Automatic measurement method, system and computer equipment for irregular oral anatomical structures

Abstract

This application proposes an automatic measurement and analysis system, method, and computer device for performing fully quantitative automatic measurement and analysis of tiny oral irregular anatomical structures. The method includes S1: obtaining historical standard sagittal sections based on the patient's historical CBCT images; S2: constructing an image database; S3: constructing a measurement and analysis database; S4: constructing an intelligent network model based on a semantic segmentation network for training, validation, and testing; S5: obtaining a standard sagittal section to be measured based on the patient's CBCT images to be measured, inputting the standard sagittal section to be measured into the intelligent network model, and obtaining the predicted area of ​​the oral irregular anatomical structure in the standard sagittal section to be measured and the coordinates of related measurement reference marks; S5: selecting an oral subspecialty measurement method and correspondingly calculating and outputting quantitative indicators of oral irregular structures. The method and device described in this application are suitable for clinical applications in different oral subspecialties.

Description

Technical Field

[0001] This patent relates to the field of medical image processing technology, specifically an automatic measurement method, system and computer equipment based on lightweight semantic segmentation involving irregular and tiny oral anatomical structures. Background Art

[0002] Oral anatomical structures such as teeth, alveolar bones, and mandibular condyles are often irregular in shape. The measurement and analysis of irregular oral anatomical structures in medical imaging is crucial for dental diagnosis and treatment. Clinical diagnosis and treatment planning often require quantitative analysis of irregular oral anatomical structures.

[0003] Due to the irregular shapes and complex structures of the irregular oral anatomy and the limited clinical time available, only representative quantitative indicators are often measured, which is not conducive to fully revealing the underlying clinical information. For example, as a compromise, the thickness of the buccal and palatal bone walls is used to indicate available bone volume and reference available bone volume. This results in the loss of much anatomical information, such as thickness at different locations, bone wall morphology, and tooth-bone positional relationships, thus affecting further treatment analysis.

[0004] In addition, the analysis and measurement of irregular oral anatomical structures are usually based on specific reference targets (such as teeth, the position of the alveolar ridge, and the root apex), such as the working length from the root apex to the tooth tip, the relationship between the periodontal bone defect and the alveolar ridge and root bifurcation, the distance between orthodontic teeth and important anatomical structures, and the estimated force arm torque of orthodontic forces based on the root apex and alveolar ridge. The measurement of the same structure may vary in different clinical subspecialties. Therefore, overcoming the limitation of only being able to measure some indicators in clinical practice and quantifying all the information contained in irregular oral anatomical structures in medical images is of great clinical significance.

[0005] With the development of algorithms and computing power, artificial intelligence (AI) may become a feasible solution to automatically place irregular tooth structures in the medical coordinate system, thereby overcoming clinical limitations and completing the entire quantitative measurement, thereby revealing deeper medical connotations and better serving clinical practice.

[0006] Existing end-to-end output artificial intelligence tools for dental quantitative data are designed to output specific indicators, but cannot solve the problem of quantifying all information about the irregular anatomical structure of the oral cavity.

[0007] In order to embed the network into the clinical system and provide real-time feedback for clinical practice, it has become a key clinical issue to develop lightweight, accurate, and efficient artificial intelligence tools for full quantitative measurement and analysis of irregular and tiny oral structures and integrate them into the clinical system. Summary of the Invention

[0008] The present invention aims to solve at least one of the technical problems existing in the above-mentioned prior art. To this end, this patent proposes to adopt a semantic segmentation method to more accurately understand and analyze oral irregular structures by identifying all segmented areas and measurement reference marks, and even the semantic categories and annotations of each pixel. After setting the measurement semantics, the program can output all quantitative indicators of oral irregular structures with medical information. Faced with different measurement procedures for different oral subspecialties, there is no need to train another network model, only the semantics of the subsequent measurement procedures need to be modified. It has strong clinical adaptability, is suitable for different subspecialties, can realize full quantitative analysis, and is an automatic measurement method for oral irregular structures with high accuracy. The technical solution adopted by the present invention is: an automatic measurement method for oral irregular anatomical structures based on semantic segmentation. Oral irregular anatomical structures include enamel, dentin, pulp cavity, labial bone wall, palatal bone wall, alveolar base bone, condyle, sinus, and nasopalatine canal, comprising the following steps:

[0009] S1. Collect historical CBCT images of multiple patients and obtain historical standard sagittal sections based on the historical CBCT images;

[0010] S2. Mark the irregular anatomical structures of the oral cavity and related measurement reference landmarks in the historical standard sagittal sections, and construct an image database using the marked historical standard sagittal sections;

[0011] S3. Measure the irregular oral anatomical structures in the image database using different oral subspecialty measurement methods, and construct a measurement analysis database using historical standard sagittal sections with markers and measurement results.

[0012] S4. Build an intelligent network model based on the semantic segmentation network, train and verify the intelligent network model using the image database, and test the intelligent network model using the image database and measurement analysis database;

[0013] S5. Obtaining a standard sagittal section to be measured based on the patient's CBCT image to be measured, inputting the standard sagittal section to be measured into the intelligent network model, and obtaining the predicted area of ​​the oral irregular anatomical structure and the coordinates of relevant measurement reference marks in the standard sagittal section to be measured;

[0014] S6. Select an oral subspecialty measurement method for calculating irregular oral anatomical structures. The oral subspecialty measurement method includes reference points, reference lines, and measurement sites. Calculate the corresponding quantitative indicators of irregular oral anatomical structures based on the selected oral subspecialty measurement method. The quantitative indicators include thickness, height, and area.

[0015] Preferably, the method of collecting historical CBCT images of multiple patients in S1 of the present invention and obtaining historical standard sagittal sections based on the historical CBCT images includes:

[0016] S1.1. Use CBCT processing software to import the patient's oral CBCT DICOM file;

[0017] S1.2. Use the jaw plane and the human midline as the horizontal and vertical reference planes;

[0018] S1.3. Select the center of the pulp cavity on the cross-section using the cementoenamel junction as the landmark point, draw the dental arch curve, obtain the standard cross-section, convert the selected cross-section into TIFF format, normalize the image and correct the grayscale to 0-255 to obtain the historical standard sagittal cross-section of the anterior teeth.

[0019] Preferably, in S2 of the present invention, the method of marking oral irregular structures and related measurement reference marks in the historical standard sagittal section and constructing the image database includes:

[0020] S2.1. Import the historical standard sagittal section into the image marking processing software;

[0021] S2.2. Use the polygon annotation tool to mark the irregular anatomical regions of the oral cavity in the historical standard sagittal section;

[0022] S2.3. Use the point marking tool to mark the relevant measurement reference landmarks in the historical standard sagittal section;

[0023] S2.4. Save the marked historical standard sagittal sections in an image format that can be recognized by a semantic segmentation convolutional neural network to build an image database.

[0024] Preferably, the method of obtaining measurement data of irregular oral anatomical structures of an image set in an image database under different oral subspecialty measurement methods in S3 of the present invention and constructing a measurement analysis database includes:

[0025] S3.1. Converting the markings of the irregular oral anatomical structures and related measurement reference landmarks in the dataset of the image database into specific coordinates;

[0026] S3.2. Derive the measurement reference baseline based on the coordinates of the relevant measurement reference marks;

[0027] S3.3. Select reference points and measurement reference baselines based on different oral subspecialty measurement methods. Use Python statements to calculate the intercepts of the selected reference baselines at different distances from the reference points with parallel lines and perpendicular lines and the irregular oral anatomical structures. Calculate the thickness and height of the irregular oral anatomical structures at multiple locations. Use Python statements to calculate the integral area of ​​the irregular oral anatomical structure region to obtain the area data of the irregular oral anatomical structures. Finally, construct a measurement and analysis database.

[0028] Preferably, in S4 of the present invention, the intelligent network model is constructed based on the semantic segmentation network, the intelligent network model is trained and verified using an image database, and the image database and the measurement analysis database are used for testing. The method for obtaining the intelligent network model includes:

[0029] S4.1. Divide the image database dataset into training set, validation set, and test set;

[0030] S4.2, build intelligent network model;

[0031] S4.3. Use the training set and validation set to supervise the training of the intelligent network model, continuously iteratively update the weights and network parameters, minimize the loss function, and obtain the optimal intelligent network model;

[0032] S4.4. Input the test set into the intelligent network model to obtain the segmentation result, verify the segmentation accuracy of the intelligent network model, and use the corresponding data set in the measurement analysis database to verify the accuracy of the measurement algorithm.

[0033] Preferably, the intelligent network model in S4 of the present invention is a lightweight semantic segmentation model, which is an autoencoder network structure, including an encoder and a decoder. The corresponding modules between the two parts are connected through a cross-connection layer, and the features of different levels in the encoder are fused with the features of the corresponding levels in the decoder to compensate for the information loss that occurs during the propagation of the image in the network;

[0034] The left side of the lightweight semantic segmentation model structure is a contraction path, which is the encoder part, including convolution units and pooling layers. Each layer on the left side first undergoes two convolution operations. After each even-numbered convolution unit, a maximum pooling layer is used for downsampling, and a cross-connection layer is used to fuse the encoded features to the corresponding upsampling layer in the decoder.

[0035] The convolution kernel size of the convolution unit is 3×3. A linear rectification function is introduced as an activation function at the output of each convolution unit to calculate the output result. A batch normalization layer and a dropout layer are introduced after each convolution unit to obtain regularized features.

[0036] The pooling layer uses a 2×2 window to perform the maximum pooling operation and then enters the next layer. In each downsampling step, the image size is reduced by half, and the number of feature channels is doubled.

[0037] The right side of the lightweight semantic segmentation model structure is an expansion path, which is the decoder part. The expansion path is used to gradually restore the downsampled image features to the original input image size. Each step includes upsampling and deconvolution of the feature map to halve the number of feature channels and double the matrix size, thereby obtaining an output result with the original input image size.

[0038] At the output end of the lightweight semantic segmentation model structure, each pixel of the input image is judged as belonging to the corresponding category with a certain probability through the Softmax classifier, that is, the image segmentation probability map is obtained;

[0039] The image segmentation probability map is processed by using a threshold function to obtain the final image segmentation result.

[0040] This application also proposes an automatic measurement system for irregular oral anatomical structures, comprising:

[0041] An image acquisition unit, used to acquire the CBCT image to be tested of the current patient;

[0042] A preprocessing unit, used for obtaining a standard sagittal section to be measured based on the CBCT image to be measured, and performing image normalization processing;

[0043] An intelligent network model is used to receive a standard sagittal section to be measured and output the standard sagittal section to be measured having coordinates of a predicted region of oral irregular structures and related measurement reference landmarks;

[0044] An oral subspecialty measurement method selection unit is used to obtain a selected oral subspecialty measurement method and send it to the automatic measurement unit, where the oral subspecialty measurement method includes reference points, reference lines, and measurement sites;

[0045] The automatic measurement unit calculates the quantitative indexes of the irregular oral anatomical structure in the standard sagittal section to be measured according to the selected oral subspecialty measurement method. The quantitative indexes include the thickness, height and area of ​​the irregular oral anatomical structure.

[0046] The present application also proposes a device, comprising: a processor and a memory, wherein the processor and the memory are capable of communicating;

[0047] The memory stores program instructions that can be executed by the processor, and the processor can execute the above method by calling the program instructions.

[0048] Beneficial effects of the present invention:

[0049] The automatic measurement method of oral irregular structures described in this application can accurately segment and fully quantitatively measure oral irregular structures.

[0050] (1) Based on the intelligent network model, the model is supervised and trained using pre-labeled samples, thereby automatically segmenting the oral irregular structures and related measurement reference landmarks in unlabeled CBCT images, achieving highly automated and high-precision segmentation with clinically applicable accuracy;

[0051] (2) Based on the automatic segmentation of oral irregular structures and the specific locations of related measurement reference landmarks in CBCT images, different oral irregular structure measurement methods can be applied to different oral subspecialties and in different scenarios;

[0052] (3) It can output the thickness, height and area of ​​the oral irregular structures of different patients under the current measurement method, assist clinicians in analysis and measurement, and provide a basis for subsequent medical diagnosis and treatment design, which has practical clinical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments with reference to the following drawings, in which:

[0054] Figure 1 Schematic diagram of obtaining historical standard sagittal sections based on historical CBCT images in S1;

[0055] 1a is a schematic diagram of historical CBCT images;

[0056] 1b is a reference diagram of the jaw plane and the human midline;

[0057] 1c is a schematic diagram of the dental arch curve;

[0058] 1d is a schematic diagram of the selected tooth position;

[0059] 1e is the standard sagittal section of the upper anterior teeth;

[0060] Figure 2 The standard sagittal section in S2 is marked with the enamel and palatal bone wall areas and the relevant measurement reference landmarks;

[0061] Figure 3 is a schematic diagram of the intelligent network model;

[0062] Figure 4 The standard sagittal section of the upper anterior teeth after the prediction of the enamel and palatal bone wall areas is output by the intelligent network model;

[0063] Figure 5 A flowchart of the automatic analysis and measurement method for oral irregular structures;

[0064] Figure 6 Schematic diagram of the module of the automatic analysis and measurement system for oral irregular structures. DETAILED DESCRIPTION

[0065] This section will describe in detail the specific embodiments of the present invention. The preferred embodiments of the present invention are shown in the accompanying drawings. The purpose of the accompanying drawings is to supplement the description of the text part of the specification with graphics, so that people can intuitively and vividly understand each technical feature and the overall technical solution of the present invention, but it should not be understood as a limitation on the scope of protection of the present invention.

[0066] In the description of this invention, "above," "below," and "within" are understood to be exclusive of the number indicated. The terms "first" and "second" are used solely to distinguish technical features and are not to be construed as indicating or implying relative importance, or as implicitly specifying the number or order of the technical features indicated.

[0067] Example 1

[0068] This embodiment provides an automatic measurement method for irregular oral anatomical structures, which include enamel, dentin, pulp cavity, labial bone wall, palatal bone wall, alveolar base bone, condyle, sinus, and nasopalatine canal. In this embodiment, the irregular oral anatomical structures include enamel and palatal bone wall. The automatic measurement method includes:

[0069] S1. Collect historical CBCT images of multiple patients and obtain the historical standard sagittal sections of anterior teeth based on the historical CBCT images, such as Figure 1 As shown, including:

[0070] S1.1. Import the patient's oral CBCT DICOM file using CoDiagnostiX software;

[0071] S1.2. Use the jaw plane and the human midline as the horizontal and vertical reference planes;

[0072] S1.3. Select the center of the pulp cavity on the transverse section using the cementoenamel junction 12-22 as the landmark, draw the dental arch curve, and obtain a standard cross-section that includes the enamel, palatal bone wall, and relevant measurement reference landmarks. Convert the selected cross-section to TIFF format, normalize the image and correct the grayscale to 0-255 to obtain the historical standard sagittal cross-section of anterior teeth.

[0073] S2. Mark the enamel, palatal bone wall, and related measurement reference landmarks in the historical standard sagittal section, and use the marked historical standard sagittal section to construct an image database. The related measurement reference landmarks include the tooth cusp point, the buccal point of the enamel-dentin junction, the palatal point of the enamel-dentin junction, the buccal alveolar ridge apex, the palatal alveolar ridge apex, the anterior nasal ridge point, the posterior nasal ridge point, the root apex, and the incisal edge apex; Figure 2 As shown, including:

[0074] S2.1. Import the historical standard sagittal section of anterior teeth into the image labeling processing software labelme;

[0075] S2.2. Use the polygon annotation tool to mark the enamel and palatal bone wall areas in the historical standard sagittal section of the anterior teeth;

[0076] S2.3. Use the point annotation tool to mark the relevant measurement reference landmarks of the irregular oral anatomical structures in the historical standard sagittal section of the anterior teeth. The relevant measurement reference landmarks include the buccal enamel-dentin junction point, the palatal enamel-dentin junction point, the buccal alveolar ridge apex, the palatal alveolar ridge apex, the anterior nasal ridge point, the posterior nasal ridge point, the root apex point, and the incisal edge apex.

[0077] S2.4. Save the marked historical standard sagittal sections of anterior teeth in an image format that can be recognized by a semantic segmentation convolutional neural network to construct an image database;

[0078] S2.5. Use the code editor Visual Studio Code to convert the marks of the enamel, palatal bone wall area, and related measurement reference landmarks into specific coordinates using Python statements.

[0079] S2.6. Use the coordinates of the relevant measurement reference markers to derive the measurement reference baselines using Python statements. The measurement reference baselines include: the cementoenamel junction line, the alveolar crest line, and the anterior and posterior nasal crest line.

[0080] S3. Measure the irregular oral anatomical structures in the image database using different oral subspecialty measurement methods, and construct a measurement analysis database using historical standard sagittal sections with markers and measurement results.

[0081] According to the measurement methods of different oral subspecialties, reference points and measurement reference baselines were selected. The intercepts of the parallel lines of the selected reference baselines at different distances from the reference points and the irregular anatomical structures of the oral cavity were calculated using Python statements. The thickness and height of the irregular anatomical structures of the oral cavity at multiple locations were calculated. The integral area of ​​the irregular anatomical structures of the oral cavity was calculated using Python statements to obtain the area data of the irregular anatomical structures of the oral cavity. Finally, a measurement and analysis database was constructed.

[0082] S4. Build an intelligent network model based on the semantic segmentation network, use the image database to train and verify the intelligent network model, and use the image database and measurement analysis database to test it;

[0083] S4.1. Divide the image database dataset into training set, validation set, and test set;

[0084] S4.2. Build an intelligent network model based on semantic segmentation network;

[0085] Semantic segmentation networks include U-net, DeepLab, and FCN. The intelligent network model in S4.2 uses a lightweight semantic segmentation model.

[0086] The lightweight semantic segmentation model has the following features: the intelligent network model is an autoencoder network structure, consisting of an encoder and a decoder. The corresponding modules in the two parts are connected through a cross-connection layer, fusing features from different levels in the encoder with features from the corresponding levels in the decoder to compensate for information loss during image propagation through the network;

[0087] The left side of the lightweight semantic segmentation model structure is a contraction path, which is the encoder part, including convolution units and pooling layers. Each layer on the left side first undergoes two convolution operations. After each even-numbered convolution unit, a maximum pooling layer is used for downsampling, and a cross-connection layer is used to fuse the encoded features to the corresponding upsampling layer in the decoder.

[0088] The convolution kernel size of the convolution unit is 3×3. A linear rectification function is introduced as an activation function at the output of each convolution unit to calculate the output result. A batch normalization layer and a dropout layer are introduced after each convolution unit to obtain regularized features.

[0089] The pooling layer uses a 2×2 window to perform the maximum pooling operation and then enters the next layer. In each downsampling step, the image size is reduced by half, and the number of feature channels is doubled.

[0090] The right side of the lightweight semantic segmentation model structure is an expansion path, which is the decoder part. The expansion path is used to gradually restore the downsampled image features to the original input image size. Each step includes upsampling and deconvolution of the feature map to halve the number of feature channels and double the matrix size, thereby obtaining an output result with the original input image size.

[0091] At the output end of the lightweight semantic segmentation model structure, each pixel of the input image is judged as belonging to the corresponding category with a certain probability through the Softmax classifier, that is, the image segmentation probability map is obtained;

[0092] The image segmentation probability map is processed by using a threshold function to obtain the final image segmentation result.

[0093] S4.3. Use the training set and validation set to supervise the intelligent network model, continuously iteratively update the weights and network parameters, use the cross entropy loss function as the loss function, minimize the loss function, and obtain the optimal intelligent network model;

[0094] The weights and network parameters can be iteratively updated using methods such as adaptive learning rate method and stochastic gradient descent method. In this embodiment, stochastic gradient descent method is used;

[0095] The test set data of the image database is input into the intelligent network model to obtain the segmentation results and the coordinates of the relevant measurement reference landmarks. The segmentation results are compared with the landmarks in the image database, and the segmentation accuracy of the intelligent network model is verified by IoU, Dice coefficient, Hausdorff distance and F1 score.

[0096] The segmentation results and the coordinates of the relevant measurement reference marks are input into the automatic measurement unit. The quantitative indicators of the oral irregular structures in the test set are calculated and output. The quantitative indicators of the oral irregular structures output by the test set are compared with the quantitative indicators of the corresponding data set in the measurement analysis database. The measurement accuracy of the intelligent network model is verified by MAE and RMSE.

[0097] S5. Obtain the standard sagittal section of the upper anterior teeth to be tested based on the patient's CBCT image to be tested, input the standard sagittal section of the upper anterior teeth to be tested into the intelligent network model, and obtain the predicted area of ​​the enamel and palatal bone wall in the standard sagittal section of the upper anterior teeth to be tested (such as Figure 4 as shown) and the coordinates of the relevant survey reference marks;

[0098] S6. Select an oral subspecialty measurement method, which includes reference points, reference lines, and measurement sites. Calculate the thickness, height, and area of ​​the corresponding irregular oral anatomical structure based on the selected oral subspecialty measurement method.

[0099] The irregular oral anatomical structures and related measurement reference landmarks in the medical images described in this embodiment are relatively small, requiring an effective region selection strategy. Otherwise, the neural network would waste most of its weights when processing the background area and would be unable to accurately locate the target area.

[0100] Moreover, the irregular structures of the oral cavity have complex shapes and complex internal textures, which are similar to the textures of the surrounding structures (for example, the palatal bone wall contains both cancellous bone signals and compact bone signals, and the compact bone signals are similar to those of adjacent teeth). Therefore, the intelligent network model described in this application requires a network with strong feature extraction capabilities.

[0101] The intelligent network model is a deep and complex model that needs to analyze multiple areas (target area and measurement auxiliary reference area), requires outstanding generalization capabilities, requires a deeper and more complex network, and consumes more computing resources.

[0102] Example 2

[0103] like Figure 6As shown, this embodiment provides an automatic analysis and measurement system for irregular oral anatomical structures, comprising:

[0104] An image acquisition unit, used to acquire the CBCT image to be tested of the current patient;

[0105] A preprocessing unit is used to obtain a standard sagittal section of the upper anterior teeth to be tested based on the CBCT image to be tested and perform image normalization processing;

[0106] An intelligent network model is used to receive a standard sagittal section of an upper anterior tooth to be measured and output the standard sagittal section of the upper anterior tooth to be measured having a predicted region and coordinates of an irregular oral anatomical structure and coordinates of related measurement landmarks;

[0107] An oral subspecialty measurement method selection unit is used to obtain a selected oral subspecialty measurement method and send it to the automatic measurement unit, where the oral subspecialty measurement method includes reference points, reference lines, and measurement sites;

[0108] Automatic measurement unit calculates the thickness, height and area of ​​irregular oral anatomical structures based on the selected oral subspecialty measurement method.

[0109] In this embodiment, the oral subspecialty measurement method selection unit allows doctors of different oral subspecialties to select different reference points, reference lines, and measurement sites, thereby outputting corresponding measurement results for different oral subspecialties to assist clinicians in measurement and analysis.

[0110] Example 3

[0111] This embodiment provides a device including: a processor and a memory, wherein the processor and the memory are capable of communicating;

[0112] The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the method described in Example 1.

[0113] Of course, the present invention is not limited to the above-mentioned embodiments. Those skilled in the art may make equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications and substitutions are all included in the scope defined by the claims of this application.

Claims

1. A method for automatically measuring irregular anatomical structures of the oral cavity, characterized in that: Automated measurement methods include: S1. Collect historical CBCT images of multiple patients and obtain historical standard sagittal sections based on the historical CBCT images; S2. Mark the irregular anatomical structures of the oral cavity and related measurement reference landmarks in the historical standard sagittal sections, and construct an image database using the marked historical standard sagittal sections; S3. Measure the irregular oral anatomical structures in the imaging database using different oral subspecialty measurement methods to obtain quantitative indices of the irregular oral anatomical structures. Construct a measurement and analysis database using historical standard sagittal sections with markers and quantitative indices of the irregular oral anatomical structures. S4. Build an intelligent network model based on the semantic segmentation network, use the image database to train and verify the intelligent network model, and use the image database and measurement analysis database to test it; S5. Obtaining a standard sagittal section to be measured based on the patient's CBCT image to be measured, inputting the standard sagittal section to be measured into the intelligent network model, and obtaining the predicted area of ​​the oral irregular anatomical structure and the coordinates of relevant measurement reference marks in the standard sagittal section to be measured; S6. Select an oral subspecialty measurement method for calculating irregular oral anatomical structures. The oral subspecialty measurement method includes reference points, reference lines, and measurement sites. Calculate quantitative indicators of the corresponding irregular oral anatomical structures based on the selected oral subspecialty measurement method. The quantitative indicators include thickness, height, and area.

2. The automatic measurement method of irregular oral anatomical structures according to claim 1, characterized in that: In S1, historical CBCT images of multiple patients were collected. The method for obtaining historical standard sagittal sections based on the historical CBCT images includes: S1.

1. Use CBCT processing software to import the patient's oral CBCT DICOM file; S1.

2. Use the jaw plane and the human midline as the horizontal and vertical reference planes; S1.

3. Select the center of the pulp cavity on the cross-section using the cementoenamel junction as the landmark point, draw the dental arch curve, obtain the standard cross-section, convert the selected cross-section into TIFF format, normalize the image and correct the grayscale to 0-255 to obtain the historical standard sagittal cross-section.

3. The automatic measurement method of irregular oral anatomical structures according to claim 1, characterized in that: In the S2, the method for marking irregular oral anatomical structures and related measurement reference landmarks in the historical standard sagittal section and constructing the image database includes: S2.

1. Import the historical standard sagittal section into the image marking processing software; S2.

2. Use the polygon annotation tool to mark the irregular anatomical regions of the oral cavity in the historical standard sagittal section; S2.

3. Use the point marking tool to mark the relevant measurement reference landmarks in the historical standard sagittal section; S2.

4. Save the marked historical standard sagittal sections in an image format that can be recognized by a semantic segmentation convolutional neural network to build an image database.

4. The automatic measurement method of irregular oral anatomical structures according to claim 1, characterized in that: The method of obtaining measurement data of irregular oral anatomical structures of an image set in the image database under different oral subspecialty measurement methods in S3 and constructing a measurement analysis database includes: S3.

1. Converting the markings of the irregular oral anatomical structures and related measurement reference landmarks in the dataset of the image database into specific coordinates; S3.

2. Derive the measurement reference baseline based on the coordinates of the relevant measurement reference marks; S3.

3. Select reference points and measurement reference baselines based on different oral subspecialty measurement methods. Use Python statements to calculate the intercepts of the selected reference baselines at different distances from the reference points with parallel lines and perpendicular lines and the irregular oral anatomical structures. Calculate the thickness and height of the irregular oral anatomical structures at multiple locations. Use Python statements to calculate the integral area of ​​the irregular oral anatomical structure region to obtain the area data of the irregular oral anatomical structures. Finally, construct a measurement and analysis database.

5. The automatic measurement method of irregular oral anatomical structures according to claim 1, characterized in that: The method of constructing an intelligent network model based on a semantic segmentation network in S4, training and verifying the intelligent network model using an image database, and testing the intelligent network model using an image database and a measurement analysis database includes: S4.

1. Divide the image database dataset into training set, validation set, and test set; S4.2, build intelligent network model; S4.

3. Use the training set and validation set to supervise the training of the intelligent network model, and continuously iteratively update the weights and network parameters to minimize the loss function and obtain the optimal intelligent network model; The test set is input into the intelligent network model to obtain the segmentation results, verify the segmentation accuracy of the intelligent network model, and use the corresponding data set in the measurement analysis database to verify the accuracy of the measurement algorithm.

6. The automatic measurement method of irregular oral anatomical structures according to claim 5, characterized in that: The intelligent network model in S4 is a lightweight semantic segmentation model, which has the following features: The lightweight semantic segmentation model is an autoencoder network structure, consisting of an encoder and a decoder. The corresponding modules in the two parts are connected through a cross-connection layer, which fuses the features of different levels in the encoder with the features of the corresponding levels in the decoder to compensate for the information loss that occurs during the image propagation process in the network. The left side of the lightweight semantic segmentation model structure is a contraction path, which is the encoder part, including convolution units and pooling layers. Each layer on the left side first undergoes two convolution operations. After each even-numbered convolution unit, a maximum pooling layer is used for downsampling, and a cross-connection layer is used to fuse the encoded features to the corresponding upsampling layer in the decoder. The convolution kernel size of the convolution unit is 3×3. A linear rectification function is introduced as an activation function at the output of each convolution unit to calculate the output result. A batch normalization layer and a dropout layer are introduced after each convolution unit to obtain regularized features. The pooling layer uses a 2×2 window to perform the maximum pooling operation and then enters the next layer. In each downsampling step, the image size is reduced by half, and the number of feature channels is doubled. The right side of the lightweight semantic segmentation model structure is an expansion path, which is the decoder part. The expansion path is used to gradually restore the downsampled image features to the original input image size. Each step includes upsampling and deconvolution of the feature map to halve the number of feature channels and double the matrix size, thereby obtaining an output result with the original input image size. At the output end of the lightweight semantic segmentation model structure, each pixel of the input image is judged as belonging to the corresponding category with a certain probability through the Softmax classifier, that is, the image segmentation probability map is obtained; The image segmentation probability map is processed by using a threshold function to obtain the final image segmentation result.

7. An automatic measurement system for irregular oral anatomical structures, characterized in that: include: An image acquisition unit, used to acquire the CBCT image to be tested of the current patient; A preprocessing unit is used to obtain a standard sagittal section to be tested based on the CBCT image to be tested and perform image normalization processing; An intelligent network model is used to receive a standard sagittal section to be measured and output the standard sagittal section to be measured having coordinates of a predicted region of an irregular anatomical structure of the oral cavity and related measurement landmarks; An oral subspecialty measurement method selection unit is used to obtain a selected oral subspecialty measurement method and send it to the automatic measurement unit, where the oral subspecialty measurement method includes reference points, reference lines, and measurement sites; The automatic measurement unit calculates the quantitative indexes of the irregular oral anatomical structure in the standard sagittal section to be measured according to the selected oral subspecialty measurement method. The quantitative indexes include the thickness, height and area of ​​the irregular oral anatomical structure.

8. A computer device comprising: a processor and a memory, the processor and the memory being capable of communicating; The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the method according to claims 1 to 6.

Images