Dental health assessment method and system based on near-infrared transillumination

Abstract

The present application relates to the technical field of tooth detection, and in particular relates to a tooth health assessment method and system based on near-infrared transmission, comprising the following steps: step S1: providing a near-infrared light source with a wavelength of 980 nm to perform transmission on teeth, and collecting a near-infrared transmission image of the teeth through a near-infrared camera; step S2: inputting the near-infrared transmission image into a pre-trained deep learning target detection model; step S3: outputting the positioning and classification results of the lesion area in the teeth through the deep learning target detection model; wherein the deep learning target detection model is a model improved based on the YOLOv8 architecture, and the improvement lies in replacing at least one standard CBS module in the neck network with a CBSv2 module; the CBSv2 module is composed of a deformable convolution DCNv2, a batch normalization BN layer and a SiLU activation function connected in sequence.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tooth detection, in particular to a tooth health assessment method and system based on near-infrared transillumination. BACKGROUND

[0002] Dental caries has become one of the most important public health problems in the world. In the early demineralization stage of caries disease, the enamel can be repaired by remineralization treatment. If no intervention is taken, the acid will penetrate into the dentin, leading to the spread of caries, tissue destruction and damage to the tooth crown.

[0003] Early treatment of dental caries is simpler, shorter in time and less expensive than late treatment. Modern studies comparing the oral conditions of diseased and healthy populations have found that dental caries can act as a source of infection to induce systemic diseases such as coronary heart disease and hypertension. Therefore, early diagnosis and detection are crucial for the treatment and control of dental caries, not only preventing further tooth damage, but also reducing treatment costs and complexity, preventing complications and improving the quality of life.

[0004] The commonly used methods for detecting dental caries are visual inspection (VI), dental probe and parallel transillumination bitewing radiography (BWR). However, visual inspection is prone to miss early caries, especially hidden caries on the adjacent surface of the teeth. The dental probe causes irreversible damage to the hard tissues of the teeth when it is used to probe the enamel by sliding over the surface of the teeth. The X-ray examination has limited sensitivity for detecting early caries, and the image is prone to overlap false images. Moreover, there is a risk of ionizing radiation, and it cannot be detected multiple times in a short period of time, especially for children and pregnant women. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a tooth health assessment method and system based on near-infrared transillumination, aiming to solve the problems of low detection accuracy and radiation risk of the existing caries detection methods.

[0006] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0007] A tooth health assessment method based on near-infrared transillumination, comprising the following steps:

[0008] Step S1: providing a near-infrared light source with a wavelength of 980 nm to transilluminate the teeth, and collecting a near-infrared transillumination image of the teeth by a near-infrared camera;

[0009] Step S2: inputting the near-infrared transillumination image into a pre-trained deep learning target detection model;

[0010] Step S3: outputting a positioning and classification result of the lesion area in the tooth through the deep learning target detection model;

[0011] The deep learning target detection model is a model improved based on a YOLOv8 architecture, and the improvement is that at least one standard CBS module in the neck network is replaced by a CBSv2 module; the CBSv2 module is composed of a deformable convolution DCNv2, a batch normalization BN layer and a SiLU activation function connected in sequence.

[0012] Further, in step S2, the training process of the deep learning target detection model includes the following steps:

[0013] Step S2.1: Data acquisition: providing a near-infrared light source with wavelengths of 850nm, 980nm, 1310nm and 1450nm, respectively, transmitting the tooth sample, and collecting the near-infrared transmission image after the tooth sample transmission through the near-infrared camera, wherein 980nm and 1310nm are the core effective wave bands, and 850nm and 1450nm are used as controls;

[0014] Step S2.2: Data processing: labeling the tooth region in the near-infrared transmission image;

[0015] Step S2.3: Data set division: dividing the labeled data set into a training set, a validation set and a test set;

[0016] Step S2.4: training the improved YOLOv8m network model;

[0017] Step S2.5: detecting the tooth using the trained improved YOLOv8m network model to obtain a tooth evaluation result;

[0018] In step S2.4, the following steps are included:

[0019] Step S2.4.1: setting training parameters: setting the training round number epoch to 1000 rounds, setting the batch training size batch_size to -1, setting the initial learning rate to 0.01, setting the training device to GPU, setting the input image size to 640x640 pixels, setting the workers to 8, i.e. using multi-thread data loading for the data loader, using the automatic selection optimizer, i.e. optimizer='auto', and adding regularization to prevent overfitting, i.e. dropout=0.1;

[0020] Step S2.4.2: inputting the near-infrared transmission image in the training set into the improved YOLOv8m network model, and using the defined loss function for back propagation and parameter updating;

[0021] Step S2.4.3: During the training process, the performance of the YOLOv8m model is evaluated periodically using the validation set, and the training parameters are adjusted according to the evaluation results to obtain the trained improved YOLOv8m network model.

[0022] Further, in step S2.2, the data processing is specifically:

[0023] Step S2.2.1: Use the annotation tool makesense to annotate the tooth region in the near-infrared transillumination image, and divide it into normal, shallow caries, medium caries, deep caries, defect and calculus;

[0024] Step S2.2.2: After annotation, save the annotation information as a YOLO format recognizable by the model.

[0025] Further, in step S2.3, specifically:

[0026] Divide the annotated data set into training set, validation set and test set in the ratio of 7:1:2.

[0027] Further, in step S2.4.2, the specific implementation of DCNv2 includes the following steps:

[0028] Step S2.4.2.1: For each position p0 on the output feature map, generate an offset field through a parallel convolution layer, and the number of channels of the offset field is 2N, where N is the total number of sampling points of the convolution kernel;

[0029] Step S2.4.2.2: Extract the offset matrix corresponding to position p0 from the offset field, which contains the two-dimensional offset Δp n of each sampling point p n of the convolution kernel;

[0030] Step S2.4.2.3: Calculate the output feature value of position p0 using the formula y(p0) = ∑ pn∈R W(p n )·x(p0+ p n +Δp n ), where R is the regular sampling grid of the convolution kernel = {(-1, -1), (-1,0), …, (0, 1), (1,1)}, W(p n ) is the convolution weight of position p n , and x(p0+ p n +Δp n ) is the input feature map.

[0031] Further, when the sampling position p=p0+ p n +Δp nWhen the non-integer coordinates, the value x(p) of the input feature map at position p is calculated by bilinear interpolation method; the bilinear interpolation formula is: x(p) =∑ q G(q, p) ·x(q), wherein q is all integer spatial positions around p, and G(q, p) is a bilinear interpolation kernel.

[0032] Further, the bilinear interpolation kernel G(q, p) can be decomposed into the product of two one-dimensional kernel functions: G(q, p)= g(q x , p x )·g(q y , p y ), wherein g(a, b) = max(0, 1-|a-b|).

[0033] Further, the step S2.4.4 of evaluating the performance of the YOLOv8m network model is further included, specifically:

[0034] Step S2.4.4.1: inputting the near-infrared transillumination images in the divided test set into the trained improved YOLOv8m network model for evaluation;

[0035] Step S2.4.4.2: calculating the precision, recall and average precision value of the YOLOv8m network model;

[0036] Step S2.4.4.3: the precision of the caries detection on the test set of 980nm wavelength is not less than 94.3%, the recall is not less than 97.2%, and the average precision value is not less than 0.979, which is excellent.

[0037] A dental health assessment system for implementing the above method, comprising:

[0038] An optical imaging module comprising a near-infrared light source with a wavelength of 980nm and a near-infrared camera, for collecting near-infrared transillumination images of teeth;

[0039] A data processing and analysis module comprising a memory and a processor, the memory storing a computer program, which when executed by the processor, can implement the dental health assessment method of any one of claims 1-7.

[0040] Further, the system is a portable device, and the optical imaging module and the data processing and analysis module are integrated, and the data processing and analysis module is a mobile terminal or an embedded device loaded with the deep learning target detection model.

[0041] The dental health assessment method and system based on near-infrared transillumination described in the present application have the following beneficial effects:

[0042] The near-infrared light completely avoids the ionizing radiation risk of traditional X-rays, so that sensitive groups such as children and pregnant women can safely and frequently carry out caries screening. The 980nm wavelength near-infrared light source is experimentally verified to be able to optimally penetrate the dental tissue, and form a very high contrast between the healthy area and the caries demineralization area, while effectively suppressing the interference of surface pigments, so as to realize high sensitivity discovery of early caries, interproximal caries and other hidden lesions. The traditional YOLO series network has always relied on two-dimensional regular convolution, but when detecting irregular defects, such convolution is easy to miss key information. Because the defect morphology is variable, the traditional convolution is difficult to accurately identify details. Therefore, the YOLOv8 architecture model is improved, the deformable convolution (DCNv2) is integrated in the feature fusion layer (SPPFv2), and the original convolution is replaced, so that the model can adaptively adjust and accurately fit the boundary of irregular shapes such as caries and calculus, so as to extract more discriminative features, fundamentally improve the detection accuracy and robustness, realize real-time and automatic positioning and classification of lesions, greatly improve the screening efficiency, reduce the dependence on professional physician experience, and lay a foundation for family and community application. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is a framework diagram of a deep learning target detection model of the present application;

[0044] Figure 2 is a Focus structure diagram of the YOLOv8 architecture of the present application;

[0045] Figure 3 is a CSPNet structure optimization schematic diagram of the YOLOv8 architecture of the present application;

[0046] Figure 4 is a CBS module structure diagram of the YOLOv5 architecture;

[0047] Figure 5 is a Bottleneck module structure diagram of the YOLOv5 architecture;

[0048] Figure 6 is a C3 module structure diagram of the YOLOv5 architecture;

[0049] Figure 7 is an ELAN module structure diagram of the YOLOv5 architecture;

[0050] Figure 8 is a C2f module structure diagram of the YOLOv8 architecture of the present application;

[0051] Figure 9 is a structure diagram of the SPP module and the SPPF module of the YOLOv8 architecture of the present application;

[0052] Figure 10 is the FPN+PAN structure diagram of the YOLOv8 architecture of the present application;

[0053] Figure 11 is the Coupled head structure diagram of the YOLOv5 architecture;

[0054] Figure 12 is the Decoupled head structure diagram of the YOLOv8 architecture of the present application;

[0055] Figure 13 is the CIoU schematic diagram of the YOLOv8 architecture of the present application;

[0056] Figure 14 is the principle diagram of the conventional convolution and the principle diagram of the deformable convolution;

[0057] Figure 15 is the implementation process of the deformable convolution of the present application;

[0058] Figure 16 is the near-infrared image of the same type of camera at different wave bands of the same tooth sample in the pre-experiment of the present application;

[0059] Figure 17 is the transillumination diagram of the same caries under the near-infrared camera and the corresponding near-infrared light source at different wavelengths in the optimized camera setting of the present application;

[0060] Figure 18 is the comparison diagram of the visible light image and the near-infrared transillumination of different types of caries of the present application;

[0061] Figure 19 is the comparison diagram of the near-infrared transillumination and the X-ray image of the present application;

[0062] Figure 20 is the near-infrared transillumination detection result diagram of the YOLOv8m model and the YOLOv5s model at 980nm wavelength and 1310nm wavelength of the present application;

[0063] Figure 21 is the flowchart of the tooth health assessment method based on near-infrared transillumination of the present application. DETAILED DESCRIPTION

[0064] The present application will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0065] As shown in Figure 1 and Figure 21 , the present application provides a tooth health assessment method based on near-infrared transillumination, comprising the following steps:

[0066] Step S1: a near-infrared light source with a wavelength of 980nm is provided to transilluminate the tooth, and a near-infrared transillumination image of the tooth is collected by a near-infrared camera;

[0067] Step S2: inputting the near-infrared transillumination image into a pre-trained deep learning target detection model;

[0068] Step S3: outputting the positioning and classification results of the lesion area in the tooth by the deep learning target detection model;

[0069] The deep learning target detection model is a model improved based on the YOLOv8 architecture, and the improvement is that at least one standard CBS module in the neck network is replaced by a CBSv2 module; the CBSv2 module is composed of a deformable convolution DCNv2, a batch normalization BN layer and a SiLU activation function connected in sequence.

[0070] The use of near-infrared light completely avoids the ionizing radiation risk of traditional X-rays, so that sensitive groups such as children and pregnant women can safely and frequently perform caries screening. The 980nm wavelength near-infrared light source has been experimentally confirmed to optimally penetrate tooth tissue and form a very high contrast between healthy areas and caries demineralization areas, while effectively suppressing the interference of surface pigments, thereby achieving high sensitivity in detecting early caries, interproximal caries and other hidden lesions. Traditional YOLO series networks have always relied on two-dimensional regular convolution, but when detecting irregular defects, such convolution is easy to miss key information. Because of the variability of defect morphology, traditional convolution cannot accurately identify details. Therefore, the YOLOv8 architecture model is improved, deformable convolution (DCNv2) is integrated into the feature fusion layer (SPPFv2), replacing the original convolution, so that the model can adaptively adjust and accurately fit the boundaries of irregular shapes such as caries and stones, thereby extracting more discriminative features, fundamentally improving detection accuracy and robustness, realizing real-time and automatic positioning and classification of lesions, greatly improving screening efficiency and reducing dependence on professional physician experience, laying a foundation for family and community applications.

[0071] The YOLOv8 architecture includes an Input layer, a Backbone layer, a Neck layer and a Head layer, the Input layer imports the caries image into the YOLOv8 network, the Backbone layer extracts features, the Neck layer fuses features, and the Head layer outputs results.

[0072] For the input layer, starting with YOLOv4, the YOLO series of algorithms gradually introduced image aliasing enhancement techniques such as Mixup and Mosaic, aiming to increase the diversity of detected backgrounds, the number of targets, and the richness of data. These image aliasing methods will be analyzed in detail in subsequent chapters along with other data augmentation techniques. Although these techniques have significant advantages, continuous use of them throughout the training process may reduce training effectiveness. Therefore, YOLOv8 chose to disable these image aliasing enhancement techniques in the final stage of training to ensure training results.

[0073] For the Backbone layer, the specific operations include the following:

[0074] Focus structure optimization: Figure 2 As shown, the Focus structure segments the original input image into four independent feature layers and stacks them through slicing operations, increasing the channel dimension by four times and thus preserving richer original information. However, while this structure reduces information loss, it increases computational complexity and places higher demands on GPU hardware performance.

[0075] CSPNet Structure: YOLOv3 introduced the Darknet53 network architecture, while YOLOv4 further combined CSPNet (Cross Stage Partial Network) with Darknet53 to propose CSPDarknet53. This combination not only enhanced learning capabilities but also effectively reduced computational cost while maintaining accuracy. YOLOv8 continues to use the CSPNet structure. To better understand the network design of YOLOv8, a detailed introduction to CSPNet will follow.

[0076] like Figure 3 As shown, CSPNet divides the underlying feature map into two parts along the channel dimension. One part is processed and output by original modules such as DenseNet, while the other part is directly fused with the output features. This design effectively reduces the computational burden on the network while improving accuracy and efficiency.

[0077] C2f module: YOLOv5 uses CSPNet in the backbone and C3 module. YOLOv8, for improvement, uses C2f to replace the C3 module, integrating C3 and ELAN modules to enhance network efficiency. Before introducing C2f, it's necessary to understand the basic principles of C3 and ELAN. This will help in a more comprehensive understanding of the design and role of C2f in YOLOv8.

[0078] CBS module: The core module of YOLOv5 convolution operation is the CBL module, which integrates convolution, batch normalization and LeakyRelu activation function. YOLOv8 updates the activation function to SiLU function, which has no upper limit and smooth non-monotonic characteristics, making the deep training effect better. The CBS module is shown in Figure 4 .

[0079] Bottleneck module: Bottleneck essentially replaces larger convolution kernels with multiple smaller ones, which not only enhances the depth of the network but also significantly reduces the parameter quantity of the network. As shown in Figure 5 .

[0080] C3 module: As shown in Figure 6 , the C3 module still follows the design philosophy of CSPNet, but uses a different correction unit, which includes three 1x1 standard convolution layers and multiple Bottleneck modules. This optimizes the network structure and significantly improves the learning performance of the network.

[0081] ELAN module: As shown in Figure 7 , the ELAN module proposes a layer aggregation architecture with efficient gradient propagation path. This architecture optimizes the stack structure in the calculation block, optimizes the gradient length of the entire network, and ensures the stability and performance of the model during scaling.

[0082] C2f module: As shown in Figure 8 , the C2f module combines the ideas of C3 module and ELAN module. It uses the gradient shunt concept to reconnect the originally connected Bottleneck modules, optimizing the module structure and effectively preventing the degradation of convergence when the network is too deep.

[0083] For the Neck layer: In the Neck layer, the original SPP module (Spatial Pyramid Pooling) is replaced by the innovative SPPF module (Spatial Pyramid Pooling-Fast), and YOLOv8 retains the FPN+PAN module combination in YOLOv5, but fine-tunes its structure, further improving the network's ability to extract features.

[0084] SPPF module: As shown in Figure 9 , the SPP module uses 1x1, 5x5, 9x9 and 13x13 convolution kernels of different sizes to extract multi-level features from the feature image. This design helps to solve the multi-scale problem, thereby significantly improving the detection accuracy of the algorithm.

[0085] The SPPF module is optimized on the basis of the SPP module. The SPPF module replaces the original three different size convolution kernels of the SPP module with three 5x5 convolution kernels. The effect of two 5x5 convolution kernels in series is equivalent to that of a 9x9 convolution kernel, and similarly, the effect of three 5x5 convolution kernels in series is equivalent to that of a 13x13 convolution kernel. Compared with a large convolution kernel, this improves the detection efficiency while reducing the amount of calculation.

[0086] The FPN+PAN module is as shown in Figure 10 .

[0087] Compared with YOLOv5, YOLOv8 not only updates the C3 module to a C2f module, but also simplifies the network structure by removing the convolution layer before upsampling. This improvement enables the output features of different stages in the backbone feature extraction network to be directly upsampled, thereby improving the detection efficiency.

[0088] For the Head layer: YOLOv8 has made important structural improvements in the Head layer. Specifically, it has abandoned the coupled head in YOLOv5 and instead adopted a decoupled head, and has realized the transition from Anchor Based to Anchor Free. These improvements help to improve detection accuracy and efficiency.

[0089] Coupled head: as shown in Figure 11 , although this module is used frequently, this structure actually has potential adverse effects on the network. Specifically, the two key tasks of classification and regression may conflict during execution, which not only affects the accuracy of the network, but also reduces its convergence speed.

[0090] Decoupled head: as shown in Figure 12 , the decoupled head decouples the regression task from the classification and divides them into independent branches. For the same detection task, using a decoupled head not only speeds up the convergence of the network, but also improves the detection accuracy to some extent.

[0091] Anchor Free network: traditional target detection algorithms all rely on anchors to achieve prediction. Although this method can improve detection accuracy when objects are similar, it also has some problems, such as the design of anchor boxes needing to be done manually, requiring the preset anchor box size to overlap with the real object size as much as possible, and for detection tasks of different scale objects, migration becomes difficult and tuning becomes complex. In contrast, YOLOv8, as an Anchor Free network, no longer predicts the offset of anchor boxes, thereby effectively solving many problems caused by anchor boxes.

[0092] Loss function: The regression loss used by YOLOv8 is in the form of CIoU Loss + DFL (Distribution Focal Loss), and the classification loss function is VFL (Varifocal Loss).

[0093] CIoU Loss function: as shown in Figure 13 . c represents the diagonal distance of the smallest closed region that can contain both the real box and the predicted box, and I = p2(b, bgt) is the Euclidean distance between the center of the real box and the predicted box. The calculation formula is shown in equations (1-1), (1-2) and (1-3).

[0094] (1-1)

[0095] (1-2)

[0096] (1-3)

[0097] where w and h are the width and height of the predicted box, wgt and hgt are the width and height of the real box, and v in the formula represents the aspect ratio difference between the real box and the predicted box, and a is the weight value of v. If the difference is greater, v will be greater; if they are consistent, v = 0, where and represent the aspect ratio of the predicted target box and the aspect ratio of the real target box, respectively.

[0098] Varifocal Loss function: The Varifocal Loss function formula is shown in equation (1-4):

[0099]

[0100] In Varifocal Loss, q represents the label value, and for positive samples, q takes the value of IoU; while for negative samples, q is 0. This design makes the loss function similar to the regular BCE (Binary Cross-Entropy) loss function for positive samples, and changes to the Focal loss function for negative samples. Varifocal Loss is not symmetric for positive and negative samples, and it treats positive samples as the main focus according to their different importance in the detection task.

[0101] Distribution Focal Loss function: To accurately reflect the semantic information of the image, a more accurate bounding box representation method is needed to solve this problem, as shown in equation (1-5), the network's predicted output Instead of simply adjusting the number of channels through convolution, the network predicts a probability distribution, which is used to weight the coordinates to obtain the final predicted coordinates.

[0102] .

[0103] In addition, the problem of not being able to directly output a continuous probability distribution needs to be solved, and discrete probability points are used here, as shown in equation (1-6):

[0104]

[0105] Direct training can lead to an excessively large probability representation space, making it difficult for the network to converge and optimize. The probability of points in the vicinity of the real coordinates is higher, so Distribution Focal Loss (DFL) is proposed, as shown in equation (1-7):

[0106]

[0107] DFL helps the new network pay more attention to values near the target y, increasing the probability of these values and thus speeding up the network convergence process.

[0108] Deformable Convolution DCNv1: Introduce deformable convolution (Deformable Convolution, DCN) to replace the traditional fixed convolution method, which can improve the model's recognition ability of geometric transformation. Deformable convolution has excellent adaptability and can dynamically adjust according to the shape and scale of the object to be measured. It overcomes the limitations of traditional fixed rectangular structure in the sampling process. This improvement makes the network more flexible and accurate when dealing with complex geometric transformations. The principle of deformable convolution is shown in Figure 14 .

[0109] Further, in step S2, the training process of the deep learning target detection model includes the following steps:

[0110] Step S2.1: Data acquisition: Provide near-infrared light sources with wavelengths of 850nm, 980nm, 1310nm, and 1450nm, respectively, to transilluminate the tooth samples, and collect the near-infrared transillumination images transmitted by the tooth samples through the near-infrared camera, wherein 980nm and 1310nm are the core effective wavebands, and 850nm and 1450nm are used as controls;

[0111] Step S2.2: Data processing: Label the tooth region in the near-infrared transillumination image;

[0112] Step S2.3: Data set division: Divide the labeled data set into a training set, a validation set, and a test set;

[0113] Step S2.4: training the improved YOLOv8m network model;

[0114] Step S2.5: detecting teeth using the trained improved YOLOv8m network model to obtain a tooth evaluation result;

[0115] In step S2.4, the following steps are performed:

[0116] Step S2.4.1: setting training parameters: setting the number of training rounds epoch to 1000, setting the batch training size batch_size to -1, setting the initial learning rate to 0.01, setting the training device to GPU, setting the input image size to 640x640 pixels, setting the workers to 8, i.e., using multi-thread data loading, using an automatic selection optimizer, i.e., optimizer='auto', and adding regularization to prevent overfitting, i.e., dropout=0.1;

[0117] Step S2.4.2: inputting the near-infrared transillumination images in the training set into the improved YOLOv8m network model and performing backpropagation and parameter updating using the defined loss function;

[0118] Step S2.4.3: during the training process, periodically evaluating the performance of the YOLOv8m model using the validation set, adjusting the training parameters according to the evaluation results, and obtaining the trained improved YOLOv8m network model.

[0119] Further, in step S2.2, the data processing is as follows:

[0120] Step S2.2.1: using the annotation tool makesense to annotate the tooth region in the near-infrared transillumination image, and classifying it into normal, shallow caries, medium caries, deep caries, defect, and calculus;

[0121] Step S2.2.2: after annotation, saving the annotation information in a YOLO format recognizable by the model.

[0122] Further, in step S2.2, the data processing is as follows:

[0123] Step S2.2.1: using the annotation tool makesense to annotate the tooth region in the near-infrared transillumination image, and classifying it into normal, shallow caries, medium caries, deep caries, defect, and calculus;

[0124] Step S2.2.2: after annotation, saving the annotation information in a YOLO format recognizable by the model.

[0125] Further, in step S2.3, the following steps are performed:

[0126] The labeled data set is divided into training set, validation set and test set in the ratio of 7:1:2.

[0127] Further, as shown in the specific implementation of introducing deformable convolution DCNv2 in the SPPFv2 module includes the following steps: Figure 15

[0128] Step S2.4.2.1: For each position p0 on the output feature map, generate an offset field through a parallel convolution layer, and the number of channels of the offset field is 2N, where N is the total number of sampling points of the convolution kernel;

[0129] Step S2.4.2.2: Extract the offset matrix corresponding to the position p0 from the offset field, which contains the two-dimensional offset Δp n of each sampling point p n of the convolution kernel;

[0130] Step S2.4.2.3: Calculate the output feature value of position p0 using the formula y(p0) = ∑ pn∈R W(p n )·x(p0+ p n +Δp n ), where R is the regular sampling grid of the convolution kernel = {(-1, -1), (-1,0), …, (0, 1), (1,1)}, W(p n ) is the convolution weight of position p n , and x(p0+ p n +Δp n ) is the input feature map.

[0131] Further, when the sampling position p = p0+ p n +Δp n is a non-integer coordinate, the value x(p) of the input feature map at position p is calculated by bilinear interpolation; the bilinear interpolation formula is: x(p) =∑ q G(q, p) ·x(q), where q is all integer spatial positions around p, and G(q, p) is a bilinear interpolation kernel.

[0132] Further, the bilinear interpolation kernel G(q, p) can be decomposed into the product of two one-dimensional kernel functions: G(q, p) = g(q x , p x )·g(q y , p y ), where g(a, b) = max(0, 1-|a-b|).

[0133] ​Further, step S2.4.4 is further included, which is evaluating the performance of the YOLOv8m network model, specifically:

[0134] Step S2.4.4.1: input the near-infrared transillumination images in the divided test set to the trained improved YOLOv8m network model for evaluation;

[0135] Step S2.4.4.2: calculate the precision, recall and average precision value of the YOLOv8m network model;

[0136] Step S2.4.4.3: the precision of caries detection on the test set of 980nm wavelength is not less than 94.3%, the recall is not less than 97.2%, and the average precision value is not less than 0.979, which is excellent.

[0137] The present application provides a dental health assessment system for implementing the above method, comprising:

[0138] An optical imaging module comprising a near-infrared light source with a wavelength of 980nm and a near-infrared camera for collecting near-infrared transillumination images of teeth;

[0139] A data processing and analysis module comprising a memory and a processor, the memory storing a computer program, which when executed by the processor, can implement the dental health assessment method of any one of claims 1-7.

[0140] Further, the system is a portable device, and the optical imaging module and the data processing and analysis module are integrated, and the data processing and analysis module is a mobile terminal or an embedded device loaded with the deep learning target detection model.

[0141] The following is the experiment and result analysis:

[0142] Experimental steps

[0143] 1.1 Experimental materials

[0144] 1.1.1 Sample

[0145] Normal teeth 322, mild caries 267, moderate caries 435, and severe caries 500.

[0146] 1.1.2 Experimental equipment

[0147] Near-infrared camera (MIDVIISION 3206IR CUT): Capture near-infrared images of different wavebands (850nm, 980nm, 1310nm, 1450nm); Near-infrared light source / torch (Beyond Optics Adjustable Light LED Research Torch): Provide light sources of 850nm, 980nm, 1310nm and 1450nm wavebands; Tooth fixing device (Nidamian G-500 Super Light Clay): Fix the tooth sample, reduce light interference, and ensure consistent shooting angle and object distance.

[0148] 1.2 Design and implementation of near-infrared transillumination experiment

[0149] 1.2.1 Waveband selection and light source matching verification

[0150] Firstly, to explore the transillumination effect of near-infrared light on teeth, we selected suitable wavebands. We conducted a pre-experiment and selected the following wavebands for study: 850nm, 980nm, 1310nm and 1450nm. We used the same model of camera with different wavebands and selected a full-spectrum torch as the same light source. We took near-infrared photos of the same tooth sample with different degrees of caries and compared their transillumination effects, as shown in Figure 1. Figure 16

[0151] Due to its proximity to the visible light waveband range, the 850nm camera has the lowest requirement for the light source and the highest imaging brightness. However, due to its proximity to visible light, the scattering is relatively strong, and the transmitted light accounts for a low proportion. Therefore, it cannot be directly used for near-infrared transillumination of caries. Secondly, the 980nm and 1310nm cameras have basically the same requirements for the light source, and both have relatively low imaging brightness. However, they both show good transillumination effects and can clearly show the internal caries of the tooth. Finally, the performance of the 1450nm camera is the weakest. Its imaging can only capture the light source, and it has a very high requirement for the light source. It needs to be equipped with a high-power light source to achieve transillumination. However, such light sources not only have high manufacturing costs, but also have significant heating problems, which are not conducive to practical application.

[0152] After that, we further conducted matching experiments between different waveband light sources (850nm, 980nm, 1310nm, 1450nm) and cameras of the same waveband. The results showed that when the wavelength matching degree of the light source and the camera is high, the transillumination effect of near-infrared light on teeth is best

[32] , and the remaining results are consistent with the previous ones.

[0153] 1.2.2 Optimization of experimental scheme and image acquisition

[0154] 1) Optimize camera settings:

[0155] ​Based on the analysis of the effects of different wavelengths of near-infrared light (850 nm band is difficult to effectively penetrate caries due to its small wavelength, and 1450 nm band requires too high a power for the light source), to comprehensively evaluate the detection performance of each band and optimize the experimental design, four bands (850 nm, 980 nm, 1310 nm, and 1450 nm) were selected, among which 980 nm and 1310 nm were the core effective bands, and 850 nm and 1450 nm were used as controls. The same type of near-infrared camera was used, and the experimental platform (including a light source support fixed by clay shaping and a tooth fixing device with adjustable aperture angle) was built to ensure that the camera of each band was accurately coupled with the power matching near-infrared LED light source of the corresponding wavelength, so as to eliminate the penetration error caused by inconsistent light sources.

[0156] 2) Image acquisition:

[0157] During shooting, the tooth sample was vertically fixed in the center of the dark box using the fixing device, and the object distance and angle were basically consistent for each shooting through three-dimensional adjustment. The corresponding band light source was turned on synchronously, and five groups of images with different exposure parameters (exposure time 50-200 ms adjustable) were collected for each sample to cover different tissue contrast scenarios. The typical penetration image is shown in Figure 17 .

[0158] Figure 17 The same caries under different wavelengths of near-infrared cameras and their corresponding near-infrared light sources is shown, and the images are analyzed again: under 850 nm wavelength, the penetration effect is poor; under 980 nm wavelength, the penetration effect is good, the pigment deposition part is basically gray-white, and the small black spots of caries are extremely obvious. Under this wavelength, the appearance features of the pigment deposition area and early caries spots can be clearly displayed; under 1310 nm wavelength, the tooth condition can be basically displayed, but the penetration effect is not as good as that of 980 nm near-infrared light. In addition, most ordinary cameras (such as CMOS or CCD) use silicon-based sensors, and the cutoff wavelength of silicon is about 1100 nm. Therefore, the sensitivity will decrease sharply when the photon energy exceeds this wavelength. In the experiment, a larger power of 1310 nm light source is required, which heats up quickly and has low efficiency; under 1450 nm wavelength, the camera has very low sensitivity to light, and even under high power, the penetration image is almost invisible after the light source penetrates the tooth. Figure 17 (d) in FIG. 1 only 1450 nm light source is shot.

[0159] 3) Establish a review mechanism:

[0160] For the problem of misjudgment caused by pigment deposition (such as tea stains and smoke stains) or enamel hypoplasia in visual classification, a near-infrared transillumination image review mechanism was established: after the preliminary classification of samples, two oral physicians independently observed the gray scale changes in the tooth interior — the healthy dentin area showed uniform high light, the caries area showed light gray (mild), dark gray (moderate), and black (severe) in turn with the decrease of mineralization degree, and the boundary clarity was positively correlated with the depth of caries. For samples with inconsistent physician determination results (about 15% of the total), after near-infrared transillumination verification, they were reclassified, and 10 groups of images were supplemented to enhance the data quantity of this type of samples.

[0161] 1.3 YOLO model training

[0162] 1.3.1 Experimental environment configuration

[0163] According to the principle of YOLO algorithm, the input layer, backbone layer, neck and head of the model are constructed as shown in Figure 1 PyTorch was selected as the deep learning framework in this study because it is more readable than TensorFlow and Caffe. The code structure of PyTorch is clear, and the network layer meaning and propagation mechanism can be understood intuitively without complex comments. Its powerful debugging function allows flexible adjustment of neural network behavior, enabling dynamic modeling. Due to the large amount of calculation in caries detection, a graphics processing unit (GPU) is needed to speed up model training. Due to the limitations of real devices, RTX4090 was selected as the graphics processing unit in the cloud computing platform AutoDL.

[0164] 1.3.2 Data preprocessing

[0165] Select typical transillumination images, use the annotation tool makesense to annotate the tooth area in the images, and divide them into normal, mild caries, moderate caries, deep caries, defect, and calculus. After annotation, save the annotation information in YOLO format that the model can recognize. Divide the annotated data set into training set, validation set and test set according to a certain proportion (7:1:2). The training set is used for model training, the validation set is used to evaluate the performance of the model during training, and the test set is used to evaluate the generalization ability of the model.

[0166] 1.3.3 Model training

[0167] Set training parameters: Parameter training settings: The number of training epochs for the model is set to 1000 (not actually trained to completion), the batch size is set to -1, which means the system will automatically determine the size based on available resources. The initial learning rate is 0.01, the training device is GPU, the input image size is 640x640 pixels, the workers setting is 8, which means the data loader uses multi-threading to load data, the optimizer is set to 'auto' for automatic selection, and dropout is set to 0.1 to prevent overfitting.

[0168] The training set is input into the YOLO model, and the defined loss function is used for backpropagation and parameter updating. During the training process, the performance of the model is evaluated periodically using the validation set, and the training parameters are adjusted based on the evaluation results.

[0169] After training, the trained model is evaluated using the test set. To better evaluate the improved model, this paper uses multiple indicators to comprehensively measure the performance of the model in the task of caries detection and classification, including precision (P), recall (R), F1 score, and mean average precision (mAP).

[0170] (1) Precision

[0171] Precision, abbreviated as P, represents the accuracy rate, which refers to the proportion of the correctly predicted part in the overall prediction results. The calculation formula is shown in (2-1):

[0172] (2-1)

[0173] (2) Recall

[0174] Recall, abbreviated as R, refers to the proportion of correctly predicted positive samples in the total positive samples. The calculation formula is shown in (2-2):

[0175] (2-2)

[0176] (3) F1 Score

[0177] The calculation formula is shown in (2-3):

[0178] (2-3)

[0179] (4) AP

[0180] The accuracy and recall rate indicators alone are insufficient to reflect the performance of the model, and a comprehensive indicator AP (Average Precision) is needed. In this paper, AP refers to the detection accuracy of each type of dental caries. It takes into account the performance of the model in different dental caries categories. The formula for calculating the AP value is shown in (2-4):

[0181] (2-4)

[0182] (5) Average precision value (mAP)

[0183] The average precision value mAP is a comprehensive indicator obtained by taking the average of the target detection accuracy value AP. Since the recall rate and accuracy rate indicators cannot fully reflect the performance of the model, mAP is needed to consider the performance of both. In this paper, it refers specifically to the average detection accuracy for all dental caries categories. The formula is shown in (2-5):

[0184] (2-5)

[0185] where R is the recall rate, P is the accuracy rate, AP is the average precision, which is the integral value of the PR curve, and mAP is the average precision for all categories. The larger the mAP value, the better the overall performance of the algorithm. TP, FP, and FN are the number of positive samples predicted as positive samples, the number of negative samples predicted as positive samples, and the number of positive samples predicted as negative samples, respectively. The PR curve, defect AP accuracy, and average precision mAP can more comprehensively reflect the performance of the model. The following will focus on these three indicators to analyze the performance of the model.

[0186] 1.4 Experimental results analysis

[0187] 1.4.1 Near-infrared transillumination

[0188] 1.4.1.1 Analysis of transillumination images

[0189] Under the near-infrared transillumination technology, the optical property differences of dental tissues are accurately presented as Figure 18The early formed micro-cavities (even covered by tea stains, smoke stains and other pigments) are shown as small dark spots with blurred boundaries in the image due to the enhanced near-infrared light absorption caused by mineral loss; while the pigmented areas only show uniform light color, forming a transition with the high-light background of healthy tooth tissue, avoiding misjudgment caused by color confusion under visible light. Calculus is rich in minerals such as hydroxyapatite, which appears as dark gray bands around the neck of the tooth, with clear boundaries and significant contrast with the surrounding tissue, effectively solving the identification problem of calculus in hidden parts such as dental crevice and gum margin. The moderate to severe caries area shows dark gray to black shadows as the mineralization degree decreases, and the irregular dark area inside corresponds to the location of the cavity, directly reflecting the depth of the lesion; even if multiple lesions coexist, the transillumination image can be clearly distinguished by the gray gradient and texture characteristics - the band-shaped structure of calculus, the progressive dark area of caries, and the uniform light area of pigments do not interfere with each other, providing a clear visual basis for subsequent quantitative analysis.

[0190] Compared with traditional visual inspection, this technology avoids the limitations of subjective judgment and structural overlap: under visible light, pit groove pigments are easily misjudged as caries, or early caries lesions are missed due to their small size and hidden location, while near-infrared transillumination can penetrate the tooth surface, remove pigment interference, and directly visualize the demineralized areas (light gray spots) at the pit groove bottom and surface pigments (uniform light yellow). For example, when dark brown patches appear on the tooth surface, transillumination can accurately distinguish between pigment deposition (uniform light color) and early caries (blurred dark area), clearly indicating whether cleaning or further diagnosis and treatment is needed. This non-invasive detection does not require a probe and is radiation-free, making it particularly suitable for sensitive groups such as children and pregnant women. It can objectively evaluate the severity of the lesion through standardized images, helping with early intervention - intervening at the demineralization stage to prevent disease progression and significantly reduce the need for complex treatment.

[0191] In Figure 18 , (a) is the visible light image of caries 1; (b) is the early caries black spot under near-infrared transillumination of caries 1; (c) is the visible light image of caries 2; (d) is the near-infrared transillumination of caries 2, where pigment deposition is not visible and calculus appears as a block-shaped dark color; (e) is the visible light image of caries 3; (f) is the near-infrared transillumination of caries 3, where pigment deposition is not visible and caries is visible; (g) is the visible light image of caries 4; (h) is the near-infrared transillumination of caries 4, where pigment deposition is not visible, caries black spots and lateral calculus are visible. Among them: red box: caries; blue box: pigment deposition; green box: calculus; yellow box: small cavity.

[0192] 1.4.1.2 Comparison of different caries detection methods:

[0193] 1) Visual inspection: doctors use probes and endoscopes to check the extent of dental lesions. This method is accurate for detecting surface caries, but it is not easy to detect interproximal caries and internal caries, and it is not convenient in daily life.

[0194] 2) X-ray imaging: by Figure 19 As shown, X-rays are not sensitive to early caries and surface decay. (a) Visible light image of a carious tooth. (b) Near-infrared radiograph at 980 nm. It can be seen that compared with visible light, the near-infrared radiograph can effectively filter out pigment deposits and only show the caries (black parts). However, (c) is a radiograph of a tooth taken by X-ray. It can be seen that the radiograph of a tooth taken by X-ray is limited by angle, root, etc., causing vertical structures to overlap, making it difficult to show the caries parts clearly.

[0195] 1.4.2 Deep Learning

[0196] 1.4.2.1 Quantitative Comparison Experiment

[0197] Under the same experimental conditions, this paper compares the experimental results before and after the model modification, as shown in Table 1. Table 1 clearly shows that the YOLOv8m model has a significant overall improvement over the YOLOv5s model. With the recall rate remaining almost unchanged, at a wavelength of 980nm, the model's precision improved by 1.6%, mAP50 by 1.9%, and mAP50-95 by 2.6%; at a wavelength of 1310nm, the model's precision improved by 4.3%, mAP50 by 3.6%, and mAP50-95 by 4.1%. These data confirm the effectiveness of the automated dental health assessment method selected in this study.

[0198] Table 1 Experimental Results of YOLOv8m & YOLOv5s Models

[0199]

[0200] Furthermore, we compared the mAP50 values ​​of the two models across various labels. Table 2 shows that the YOLOv8m model outperforms the YOLOv5s model on multiple key labels, especially in the "Loss" and "Stone" categories, where the data volume is relatively small and the performance is significantly better. This indicates that the YOLOv8m model has significant advantages in detection accuracy and robustness, effectively reducing false positives and false negatives, thus providing higher detection performance in practical applications.

[0201] Table 2. mAP50 values ​​of each tag in YOLOv8m & YOLOv5s

[0202]

[0203] 1.4.2.2 Qualitative comparison experiment

[0204] In addition to quantitative analysis, qualitative comparison experiments are also conducted in this paper to more intuitively show the performance difference before and after the model improvement. By selecting representative test images, the detection results of YOLOv8m model and YOLOv5s model under different wavelengths are compared and analyzed. Figure 20 The detection results of the two models on caries under 980nm and 1310nm wavelengths are shown. Among them, (a) YOLOv5s detection result under 980nm; (b) YOLOv8m detection result under 980nm; (c) YOLOv5s detection result under 1310nm; (d) YOLOv8m detection result under 1310nm.

[0205] From Figure 20 (a) and (b) in the above figure, it can be seen that under 980nm wavelength, both YOLOv8m model and YOLOv5s model can identify and frame the areas of caries, calculus and defect, while compared with YOLOv5s model, YOLOv8m model has more accurate framing of calculus and defect, with higher confidence, which is 10 and 6 percentage points higher respectively, and can more effectively reduce the false detection and missed detection.

[0206] Under 1310nm wavelength, YOLOv8m model also shows higher detection accuracy. As shown in (c) and (d) in the above figure, under the condition of similar high confidence on caries, the detection confidence of YOLOv8m model on calculus is 20 percentage points higher than that of YOLOv5s model. Figure 20

[0207] 1.4.2.3 Comprehensive performance evaluation

[0208] ​By comparing the detection accuracy, key label detection performance, robustness and generalization ability of YOLOv8m and YOLOv5s models under different wavelengths, it is found that YOLOv8m has obvious advantages in caries detection task: under 980nm and 1310nm wavelengths, YOLOv8m is superior to YOLOv5s in precision, recall rate and mAP indicators, especially in dealing with complex and small data volume labels, the detection accuracy and robustness are higher. Specifically, the precision of YOLOv8m for caries detection under 980nm wavelength is 94.3%, the recall rate is 97.2%, and the mAP50 value is 0.979, which is 1.6%, 0.5% and 1.9% higher than YOLOv5s respectively. Therefore, combined with the transillumination effect of different waveband near-infrared light sources in near-infrared transillumination experiment, it is proposed that the combination of 980nm wavelength light source and YOLOv8m model is the optimal scheme to realize the evaluation of dental health.

[0209] The present application designs a dental health evaluation method based on near-infrared transillumination, which collects multi-waveband near-infrared transillumination images such as 980nm, 1310nm and 1450nm, and compares the caries detection performance of YOLOv5s and improved YOLOv8m models. A multi-waveband light source matching optimization scheme is innovatively designed, the C2f module and SPPFv2 structure are introduced to strengthen feature extraction, and the classification accuracy is improved by using decoupling detection head and Anchor-Free design. Experiments show that the imaging effect of 980nm band is the best, the YOLOv8m model realizes 94.3% precision, 97.2% recall rate and 0.979 mAP50 under 980nm wavelength, which is 1.6%, 0.5% and 1.9% higher than YOLOv5s respectively, and performs better on complex labels (such as calculus and deep caries).

[0210] The method breaks through the radiation limitation of traditional X-ray detection, providing a safe and efficient caries screening tool for sensitive groups such as children and pregnant women; the near-infrared light source and imaging module can be integrated into a portable device, and the YOLOv8m model can be loaded on a mobile phone APP, the low cost and easy operation of the instrument make it have advantages in realizing regular monitoring of dental health at home. At the same time, through the accumulation of long-term family monitoring data, personal oral health time series archives can be formed, combined with AI analysis of caries risk factors, personalized prevention suggestions can be pushed, providing technical support for early caries precision intervention, and promoting the popularization of family oral health management from passive treatment to active prevention; in addition, for areas with insufficient medical resources, low-cost screening equipment combined with remote consultation and AI-assisted diagnosis can break through the difficulties of not being able to seek medical treatment in time due to economic or geographical restrictions, so that early intervention measures can benefit more people and narrow the oral health gap between different groups.

[0211] The above merely describes preferred embodiments of the present application, and is not intended to limit the technical scope of the present application in any way. Any minor modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application shall still fall within the technical scope of the present application.

Claims

1. A method for assessing dental health based on near-infrared transillumination, characterized in that, Includes the following steps: Step S1: Provide a near-infrared light source with a wavelength of 980nm to irradiate the teeth, and acquire near-infrared irradiated images of the teeth using a near-infrared camera; Step S2: Input the near-infrared illumination image into a pre-trained deep learning object detection model; Step S3: Using the deep learning target detection model, output the localization and classification results of the lesion area in the tooth; The deep learning object detection model is an improved model based on the YOLOv8 architecture. The improvement is that at least one standard CBS module in the neck network is replaced with a CBSv2 module. The CBSv2 module is composed of deformable convolutional DCNv2, batch normalized BN layer and SiLU activation function connected in sequence.

2. The method for dental health assessment based on near-infrared transillumination according to claim 1, characterized in that, In step S2, the training process of the deep learning object detection model includes the following steps: Step S2.1: Data acquisition: Near-infrared light sources with wavelengths of 850nm, 980nm, 1310nm and 1450nm are provided to transmit light to the tooth samples. Near-infrared images transmitted through the tooth samples are acquired by a near-infrared camera. Among them, 980nm and 1310nm are the core effective wavelengths, and 850nm and 1450nm are used as controls. Step S2.2: Data processing: Annotate the tooth regions in the near-infrared transillumination image; Step S2.3: Dataset partitioning: Divide the labeled dataset into training set, validation set and test set; Step S2.4: Train the improved YOLOv8m network model; Step S2.5: Use the trained improved YOLOv8m network model to detect teeth and obtain tooth evaluation results; Specifically, step S2.4 is as follows: Step S2.4.1: Set training parameters: the number of training epochs is set to 1000 epochs, the batch size is set to -1, the initial learning rate is 0.01, the training device is GPU, the input image size is 640×640 pixels, the workers are set to 8, that is, the data loader uses multi-threading to load data, the optimizer is automatically selected, i.e., optimizer='auto', and regularization is added to prevent overfitting, i.e., dropout=0.1; Step S2.4.2: Input the near-infrared transmission images from the training set into the improved YOLOv8m network model, and use the predefined loss function for backpropagation and parameter update; Step S2.4.3: During the training process, the performance of the YOLOv8m model is evaluated periodically using the validation set. The training parameters are adjusted according to the evaluation results to obtain the trained improved YOLOv8m network model.

3. The method for dental health assessment based on near-infrared transillumination according to claim 2, characterized in that, In step S2.2, the data processing specifically involves: Step S2.2.1: Use the annotation tool Makesense to annotate the tooth areas in the near-infrared transillumination image and classify them into normal, superficial caries, moderate caries, deep caries, defects, and calculus; Step S2.2.2: After the annotation is completed, save the annotation information in YOLO format that the model can recognize.

4. The method for dental health assessment based on near-infrared transillumination according to claim 2, characterized in that, In step S2.3, specifically: The labeled dataset is divided into training, validation, and test sets in a ratio of 7:1:

2.

5. The method for dental health assessment based on near-infrared transillumination according to claim 2, characterized in that, In step S2.4.2, the specific implementation of DCNv2 includes the following steps: Step S2.4.2.1: For each position p0 on the output feature map, an offset field is generated through a parallel convolutional layer. The number of channels in the offset field is 2N, where N is the total number of sampling points of the convolutional kernel. Step S2.4.2.2: Extract the offset matrix corresponding to position p0 from the offset field. This matrix contains the offset matrix for each sampling point p of the convolution kernel. n Two-dimensional offset Δp n ; Step S2.4.2.3: Use the formula y(p0) = ∑ pn∈R W(p n )·x(p0+ p n +Δp n Calculate the output feature value at position p0, where R is the regular sampling grid of the convolution kernel = {(-1, -1), (-1,0), … , (0, 1), (1, 1)}, W(p n ) is p n Convolution weights at position, x(p0+ p n +Δp n ) is the input feature map.

6. The method for dental health assessment based on near-infrared transillumination according to claim 5, characterized in that, When the sampling position p = p0 + p n +Δp n When the coordinates are non-integer, the value x(p) of the input feature map at position p is calculated using bilinear interpolation; the bilinear interpolation formula is: x(p) = ∑ q G(q, p) ·x(q), where q represents all integer spatial locations around p, and G(q,p) is the bilinear interpolation kernel.

7. The method for dental health assessment based on near-infrared transillumination according to claim 6, characterized in that, The bilinear interpolation kernel G(q, p) can be decomposed into the product of two one-dimensional kernel functions: G(q, p) = g(q x , p x )·g(q y , p y ), where g(a, b) = max(0, 1-|ab|).

8. The method for dental health assessment based on near-infrared transillumination according to claim 2, characterized in that, It also includes step S2.4.4, evaluating the performance of the YOLOv8m network model, specifically: Step S2.4.4.1: Input the near-infrared transmission images from the divided test set into the trained improved YOLOv8m network model for evaluation; Step S2.4.4.2: Calculate the precision, recall, and mean precision of the YOLOv8m network model; Step S2.4.4.3: If the accuracy of caries detection on the 980nm wavelength test set is not less than 94.3%, the recall is not less than 97.2%, and the mean precision is not less than 0.979, then it is considered excellent.

9. A dental health assessment system for implementing the method according to any one of claims 1 to 8, characterized in that, include: The optical imaging module includes a near-infrared light source with a wavelength of 980nm and a near-infrared camera for acquiring near-infrared transillumination images of teeth. The data processing and analysis module includes a memory and a processor. The memory stores a computer program that, when executed by the processor, enables the implementation of the dental health assessment method as described in any one of claims 1-8.

10. The dental health assessment system according to claim 9, characterized in that, The system is a portable device, and the optical imaging module and the data processing and analysis module are integrated into one unit. The data processing and analysis module is a mobile terminal or embedded device equipped with the deep learning target detection model.

Images