Finally, through computational optimization strategies, the classification module processes only a small number of ROI regions from the detection output. This mechanism structurally aligns with the clinical approach of "localization first, diagnosis later" in oral medical imaging, while simultaneously achieving professional decoupling of the model structure and synergistic optimization of computational efficiency.

[0032] In the ROI localization stage, target recognition is challenging due to the large area of ​​the curved tomographic slice, the small size of the supernumerary tooth region, and the complex background. Therefore, in step 1 of this invention, YOLOv8 is used to adapt to the small target detection task of supernumerary teeth in children. YOLOv8 significantly improves the detection capability of small, blurred-edge targets by utilizing multi-scale feature fusion and dynamic receptive field mechanisms. A good balance is achieved between detection accuracy, efficiency, and resource consumption, providing a stable and reliable foundation for subsequent classification tasks.

[0033] While YOLOv8 possesses strong basic localization capabilities, its default traditional loss function still has limitations when dealing with complex targets such as supernumerary teeth in children. Common object detection loss functions, such as GIoU and DioU, primarily focus on the overlap between the predicted and ground truth bounding boxes during calculation, failing to comprehensively characterize the geometric relationships between bounding boxes. This is especially problematic when the target has an irregular shape or small size, easily leading to decreased localization accuracy. In pediatric curved tomographic images, supernumerary teeth often exhibit irregular structures such as conical, curved, or malformed shapes. Traditional IoU-based loss functions fail to adequately consider factors such as center point offset and aspect ratio differences, resulting in biased detection results. To address this, this invention introduces MPDIoU (Multi-Perspective Distance-IoU) as an alternative loss function, fusing spatial relationships and scale differences between bounding boxes from multiple perspectives to effectively improve the model's localization accuracy on complex structural targets.

[0034] The MPDIoU loss function calculates the distances between the predicted bounding box and the ground truth bounding box in the horizontal, vertical, and diagonal directions, achieving multi-directional distance perception and enhancing adaptability to irregular shapes. It introduces an aspect ratio penalty term to ensure the predicted bounding box's proportions more closely resemble the ground truth bounding box, achieving pixel-level alignment. Furthermore, it smooths the gradients in non-overlapping regions, avoiding training oscillations. Its formula is:

[0035]

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

[0037] Accordingly, in step 1 of the present invention, the ROI feature map is detected from the global image based on the YOLOv8 network using the MPDIoU loss function as the localization module.

[0038] In the classification stage, the classification module of this invention selects ResNeXt as the backbone network. The input of the classification module is a global image or a local image. That is, the supernumerary teeth are classified with the global image and the local image as input respectively, and the first classification result and the second classification result are obtained.

[0039] ResNeXt enhances feature representation through a grouped convolution mechanism, efficiently capturing subtle differences in morphology and texture between supernumerary teeth and normal teeth. It maintains strong feature extraction stability, especially in small sample scenarios, making it suitable for classification tasks requiring fine-grained discrimination, such as those involving supernumerary teeth in children. Its main structure is... Figure 4 This can be seen in the fact that it includes multiple residual blocks.

[0040] However, ResNeXt struggles to focus on target areas when processing large-format panoramic imagery. In the classification of supernumerary teeth in children, the complex and varied tooth morphology, along with the highly complex background structure in curved tomographic sections, presents significant challenges to feature extraction and accurate classification. Its limited ability to capture minute structures and edge details restricts its performance in practical applications.

[0041] To address this, this invention designs a dual attention mechanism that integrates spatial priors and tooth morphological features—the Morphology-Fused Spatial Adaptive Attention Module (MFSA)—and specifically places it within some or all of the residual blocks of ResNeXt to enhance the model's ability to focus on supernumerary tooth regions. Specifically, the MFSA module can be placed after the last convolutional block of the residual blocks.

[0042] refer to Figure 4 The MFSA module of this invention mainly performs the following steps:

[0043] First, for the input tooth feature map F∈R C×H×W Average pooling and max pooling are performed along the channel axis (C-axis) to obtain two two-dimensional feature maps. and It can capture the spatial information of curved tomographic slices and stitch these two two-dimensional feature maps together along the channel axis to form a new feature map.

[0044]

[0045] Wherein, the tooth feature map F is the output of the last convolutional block of the residual block, such as... Figure 4 There are 3 convolutional blocks in the residual block. The MFSA module is placed at the output position of the last 1×1 convolutional block. The output of the 1×1 convolutional block is the tooth feature map F. The output of the MFSA module is then summed element-wise with the input of the residual block to obtain the final output of the residual block.

[0046] Secondly, for Apply a convolutional kernel to generate a spatial attention map M. S .

[0047] Specifically, the spliced ​​feature map can be... Apply a 7×7 convolution kernel to generate a spatial attention map M. S ∈R 1×H×W , means as follows:

[0048]

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

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

[0051] Specifically, using 1×1 convolution pairs F gap F is obtained by performing feature mapping. r By reducing the number of channels in intermediate features, the number of parameters is reduced from C×C to C×(C / r), where C is the number of channels in the tooth feature map F, and r is a predefined compression ratio. This design significantly reduces the computational burden while maintaining feature representation capabilities. (Obtain F) gap and F r The process can be described as follows:

[0052] F gap =GAP(F)∈R C×1×1

[0053] F r =Conv 1×1 (F gap )∈R C×1×1

[0054] Finally, based on the adjacency matrix A, the enhanced feature map F is... r Perform dynamic graph convolution operation to obtain feature weights F out and map it back to F gap The results obtained are consistent with F gao The recalibrated feature map f is obtained by splicing. refined This is the output of the MFSA module.

[0055] The core of the classification module in this invention lies in adjusting the weights between feature vertices through dynamic graph convolution operations. This process is based on the adjacency matrix A = A0 × A1 + A2, which consists of three parts:

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

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

[0058] 3) Learnable adjacency matrix A2: Represents the relationship between feature vertices. Through backpropagation and dynamic learning, it can capture the complex relationships between feature vertices.

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

[0060]

[0061] Where W is the weight of the 1D convolutional layer. It is a normalized adjacency matrix. The normalization operation is implemented through the degree matrix D, ensuring that the sum of the weights of each feature vertex is 1.

[0062]

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

[0064] The feature weights F generated by the graph convolution operation out It was then mapped back to F. gap , as the weights for channel attention. This process is achieved through a linear embedding function: F out The process involves a 1×1 convolutional layer, followed by normalization to the [0,1] range using the sigmoid function, and finally, the result is compared with F. gao Perform element-wise multiplication.

[0065] Feature map F after recalibration according to the present invention refined This can more effectively suppress redundant information and enhance key features.

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

[0067]

[0068] In the classification of supernumerary teeth in children, due to significant differences in their morphological structure, the extracted tooth features are highly complex and diverse, leading to complex feature distribution and nonlinear enhancement due to class imbalance. Traditional fully connected layers, when mapping these complex features to specific diagnostic categories, are limited by linear transformation characteristics and struggle to capture the nonlinear relationships between features, thus affecting classification performance. To address this issue, this invention introduces Efficient-KAN to replace the traditional fully connected layer, further enhancing ResNeXt's ability to fit nonlinear relationships to complex input features, thereby improving overall classification performance. Specifically, this manifests as enhanced ResNeXt's ability to express and discriminate complex features of supernumerary teeth, more effectively modeling nonlinear relationships between features, and thus improving classification accuracy and robustness. The specific implementation steps are shown in the algorithm below.

[0069] Step 001: Construct the spline node grid: Based on the input range grid range = [a,b] and spacing Construct nodes: grid = [a-kh,...,b+kh] (k = spline_order);

[0070] Step 002: Initialize the linear weights base_weight using the Kaiming method:

[0071] Step 003: Initialize spline weights weight ; The spline coefficients are obtained:

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

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

[0074] Step 006: Map the activated input using linear weights: y_base = x′·W_base T ;

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

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

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

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

[0079] Step 011: Spline weight refit. Using the current input and the old output, refit the spline weights: W_spline = curve2coeff(x,y_spline);

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

[0081] Step 013: Multiple KANLinear layers are stacked and connected in series to form a complete network:

[0082] The original KAN requires expanding the input into a high-dimensional tensor to achieve a non-linear expression of the activation function. While it possesses some modeling capability, it suffers from high memory overhead, low computational efficiency, and limited ability to fit features from complex distributions. To address this bottleneck, Efficient-KAN introduces a B-spline function as a linear combination of fixed basis functions, significantly improving the model's ability to express non-linear relationships while avoiding redundant high-dimensional expansion operations and achieving efficient forward and backward propagation. Specifically, the B-spline function in Efficient-KAN is calculated as follows:

[0083]

[0084] Among them, B k (x) is a B-spline basis function, c k These are the corresponding coefficients, and K is the number of basis functions. Compared to the complex tensor construction of the original KAN, this method completes the characterization of the nonlinear mapping relationship between input features and output with lower computational cost. It is particularly suitable for highly nonlinear structures in supernumerary teeth images in the mixed dentition stage of children, such as ectopic tooth growth, overlapping arrangement, morphological variations and blurred boundaries.

[0085] For sparsity regularization, the L1 regularization on the original input samples is replaced with the weight L1 regularization commonly used in neural networks. The regularization loss of the original KAN is:

[0086]

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

[0088]

[0089] Where, ω j This refers to network weights. This strategy not only reduces computational overhead but also effectively suppresses non-target region interference in the detection of supernumerary teeth in children, highlighting a small number of clinically valuable lesions, thereby improving the discriminativeness and interpretability of the test results.

[0090] Furthermore, 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, specifically expressed as follows:

[0091] enable_spline_weight=spline_weight×spline_scaler

[0092] Here, `spline_weight` represents the initial weights corresponding to the B-spline basis functions, and `spline_scaler` is a learnable parameter used to adjust the scale of the B-spline weights. This further enhances the model's adaptability to differences in feature scales, maintaining stable and effective nonlinear modeling capabilities even when dealing with significant grayscale variations or size differences in supernumerary tooth target regions. The final effective weights, `enable_spline_weight`, are obtained by multiplying the initial weights by the learnable scale factor `spline_scaler`.

[0093] In the clinical diagnosis of supernumerary teeth in children, doctors need to focus on observing the number of supernumerary teeth (single or multiple), their arrangement (whether they encroach on the space of the normal dentition), and their growth direction (vertical, horizontal, or inverted). A judgment is made through a linkage analysis of the overall dentition structure and local lesion characteristics. However, the growth and development characteristics of children's dentition lead to diverse tooth morphologies, and imaging is often interfered with by complex backgrounds such as permanent tooth germs and unerupted teeth. Classification models are prone to misjudgment and missed detection, especially in the identification of fine-grained lesion features.

[0094] To address this issue, this invention proposes a "global-local" collaborative mechanism: global features focus on the overall structure of the child's dentition, such as the symmetry of the dental arch during the transition from deciduous to permanent teeth and changes in the interdental spaces, which can help determine the impact of the number and arrangement of supernumerary teeth on the overall dentition. This avoids misjudgments caused by interference from local structures at the macroscopic level, consistent with the approach of doctors to first grasp global abnormalities through panoramic radiographs. Local features delve into the microscopic details of supernumerary teeth, accurately identifying typical morphologies such as conical crowns and curved roots, while capturing subtle differences in growth direction, effectively identifying concealed impacted supernumerary teeth, compensating for the shortcomings of the global model in judging details of minute lesions. Finally, through a probabilistic fusion strategy at the label layer, a more stable and comprehensive classification decision is achieved.

[0095] In the task of classifying supernumerary teeth, due to the high individual differences in the number, arrangement, and growth direction of teeth, and the complexity of image backgrounds, different base classifiers often exhibit significant differences in performance on different samples, leading to blurred decision boundaries between categories. Against this backdrop, a fusion strategy based on category label probabilities (i.e., soft voting) is not only a conventional result integration method but also a highly suitable and optimized solution for the task characteristics. Compared to hard voting, which relies solely on majority vote, soft voting, by weighted averaging of the category probabilities output by each base classifier and selecting the category with the highest probability as the final prediction, can fully utilize the model's confidence information to achieve more refined and stable discrimination. This strategy effectively coordinates the information trade-off between global structure and local details, mitigates the fluctuations in model performance caused by scene differences, significantly improves the generalization ability and prediction accuracy of the ensemble model, and outperforms other fusion methods overall.

[0096] For each category, calculate the average probability of the two predicted outcomes, the first classification result and the second classification result, to obtain the average disease probability. and average health probability

[0097]

[0098] Subsequently and By comparing the results, the category with the higher probability is selected as the final classification result:

[0099] like Then it is abnormal;

[0100] like Then it is normal;

[0101] In the verification experiment of this invention, the dataset (CSTD) was obtained from the Department of Pediatrics, Stomatological Hospital of Xi'an Jiaotong University. Curved surface radiographs were taken using a cone-beam computed tomography (CBCT) system (KaVo Shanghai Dental Medical Instruments Co., Ltd., model: OP300-1). The data consisted of curved surface radiographs of children aged 5 to 8 years. This age group is in the mixed dentition stage, a critical period where deciduous and permanent teeth coexist and are gradually replaced. Dental characteristics are relatively complex during this period, and it is the optimal time for early detection of supernumerary teeth. The data collection spanned from 2019 to 2023, covering 406 cases of supernumerary teeth and 2498 normal controls.

[0102] In this study, the experimental environment was based on PyTorch 2.3.0 and Python 3.12 (running on Ubuntu 22.04), and the hardware configuration included an RTX 3080 GPU (10GB VRAM) and a 12-core Intel(R) Xeon(R) Platinum 8255C CPU, coupled with the CUDA 12.1 acceleration framework. During model training, the Cross-Entropy Loss function was used, which is a standard loss function widely used in classification tasks. Furthermore, 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 a metric for evaluating the consistency of a classification model, taking into account potential random consistency factors in the classification results. Its calculation formula is as follows:

[0104]

[0105] Where, p o It is the sum of the number of correctly classified samples in each class divided by the total number of samples, which is the overall classification accuracy. Let's assume the number of true samples in each class is a1, a2, ..., a... c The predicted number of samples for each class are b1, b2, ..., b c If the total number of samples is n, then we have

[0106]

[0107] The confusion matrix, a core tool for evaluating the performance of classification models, visually represents the mapping relationship between the model and the predicted class in matrix form. Matrix elements represent the matching relationship between the actual and predicted classes, and its performance is primarily evaluated based on metrics such as accuracy, precision, recall, and F1 score. Specifically:

[0108] (1) Accuracy reflects the overall classification ability of the model and is 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 samples predicted as supernumerary teeth that are actually true positives.

[0110] (3) The recall rate assessment model is used to evaluate the detection capability of the target lesion and calculate the proportion of correctly identified teeth in actual supernumerary tooth samples.

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

[0112] By comprehensively analyzing multi-dimensional indicators, the diagnostic efficacy of the model in the task of detecting supernumerary teeth can be systematically evaluated. The performance metrics are calculated using the following formula:

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

[0114]

[0115] TP represents true positive, TN represents true negative, FP represents false positive, and FN represents the false negative count of case testing.

[0116] To verify the impact of various network components on the performance of recognizing supernumerary teeth in children, the following ablation experiment was designed. The effectiveness of each module and the modified loss function was verified one by one using the controlled variable method. Based on the original ResNeXt classification model, each component was gradually introduced. The control variables and corresponding experimental groups (e.g., E1 represents experimental group 1) are shown in Table 1.

[0117] Table 1 Ablation Experiment

[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. System verification of the proposed dual-cooperation mechanism and its key modules shows that as the network gradually integrates its components, various performance indicators continuously optimize, demonstrating a good performance improvement trend. Through... Figure 5The recall rate trend chart shown demonstrates that after introducing the "location-classification" collaborative mechanism, the recall rate jumped from 0.7595 to 0.9114, an increase of 15.19 percentage points, compared to E4 and E5. This verifies the effectiveness of the mechanism in enhancing key feature representation and reducing false negatives. Under the synergistic effect of the dual collaborative mechanism, the recall rate significantly improved compared to E6 and E7, and the F1 score also improved markedly.

[0122] It is worth noting that, due to the improvement in the fully connected layer structure, the model tends to prioritize the recall rate of the positive class during training. This characteristic has significant clinical value in reducing missed diagnoses of supernumerary teeth, thus resulting in a slight decrease in precision. In summary, the results demonstrate that the proposed method achieves substantial breakthroughs in accurate ROI localization and key feature extraction. Each module in the mechanism plays a crucial role in the supernumerary tooth recognition task, synergistically driving a significant improvement in overall classification performance.

[0123] To comprehensively verify the advantages of the proposed method for classifying supernumerary teeth in children in terms of accuracy, recall, and overall model performance, a comparative experiment was designed, selecting current mainstream models and representative methods from relevant literature as references. The experimental results are shown in Table 3, where methods are sorted from low to high recall for easy comparison of differences in missed diagnoses.

[0124] Table 3 Comparative Experiments

[0125]

[0126]

[0127] The performance differences of the models in the table show that the performance of different network structures in the children's supernumerary tooth classification task is closely related to their feature extraction ability and adaptability to the task scenario. Lightweight networks such as SqueezeNet and AlexNet, due to their fewer network layers and smaller parameter scale, have limited feature extraction capabilities and are unable to capture the subtle features of supernumerary teeth. Therefore, their accuracy and precision are at a lower level.

[0128] Classic deep networks such as EfficientNetV2 and Inception v3 have improved feature extraction capabilities by deepening the network or optimizing convolution methods, but they are not well adapted to targets with high morphological variability, such as supernumerary teeth. Although some models can cover more samples (with higher recall), their accuracy is low due to their weak ability to distinguish similar features, and their overall performance is at a medium level.

[0129] High-performance backbone networks such as ConvNeXt, ResNeXt, and ResNet, with their deeper network structures, residual connections, or grouped convolutions, can extract richer hierarchical features and model the complex morphology of supernumerary teeth and their relationships with surrounding tissues more accurately, thus achieving an accuracy exceeding 0.84. Among them, ResNeXt enhances feature diversity and improves accuracy through grouped convolutions, while ResNet's residual structure focuses more on the stable propagation of features, resulting in better recall.

[0130] In the comparative analysis, the method proposed in this invention demonstrates superior recognition performance for some supernumerary tooth samples that existing networks failed to accurately identify or completely missed, and can successfully identify them. Specific examples include... Figure 6 .

[0131] In conclusion, the method of this invention improved the ability to identify malformed supernumerary teeth. Figure 2 As shown in the example, the method of the present invention can correctly determine, and further combined with Figure 7 The heatmap shown intuitively demonstrates that this performance improvement stems from the model's effective capture of key features, and the accuracy of its feature extraction is visually confirmed in the visualization results. This method not only improves the accuracy of identifying malformed supernumerary teeth but also reduces the missed detection of variant samples, ultimately achieving a significant lead in accuracy (0.9851) and recall (0.962), fully validating its suitability for the task of classifying supernumerary teeth in children.

[0132] In summary, this invention addresses the challenges in classifying pediatric supernumerary teeth, including complex image backgrounds, difficulty in identifying local feature differences, inability to focus on the Region of Interest (ROI), and the need for clinical diagnostic assistance. Inspired by clinical knowledge, it proposes a feature-adaptive dual-collaboration mechanism for classifying pediatric supernumerary teeth. This method effectively improves diagnostic performance by optimizing feature extraction and constructing a dual-collaboration strategy of "localization-classification" and "global-local." Experimental results show that the model performs excellently in terms of accuracy (98.51%), precision (93.83%), and recall (96.2%), significantly overcoming the bottleneck of traditional methods that struggle to focus on the ROI of pediatric supernumerary teeth, demonstrating strong potential for clinical application.

[0133] From a clinical perspective, the kappa coefficient reached 0.94, demonstrating a high degree of consistency with expert diagnoses. This effectively reduces subjective differences in human diagnosis and significantly improves the standardization of diagnosis in primary healthcare institutions. Based on this, the method's accurate identification of supernumerary teeth not only assists clinicians in making more accurate diagnostic decisions but also lays a solid foundation for subsequent multi-classification tasks such as supernumerary tooth typing. Future research will delve deeper into multi-classification tasks, covering the determination of the number of single and multiple supernumerary teeth, as well as the identification of different growth directions such as upright and inverted positions. This will drive the model towards a higher dimension of "precise typing diagnosis," thereby providing more comprehensive and detailed intelligent support for clinical decision-making.

Claims

1. A method for classifying supernumerary teeth in children based on a dual-cooperative mechanism, characterized in that, Includes the following steps: Step 1: Using the panoramic tomographic images of children's teeth as global images, locate the locations of multiple supernumerary teeth in the images to obtain ROI feature maps as local images; Step 2: Use the global image and local image to classify supernumerary teeth respectively, and obtain the first classification result and the second classification result; Step 3: Using a fusion strategy based on category label probabilities, the first classification result and the second classification result are fused to obtain the final classification result.

2. The method for classifying supernumerary teeth in children based on a dual-cooperative mechanism according to claim 1, characterized in that, Step 1 involves detecting local images from the global image using a YOLOv8 network employing the MPDIoU loss function.

3. The method for classifying supernumerary teeth in children based on a dual-cooperative mechanism according to claim 1, characterized in that, Step 2 involves classifying the supernumerary teeth using a global or local image as input to the classification module. The classification module employs ResNeXt as its backbone network and introduces a morphological fusion spatial adaptive attention module into some or all of its residual blocks to enhance its ability to focus on the supernumerary tooth region. The morphological fusion spatial adaptive attention module performs the following steps: First, the input tooth feature map F is subjected to average pooling and max pooling 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 F is the output of the last convolutional block of the residual block; Secondly, for Apply a convolutional kernel to generate a spatial attention map M. S ; Then, M S Perform global average pooling to compress its spatial dimensions into a single feature vector F. gap , will F gap Feature mapping is performed using a linear embedding function to enhance feature representation, resulting in the enhanced feature map F. r ; Finally, based on the adjacency matrix A, the enhanced feature map F is... r Perform dynamic graph convolution operation to obtain feature weights F out and map it back to F gap The results obtained are consistent with F gap The recalibrated feature map F is obtained by splicing. refined This is the output of the MFSA module.

4. The method for classifying supernumerary teeth in children based on a dual-cooperative mechanism according to claim 3, characterized in that, right Apply a 7×7 convolution kernel to generate a spatial attention map M. S , means as follows: Here, σ is the Sigmoid activation function, which is used to normalize the values ​​of the attention map to the range [0,1].

5. The method for classifying supernumerary teeth in children based on a dual-cooperative mechanism according to claim 3, characterized in that, F is obtained through a 1×1 convolution pair gap F is obtained by performing feature mapping. r The number of parameters is reduced from C×C to C×(C / r), where C is the number of channels in the tooth feature map F and r is a predefined compression ratio.

6. The method for classifying supernumerary teeth in children based on a dual-cooperative mechanism according to claim 3, characterized in that, The adjacency matrix A = A0 × A1 + A2, where A0 is the identity matrix, representing the weight of each feature vertex to ensure that each feature vertex has the same weight in the initial state; A1 is the self-attention mechanism, which uses the softmax function to adjust the weight of each feature vertex. r Normalize to a probability distribution, then diagonalize the result to form a diagonal matrix, A1, which reflects the importance of each feature vertex; A2 is a learnable adjacency matrix that represents the relationship between feature vertices. Through backpropagation and dynamic learning, it can capture the complex relationships between feature vertices.

7. The method for classifying supernumerary teeth in children based on a dual-cooperative mechanism according to claim 3, characterized in that, The dynamic graph convolution operation is represented as follows: Where W is the weight of the 1D convolutional layer. It is a normalized adjacency matrix. The normalization operation is implemented through the degree matrix D, ensuring that the sum of the weights of each feature vertex is 1. Where each element D of the degree matrix D ii Defined as ∑ j A ij +∈, where ∈ is a very small value used to avoid division by zero, A ij It is the element in the i-th row and j-th column of the adjacency matrix A, representing the correlation strength between the two channels.

8. The method for classifying supernumerary teeth in children based on a dual-cooperative mechanism according to claim 3, characterized in that, The feature weights F are embedded using a linear embedding function. out Map back to F gap The method is as follows: F out The process involves a 1×1 convolutional layer, followed by normalization to the [0,1] range using the sigmoid function, and finally, the result is compared with F. gap Perform element-wise multiplication.

9. The method for classifying supernumerary teeth in children based on a dual-cooperative mechanism according to any one of claims 3 to 8, characterized in that, The backbone network uses Efficient-KAN as the fully connected layer, and the enable_standalone_scale_spline option is enabled by default to control the learnable scale in the activation function. The final effective weights enable_spline_weight are specifically represented as follows: enable_spline_weight=spline_weight×spline_scaler Here, 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 weights.

10. The method for classifying supernumerary teeth in children based on a dual-cooperative mechanism according to any one of claims 1 to 8, characterized in that, The fusion strategy based on category label probabilities calculates the average disease probability of the first classification result and the second classification result. Calculate the average health probability of the first classification result and the second classification result. Subsequently and The categories with higher probabilities are compared, and the category with the higher probability is selected as the final classification result.

Images