Intelligent grading evaluation method and system for inlay repair quality

Abstract

The invention relates to the technical field of tooth inlay repair, in particular to an inlay repair quality intelligent grading evaluation method and system. Clinical examination data at different times after a target tooth is repaired and an SEM image under a multi-angle light source are obtained, a microscopic 3D image is constructed based on an image recognition and segmentation technology, key microscopic parameters are calculated, then the microscopic parameters and the clinical data are fused into feature vectors, and the feature vectors are used as feature vectors. Comprehensive evaluation and risk classification of the future health state of the inlay are achieved through a time sequence prediction model, the repair quality grade is finally output, micro-morphology data and macroscopic clinical evaluation indexes are systematically fused, the problem that evaluation is inconsistent due to the fact that traditional methods depend on subjective experience of doctors is solved, and the method is suitable for popularization and application. Through quantitative analysis, the objectivity of evaluation is remarkably improved, and manual operation deviation is reduced; the technical problems that an existing inlay repair quality intelligent evaluation method has subjectivity and is difficult to quantify, and microscopic and macroscopic information is not fused are solved.

Description

Technical Field

[0001] This invention relates to the field of dental inlay restoration technology, and in particular to an intelligent grading and evaluation method and system for inlay restoration quality. Background Technology

[0002] Inlay restoration is a key treatment for restoring the morphology and function of tooth defects. Its long-term success rate is highly dependent on the quality of the inlay, especially the marginal integrity of the bonding interface between the inlay and the tooth structure. Poor interface integrity can lead to microleakage, which in turn can cause a series of clinical complications such as marginal staining, secondary caries, pulpitis, and restoration failure. Therefore, accurate and objective evaluation of the quality of inlay restoration is crucial.

[0003] Currently, the assessment of inlay restoration quality in clinical and research settings suffers from the following core deficiencies: 1. Clinical assessment standards rely heavily on physician experience, leading to significant fluctuations in results; 2. There is a lack of precise quantitative indicators for the inlay's condition, particularly its microscopic state; 3. SEM analysis struggles to obtain key three-dimensional geometric parameters of interface defects; 4. Microscopic SEM observations are separated from macroscopic clinical assessment data, failing to establish a unified and mutually verifiable quantitative assessment system. Therefore, existing technologies lack an intelligent assessment method for inlay restoration quality that overcomes subjectivity, achieves quantification, and integrates both microscopic and macroscopic information. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent grading and evaluation method and system for inlay restoration quality, which solves the technical problems of existing intelligent evaluation methods for inlay restoration quality being subjective, difficult to quantify, and failing to integrate microscopic and macroscopic information.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a smart grading assessment method for inlay repair quality, which specifically includes the following steps:

[0006] S1. Acquire tooth data at different times after the target tooth restoration, the tooth data including clinical examination data and several SEM images of the target tooth under the same shooting angle and different light sources;

[0007] S2. Identify and segment the regions corresponding to different materials based on the SEM images. Construct a microscopic 3D image of the target tooth based on the segmentation results and the albedo of each region, and calculate several microscopic parameters.

[0008] S3. By splicing together clinical examination data and microscopic parameters to obtain a fusion feature vector, and based on the fusion feature vectors at different times after the target tooth restoration, predict the inlay's prediction results at any future moment, including the comprehensive health score and risk classification results.

[0009] S4. Calculate the inlay repair quality level based on the fused feature vector, comprehensive health status coding, and prediction results.

[0010] Preferably, step S2 specifically includes the following steps:

[0011] S21. Preprocess the SEM image to enhance local contrast and highlight edge information to obtain a preprocessed image;

[0012] S22. Add a self-attention mechanism module between the encoder and decoder of the U-Net deep learning model to obtain an improved U-Net deep learning model, and use the pre-trained improved U-Net deep learning model to perform semantic segmentation on the preprocessed image so as to divide each pixel in the preprocessed image into several classification labels.

[0013] S23. Based on the classification labels corresponding to each pixel in the preprocessed image, obtain the classification labels corresponding to different pixels in the corresponding SEM image, as well as the material and original albedo corresponding to the classification labels.

[0014] S24. Construct a global energy function for joint optimization of normal vector and albedo based on several SEM images of the target tooth taken at the same time after tooth restoration under different light sources at the same shooting angle.

[0015] S25. Use an iterative algorithm to minimize the global energy function, and in each iteration, alternately minimize the albedo and normal vector to update the normal vector and albedo respectively.

[0016] S26. Calculate the height of each pixel based on the final normal vector of each pixel in the SEM image to construct a microscopic 3D image.

[0017] S27. Calculate several micro parameters based on the final albedo of each pixel in the SEM image and the micro 3D image.

[0018] Preferably, in step S24, the expression for the global energy function is:

[0019]

[0020]

[0021]

[0022]

[0023] In the above formula, Represents the global energy function. For data fidelity items, Represents the smoothing term of the normal vector. and Let these represent the normal vector hyperparameter and the albedo hyperparameter, respectively. Represents the midpoint of the SEM image under the i-th light source. grayscale value, Representing coordinates The albedo, i.e., the albedo field, Representing coordinates The normal vector at that point, i.e., the normal vector field. Let represent the unit vector of the light source direction in the SEM image under the i-th light source. There are k different light sources. Represents the gradient operator. Representing coordinates Gradient of the normal vector at that point Represents the albedo smoothing term. This represents the p-th material region. Let the set of pixels belonging to the same material region as any pixel within the p-th material region be defined. This represents the reflectance of any pixel within the p-th material region. It represents the reflectance of the pixels in the set of neighboring pixels of the pixel in the p-th material region.

[0024] Preferably, in step S27, the microscopic parameters include the total volume of inlay microcracks in the target tooth, the number of inlay pores, the degree of degradation of the adhesive layer, the average albedo variation in the adhesive region, the standard deviation of albedo in the inlay region, and the average gap width, maximum gap width, and critical index of marginal defects between the inlay and the natural tooth tissue. The formula for calculating the degree of degradation of the adhesive layer is:

[0025]

[0026] The formula for calculating the average albedo change in the adhesive region is:

[0027]

[0028] The formula for calculating the critical index of edge defects is:

[0029]

[0030] In the above formula, Indicates the degree of degradation of the adhesive layer. This represents the total volume of pores within the adhesive layer. This indicates the number of interconnected pore clusters within the adhesive layer. and These represent the weighting coefficients of pore volume and the number of interconnected pores on the degree of adhesive layer degradation, respectively. This represents the change in the average albedo of adhesive region A. This indicates the number of pixels in adhesive region A. Represents any coordinate position in adhesive region A The final albedo of the pixel, This represents the original albedo of the material corresponding to adhesive region A. This indicates the critical index for edge defects. This represents the gap width at the i-th sampling point in the gap between the inlay and the natural tooth structure. A total of N sampling points were set.

[0031] Preferably, the weighting coefficient of the pore volume on the degree of adhesive layer degradation is... The weighting coefficient of the number of interconnected pores on the degree of adhesive layer degradation The specific steps are as follows:

[0032] S271. Obtain several tooth samples containing inlays, as well as the total volume of pores in the adhesive layer and the number of interconnected pore clusters in the adhesive layer in each tooth sample, and perform Z-score normalization.

[0033] S272. Classify tooth samples according to their appearance and set corresponding adhesive layer degradation scores; the tooth sample classification types include completely successful, slightly degraded, moderately degraded, and severely degraded / failed, wherein:

[0034] The fully successful tooth sample had no marginal staining, no secondary caries, and the inlay was intact, with a corresponding adhesive layer degradation score of j0;

[0035] The slightly degenerated tooth sample had slight edge staining and no secondary caries, and the corresponding adhesive layer degradation score was j1;

[0036] The moderately degraded tooth sample had obvious marginal staining and mild secondary caries, and the corresponding adhesive layer degradation score was j2.

[0037] The severely degraded and failed tooth samples had inlays that fell off and had severe secondary caries, with the corresponding adhesive layer degradation score being j3;

[0038] And j0 < j1 < j2 < j3;

[0039] S273. Construct the influence function and likelihood function between the adhesive layer degradation score and the total volume of pores in the adhesive layer and the number of interconnected pore clusters in the adhesive layer;

[0040] S274. Estimate the weighting coefficient of pore volume on the degree of adhesive layer degradation using maximum likelihood estimation. The weighting coefficient of the number of interconnected pores on the degree of adhesive layer degradation .

[0041] Preferably, in step S273, the expression for the influence function is:

[0042]

[0043] The expression for the likelihood function is:

[0044]

[0045]

[0046] In the above formula, This represents an ordered logistic regression function. This represents the probability that the adhesive layer degradation score does not exceed j. This represents the j-th tangent parameter. and The regression coefficients represent the total volume of pores within the standardized adhesive layer and the number of interconnected pore clusters within the adhesive layer, respectively. This is an indicator function; it takes the value 1 when the adhesive layer degradation score of sample i is j, and 0 otherwise. , and Let represent the probability that the i-th sample has an adhesive layer degradation score of j, the probability that the i-th sample has an adhesive layer degradation score of j less than or equal to j, and the probability that the i-th sample has an adhesive layer degradation score of j-1 less than or equal to j, respectively. The number of samples is n.

[0047] Preferably, step S3 specifically includes the following steps:

[0048] S31. The clinical examination data and microscopic parameters at different times after the target tooth restoration are concatenated to obtain the fused feature vector at the corresponding time, the expression of which is:

[0049]

[0050] In the above formula, This represents the fused feature vector at time t. This represents the Nth parameter in the fused feature vector;

[0051] S32. Combine the fused feature vectors from different times in chronological order and standardize them with Z-score to obtain a temporal feature vector sequence, and calculate the time interval between each element in the temporal feature vector sequence and the first element to construct a time interval vector sequence.

[0052] S33. Construct a Transformer model, input the time-series feature vector sequence and time interval vector sequence into the Transformer model, and obtain the Transformer model encoder at the last time point. The output is used as a comprehensive health status code;

[0053] S34. Predict the comprehensive health score at any future time based on the comprehensive health status code, and classify the corresponding risk categories based on the comprehensive health score to obtain a prediction result that includes the comprehensive health score and risk classification results.

[0054] Preferably, step S33 specifically includes the following steps:

[0055] S331. The fused feature vectors at each time point are sequentially mapped to a high-dimensional space through a fully connected neural network to obtain the initial feature vectors at each time point, the expression of which is:

[0056]

[0057] In the above formula, Indicates the first The initial feature vector of the fused feature vector at each time point. Indicates the first The fused feature vectors at each time are input into a fully connected neural network MLP to map to a high-dimensional space;

[0058] S332. Map any time interval to an initial time vector with the same dimension as the initial feature embedding, expressed as:

[0059]

[0060] In the above formula, For the first The initial time vector of time intervals. For another fully connected neural network;

[0061] S333. Add the initial feature vector and the initial time vector at each time point to obtain the fusion vector at each time point;

[0062] S334. Construct a Transformer model and input the fused vector at each time step into the encoder of the Transformer model;

[0063] S335. Obtain the output of the Transformer model encoder at the last time point as the comprehensive health status code.

[0064] Preferably, step S4 specifically includes the following steps:

[0065] S41. Input the comprehensive health status code into the fully connected layer to calculate the current health score. The calculation formula is as follows:

[0066]

[0067] In the above formula, Indicates the current health score. As an activation function, it compresses the output to the (0,1) interval. and These represent the weight parameters and bias parameters, respectively.

[0068] S42. Calculate the future risk index based on the forecast results. The calculation formula is as follows:

[0069]

[0070] In the above formula, Indicates a future risk index. Represents any future moment Comprehensive health score at the time, Weighting coefficients for risk categories corresponding to the overall health score;

[0071] S43. Determine the inlay repair quality level based on the future risk index, fusion feature vector, current health score, and preset rules.

[0072] The present invention also provides an intelligent grading and evaluation system for inlay repair quality, including a processor and a memory, wherein the memory is used to store a computer program, and the computer program is executed by the processor to implement the grading and evaluation method.

[0073] By employing the above technical solutions, the present invention provides a method and system for intelligent grading and evaluating the quality of inlay repairs, which has at least the following beneficial effects:

[0074] 1. This invention acquires clinical examination data and SEM images under multi-angle light sources at different times after the target tooth restoration, constructs a microscopic 3D image based on image recognition and segmentation technology, calculates key microscopic parameters, and then integrates the microscopic parameters with clinical data into a feature vector. Using a time-series prediction model, it achieves a comprehensive assessment and risk classification of the future health status of the inlay, and finally outputs the restoration quality level. This method systematically integrates microscopic morphological data with macroscopic clinical assessment indicators, overcomes the problem of inconsistent assessment caused by traditional reliance on doctors' subjective experience, significantly improves the objectivity of the assessment through quantitative analysis, and has a high degree of automation throughout the process, reducing human error and making it suitable for large-scale clinical data analysis.

[0075] 2. This invention employs an improved U-Net deep learning model, introducing a self-attention mechanism between the encoder and decoder to achieve high-precision semantic segmentation of SEM images. It can accurately distinguish different material regions such as inlays, adhesive layers, natural teeth, and gaps, significantly improving the ability to identify complex boundaries and minute defects. It is especially suitable for scenarios where the optical properties of material surfaces change due to staining, wear, etc. in the oral environment.

[0076] 3. This invention combines multi-angle light source images and global energy functions to jointly optimize the normal vector and albedo, enabling the reconstruction of high-precision microscopic 3D morphology under the premise of allowing slow changes in albedo. It effectively overcomes the reconstruction errors caused by material aging, staining and other factors in the intraoral environment of traditional photometric stereo methods. This method not only provides three-dimensional geometric parameters of defects such as pores and cracks, but also lays a reliable data foundation for subsequent degradation degree calculation and risk prediction.

[0077] 4. This invention constructs a quantitative model of adhesive layer degradation, combining microscopic parameters such as total pore volume and number of connected pore clusters, and uses ordered logistic regression and maximum likelihood estimation to determine weighting coefficients, achieving an objective and quantitative assessment of the adhesive interface degradation state. This method overcomes the limitation of traditional clinical assessments in quantifying early interface degradation, and can automatically calculate degradation scores based on microscopic 3D image data, establishing a correlation with macroscopic clinical manifestations (such as edge staining and secondary caries). By introducing novel indicators such as edge defect critical index and albedo changes, it further enriches the evaluation dimensions of interface integrity, providing a more comprehensive predictive basis for the long-term durability of inlay restorations, and helping clinicians to identify high-risk cases in advance and develop personalized maintenance strategies.

[0078] 5. The microscopic parameter system proposed in this invention covers multiple dimensions such as the total volume of microcracks in the inlay, the number of pores, the degree of degradation of the adhesive layer, the change in albedo, and the gap width. It can comprehensively reflect the structural integrity and material stability of the restoration. These parameters are calculated based on microscopic 3D images and have clear physical meaning and statistical basis, which makes up for the shortcomings of traditional SEM analysis, which can only provide two-dimensional information. Attached Figure Description

[0079] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0080] Figure 1 This is a flowchart of the intelligent grading and evaluation method for inlay repair quality according to the present invention. Detailed Implementation

[0081] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0082] Currently, the assessment of inlay restoration quality in clinical and research settings mainly relies on the following two methods:

[0083] 1. Clinical Subjective Assessment Standards: The currently accepted international assessment method uses a series of standardized assessment criteria, such as the modified US Public Health Service (USPHS) standards. These standards assess multiple indicators of the restoration, including retention and integrity, marginal fit, adjacency, and color matching, through methods such as visual examination and probe examination (e.g., Alpha, Bravo, Charlie). While widely used, these standards are qualitative or semi-quantitative in nature and highly dependent on the examiner's personal experience, subjective judgment, and technique. Different physicians may yield significantly different assessments of the same restoration, indicating poor consistency and repeatability. Furthermore, this subjective assessment struggles to detect early, microscopic interface degradation, often only identifying failure when problems become visible or clinical symptoms appear, thus missing the optimal window for early intervention.

[0084] 2. Microscopic Morphology Observation Techniques: Scanning electron microscopy (SEM) is widely used in in vitro research and clinical sampling analysis to more accurately study adhesive interfaces. SEM can provide high-resolution microscopic images of interfaces, revealing defects such as microcracks, gaps, and bubbles. However, traditional SEM image analysis has the following significant limitations: First, the analysis process mainly relies on the researcher's subjective visual interpretation and manual measurement, lacking unified and quantitative analytical standards, resulting in relatively low efficiency and susceptibility to human bias; second, traditional SEM images are two-dimensional and cannot provide crucial three-dimensional geometric information about defects (such as crack depth and gap morphology), which greatly limits the in-depth understanding and accurate quantification of interface failure mechanisms.

[0085] In recent years, although some studies have attempted to apply computer image processing technology to dental material analysis, most of these efforts remain at the level of simple image enhancement, binarization, or basic feature measurements such as length and area. They have failed to fundamentally solve the core technical challenge of extracting three-dimensional morphological information from two-dimensional images, nor have they achieved intelligent and automated integration with clinical multidimensional assessment standards (such as the modified USPHS). For example, existing technologies cannot automatically and directly and quantitatively correlate nanoscale microcracks observed in SEM images with the "edge fit" degradation in clinical assessment.

[0086] Therefore, to address the technical problems of existing intelligent assessment methods for inlay restoration quality, such as subjectivity, difficulty in quantification, and lack of integration of microscopic and macroscopic information, this invention provides an intelligent grading assessment method for inlay restoration quality. This method should be able to automatically extract key three-dimensional morphological features based on objective SEM images and combine them with clinical assessment dimensions, ultimately outputting a consistent, accurate, and predictable comprehensive quality grade to assist clinicians in making precise decisions. Figure 1 As shown, this grading assessment method specifically includes the following steps:

[0087] S1. Obtain tooth data at different times after the restoration of the target tooth, such as 6 months, 12 months, and 24 months after restoration. The tooth data includes clinical examination data and several SEM images of the target tooth under the same shooting angle but different light sources. The clinical examination data is based on the modified USPHS standard, which records the grade or quantitative observation results of dimensions such as restoration retention and integrity, marginal fit, proximal relationship, tooth integrity, secondary caries, color stability, and transparency. The SEM images are high-resolution microscopic images of the interface between the inlay, adhesive, and original tooth tissue obtained by scanning electron microscopy. During the SEM image taking process, a shooting angle is selected and taken multiple times under the same shooting angle by changing the light source at different angles, thereby obtaining several SEM images.

[0088] S2. Identify and segment the regions corresponding to different materials based on the SEM image. Construct a microscopic 3D image of the target tooth based on the segmentation results and the albedo of each region, and calculate several microscopic parameters. Before calculating the microscopic parameters, it is necessary to accurately identify the inlays, adhesive layers, and the user's original natural tooth structure in the SEM image. The specific steps include the following:

[0089] S21. First, the SEM image is preprocessed to enhance local contrast and highlight edge information to obtain a preprocessed image. For example, the SEM image can be divided into several small regions and histogram equalization can be performed using the CLAHE algorithm to enhance local contrast, thereby highlighting details such as the edges of microcracks in the tooth and suppressing the overall background noise of the SEM image. Then, the uniform region is smoothed by nonlinear anisotropic diffusion filtering. The uniform region is the area on the tooth that is made of the same material. At the same time, edge information with large gradients is retained and enhanced to identify the boundaries between different materials, thereby better preparing for segmentation.

[0090] S22. A self-attention mechanism module is added between the encoder and decoder of the U-Net deep learning model to obtain an improved U-Net deep learning model. The pre-trained improved U-Net deep learning model is used to perform semantic segmentation on the preprocessed image to divide each pixel in the preprocessed image into several classification labels, including inlay C, adhesive layer A, natural tooth D, and gap and background G. The introduction of the self-attention mechanism enables the model to better understand the global context of the image. For example, it can more accurately determine that a continuous dark line is a "crack" rather than an adhesive layer with "uneven resin coloring" based on the shape of the adhesive, thereby significantly improving the segmentation accuracy, especially for complex and blurry boundaries.

[0091] S23. Based on the classification labels corresponding to each pixel in the preprocessed image, obtain the classification labels corresponding to different pixels in the corresponding SEM image, as well as the materials and original albedo corresponding to the classification labels. For the materials and original albedo corresponding to the classification labels, images of a large number of known pure material samples used in different inlay restoration methods, such as pure porcelain pieces, pure resin adhesive pieces, and natural tooth pieces, can be collected in advance under the same and standard SEM imaging parameters. Then, by measuring the gray values ​​in the images at different angles, the average albedo of each material can be calculated to obtain the value of the original albedo. Among them, for the gap and background G region, an empirically low albedo can be assigned to directly distinguish it from the inlay material and avoid interference in the subsequent calculation process.

[0092] S24. Construct a global energy function based on several SEM images of the target tooth taken at the same time after tooth restoration, taken from the same shooting angle but with different light source angles. That is, the several SEM images are taken at the same time after tooth restoration and from the same shooting angle, but with different light source angles.

[0093] The basic equation of traditional photometric stereochemistry is a functional expression related to four factors: grayscale value, normal vector, albedo, and light source angle. Grayscale value, albedo, and light source angle are all known quantities. The normal vector is solved to obtain the normal vector field of the object's surface. After solving for the normal vector field, the Frankot-Chellappa algorithm or similar methods are used to integrate the normal vector field to obtain the height value of each pixel, thus generating a microscopic 3D image. However, in this method, the restored tooth contains inlays, adhesive layers, and natural tooth structure. Over time, the oral environment in the patient's mouth causes the following changes to the material surface, affecting its albedo: 1. Coffee, tea, etc., may darken the tooth color and reduce albedo; 2. Tooth wear leads to a decrease in surface smoothness, becoming rougher, resulting in increased light scattering and changes in albedo and specular reflection components; 3. Saliva film and plaque adhere to the tooth, forming an additional film that may alter the optical properties of the tooth surface. There may even be other variations that mean that, after a period of service in the oral cavity, the albedo of inlay C, adhesive layer A, and natural tooth D may deviate from the original albedo we measured on the sample beforehand, i.e., the original albedo assigned to each material in step S23.

[0094] Therefore, the albedo in the photometric stereo normal function cannot be simply regarded as a constant quantity. It is necessary to allow for continuous and slow changes in the albedo of pixels within the same material region, rather than a constant. That is, it needs to be able to tolerate and even infer changes in albedo. Here, the original albedo corresponding to each material is used as an initial albedo value to construct a simultaneously optimized normal vector field. and albedo field The global energy function is expressed as:

[0095]

[0096]

[0097]

[0098]

[0099] In the above formula, Represents the global energy function. For data fidelity items, This represents the smoothing term of the normal vector, ensuring a smooth surface shape. and Let these represent the normal vector hyperparameter and the albedo hyperparameter, respectively. Represents the midpoint of the SEM image under the i-th light source. grayscale value, Representing coordinates The albedo, i.e., the albedo field, Representing coordinates The normal vector at that point, i.e., the normal vector field. Let represent the unit vector of the light source direction in the SEM image under the i-th light source. There are k different light sources. Represents the gradient operator. Representing coordinates Gradient of the normal vector at that point This represents the albedo smoothing term, which mandates that albedo variations are smooth within the same material region, while allowing abrupt changes in albedo between different materials to reflect physical processes such as tooth staining and wear within the oral cavity. This represents the p-th material region. Let the set of pixels belonging to the same material region as any pixel within the p-th material region be defined. This represents the reflectance of any pixel within the p-th material region. This represents the reflectance of the pixels within the set of pixels in the neighborhood of the p-th material region.

[0100] S25. Use an iterative algorithm to minimize the global energy function, and alternately minimize the albedo and normal vector in each iteration to update the normal vector and albedo respectively. That is, first fix the currently estimated albedo to update the normal vector to keep the object surface in the SEM image smooth, and then fix the currently estimated normal vector to update the albedo to keep the albedo in the same material region smooth. Repeat this alternation until the global energy function converges to obtain the final normal vector and final albedo of each pixel in the SEM image.

[0101] This algorithm no longer forces all pixels in the same material region to have exactly the same albedo value. It allows certain areas of the same material to vary due to staining or other factors, as long as the variation is smooth (i.e., it doesn't jump drastically within a few pixels). Through joint optimization, the algorithm can correctly decompose grayscale variations into components originating from shape (i.e., the normal vector) and components originating from albedo. For example, a dark area accompanied by a drastic change in the normal vector is interpreted as a "shadow" (i.e., a pit or crack); if its normal vector is flat, it is interpreted as "surface staining" (i.e., reduced albedo). This significantly reduces the microscopic 3D image reconstruction error caused by changes in the optical properties of the material surface due to the oral cavity environment.

[0102] S26. Calculate the height of each pixel based on its final normal vector in the SEM image to construct a microscopic 3D image. The formula for calculating the height of each pixel is as follows:

[0103]

[0104]

[0105] In the above formula, x, y, and z are the x-axis, y-axis, and z-axis coordinates of any pixel. Let be the final normal vector of any pixel, whose components along the x-axis, y-axis, and z-axis are respectively... , and , For the Laplace operator.

[0106] S27. Based on the final albedo of each pixel in the SEM image and the microscopic 3D image, calculate several microscopic parameters. These parameters include the total volume of inlay microcracks in the target tooth, the number of inlay pores, the degree of degradation of the adhesive layer, the average albedo variation in the adhesive region, the standard deviation of the albedo in the inlay region, and the average gap width, maximum gap width, and critical index of edge defects between the inlay and the natural tooth tissue. The formula for calculating the degree of degradation of the adhesive layer is as follows:

[0107]

[0108] The formula for calculating the average albedo change in the adhesive region is:

[0109]

[0110] The formula for calculating the critical index of edge defects is:

[0111]

[0112] In the above formula, Indicates the degree of degradation of the adhesive layer. This represents the total volume of pores within the adhesive layer. The number of interconnected pore clusters within the adhesive layer can be obtained through 3D connectivity component analysis. and These represent the weighting coefficients of pore volume and the number of interconnected pores on the degree of adhesive layer degradation, respectively. This represents the change in the average albedo of adhesive region A. This indicates the number of pixels in adhesive region A. This indicates that any coordinate in adhesive region A is... The final albedo of the pixel, This represents the original albedo of the material corresponding to adhesive region A. This indicates the critical index for edge defects. This represents the gap width at the i-th sampling point in the gap between the inlay and the natural tooth structure. A total of N sampling points were set.

[0113] In addition, the total volume of microcracks in the inlay, the number of pores in the inlay, the average gap width between the inlay and the natural tooth structure, and the maximum gap width can be directly read from the microscopic 3D image. The albedo standard deviation of the inlay area is obtained by obtaining the albedo of each pixel in that area and calculating its standard deviation.

[0114] In addition, to more objectively and accurately determine the weighting coefficient of pore volume on the degree of adhesive layer degradation The weighting coefficient of the number of interconnected pores on the degree of adhesive layer degradation The values ​​are used to ensure the accuracy of the adhesive layer degradation calculation, and the weighting coefficients are... and The specific steps are as follows:

[0115] S271. Obtain several tooth samples containing inlays, as well as the total volume of pores in the adhesive layer and the number of interconnected pore clusters in the adhesive layer in each tooth sample, and perform Z-score normalization.

[0116] S272. Classify tooth samples according to their appearance and set corresponding adhesive layer degradation scores; the tooth sample classification types include completely successful, slightly degraded, moderately degraded, and severely degraded / failed, wherein:

[0117] The fully successful tooth sample had no marginal staining, no secondary caries, and the inlay was intact, with a corresponding adhesive layer degradation score of j0;

[0118] The slightly degenerated tooth sample had slight edge staining and no secondary caries, and the corresponding adhesive layer degradation score was j1;

[0119] The moderately degraded tooth sample had obvious marginal staining and mild secondary caries, and the corresponding adhesive layer degradation score was j2.

[0120] The severely degraded and failed tooth samples had inlays that fell off and had severe secondary caries, with the corresponding adhesive layer degradation score being j3;

[0121] And j0 < j1 < j2 < j3. For the convenience of subsequent calculations, j0, j1, j2 and j3 can be set to 0, 1, 2 and 3 respectively.

[0122] S273. Construct the influence function and likelihood function between the adhesive layer degradation score and the total volume of pores within the adhesive layer and the number of interconnected pore clusters within the adhesive layer. The expression for the influence function is:

[0123]

[0124] The expression for the likelihood function is:

[0125]

[0126]

[0127] In the above formula, This represents an ordered logistic regression function. This represents the probability that the adhesive layer degradation score does not exceed j. This represents the j-th tangent parameter. and The regression coefficients represent the total volume of pores within the standardized adhesive layer and the number of interconnected pore clusters within the adhesive layer, respectively. This is an indicator function; it takes the value 1 when the adhesive layer degradation score of sample i is j, and 0 otherwise. , and Let represent the probability that the i-th sample has an adhesive layer degradation score of j, the probability that the i-th sample has an adhesive layer degradation score of j less than or equal to j, and the probability that the i-th sample has an adhesive layer degradation score of j-1 less than or equal to j, respectively. The number of samples is n.

[0128] S274. Estimate the weighting coefficient of pore volume on the degree of adhesive layer degradation using maximum likelihood estimation. The weighting coefficient of the number of interconnected pores on the degree of adhesive layer degradation .

[0129] S3. The clinical examination data and microscopic parameters are concatenated to obtain a fusion feature vector. Based on the fusion feature vectors at different times after the target tooth restoration, the prediction result of the inlay at any future moment, including the comprehensive health score and risk classification results, is predicted. Step S3 specifically includes the following steps:

[0130] S31. The clinical examination data and microscopic parameters at different times after the target tooth restoration are concatenated to obtain the fused feature vector at the corresponding time, the expression of which is:

[0131]

[0132] In the above formula, This represents the fused feature vector at time t. This represents the Nth parameter in the fusion feature vector, thereby enabling the splicing of micro and macro features to form a fusion feature vector at each time point (such as 6 months and 12 months after inlay repair).

[0133] S32. Combine the fused feature vectors from different times in chronological order and standardize them using Z-score to obtain a temporal feature vector sequence. This prevents significant differences in the dimensions and orders of magnitude of the features. Calculate the time interval between each element and the first element in the temporal feature vector sequence to construct a time interval vector sequence. The expression for the temporal feature vector sequence is: The expression for the time interval vector sequence is: ,in, Indicates in The fused feature vector at each time step.

[0134] S33. Construct a Transformer model, input the time-series feature vector sequence and time interval vector sequence into the Transformer model, and obtain the Transformer model encoder at the last time point. The output is used as a comprehensive health status code, and its implementation method is further detailed below:

[0135] S331. The fused feature vectors at each time point are sequentially mapped to a high-dimensional space through a fully connected neural network to obtain the initial feature vectors at each time point, the expression of which is:

[0136]

[0137] In the above formula, Indicates the first The initial feature vector of the fused feature vector at each time point. Indicates the first The fused feature vectors at each time step are input into a fully connected neural network (MLP) to map to a high-dimensional space.

[0138] S332. Map any time interval to an initial time vector with the same dimension as the initial feature embedding, expressed as:

[0139]

[0140] In the above formula, For the first The initial time vector of time intervals. For another fully connected neural network.

[0141] S333. Add the initial feature vector and the initial time vector at each time point to obtain the fusion vector at each time point.

[0142] S334. Construct a Transformer model and input the fused vector at each time step into the encoder of the Transformer model.

[0143] S335. Obtain the output of the Transformer model encoder at the last time point as the comprehensive health status code.

[0144] S34. Predict the comprehensive health score at any future time based on the comprehensive health status code, and classify the corresponding risk categories based on the comprehensive health score. Each risk category corresponds to a comprehensive health score within a fixed numerical range, so as to obtain a prediction result that includes the comprehensive health score and risk classification results.

[0145] S4. Calculate the inlay repair quality level based on the fused feature vector, comprehensive health status encoding, and prediction results. Step S4 specifically includes the following steps:

[0146] S41. Input the comprehensive health status code into the fully connected layer to calculate the current health score. The calculation formula is as follows:

[0147]

[0148] In the above formula, Indicates the current health score. As an activation function, it compresses the output to the (0,1) interval. and These represent the weight parameters and bias parameters, respectively.

[0149] S42. Calculate the future risk index based on the forecast results. The calculation formula is as follows:

[0150]

[0151] In the above formula, Indicates a future risk index. Represents any future moment Comprehensive health score at the time, The weighting coefficients for risk categories corresponding to the comprehensive health score can be preset in advance. The higher the risk, the higher the weighting coefficient can be assigned. When calculating the future risk index at the selected reference time, it can be determined based on the quality requirements within a certain period of time. For example, if the quality requirements within 24 months need to reach certain indicators, then the future risk index at 24 months after inlay repair can be calculated.

[0152] S43. Determine the inlay repair quality level based on the future risk index, fusion feature vector, current health score, and preset rules. Different inlay repair quality levels exist in the preset rules. Each inlay repair quality level corresponds to different requirements for the future risk index, fusion feature vector, and current health score. Thus, the inlay repair quality level can be directly determined based on the preset rules.

[0153] The present invention also provides an intelligent grading and evaluation system for inlay repair quality, including a processor and a memory, wherein the memory is used to store a computer program, and the computer program is executed by the processor to implement an intelligent grading and evaluation method for inlay repair quality.

[0154] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0155] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Since the above embodiments are substantially similar to the method embodiments, their descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0156] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for intelligent grading and evaluating the quality of inlay restorations, characterized in that, The grading assessment method specifically includes the following steps: S1. Acquire tooth data at different times after the target tooth restoration, the tooth data including clinical examination data and several SEM images of the target tooth under the same shooting angle and different light sources; S2. Identify and segment the regions corresponding to different materials based on the SEM images. Construct a microscopic 3D image of the target tooth based on the segmentation results and the albedo of each region, and calculate several microscopic parameters. S3. By splicing together clinical examination data and microscopic parameters to obtain a fusion feature vector, and based on the fusion feature vectors at different times after the target tooth restoration, predict the inlay's prediction results at any future moment, including the comprehensive health score and risk classification results. S4. Calculate the inlay repair quality level based on the fused feature vector, comprehensive health status coding, and prediction results.

2. The grading assessment method according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21. Preprocess the SEM image to enhance local contrast and highlight edge information to obtain a preprocessed image; S22. Add a self-attention mechanism module between the encoder and decoder of the U-Net deep learning model to obtain an improved U-Net deep learning model, and use the pre-trained improved U-Net deep learning model to perform semantic segmentation on the preprocessed image so as to divide each pixel in the preprocessed image into several classification labels. S23. Based on the classification labels corresponding to each pixel in the preprocessed image, obtain the classification labels corresponding to different pixels in the corresponding SEM image, as well as the material and original albedo corresponding to the classification labels. S24. Construct a global energy function for joint optimization of normal vector and albedo based on several SEM images of the target tooth taken at the same time after tooth restoration under different light sources at the same shooting angle. S25. Use an iterative algorithm to minimize the global energy function, and in each iteration, alternately minimize the albedo and normal vector to update the normal vector and albedo respectively. S26. Calculate the height of each pixel based on the final normal vector of each pixel in the SEM image to construct a microscopic 3D image. S27. Calculate several micro parameters based on the final albedo of each pixel in the SEM image and the micro 3D image.

3. The grading assessment method according to claim 2, characterized in that, In step S24, the expression for the global energy function is: ； ； ； ； In the above formula, Represents the global energy function. For data fidelity items, Represents the smoothing term of the normal vector. and Let these represent the normal vector hyperparameter and the albedo hyperparameter, respectively. Represents the midpoint of the SEM image under the i-th light source. grayscale value, Representing coordinates The albedo, i.e., the albedo field, Representing coordinates The normal vector at that point, i.e., the normal vector field. Let represent the unit vector of the light source direction in the SEM image under the i-th light source. There are k different light sources. Represents the gradient operator, Representing coordinates Gradient of the normal vector at that point Represents the albedo smoothing term. This represents the p-th material region. Let the set of pixels belonging to the same material region as any pixel within the p-th material region be defined. This represents the reflectance of any pixel within the p-th material region. It represents the reflectance of the pixels in the set of neighboring pixels of the pixel in the p-th material region.

4. The grading assessment method according to claim 2, characterized in that, In step S27, the microscopic parameters include the total volume of inlay microcracks in the target tooth, the number of inlay pores, the degree of degradation of the adhesive layer, the average albedo variation in the adhesive region, the standard deviation of albedo in the inlay region, and the average gap width, maximum gap width, and critical index of marginal defects between the inlay and the natural tooth tissue. The formula for calculating the degree of degradation of the adhesive layer is: ； The formula for calculating the average albedo change in the adhesive region is: ； The formula for calculating the critical index of edge defects is: ； In the above formula, Indicates the degree of degradation of the adhesive layer. This represents the total volume of pores within the adhesive layer. This indicates the number of interconnected pore clusters within the adhesive layer. and These represent the weighting coefficients of pore volume and the number of interconnected pores on the degree of adhesive layer degradation, respectively. This represents the change in the average albedo of adhesive region A. This indicates the number of pixels in adhesive region A. This indicates that any coordinate in adhesive region A is... The final albedo of the pixel, This represents the original albedo of the material corresponding to adhesive region A. This indicates the critical index for edge defects. This represents the gap width at the i-th sampling point in the gap between the inlay and the natural tooth structure. A total of N sampling points were set.

5. The grading assessment method according to claim 4, characterized in that, The weighting factor of the pore volume on the degree of adhesive layer degradation The weighting coefficient of the number of interconnected pores on the degree of adhesive layer degradation The specific steps are as follows: S271. Obtain several tooth samples containing inlays, as well as the total volume of pores in the adhesive layer and the number of interconnected pore clusters in the adhesive layer in each tooth sample, and perform Z-score normalization. S272. Classify tooth samples according to their appearance and set corresponding adhesive layer degradation scores; the tooth sample classification types include completely successful, slightly degraded, moderately degraded, and severely degraded / failed, wherein: The fully successful tooth sample had no marginal staining, no secondary caries, and the inlay was intact, with a corresponding adhesive layer degradation score of j0; The slightly degenerated tooth sample had slight edge staining and no secondary caries, and the corresponding adhesive layer degradation score was j1; The moderately degraded tooth sample had obvious marginal staining and mild secondary caries, and the corresponding adhesive layer degradation score was j2. The severely degraded and failed tooth samples had inlays that fell off and had severe secondary caries, with the corresponding adhesive layer degradation score being j3; And j0 < j1 < j2 < j3; S273. Construct the influence function and likelihood function between the adhesive layer degradation score and the total volume of pores in the adhesive layer and the number of interconnected pore clusters in the adhesive layer; S274. Estimate the weighting coefficient of pore volume on the degree of adhesive layer degradation using maximum likelihood estimation. The weighting coefficient of the number of interconnected pores on the degree of adhesive layer degradation .

6. The grading assessment method according to claim 5, characterized in that, In step S273, the expression for the influencing function is: ； The expression for the likelihood function is: ； ； In the above formula, This represents an ordered logistic regression function. This represents the probability that the adhesive layer degradation score does not exceed j. This represents the j-th tangent parameter. and The regression coefficients represent the total volume of pores within the standardized adhesive layer and the number of interconnected pore clusters within the adhesive layer, respectively. This is an indicator function; it takes the value 1 when the adhesive layer degradation score of sample i is j, and 0 otherwise. , and Let represent the probability that the i-th sample has an adhesive layer degradation score of j, the probability that the i-th sample has an adhesive layer degradation score of j less than or equal to j, and the probability that the i-th sample has an adhesive layer degradation score of j-1 less than or equal to j, respectively. The number of samples is n.

7. The grading assessment method according to claim 1, characterized in that, Step S3 specifically includes the following steps: S31. The clinical examination data and microscopic parameters at different times after the target tooth restoration are concatenated to obtain the fused feature vector at the corresponding time, the expression of which is: ； In the above formula, This represents the fused feature vector at time t. This represents the Nth parameter in the fused feature vector; S32. Combine the fused feature vectors from different times in chronological order and standardize them with Z-score to obtain a temporal feature vector sequence, and calculate the time interval between each element in the temporal feature vector sequence and the first element to construct a time interval vector sequence. S33. Construct a Transformer model, input the time-series feature vector sequence and time interval vector sequence into the Transformer model, and obtain the Transformer model encoder at the last time point. The output is used as a comprehensive health status code; S34. Predict the comprehensive health score at any future time based on the comprehensive health status code, and classify the corresponding risk categories based on the comprehensive health score to obtain a prediction result that includes the comprehensive health score and risk classification results.

8. The grading assessment method according to claim 7, characterized in that, Step S33 specifically includes the following steps: S331. The fused feature vectors at each time point are sequentially mapped to a high-dimensional space through a fully connected neural network to obtain the initial feature vectors at each time point, the expression of which is: ； In the above formula, Indicates the first The initial feature vector of the fused feature vector at each time point. Indicates the first The fused feature vectors at each time are input into a fully connected neural network MLP to map to a high-dimensional space; S332. Map any time interval to an initial time vector with the same dimension as the initial feature embedding, expressed as: ； In the above formula, For the first The initial time vector of time intervals. For another fully connected neural network; S333. Add the initial feature vector and the initial time vector at each time point to obtain the fusion vector at each time point; S334. Construct a Transformer model and input the fused vector at each time step into the encoder of the Transformer model; S335. Obtain the output of the Transformer model encoder at the last time point as the comprehensive health status code.

9. The grading assessment method according to claim 1, characterized in that, Step S4 specifically includes the following steps: S41. Input the comprehensive health status code into the fully connected layer to calculate the current health score. The calculation formula is as follows: ； In the above formula, Indicates the current health score. As an activation function, it compresses the output to the (0,1) interval. and These represent the weight parameters and bias parameters, respectively. S42. Calculate the future risk index based on the forecast results. The calculation formula is as follows: ； In the above formula, Indicates a future risk index. Represents any future moment Comprehensive health score at the time, Weighting coefficients for risk categories corresponding to the overall health score; S43. Determine the inlay repair quality level based on the future risk index, fusion feature vector, current health score, and preset rules.

10. A system for implementing the grading evaluation method according to any one of claims 1-9, characterized in that, It includes a processor and a memory, the memory being used to store a computer program that, when executed by the processor, implements the hierarchical evaluation method as described in any one of claims 1-9.

Images