阅读说明：本技术 一种基于ai算法的龋病和牙菌斑检测其分布方法 (AI algorithm-based caries and dental plaque detection and distribution method ) 是由 不公告发明人 于 2021-07-29 设计创作，主要内容包括：本发明涉及临床口腔医学技术领域,公开了一种基于AI算法的龋病和牙菌斑检测其分布方法,实现方法包括以下步骤：S10、龋病和牙菌斑数据集构建；S20、龋病和牙菌斑检测模型训练；S30、龋病和牙菌斑检测模型应用。本发明通过龋病和牙菌斑数据集构建、龋病和牙菌斑检测模型训练和龋病和牙菌斑检测模型应用训练AI算法模型,在不需要人工介入的情况下,检测龋病和牙菌斑的分布,检测的结果符合龋病诊断金标准(组织切片)和牙菌斑诊断标准的判断,并且在特定的牙面判断上灵敏度和特异性甚至优于经验医生的视诊和探诊,从而避免了其他设备由于不同外形、光照和拍摄角度带来的检测结果错误,同时也解放了医生和专业人员。(The invention relates to the technical field of clinical oral medicine, and discloses a distribution method for detecting caries and dental plaque based on AI algorithm, which comprises the following steps: s10, constructing a caries and dental plaque data set; s20, training caries and dental plaque detection models; s30, caries and dental plaque detection model application. According to the invention, through the construction of a caries and dental plaque data set, the training of a caries and dental plaque detection model and the application and training of an AI algorithm model, the distribution of caries and dental plaque is detected without manual intervention, the detection result meets the judgment of caries diagnosis golden standard (tissue slice) and dental plaque diagnosis standard, and the sensitivity and specificity are even better than those of visual diagnosis and exploratory diagnosis of an experienced doctor in specific dental surface judgment, so that the detection result errors of other equipment caused by different shapes, illumination and shooting angles are avoided, and the doctor and the professional are liberated.)

1. A distribution method for detecting caries and dental plaque based on AI algorithm is characterized by comprising the following steps:

s10, caries and plaque dataset construction: under the state of a special light source and natural light, tooth images of the whole tooth position, caries and dental plaque degrees and distribution data are collected, the collected data are classified according to the caries degree and dental plaque degree, the data are labeled according to the classification and the position, the data are cleaned according to the image quality, and the construction of a data set is completed;

s20, training caries and dental plaque detection models: randomly dividing a data set into a caries and dental plaque detection model training set, a caries and dental plaque detection model verification set and a caries and dental plaque detection model test set according to a proportion, constructing a caries and dental plaque detection pre-training model, inputting the data set into the pre-training model, and performing data training, verification and test to obtain a caries and dental plaque detection model;

s30, caries and plaque detection model application: the image to be measured is used as input, and the caries degree, the plaque degree and the position distribution condition of the corresponding tooth position can be output.

2. The AI algorithm-based caries and plaque detection distribution method according to claim 1, wherein in the S10 step, when acquiring tooth images of the entire dental position, each dental position acquires images of the maxillofacial surface, the buccal surface, the lingual surface and the adjacent surface, respectively, so that each dental position image has a complete tooth; when tooth data of a whole dental position are collected, a dentist inspects the whole dental position and records the ICDAS score, the plaque index and the position information, in the step S10, a dental caries and plaque data set is constructed, and different tooth surface images of each dental position need to be classified according to the recorded ICDAS score and plaque index; labeling the image according to the classification information and the position information by using a labeling tool; reviewing the label after the label is finished, deleting repeated information and correcting error information; and simultaneously, removing images with inaccurate focusing and serious color deviation.

3. The AI algorithm-based caries and plaque detection distribution method according to claim 1, wherein said data set classification in S10 includes caries level, caries site, plaque level and plaque attachment site, wherein:

degree of caries: according to the ICDAS grade, the Caries degree is divided into 0-6 grades, and 7 grades are divided into Caries _ 0-Caries _6, wherein Caries _0 represents no Caries;

carious lesion location: when the caries degree is more than 0, marking the caries lesion position;

plaque level: dividing the plaque degree into 0-3 according to the plaque index, and dividing the plaque degree into Qlaque _ 0-Qlaque _3 in 4 grades, wherein Qlaque _0 represents no plaque;

plaque attachment site: when the plaque level is greater than 0, the plaque location is marked.

4. The AI algorithm-based caries and plaque detection distribution method according to claim 1, wherein the AI algorithm predicts caries level, plaque level and location distribution of each dental site in S30, and when the level is greater than 0, outputs a complete dental site, caries location or plaque attachment location bounding box and prints the confidence level; the bounding box comprises any one of a rectangle, a circle or a polygon; the AI algorithm comprises model detection, an evaluation algorithm and a judgment algorithm; the model detection is that the dental position image to be detected is input into a caries and dental plaque detection model, and the caries degree, the dental plaque degree and the position distribution condition are output through the model detection; the evaluation algorithm is that after the teeth positions of the whole mouth are predicted, a diagnosis report and an oral health comprehensive index score are output according to the statistics of the whole mouth condition, and the oral health comprehensive index calculation formula is as follows:

wherein H represents the number of healthy teeth, CQ represents the number of teeth with caries or dental plaque, and A represents the number of teeth in the whole mouth;

the judgment algorithm is that when the caries degree or the plaque degree of a certain tooth position is more than 0, the boundary frames of the caries position or the plaque attachment position are possibly overlapped; when the two bounding boxes are overlapped, calculating the overlapping degree of the two bounding boxes, and reserving the bounding box with the large overlapping degree; the calculation formula of the overlapping degree is as follows:

wherein I_(b1,b2)Is the intersection area, U, of two bounding boxes_(bi)The area of the bounding box is shown, and the Confidence is the Confidence of the bounding box;

when the caries degree or the plaque degree of a certain tooth position is equal to 0, the caries position or the plaque attachment position predicted by the model is judged to be invalid by the algorithm;

the algorithm is judged to be invalid when the model-predicted caries location or plaque attachment location bounding box appears outside the complete tooth location bounding box.

5. The AI algorithm-based caries and plaque detection distribution method according to claim 1, wherein the data set construction process in step S10 requires the collection of sample data set, and before the collection of sample data set, the sample size is estimated, the sample size is related to significance level α, tolerance error δ, classification category c, and sensitivity or specificity estimation value pdarameter, wherein the smaller the α value, the larger the sample size is required; the smaller the value of δ, the larger the required sample size; the larger the value of c, the larger the required sample size; calculating the positive sample volume by using the estimated value of the sensitivity, and calculating the negative sample volume by using the estimated value of the specificity;

the calculation formula of the sample size is as follows:

wherein, U_αIs a cumulative probability of being normally distributedThe value of U.

6. An AI algorithm-based caries and plaque detection distribution method according to claim 1, wherein the image acquired in the S10 step includes: an image collected using a fiber optic transmission illumination method; the image collected by using an X-ray imaging technology and the image collected by using an infrared light scattering characteristic technology; fluorescence images of teeth excited using a special light source.

7. An AI algorithm based caries and plaque detection distribution method according to claim 1, wherein the model used in the S20 step includes: linear regression, logistic regression, linear discriminant analysis, decision trees, naive Bayes, K-nearest neighbor algorithm, learning vector quantization, support vector machine, bagging and random forest and depth neural network; when a deep neural network model is selected, the pre-training weight of the model is required to be used, and the deep neural network model is realized by the following steps:

s11, performing model training by using a large amount of resources in advance to enable the model to obtain better performance;

and S21, training a model by using the image data set of the caries and the dental plaque collected in the S10 step, and generating model weights suitable for caries and dental plaque detection.

Technical Field

The invention relates to the technical field of clinical oral medicine, in particular to a distribution method for detecting caries and dental plaque based on an AI algorithm.

Background

Dental caries and periodontal disease are two major intrinsic diseases of the oral cavity; caries is a bacterial disease caused by multiple factors, and lesions range from shallow to deep and can affect enamel, dentin and cementum; if the disease is not treated in time, pathogenic bacteria can further invade dental pulp, pulpitis or periapical periodontitis can be caused, and even tooth loss can be caused in serious cases; dental plaque is an initiating factor for the onset of periodontal disease, and whether the dental plaque is controlled in time can directly influence the onset and prognosis of periodontal disease; periodontal diseases include gingival disease and periodontitis, and the lesions of periodontal ligament, alveolar bone and cementum spread from gingiva and deep periodontal tissue can cause loose teeth, loss of teeth and even loss of masticatory function. In addition to the physiological discomfort brought to the patient, the dental caries and periodontal disease also seriously affect the aesthetic appearance of the patient, causing psychological problems such as apprehension of inferior society.

The traditional method for detecting the decayed tooth is to identify the boundary of the decayed tissue through the clinical examination and probing of doctors, but the method depends on the experience of doctors, different doctors have different subjective evaluation standards, even the diagnosis of the same doctor is influenced by different environments and apparatuses, the existing decayed tooth and dental plaque detection equipment abroad uses the fluorescence effect of teeth, adopts the quantitative light-induced fluorescence technology (QLF), uses a special camera to receive the reflected fluorescence image, obtains the fluorescence loss by reconstructing the fluorescence of healthy enamel, calculates the percentage difference between the actual surface and the reconstructed surface to determine the reduction of the fluorescence, any area with the fluorescence reduction exceeding 5 percent is considered as the lesion, the existing decayed tooth and dental plaque detection equipment is not only expensive, but also has no intelligent diagnosis output and only can be used as an auxiliary equipment, manual intervention is also required to make an effective diagnosis.

Disclosure of Invention

The invention aims to provide a distribution method for detecting caries and dental plaque based on AI algorithm, so as to solve the problems in the background technology.

In order to achieve the purpose, the invention provides the following technical scheme:

a distribution method for detecting caries and dental plaque based on AI algorithm comprises the following steps:

s10, caries and plaque dataset construction: under the state of a special light source and natural light, tooth images of the whole tooth position, caries and dental plaque degrees and distribution data are collected, the collected data are classified according to the caries degree and dental plaque degree, the data are labeled according to the classification and the position, the data are cleaned according to the image quality, and the construction of a data set is completed;

s20, training caries and dental plaque detection models: randomly dividing a data set into a caries and dental plaque detection model training set, a caries and dental plaque detection model verification set and a caries and dental plaque detection model test set according to a proportion, constructing a caries and dental plaque detection pre-training model, inputting the data set into the pre-training model, and performing data training, verification and test to obtain a caries and dental plaque detection model;

s30, caries and plaque detection model application: the image to be measured is used as input, and the caries degree, the plaque degree and the position distribution condition of the corresponding tooth position can be output.

As a still further scheme of the invention: when the tooth images of the whole tooth positions are collected in the step S10, the images of the jaw face, the cheek face, the tongue face and the adjacent face of each tooth position are respectively collected, so that each tooth position image has a complete tooth; when tooth data of a whole dental position are collected, a dentist inspects the whole dental position and records the ICDAS score, the plaque index and the position information, in the step S10, a dental caries and plaque data set is constructed, and different tooth surface images of each dental position need to be classified according to the recorded ICDAS score and plaque index; labeling the image according to the classification information and the position information by using a labeling tool; reviewing the label after the label is finished, deleting repeated information and correcting error information; and simultaneously, removing images with inaccurate focusing and serious color deviation.

As a still further scheme of the invention: the data set classification in step S10 includes caries level, caries site, plaque level and plaque attachment site, wherein:

degree of caries: according to the ICDAS grade, the Caries degree is divided into 0-6 grades, and 7 grades are divided into Caries _ 0-Caries _6, wherein Caries _0 represents no Caries;

carious lesion location: when the caries degree is more than 0, marking the caries lesion position;

plaque level: dividing the plaque degree into 0-3 according to the plaque index, and dividing the plaque degree into Qlaque _ 0-Qlaque _3 in 4 grades, wherein Qlaque _0 represents no plaque;

plaque attachment site: when the plaque level is greater than 0, the plaque location is marked.

As a still further scheme of the invention: in the step S30, the caries degree, the plaque degree and the position distribution condition of each dental position are predicted through an AI algorithm, and when the degree is more than 0, a complete dental position, a caries position or a plaque attachment position bounding box is output, and the confidence coefficient is printed; the bounding box comprises any one of a rectangle, a circle or a polygon; the AI algorithm comprises model detection, an evaluation algorithm and a judgment algorithm; the model detection is that the dental position image to be detected is input into a caries and dental plaque detection model, and the caries degree, the dental plaque degree and the position distribution condition are output through the model detection; the evaluation algorithm is that after the teeth positions of the whole mouth are predicted, a diagnosis report and an oral health comprehensive index score are output according to the statistics of the whole mouth condition, and the oral health comprehensive index calculation formula is as follows:

wherein H represents the number of healthy teeth, CQ represents the number of teeth with caries or dental plaque, and A represents the number of teeth in the whole mouth;

the judgment algorithm is that when the caries degree or the plaque degree of a certain tooth position is more than 0, the boundary frames of the caries position or the plaque attachment position are possibly overlapped; when the two bounding boxes are overlapped, calculating the overlapping degree of the two bounding boxes, and reserving the bounding box with the large overlapping degree; the calculation formula of the overlapping degree is as follows:

wherein I_(b1,b2)Is the intersection area, U, of two bounding boxes_(bi)The area of the bounding box is shown, and the Confidence is the Confidence of the bounding box;

when the caries degree or the plaque degree of a certain tooth position is equal to 0, the caries position or the plaque attachment position predicted by the model is judged to be invalid by the algorithm;

the algorithm is judged to be invalid when the model-predicted caries location or plaque attachment location bounding box appears outside the complete tooth location bounding box.

As a still further scheme of the invention: in the step S10, a sample data set needs to be collected in the process of constructing the data set, and before the sample data set is collected, a sample size is estimated, where the sample size is related to a significance level α, an allowable error δ, a classification category c, and an estimated value P parameter of sensitivity or specificity, and the smaller the α value is, the larger the required sample size is; the smaller the value of δ, the larger the required sample size; the larger the value of c, the larger the required sample size; calculating the positive sample volume by using the estimated value of the sensitivity, and calculating the negative sample volume by using the estimated value of the specificity;

the calculation formula of the sample size is as follows:

wherein, U_αIs a cumulative probability of being normally distributedThe value of U.

As a still further scheme of the invention: the image acquired in the step S10 includes: an image collected using a fiber optic transmission illumination method; the image collected by using an X-ray imaging technology and the image collected by using an infrared light scattering characteristic technology; fluorescence images of teeth excited using a special light source.

As a still further scheme of the invention: the model used in the step S20 includes: linear regression, logistic regression, linear discriminant analysis, decision trees, naive Bayes, K-nearest neighbor algorithm, learning vector quantization, support vector machine, bagging and random forest and depth neural network; when a deep neural network model is selected, the pre-training weight of the model is required to be used, and the deep neural network model is realized by the following steps:

s11, performing model training by using a large amount of resources in advance to enable the model to obtain better performance;

and S21, training a model by using the image data set of the caries and the dental plaque collected in the S10 step, and generating model weights suitable for caries and dental plaque detection.

Compared with the prior art, the invention has the beneficial effects that:

according to the invention, through the construction of a caries and dental plaque data set, the training of a caries and dental plaque detection model and the application and training of an AI algorithm model, the distribution of caries and dental plaque is detected without manual intervention, the detection result meets the judgment of caries diagnosis golden standard (tissue slice) and dental plaque diagnosis standard, and the sensitivity and specificity are even better than those of visual diagnosis and exploratory diagnosis of an experienced doctor in specific dental surface judgment, so that the detection result errors of other equipment caused by different shapes, illumination and shooting angles are avoided, and the doctor and the professional are liberated.

Drawings

FIG. 1 is a schematic flow chart of a method for detecting distribution of caries and dental plaque based on AI algorithm;

FIG. 2 is a schematic diagram of the detection results of a distribution method for caries and dental plaque detection based on AI algorithm;

FIG. 3 is a graph of the loss of loss function during training of a method for detecting distribution of caries and plaque based on AI algorithm;

FIG. 4 is a graph of the variation of mAP value with iteration number during training of a distribution method for caries and dental plaque detection based on AI algorithm;

FIG. 5 is a comparison graph of model optimization for a distribution method of caries and plaque detection based on AI algorithm.

Detailed Description

In the embodiment of the invention, an AI algorithm-based distribution method for detecting caries and dental plaque comprises the following steps:

s10, caries and plaque dataset construction: under the state of a special light source and natural light, tooth images of the whole tooth position, caries and dental plaque degrees and distribution data are collected, the collected data are classified according to the caries degree and dental plaque degree, the data are labeled according to the classification and the position, the data are cleaned according to the image quality, and the construction of a data set is completed;

s20, training caries and dental plaque detection models: randomly dividing a data set into a caries and dental plaque detection model training set, a caries and dental plaque detection model verification set and a caries and dental plaque detection model test set according to a proportion, constructing a caries and dental plaque detection pre-training model, inputting the data set into the pre-training model, and performing data training, verification and test to obtain a caries and dental plaque detection model;

s30, caries and plaque detection model application: the image to be measured is used as input, and the caries degree, the plaque degree and the position distribution condition of the corresponding tooth position can be output.

Preferably, in the step S10, when the dental images of the entire dental position are acquired, the images of the maxillofacial surface, the buccal surface, the lingual surface and the adjacent surface are acquired for each dental position, so that each dental position image has a complete tooth; when tooth data of a whole dental position are collected, a dentist inspects the whole dental position, records ICDAS score, dental plaque index and position information, a dental caries and dental plaque data set is constructed in the step S10, and different tooth surface images of each dental position need to be classified according to the recorded ICDAS score and dental plaque index; labeling the image according to the classification information and the position information by using a label img labeling tool, labeling the position and the class name of each class by using a rectangular frame according to a dense labeling mode, and storing the position and the class name into a labeling document; reviewing the label after the label is finished, deleting repeated information and correcting error information; and simultaneously, removing images with inaccurate focusing and serious color deviation.

Preferably, the data set classification in step S10 includes caries level, caries site, plaque level and plaque attachment site, wherein:

degree of caries: according to the ICDAS grade, the Caries degree is divided into 0-6 grades, and 7 grades are divided into Caries _ 0-Caries _6, wherein Caries _0 represents no Caries;

carious lesion location: when the caries degree is more than 0, marking the caries lesion position;

plaque level: dividing the plaque degree into 0-3 according to the plaque index, and dividing the plaque degree into Qlaque _ 0-Qlaque _3 in 4 grades, wherein Qlaque _0 represents no plaque;

plaque attachment site: when the plaque degree is greater than 0, marking the plaque position;

wherein each dentition image comprises at least one of the four types.

Preferably, in the step S30, the caries level, plaque level and position distribution of each dental site are predicted by an AI algorithm, and when the level is greater than 0, a complete dental site, caries position or plaque attachment position bounding box is output, and the confidence level is printed; the bounding box comprises any one of a rectangle, a circle or a polygon; the AI algorithm comprises model detection, an evaluation algorithm and a judgment algorithm; the model detection is that the dental position image to be detected is input into a caries and dental plaque detection model, and the caries degree, the dental plaque degree and the position distribution condition are output through the model detection; the evaluation algorithm is that after the teeth positions of the whole mouth are predicted, a diagnosis report and the oral health comprehensive index score are output according to the statistics of the whole mouth condition, and the oral health comprehensive index calculation formula is as follows:

wherein H represents the number of healthy teeth, CQ represents the number of teeth with caries or dental plaque, and A represents the number of teeth in the whole mouth;

the bounding box contains 5 parameters (x, y, w, h, Confidence), where (x, y) is the offset of the bounding box center relative to the top left corner of the image, (w, h) is the width and height of the bounding box, and Confidence is the Confidence of the bounding box;

the judgment algorithm is that when the caries degree or the plaque degree of a certain tooth position is more than 0, the boundary frames of the caries position or the plaque attachment position are possibly overlapped; when the two bounding boxes are overlapped, calculating the overlapping degree of the two bounding boxes, and reserving the bounding box with the large overlapping degree; the calculation formula of the overlapping degree is as follows:

wherein I_(b1,b2)Is the intersection area, U, of two bounding boxes_(bi)The area of the bounding box is shown, and the Confidence is the Confidence of the bounding box;

when the caries degree or the plaque degree of a certain tooth position is equal to 0, the caries position or the plaque attachment position predicted by the model is judged to be invalid by the algorithm;

the algorithm is judged to be invalid when the model-predicted caries location or plaque attachment location bounding box appears outside the complete tooth location bounding box.

Preferably, in the step S10, a sample data set needs to be collected during the process of constructing the data set, and before collecting the sample data set, a sample size is estimated, where the sample size is related to the significance level α, the allowable error δ, the classification category c, and the sensitivity or specificity estimation value P parameter, and the smaller the α value is, the larger the required sample size is; the smaller the value of δ, the larger the required sample size; the larger the value of c, the larger the required sample size; calculating the positive sample volume by using the estimated value of the sensitivity, and calculating the negative sample volume by using the estimated value of the specificity;

the formula for calculating the sample size is as follows:

wherein, U_αIs a cumulative probability of being normally distributedThe value of U when, for example: get U_0.05＝1.960,U_0.012.576, when the significance level alpha is 0.05, the tolerance delta is 0.05, and the sensitivity and specificity estimates are both 90%, the estimated sample content is calculated

As shown in fig. 3, in the initial part of the iterative process of model training, the loss value decreases rapidly, and the model learns and fits rapidly; the loss of the rear part is slowly reduced and approaches to 0; the minimum value is 0.008, which shows that the model has good convergence effect.

As shown in fig. 4, the value of the mapp in the model training process changes with the number of iterations, and the mapp in the graph represents an average precision mean value, and the calculation formula is as follows:

wherein TP represents the number of the intersection ratio (IoU) of the prediction frame and the real frame being equal to or larger than 0.5, FP represents the number of the intersection ratio (IoU) of the prediction frame and the real frame being smaller than 0.5, and as can be seen from FIG. 4, the 27 th epochs, mAP value reaches 95.19, so the model weight parameter of the 27 th epochs is selected as the optimal model parameter.

Correctly predicting whether the caries degree or the plaque degree is greater than 0 to represent True Positive (TP), incorrectly predicting whether the caries degree or the plaque degree is equal to 0 to represent False Positive (FP), correctly predicting whether the caries degree or the plaque degree is equal to 0 to represent True Negative (TN), and incorrectly predicting whether the caries degree or the plaque degree is greater than 0 to represent False Negative (FN), and performing performance verification by using test set data to obtain the following indexes:

accuracy (Accuracy): (TP + TN)/(TP + TN + FP + FN) ═ 0.895;

precision (Precision): TP/(TP + FP) ═ 0.913;

recall (Recall): TP/(TP + FN) ═ 0.905;

specificity (Specificity): TN/(TN + FP) ═ 0.915.

Preferably, the image acquired in the S10 step includes: an image collected using a fiber optic transmission illumination method; the image collected by using an X-ray imaging technology and the image collected by using an infrared light scattering characteristic technology; the tooth fluorescence image excited by the special light source and the tooth fluorescence image excited by the light with the wavelength of 390-430 nm of the special light source are used as the input of the model, the characteristics of caries and dental plaque are more obvious by using the image of the technology, and the model with better sensitivity, specificity and accuracy can be obtained.

Preferably, the model used in step S20 includes: linear regression, logistic regression, linear discriminant analysis, decision trees, naive Bayes, K-nearest neighbor algorithm, learning vector quantization, support vector machine, bagging and random forest and depth neural network;

when the deep neural network model YOLOv5S is used in the step S20, compared with other network models YOLOv5S, the deep neural network model YOLOv5 has the advantages of high speed, high precision and small volume, and is easy to deploy to a mobile device; when the method is deployed to a mobile device, the following conversion and optimization need to be performed on the trained network model:

model simplification: simplifying the nodes such as cast, Identity and the like by using an onnx-simplifer simplified model;

model conversion: converting the model to a ncnn model using an onnx2ncnn tool;

model optimization: as shown in fig. 5, a custom Focus node is used to replace the Focus layer of the original model Backbone part which is not supported by the ncnn model; and modifying the grid parameter of a Reshape layer in front of the Permutee output layer to be-1, so that the model can be output in a self-adaptive mode and cannot be influenced by the size of an input image.

When the deep neural network model is selected, the pre-training weight of the model is required to be used, and the deep neural network model is realized by the following steps:

s11, performing model training by using a large amount of resources in advance to enable the model to obtain better performance;

s21, training a model by using the image data set of the caries and the dental plaque collected in the S10 step, and generating a model weight suitable for caries and dental plaque detection; the pre-training weight can be the pre-training weight of the fast-RCNN, the pre-training weight of the Mask-RCNN or the pre-training weight of the YOLO.

The above embodiments are only preferred embodiments of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art should be considered to be within the scope of the present invention, and the technical solutions and the inventive concepts thereof according to the present invention should be equally replaced or changed within the scope of the present invention.