An oral diagnosis method for establishing an oral diagnosis neural network model

Abstract

The application relates to the technical field of oral cavity diagnosis, in particular to an oral cavity diagnosis method for establishing an oral cavity diagnosis neural network model, which comprises the following steps: collecting oral cavity CT, endoscope hard tissue three-dimensional coordinates, soft tissue color parameters, tooth surface texture gray value parameters, checking and removing abnormal values to generate an oral cavity standardized original data set, extracting features and assigning weights to generate a weighted signal set, iteratively training to obtain diagnosis network weight parameters, calculating lesion matching degrees, dividing grades to establish oral cavity lesion quantitative grading data, in the application, oral cavity hard tissue three-dimensional coordinates, soft tissue color parameters and tooth surface texture gray value parameters are collected, values exceeding the normal physiological range are checked and removed to generate a standardized data set, features are extracted, converted into discrete signals and assigned with degree of distinction weights, stable diagnosis criteria are established through hierarchical transmission calculation and iterative training, accurate positioning and quantitative grading of oral cavity lesions are realized, and the diagnosis precision, efficiency and result consistency are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of oral diagnostic technology, and in particular to an oral diagnostic method for establishing an oral diagnostic neural network model. Background Technology

[0002] The field of oral diagnostic technology covers core aspects such as oral disease detection, lesion identification, image analysis, and data processing. The diagnostic process is completed by collecting data on the internal tissue status of the oral cavity, extracting features, and integrating information. Traditional oral diagnostic methods rely on manual visual examination, probe examination, X-ray image interpretation, and pathological slide observation to judge the oral condition and identify diseases. They rely on human experience to complete data comparison and result determination, and form diagnostic basis by manually recording and organizing oral examination information.

[0003] Current oral diagnostic technologies rely heavily on manual operation and experience-based judgment. Data acquisition is limited to manual visual inspection, probe examination, X-ray image interpretation, and pathological slide observation. This limits the dimensions of oral physiological information that can be covered, failing to simultaneously acquire multi-dimensional physiological data such as the three-dimensional morphology of hard tissues, dynamic color changes of soft tissues, and subtle differences in tooth surface texture. The comprehensiveness of data acquisition is inherently limited. Furthermore, the manual operation process is directly affected by the operator's technique, observation angle, and equipment proficiency, making it prone to data acquisition bias and the omission of crucial lesion information. For example, manual visual inspection and routine X-ray examinations are highly likely to miss early small defects in tooth hard tissues and early periodontal tissue resorption, directly delaying lesion detection and missing the optimal intervention window. The diagnostic judgment relies entirely on the operator's personal clinical experience, lacking standardized judgment criteria. Operators with different years of experience and professional levels can have significantly different judgments on the same oral condition. For instance, different dentists may reach completely different conclusions about early periapical lesions based on the same oral X-ray image, failing to guarantee the consistency and reproducibility of diagnostic results. The manual recording and organization of diagnostic information makes it impossible to achieve standardized management and quantitative analysis of oral data. The output diagnostic results are mostly qualitative descriptions, lacking accurate lesion grading and quantitative data support. This is not only detrimental to the long-term dynamic tracking and comparison of oral lesions, but also prone to problems such as non-standard diagnostic records and missing key information, bringing uncertainty to the formulation of subsequent clinical treatment plans. At the same time, the efficiency of the manual operation mode is directly limited by the number of operators, working hours and mental state, making it impossible to achieve rapid and standardized diagnosis of large-scale oral samples and failing to meet the application needs of large-scale oral health screening. Summary of the Invention

[0004] To address the technical problems existing in the prior art, this invention provides a method for establishing an oral diagnostic neural network model, comprising the following steps: To achieve the above objectives, the present invention adopts the following technical solution: an oral diagnosis method for establishing an oral diagnosis neural network model, comprising the following steps: S1: Collect the three-dimensional coordinates of oral hard tissues, soft tissue color parameters, and tooth surface texture gray values ​​from oral CT equipment and oral endoscopes; verify the values ​​of each parameter; remove values ​​that exceed the normal physiological range; and generate a standardized raw dataset of oral cavity. S2: Based on the standardized original dataset of the oral cavity, extract the hard tissue coordinate difference, color parameter ratio, and texture gray value change, convert them into discrete feature signals, assign feature discrimination weights, and generate a set of oral cavity feature weighted signals. S3: Call the oral cavity feature weighted signal set, build the input layer, hidden layer and output layer structure, pass the feature signals according to the level, calculate the weighted sum of each level, compare with the activation threshold, retain the signals that meet the conditions, iteratively train with known lesion feature data, calculate the deviation value between the output result and the actual lesion, adjust the feature weights until the deviation value is lower than the diagnostic accuracy threshold, and obtain the initial weight parameters of the oral cavity diagnosis network. S4: For the standardized original oral dataset, call the initial weight parameters of the oral diagnosis network, calculate the matching degree between the features of each oral region and the lesion features, and generate the matching degree value of the oral region lesion. S5: Based on the matching degree value of oral region lesions, compare it with the preset lesion judgment threshold, filter the regions with matching degree values ​​higher than the threshold, classify the lesion level, and establish quantitative grading data of oral lesions.

[0005] As a further aspect of the present invention, the oral standardization raw dataset includes verified three-dimensional coordinates, qualified color parameters, and effective tooth surface texture gray values.

[0006] As a further aspect of the present invention, the oral cavity feature weighted signal set includes a weighted hard tissue difference signal, a weighted color proportion signal, and a weighted texture change signal.

[0007] As a further aspect of the present invention, the initial weight parameters of the oral diagnostic network include input layer weights, hidden layer weights, output layer weights, and an iterative optimization threshold; The matching degree values ​​for oral region lesions include hard tissue matching degree, soft tissue matching degree, and tooth surface texture matching degree.

[0008] As a further aspect of the present invention, the quantitative grading data for oral lesions includes the lesion area, lesion grade, and quantitative grading index.

[0009] As a further aspect of the present invention, the specific steps of S1 are as follows: S101: Collect the three-dimensional coordinates of oral hard tissues, soft tissue color parameters, and tooth surface texture gray values ​​from oral CT equipment and oral endoscopes. Record the three-dimensional coordinates of oral hard tissues, soft tissue color parameters, and tooth surface texture gray values ​​item by item, integrate the above three types of parameters to form a unified data set, and generate the original set of oral multi-source parameters. S102: Based on the original set of oral multi-source parameters, the three-dimensional coordinates of oral hard tissues, soft tissue color parameters, and tooth surface texture gray values ​​are compared with the normal physiological range values ​​to determine the range of the parameters, mark the parameter items that exceed the normal physiological range, and obtain the abnormal parameter identifier set. S103: Based on the original set of oral multi-source parameters and the set of abnormal parameter identifiers, remove the parameter items corresponding to the abnormal parameter identifiers, retain the parameter items within the normal physiological range, and arrange the retained parameters in a regular manner to generate a standardized original dataset of oral parameters.

[0010] As a further aspect of the present invention, the specific steps of S2 are as follows: S201: Based on the standardized raw dataset of the oral cavity, the three-dimensional coordinates of hard tissue are extracted to calculate the difference, the proportion of soft tissue color parameters is statistically analyzed, the change in gray value of tooth surface texture is measured, and the three types of data are integrated to form a set of feature indicators, thus obtaining the core feature indicator group of the oral cavity. S202: Call the core feature index group of the oral cavity, convert the hard tissue coordinate difference, color parameter ratio, and texture gray value change into discrete numerical signals one by one, sort out the signal arrangement logic, and generate a set of discrete feature signals of the oral cavity. S203: For the set of discrete feature signals in the oral cavity, analyze the discrimination of each discrete feature signal, assign corresponding weight values ​​according to the discrimination of high and low, and form a complete data set by associating the signals and weights to generate a weighted signal set of oral cavity features.

[0011] As a further aspect of the present invention, the specific steps of S3 are as follows: S301: Call the oral cavity feature weighted signal set, build the input layer, hidden layer and output layer structure, pass the feature weighted signal in sequence according to the level, calculate the weighted sum of the signals of each level, compare the weighted sum with the activation threshold, retain the signals that meet the activation conditions, and generate the oral cavity level transmission effective signal group. S302: Based on the effective signal group transmitted at the oral cavity level, known lesion feature data is introduced to perform iterative calculation, calculate the difference between the network output result and the actual lesion, and obtain the oral cavity diagnosis deviation value; S303: Based on the oral diagnosis deviation value, compare the deviation value with the diagnostic accuracy threshold, adjust the corresponding feature weights according to the comparison results, repeat the weight adjustment operation, and obtain the initial weight parameters of the oral diagnosis network.

[0012] As a further aspect of the present invention, the specific steps of S4 are as follows: S401: For the standardized original dataset of the oral cavity, divide the oral cavity into independent regions, extract the hard tissue coordinates, soft tissue color, and tooth surface texture gray values ​​of each region, integrate them to form the corresponding region feature set, and obtain the oral cavity partition feature parameter set; S402: Call the oral cavity partition feature parameter set and the initial weight parameters of the oral diagnosis network, substitute the partition feature parameters into the weight parameters to perform the calculation, extract the baseline values ​​of lesion features, and generate oral lesion feature reference values; S403: Based on the oral cavity partition feature parameter set and oral lesion feature reference value, calculate the degree of correspondence between the two, quantify the degree of correspondence and assign values, classify the assignment results by region, and generate the oral cavity region lesion matching degree value.

[0013] As a further aspect of the present invention, the specific steps of S5 are as follows: S501: Based on the lesion matching degree values ​​of oral regions, retrieve the preset lesion judgment threshold, compare the matching degree values ​​of each region with the lesion judgment threshold one by one, record the comparison results, mark the regions with values ​​higher than the threshold, and obtain the lesion suspected region marking table. S502: Call the suspected lesion area marking table, extract the matching degree value of each suspected area, divide the interval according to the value, assign corresponding grade number to different intervals, associate the area with the grade number, and generate an oral lesion grade correspondence table. S503: Based on the suspected lesion area marking table and the oral lesion grade correspondence table, integrate the regional information matching degree numerical grade number, sort out the data association logic, form a complete quantitative system, and establish oral lesion quantitative grading data.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, three-dimensional coordinates of oral hard tissues, color parameters of soft tissues, and gray values ​​of tooth surface texture are collected. Values ​​exceeding the normal physiological range are verified and eliminated to generate a standardized dataset. Features are extracted, converted into discrete signals, and assigned discrimination weights. Through hierarchical transfer calculation and iterative training, a stable diagnostic benchmark is established to achieve accurate localization and quantitative grading of oral lesions, significantly improving diagnostic accuracy, efficiency, and consistency of results. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0019] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0020] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0021] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0022] Please see Figure 1 This invention provides a method for establishing an oral diagnostic neural network model, comprising the following steps: S1: Collect the three-dimensional coordinates of oral hard tissues, soft tissue color parameters, and tooth surface texture gray values ​​from oral CT equipment and oral endoscopes; verify the values ​​of each parameter; remove values ​​that exceed the normal physiological range; and generate a standardized raw dataset of oral cavity. The standardized raw dataset for oral cavity includes verified three-dimensional coordinates, qualified color parameters, and valid grayscale values ​​of tooth surface texture.

[0023] S2: Based on the standardized original dataset of the oral cavity, extract the hard tissue coordinate difference, color parameter ratio, and texture gray value change, convert them into discrete feature signals, assign feature discrimination weights, and generate a set of oral cavity feature weighted signals. The oral cavity feature weighted signal set includes weighted hard tissue difference signal, weighted color proportion signal, and weighted texture change signal.

[0024] S3: Call the oral cavity feature weighted signal set, build the input layer, hidden layer and output layer structure, pass the feature signals according to the level, calculate the weighted sum of each level, compare with the activation threshold, retain the signals that meet the conditions, iteratively train with known lesion feature data, calculate the deviation value between the output result and the actual lesion, adjust the feature weights until the deviation value is lower than the diagnostic accuracy threshold, and obtain the initial weight parameters of the oral cavity diagnosis network. The initial weight parameters of the oral diagnostic network include the input layer weights, hidden layer weights, output layer weights, and the iterative optimization threshold.

[0025] S4: For the standardized original oral dataset, call the initial weight parameters of the oral diagnosis network, calculate the matching degree between the features of each oral region and the lesion features, and generate the matching degree value of the oral region lesion. The matching degree of oral region lesions includes hard tissue matching degree, soft tissue matching degree, and tooth surface texture matching degree.

[0026] S5: Based on the matching degree value of oral region lesions, compare it with the preset lesion judgment threshold, filter the regions with matching degree values ​​higher than the threshold, classify the lesion level, and establish quantitative grading data of oral lesions. The quantitative grading data for oral lesions includes the lesion area, lesion grade, and quantitative grading indicators.

[0027] Please see Figure 2 The specific steps of S1 are as follows: S101: Collect the three-dimensional coordinates of oral hard tissues, soft tissue color parameters, and tooth surface texture gray values ​​from oral CT equipment and oral endoscopes. Record the three-dimensional coordinates of oral hard tissues, soft tissue color parameters, and tooth surface texture gray values ​​item by item, integrate the above three types of parameters to form a unified data set, and generate the original set of oral multi-source parameters. The three-dimensional coordinates of oral hard tissues, soft tissue color parameters, and tooth surface texture grayscale values ​​were acquired using oral CT equipment and oral endoscopes. Each of these parameters was recorded individually. For hard tissue three-dimensional coordinates, the coordinates of the enamel layer, dentin layer, and periodontal ligament layer were selected, with at least 15 coordinate points selected for each tooth. For example, for the central incisor, labial / lingual, mesiodistal, and vertical coordinates were selected. For soft tissue color parameters, RGB three-channel values ​​were selected, with R values ​​of 180-220, G values ​​of 100-140, and B values ​​of 80-110 for the gingival region. Tooth surface texture grayscale values ​​were selected as integer values ​​within the range of 0-255. Each type of parameter was entered line by line according to the acquisition sequence. The same parameter was acquired three times, and the arithmetic mean was calculated. Single acquisitions with a deviation from the mean exceeding 10 were discarded. The abnormal collected values ​​are classified according to the corresponding region numbering of hard tissue coordinates, color parameters and gray values. The three types of parameters in the same region are arranged and combined in a fixed order to form a continuous data segment. All regional data segments are uniformly spliced ​​to form a complete and coherent data set, and the original set of oral multi-source parameters is generated. S102: Based on the original set of oral multi-source parameters, the three-dimensional coordinates of oral hard tissues, soft tissue color parameters, and tooth surface texture gray values ​​are compared with the normal physiological range values ​​to determine the range of the parameters, mark the parameter items that exceed the normal physiological range, and obtain the abnormal parameter identifier set. Based on the original set of multi-source oral parameters, the normal physiological range values ​​of hard tissue three-dimensional coordinates, soft tissue color parameters, and tooth surface texture gray values ​​were determined. The normal range of hard tissue three-dimensional coordinates was set as X-axis -20mm to 20mm, Y-axis -15mm to 15mm, and Z-axis 0mm to 30mm. The normal range of soft tissue color parameters was set as R170-230, G90-150, and B70-120. The normal range of tooth surface texture gray values ​​was set as 40-220. Each parameter was compared with its corresponding normal physiological range value to determine whether the parameter value fell within the above range. Parameters with coordinates exceeding ±20mm, color channels exceeding the corresponding upper and lower limits, or gray values ​​lower than 40 or higher than 220 were marked. Each marked parameter was assigned a unique region number and parameter type code. The marked information was summarized in order to form independent data items. All marked data items were stored together to obtain an abnormal parameter identifier set. S103: Based on the original set of oral multi-source parameters and the abnormal parameter identifier set, remove the parameter items corresponding to the abnormal parameter identifier set, retain the parameter items within the normal physiological range, and arrange the retained parameters in a regular manner to generate a standardized original dataset of oral cavity. Based on the original set of oral multi-source parameters and the set of abnormal parameter identifiers, the markers carrying region numbers and parameter type codes in the abnormal parameter identifier set are read. The parameter positions matching the markers in the original set of oral multi-source parameters are located, and the located parameters are removed from the original data sequence. Unmarked parameters within the normal physiological range are retained. The retained parameters are reordered in ascending order of region number. Within the same region, they are arranged in a fixed order of hard tissue 3D coordinates, soft tissue color parameters, and tooth surface texture gray values. The arranged data are formatted uniformly, with coordinates retained to two decimal places and color parameters and gray values ​​retained as integers. Adjacent region data are seamlessly connected to remove redundant intervals between data, forming a structurally regular and formatted continuous data set, generating a standardized oral raw dataset.

[0028] Please see Figure 3 The specific steps of S2 are as follows: S201: Based on the standardized raw dataset of the oral cavity, the three-dimensional coordinates of hard tissue are extracted to calculate the difference, the proportion of soft tissue color parameters is statistically analyzed, the change in gray value of tooth surface texture is measured, and the three types of data are integrated to form a set of feature indicators, thus obtaining the core feature indicator group of the oral cavity. Based on the standardized oral dataset, the three-dimensional coordinates of hard tissue in individual teeth and corresponding periodontal regions were extracted. The X, Y, and Z axis values ​​of adjacent points on the same tooth were selected, and the absolute difference was calculated by subtracting the value of the preceding point from the value of the posterior point. For example, the X-axis coordinates of adjacent points of the central incisor were 12.34 mm and 12.56 mm, with a difference of 0.22 mm. The coordinate difference was calculated region by region for all teeth. Then, the R, G, and B color parameters from the soft tissue were extracted. The channel value is calculated by dividing the single channel value by the sum of the three channel values ​​to obtain the single channel percentage. For example, in the gingival region, R=200, G=120, B=100, the total is 420, and the R channel percentage is 200÷420≈0.476. Then, five consecutive gray values ​​are selected on the same texture line of the tooth surface, and the change is obtained by subtracting the previous gray value from the next gray value. For example, gray values ​​82, 85, 87, 84, 88, the adjacent changes are 3, 2, -3, 4 respectively. The calculated coordinate difference, color parameter percentage, and gray value change are uniformly classified according to the region number. The three types of indicators are combined into independent feature items in a fixed order. The feature items of all regions of the whole mouth are summarized and arranged to obtain the core feature index group of the oral cavity. S202: Call the core feature index group of the oral cavity, convert the hard tissue coordinate difference, color parameter ratio, and texture gray value change into discrete numerical signals one by one, sort out the signal arrangement logic, and generate a set of discrete feature signals of the oral cavity. The system calls upon the core oral feature index group, sequentially reading three types of values: hard tissue coordinate difference, soft tissue color parameter ratio, and tooth surface texture grayscale value change. The coordinate difference is converted to an integer with a minimum unit of 0.01mm (e.g., 0.22mm is converted to 22). The color parameter ratio is multiplied by 1000 and rounded (e.g., 0.476 is converted to 476). The grayscale value change is directly retained as an original integer value, completing the conversion from continuous to discrete values. A fixed type identifier is assigned to each type of discrete value: coordinate difference is identified as 1, color ratio as 2, and grayscale change as 3. These identifiers are arranged in ascending order of region number, and within the same region, they are arranged in the order of identifier 1, identifier 2, and identifier 3. All discrete values ​​are combined with their corresponding identifiers into an ordered signal sequence. The signal sequence is then processed to ensure consistent signal length and no redundant data, generating an oral discrete feature signal set. S203: For the set of discrete feature signals in the oral cavity, analyze the discrimination of each discrete feature signal, assign corresponding weight values ​​according to the discrimination of high and low, and form a complete data set by associating the signals and weights to generate a weighted signal set of oral cavity features; For the discrete feature signal set of the oral cavity, three types of signals—hard tissue coordinate difference, soft tissue color parameter proportion, and tooth surface texture gray value change—are compared with preset discrimination benchmarks. The discrimination benchmarks are set as follows: coordinate difference 50, color proportion 300, and gray value change 10. Signal values ​​higher than the corresponding benchmarks are judged as high discrimination, and those lower than the corresponding benchmarks are judged as low discrimination. A weight of 0.4 is assigned to high discrimination signals, 0.35 to medium discrimination signals, and 0.25 to low discrimination signals. Each discrete feature signal is bound one-to-one with its corresponding weight value, and the data items containing the weights are recombined according to the original region and signal order. All weighted data items are uniformly arranged to ensure that the data structure and signal sequence maintain a corresponding relationship without misalignment or missing data, thus generating a weighted signal set of oral cavity features.

[0029] Please see Figure 4 The specific steps of S3 are as follows: S301: Call the oral cavity feature weighted signal set, build the input layer, hidden layer and output layer structure, pass the feature weighted signal in sequence according to the level, calculate the weighted sum of the signals of each level, compare the weighted sum with the activation threshold, retain the signals that meet the activation conditions, and generate the oral cavity level transmission effective signal group. A three-layer structure (input layer, hidden layer, and output layer) is constructed using a weighted signal set of oral cavity features. The input layer has 30 nodes, corresponding to the 30 feature classes in the weighted signal set. The hidden layer has 15 nodes, and the output layer has 5 nodes, corresponding to different oral lesion types. A simple transfer program is written in Python to sequentially input the weighted signals of the oral cavity features into the input layer nodes according to their regions. Each input node signal is transferred to the hidden layer according to its corresponding weight. The weighted sum of all hidden layer nodes is calculated using the formula S = Σ(xi × wi), where xi is the signal value of the i-th node in the input layer, and wi is the connection weight from the input layer to the i-th node in the hidden layer. wi is initially set to a random number between 0.01 and 0.05. For example, in input layer 3... The node signals are 476, 22, and 4, with corresponding weights of 0.02, 0.03, and 0.04. The weighted sum S = 476 × 0.02 + 22 × 0.03 + 4 × 0.04 = 9.52 + 0.66 + 0.16 = 10.34. The activation threshold is set to 8.5, which is determined by referring to the average of the weighted sums of normal and pathological characteristic signals in the oral cavity. The calculated weighted sums of each level are compared with the activation threshold of 8.5. Signals with a weighted sum greater than 8.5 are retained, and signals with a weighted sum less than or equal to 8.5 are removed. The retained effective signals are marked by level and region, and summarized to form an ordered signal set, generating an effective signal group for oral cavity hierarchical transmission. S302: Based on the effective signal group transmitted at the oral cavity level, known lesion feature data is introduced to perform iterative calculation, calculate the difference between the network output result and the actual lesion, and obtain the oral cavity diagnosis deviation value; Based on the oral cavity hierarchical transmission effective signal set, known lesion feature data is introduced. This data consists of 100 clinically diagnosed caries, periodontitis, and periapical periodontitis samples. Each sample contains the characteristic signal of the corresponding lesion area and the actual lesion type identifier. An iterative calculation table is built using Excel. The signal values ​​from the oral cavity hierarchical transmission effective signal set are input into the calculation table and matched with the known lesion feature data. Each iteration selects 20 samples and calculates the difference between the network output and the actual lesion type identifier. The difference is calculated using the formula E = |Yj - Yj'|, where Yj is the lesion type identifier value output by the network, and Yj' is the actual lesion type identifier value. For example, if the network output identifier for a certain region is 2 and the actual lesion identifier is 3, the difference E = |2 - 3| = 1. This iterative calculation is repeated for all 100 samples. After each iteration, the difference data for all regions is recorded. The arithmetic mean of the differences from multiple iterations for the same region is taken. For example, if a region is 5... The iteration differences were 1, 0.8, 1.2, 0.9, and 1.1, respectively, and the average value was (1+0.8+1.2+0.9+1.1)÷5=1.0. The average difference of all regions was summed to obtain the oral diagnosis deviation value. S303: Based on the oral diagnosis deviation value, compare the deviation value with the diagnostic accuracy threshold, correct the corresponding feature weights according to the comparison results, repeat the weight correction operation, and obtain the initial weight parameters of the oral diagnosis network. Based on the oral diagnosis deviation values, the diagnostic accuracy threshold is set to 0.5. This threshold is determined with reference to the allowable error range of clinical oral diagnosis. An Excel function is used to compare the oral diagnosis deviation value of each region with the diagnostic accuracy threshold of 0.5. If the deviation value is greater than 0.5, it indicates that the corresponding feature weights are unreasonable and need to be corrected. The weight correction formula is wi'=wi-η×(E / xi), where wi' is the corrected weight, wi is the original weight, η is the learning rate (set to 0.01), E is the diagnostic deviation value for that region, and xi... For the corresponding feature signal values, for example, a feature signal xi=476, the weight wi=0.02 before correction, the deviation value E=1.2, and the weight wi'=0.02-0.01×(1.2 / 476)≈0.02-0.000025≈0.019975 after correction. After correction, the diagnostic deviation value of the region is recalculated. If the deviation value is still greater than 0.5, the weight correction operation is repeated until the diagnostic deviation value of all regions is lower than 0.5. The weight values ​​of all feature signals at this time are recorded, and they are classified and organized according to feature type and region to form a complete set of weight parameters, thus obtaining the initial weight parameters of the oral diagnostic network.

[0030] Please see Figure 5 The specific steps of S4 are as follows: S401: For the standardized original dataset of the oral cavity, divide the oral cavity into independent regions, extract the hard tissue coordinates, soft tissue color, and tooth surface texture gray values ​​of each region, integrate them to form the corresponding region feature set, and obtain the oral cavity partition feature parameter set; For the standardized oral cavity dataset, the entire mouth was divided into 32 independent oral regions according to tooth position and anatomical structure. Each region corresponds to a single tooth and surrounding periodontal tissue. MATLAB was used to locate and segment the dataset into regions. Each independent region was assigned a unique number from 1 to 32. Three-dimensional coordinate data of hard tissue was read region by region. 15 sets of X, Y, and Z axis coordinates were extracted from each region. The R, G, and B channel values ​​of soft tissue color parameters were read. 20 consecutive values ​​of tooth surface texture grayscale along the long axis of the tooth surface were read. The hard tissue coordinates, color parameters, and grayscale values ​​within the same region were summarized in a fixed order. Redundant values ​​were removed, and valid data items were retained. The arithmetic mean of similar parameters within the same region was taken. For example, the average X-axis coordinate of a certain region was 12.46 mm, the average Y-axis coordinate was -3.72 mm, and the average Z-axis coordinate was 12.46 mm. The average axis value is 18.93 mm, the color parameters are R=202, G=124, B=102, and the average gray value is 86. The average data of each region after processing are encapsulated into independent parameter units and arranged in the order of region number to form a parameter set with a unified structure and clear partitioning, thus obtaining the oral cavity partition feature parameter set. S402: Call the oral cavity partition feature parameter set and the initial weight parameters of the oral diagnosis network, substitute the partition feature parameters into the weight parameters to perform the calculation, extract the baseline values ​​of lesion features, and generate oral lesion feature reference values; The oral cavity region feature parameter set and the initial weight parameters of the oral diagnostic network were used. An Excel spreadsheet was created, and the feature parameters of the 32 regions were entered row by row. The corresponding initial weight parameters of the oral diagnostic network were entered according to feature type. The weight for hard tissue coordinates was set to 0.42, the weight for soft tissue color was set to 0.35, and the weight for tooth surface texture grayscale value was set to 0.23. A weighted summation operation was performed on each region using the formula F = x1・w1 + x2・w2. +x3・w3, where x1 is the comprehensive value of hard tissue coordinates, x2 is the comprehensive value of color parameters, x3 is the comprehensive value of grayscale values, and w1, w2, and w3 are the corresponding weights. For example, for a certain region, x1=15.62, x2=142, x3=86, after weighted calculation, F=15.62×0.42+142×0.35+86×0.23=6.5604+49.7+19.78=76.0404. The weighted calculation result of each region is used as the baseline value of the lesion characteristics of that region. The baseline values ​​of the lesion characteristics of the 32 regions are uniformly classified and arranged into a continuous numerical sequence according to the region number order. Four decimal places are retained to ensure data accuracy, and oral lesion characteristic reference values ​​are generated. S403: Based on the oral cavity partition feature parameter set and oral lesion feature reference value, calculate the degree of correspondence between the two, quantify the degree of correspondence and assign values, classify the assignment results by region, and generate oral cavity region lesion matching degree value. Based on the oral cavity region feature parameter set and oral lesion feature reference values, a numerical matching program was written in Python to calculate the degree of fit between the feature parameters and lesion feature reference values ​​for each region. The degree of fit was calculated using the absolute value of the difference normalization method, and the formula is M = 1 - |F a - Fᵦ| / Fᵦ, where F a Fᵦ represents the comprehensive value of regional characteristic parameters, and Fᵦ represents the reference value of the lesion characteristics in the corresponding region. For example, Fᵦ represents the comprehensive value of regional characteristic parameters in a certain region. a =72.36, Fᵦ=76.04, M=1-|72.36-76.04| / 76.04=1-3.68 / 76.04≈1-0.0484=0.9516. Multiply the calculated fit value by 100 to convert it into a percentage quantification value. The value assigned to this region is 95.16. The quantification and assignment operation of 32 regions is completed in sequence. The number of each region is bound to the corresponding matching value. The values ​​are arranged in ascending order of the numbers to form a complete set of values, ensuring that each region corresponds to only one matching value, without duplication or omission, and generating the matching value of oral region lesions.

[0031] Please see Figure 6 The specific steps of S5 are as follows: S501: Based on the lesion matching degree values ​​of oral regions, retrieve the preset lesion judgment threshold, compare the matching degree values ​​of each region with the lesion judgment threshold one by one, record the comparison results, mark the regions with values ​​higher than the threshold, and obtain the lesion suspected region marking table. Based on the matching degree values ​​of oral region lesions, a preset lesion judgment threshold was retrieved. This threshold was determined with reference to the average matching degree of 1000 clinical oral normal and lesion samples, calculated using the formula T = (T1 + T2) / 2, where T is the lesion judgment threshold, T1 is the average matching degree of normal samples (65.3), and T2 is the average matching degree of lesion samples (78.6). The calculated T = (65.3 + 78.6) / 2 = 71.95, and the final lesion judgment threshold was set to 72.0. The matching degree values ​​of 32 regions were imported into an Excel spreadsheet, and the matching degree value of each region was compared with 72.0 row by row. The comparison formula was C = AB (A is the region matching degree value, B is the lesion judgment threshold). If C > 0, it was marked as a suspected lesion region; if C ≤ 0, it was marked as a normal region. For example, region 1 has a matching degree of 89.2, 89.2 - 72.0 = 17.2 > 0, so it was marked as a suspected lesion region. Region 2... A match score of 68.5 and a match score of 68.5-72.0=-3.5≤0 are marked as normal. The comparison is completed region by region, and the comparison results of each region are recorded. The region numbers marked as suspected lesions, the match score values, and the comparison results are compiled into a table to ensure that no regions are missed or marked incorrectly, thus obtaining a suspected lesion region marking table.

[0032] S502: Call the suspected lesion area marking table, extract the matching degree value of each suspected area, divide the interval according to the value, assign corresponding grade number to different intervals, associate the area with the grade number, and generate an oral lesion grade correspondence table. The suspected lesion area labeling table is accessed, and the matching degree values ​​of all suspected lesion areas in the table are extracted. Assuming a total of 8 suspected areas are extracted, with values ​​of 89.2, 85.7, 79.3, 76.8, 74.5, 73.2, 72.8, and 72.1, these values ​​are sorted using WPS Spreadsheet from largest to smallest: 89.2, 85.7, 79.3, 76.8, 74.5, 73.2, 72.8, and 72.1. The values ​​are then divided into three intervals: Interval 1 (72.0-75.0, inclusive, exclusive), Interval 2 (75.0-85.0, inclusive, exclusive), and Interval 3 (85.0 and above). A corresponding level number is assigned to each interval: Interval 1 is assigned level 1, Interval 2 is assigned level 2, and Interval 3 is assigned level 3. The number of each suspected area is then associated with its corresponding level number. For example, area 1... Matching degree 89.2 belongs to interval 3, association level 3, region 3. Matching degree 79.3 belongs to interval 2, association level 2, region 8. Matching degree 72.1 belongs to interval 1, association level 1. Organize all association relationships into an ordered table, clarify the correspondence between region number, matching degree value, interval, and level number, and generate an oral lesion level correspondence table.

[0033] S503: Based on the suspected lesion area marking table and the oral lesion grade correspondence table, integrate the regional information matching degree numerical grade number, sort out the data association logic, form a complete quantitative system, and establish oral lesion quantitative grading data; Based on the suspected lesion area marking table and the oral lesion grade correspondence table, Python tools were used to read data from both tables. The region number, matching degree value, and comparison result were extracted from the suspected lesion area marking table, and the region number and grade number were extracted from the oral lesion grade correspondence table. Data association functions were used to match the two tables by region number, ensuring accurate correspondence of data within the same region without misalignment or missing information. The region number, matching degree value, comparison result, and grade number of each suspected region were integrated, and the data association logic was clarified. Regions were categorized according to grade number from 1 to 3: Grade 1 includes regions in the 72.0-75.0 range, Grade 2 includes regions in the 75.0-85.0 range, and Grade 3 includes regions in the 85.0 range and above. The regional data for each grade were summarized, and basic parameters of hard tissue, soft tissue, and tooth surface texture for each region were added to form a complete quantitative grading system. This system includes four main categories of data: region information, matching degree value, lesion grade, and basic parameters, ensuring a well-structured and logically clear data structure, thus establishing quantitative grading data for oral lesions.

[0034] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for establishing a neural network model for oral diagnosis, characterized in that, Includes the following steps: S1: Collect the three-dimensional coordinates of oral hard tissues, soft tissue color parameters, and tooth surface texture gray values ​​from oral CT equipment and oral endoscopes; verify the values ​​of each parameter; remove values ​​that exceed the normal physiological range; and generate a standardized raw dataset of oral cavity. S2: Based on the standardized original dataset of the oral cavity, extract the hard tissue coordinate difference, color parameter ratio, and texture gray value change, convert them into discrete feature signals, assign feature discrimination weights, and generate a set of oral cavity feature weighted signals. S3: Call the oral cavity feature weighted signal set, build the input layer, hidden layer and output layer structure, pass the feature signals according to the level, calculate the weighted sum of each level, compare with the activation threshold, retain the signals that meet the conditions, iteratively train with known lesion feature data, calculate the deviation value between the output result and the actual lesion, adjust the feature weights until the deviation value is lower than the diagnostic accuracy threshold, and obtain the initial weight parameters of the oral cavity diagnosis network. S4: For the standardized original oral dataset, call the initial weight parameters of the oral diagnosis network, calculate the matching degree between the features of each oral region and the lesion features, and generate the matching degree value of the oral region lesion. S5: Based on the matching degree value of oral region lesions, compare it with the preset lesion judgment threshold, filter the regions with matching degree values ​​higher than the threshold, classify the lesion level, and establish quantitative grading data of oral lesions.

2. The oral diagnosis method for establishing an oral diagnosis neural network model according to claim 1, characterized in that, The standardized raw dataset for oral cavity includes verified three-dimensional coordinates, qualified color parameters, and valid grayscale values ​​of tooth surface texture.

3. The oral diagnosis method for establishing an oral diagnosis neural network model according to claim 1, characterized in that, The oral cavity feature weighted signal set includes weighted hard tissue difference signal, weighted color proportion signal, and weighted texture change signal.

4. The oral diagnosis method for establishing an oral diagnosis neural network model according to claim 1, characterized in that, The initial weight parameters of the oral diagnostic network include input layer weights, hidden layer weights, output layer weights, and iterative optimization thresholds; The matching degree values ​​for oral region lesions include hard tissue matching degree, soft tissue matching degree, and tooth surface texture matching degree.

5. The oral diagnosis method for establishing an oral diagnosis neural network model according to claim 1, characterized in that, The quantitative grading data for oral lesions includes the lesion area, lesion grade, and quantitative grading indicators.

6. The oral diagnosis method for establishing an oral diagnosis neural network model according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: Collect the three-dimensional coordinates of oral hard tissues, soft tissue color parameters, and tooth surface texture gray values ​​from oral CT equipment and oral endoscopes. Record the three-dimensional coordinates of oral hard tissues, soft tissue color parameters, and tooth surface texture gray values ​​item by item, integrate the above three types of parameters to form a unified data set, and generate the original set of oral multi-source parameters. S102: Based on the original set of oral multi-source parameters, the three-dimensional coordinates of oral hard tissues, soft tissue color parameters, and tooth surface texture gray values ​​are compared with the normal physiological range values ​​to determine the range of the parameters, mark the parameter items that exceed the normal physiological range, and obtain the abnormal parameter identifier set. S103: Based on the original set of oral multi-source parameters and the set of abnormal parameter identifiers, remove the parameter items corresponding to the abnormal parameter identifiers, retain the parameter items within the normal physiological range, and arrange the retained parameters in a regular manner to generate a standardized original dataset of oral parameters.

7. The oral diagnosis method for establishing an oral diagnosis neural network model according to claim 1, characterized in that, The specific steps of S2 are as follows: S201: Based on the standardized raw dataset of the oral cavity, the three-dimensional coordinates of hard tissue are extracted to calculate the difference, the proportion of soft tissue color parameters is statistically analyzed, the change in gray value of tooth surface texture is measured, and the three types of data are integrated to form a set of feature indicators, thus obtaining the core feature indicator group of the oral cavity. S202: Call the core feature index group of the oral cavity, convert the hard tissue coordinate difference, color parameter ratio, and texture gray value change into discrete numerical signals one by one, sort out the signal arrangement logic, and generate a set of discrete feature signals of the oral cavity. S203: For the set of discrete feature signals in the oral cavity, analyze the discrimination of each discrete feature signal, assign corresponding weight values ​​according to the discrimination of high and low, and form a complete data set by associating the signals and weights to generate a weighted signal set of oral cavity features.

8. The oral diagnosis method for establishing an oral diagnosis neural network model according to claim 1, characterized in that, The specific steps for S3 are as follows: S301: Call the oral cavity feature weighted signal set, build the input layer, hidden layer and output layer structure, pass the feature weighted signal in sequence according to the level, calculate the weighted sum of the signals of each level, compare the weighted sum with the activation threshold, retain the signals that meet the activation conditions, and generate the oral cavity level transmission effective signal group. S302: Based on the effective signal group transmitted at the oral cavity level, known lesion feature data is introduced to perform iterative calculation, calculate the difference between the network output result and the actual lesion, and obtain the oral cavity diagnosis deviation value; S303: Based on the oral diagnosis deviation value, compare the deviation value with the diagnostic accuracy threshold, adjust the corresponding feature weights according to the comparison results, repeat the weight adjustment operation, and obtain the initial weight parameters of the oral diagnosis network.

9. The oral diagnosis method for establishing an oral diagnosis neural network model according to claim 1, characterized in that, The specific steps of S4 are as follows: S401: For the standardized original dataset of the oral cavity, divide the oral cavity into independent regions, extract the hard tissue coordinates, soft tissue color, and tooth surface texture gray values ​​of each region, integrate them to form the corresponding region feature set, and obtain the oral cavity partition feature parameter set; S402: Call the oral cavity partition feature parameter set and the initial weight parameters of the oral diagnosis network, substitute the partition feature parameters into the weight parameters to perform the calculation, extract the baseline values ​​of lesion features, and generate oral lesion feature reference values; S403: Based on the oral cavity partition feature parameter set and oral lesion feature reference value, calculate the degree of correspondence between the two, quantify the degree of correspondence and assign values, classify the assignment results by region, and generate the oral cavity region lesion matching degree value.

10. The oral diagnosis method for establishing an oral diagnosis neural network model according to claim 1, characterized in that, The specific steps of S5 are as follows: S501: Based on the lesion matching degree values ​​of oral regions, retrieve the preset lesion judgment threshold, compare the matching degree values ​​of each region with the lesion judgment threshold one by one, record the comparison results, mark the regions with values ​​higher than the threshold, and obtain the lesion suspected region marking table. S502: Call the suspected lesion area marking table, extract the matching degree value of each suspected area, divide the interval according to the value, assign corresponding grade number to different intervals, associate the area with the grade number, and generate an oral lesion grade correspondence table. S503: Based on the suspected lesion area marking table and the oral lesion grade correspondence table, integrate the regional information matching degree numerical grade number, sort out the data association logic, form a complete quantitative system, and establish oral lesion quantitative grading data.

Images