Method for deducing gender and age based on maxillary sinus, model establishing method and system

Abstract

The invention relates to the technical field of biology and new medicine, in particular to a method for deducing gender and age based on maxillary sinus and a model establishment method and system. The method for deducing the gender and age of the maxillary sinus comprises the steps of obtaining skull CBCT images of all target samples, establishing an automatic segmentation model, performing automatic segmentation of the maxillary sinus on all unlabeled skull CBCT images, and performing automatic measurement to obtain linear data and volumes of all maxillary sinus segmentation images; the selected target samples are grouped according to preset age intervals, difference and correlation test is carried out on various data obtained through measurement, gender difference parameters and age correlation parameters of all the age intervals are obtained respectively, and a gender inference model and an age inference model are established. Compared with previous manual data acquisition, the method is faster, more accurate and more objective, related data of the maxillary sinus can be automatically segmented and measured only by inputting the skull CBCT image subsequently, and gender and age deduction results are obtained on the basis of the related data.

Description

Technical Field

[0001] The present invention relates to the field of biology and new medical technology, and in particular to a method for inferring gender and age based on maxillary sinus, a model building method and a system. Background Art

[0002] Individual identification is one of the main research directions of forensic medicine, and the inference of gender and age is an important part of it. When the body is in a more extreme situation, such as chemical corrosion, fire, explosion and vicious criminal cases, the remains often become unrecognizable. At this time, how to infer the gender and age through the incomplete remains and then identify the individual identity becomes a difficult problem.

[0003] There are many methods available for gender inference, such as bone measurement, dental measurement, and DNA testing. Among them, gender inference methods based on DNA testing have been widely used. However, in some areas with relatively backward economic levels and relatively harsh environmental conditions, the requirements of DNA testing for equipment, sample quality, and time are difficult to meet under certain conditions, making testing difficult. In some cases, it may be impossible to effectively extract DNA from body remains. Therefore, it is necessary to develop simpler, more efficient, and less equipment-demanding testing methods to supplement conventional methods.

[0004] As for age inference, the methods commonly used in forensic practice still have room for improvement and need to be optimized. Although the method of using X-rays to determine bone age is highly feasible, its accuracy in the adult population is relatively low. Many emerging age inference technologies have been developed, such as age estimation based on DNA methylation levels and age estimation based on clavicular sternal epiphysis staging. Although these technologies have high accuracy, they often have certain limitations in practical application and have large errors when inferring the age of adults. Among them, DNA methylation level detection requires the use of professional equipment with high equipment costs, and clavicular sternal epiphysis staging requires experts in this professional field to accurately stage the tissue structure first. These factors limit the promotion and application of related methods.

[0005] Bones are relatively stable tissues in the human body structure, and a large number of studies have been conducted on gender and age inference based on bones. For gender inference, the accuracy of inference using the skull and pelvis is over 90%, and there are also some methods using long bones that have higher accuracy. For age inference, there are but are not limited to the two methods mentioned above, but these larger bones may not be able to guarantee integrity when conducting forensic identification. In many cases, even if other bone structures have been severely damaged, the maxillary sinus can still maintain its integrity, so the maxillary sinus has great potential in forensic identification. In addition, studies have shown that the various linear data and volumes in the maxillary sinus have gender differences and are linearly correlated with age, so gender and age inferences can be made by measuring the anatomical structure of the maxillary sinus.

[0006] In recent years, with the gradual promotion of three-dimensional imaging technology in medical practice, there have been related studies at home and abroad using computed tomography (CT), micro computed tomography (micro-CT) and cone beam CT (CBCT) technologies to detect various bone structures for forensic practice. At the same time, the application of machine learning in medical research is becoming more and more extensive. For example, machine learning is used to fit regression models to evaluate acute kidney injury. There are also studies showing that the interaction between machine learning and imaging omics can better stage cancer. There is no doubt that the application of machine learning in the field of medical research will play a more important role.

[0007] At present, most studies at home and abroad on gender and age inference based on the anatomical structure of the maxillary sinus are based on the linear data of the bilateral maxillary sinus. The data obtained are not comprehensive enough. If the volume of the maxillary sinus is to be collected, it is necessary to face the problem of cumbersome steps, long time consumption, high technical threshold, and data analysis for each sample takes about 2 hours. In addition, the data acquisition methods in existing studies are all through manual measurement. When only linear data that is easier to obtain is used, the accuracy cannot be guaranteed. When data that is more difficult to obtain, such as volume, is used, a lot of time will be spent on data collection. Secondly, the models in these studies are all traditional multivariate linear regression and logistic regression models (logistic). Based on the superiority of machine learning in model fitting, the use of machine learning for model fitting can obtain higher accuracy. In addition, the sample size of most existing studies does not exceed 200 cases, and the model fitted is not representative. In addition, since the data collection method is manual measurement, on the one hand, it increases the workload and time spent. On the other hand, the errors caused by the subjective judgment of different measurement personnel during measurement and the difference in professional knowledge level will also affect the accuracy and stability of the data, which will also affect the accuracy of the inference results, making the model difficult to put into use. Therefore, with regard to the use of the maxillary sinus for age inference, few researchers have attempted to use it as a basis for age inference, resulting in a large gap in current research in this area. Summary of the invention

[0008] The purpose of the present invention is to provide a method for inferring gender and age based on the maxillary sinus, a model building method and a system, so as to solve the problem in the prior art that when other bones and remaining tissues are lost, it is difficult for forensic doctors to make accurate and effective judgments on the gender and age of the corpse.

[0009] In order to achieve the above object, the present invention adopts the following technical solution:

[0010] The present invention discloses a method for inferring gender and age based on maxillary sinus, comprising:

[0011] S1, selecting a preset number of target samples and obtaining skull CBCT images of all target samples;

[0012] S2, selecting skull CBCT images of some target samples to label and segment the maxillary sinus, obtaining labeled images, and establishing an automatic segmentation model based on the labeled images and unlabeled skull CBCT images of some target samples;

[0013] S3, automatically segmenting the maxillary sinus of all unlabeled skull CBCT images according to the automatic segmentation model to obtain maxillary sinus segmentation images, and obtaining the linear data and volume of all maxillary sinus segmentation images by measurement;

[0014] S4, grouping the selected target samples according to preset age ranges, performing difference tests between the measured data and the real gender, and obtaining gender difference parameters of each age range, and performing correlation tests between the measured data and the real age, and obtaining age correlation parameters of each age range;

[0015] S5. Establishing a gender inference model based on the gender difference parameters of each age interval, and establishing an age inference model based on the age correlation parameters of each age interval;

[0016] S6. Input the skull CBCT image to be detected into the automatic segmentation model to obtain the linear data and volume of the maxillary sinus of the skull CBCT image to be detected, input the data of the maxillary sinus into the gender inference model and the age inference model for inference, and output the gender and age corresponding to the skull CBCT image to be detected respectively.

[0017] Optionally, the method of selecting a portion of the target sample's skull CBCT images for maxillary sinus annotation and segmentation includes:

[0018] The skull CBCT images of all target samples were imported into ITK-SNAP software;

[0019] The skull CBCT images of some selected target samples were retrieved one by one, and the range of the maxillary sinus was manually marked using the ROI tool in each slice where the maxillary sinus existed in the retrieved skull CBCT images;

[0020] According to the maxillary sinus range marked on the skull CBCT image, select the segmentation tool to perform maxillary sinus segmentation on the retrieved skull CBCT image to obtain a marked image containing only the maxillary sinus;

[0021] The annotated images and other unannotated skull CBCT images are stored in nii.gz format for future use.

[0022] Optionally, the method for establishing an automatic segmentation model based on the annotated image and the unannotated skull CBCT image of a portion of the target sample comprises:

[0023] The VNet architecture based on 3D convolution and 3D deconvolution is selected as the basic model, and the Student training model and Teacher auxiliary model are established. The UNet architecture with cross-layer links is added between the Encoder part of the Student training model and the Decoder part of the Teacher auxiliary model.

[0024] Extracting labeled data from the labeled images, and using the labeled data to train the Student training model and the Teacher auxiliary model respectively, the Teacher auxiliary model assists the Student training model to learn the labeling features and segmentation patterns of the maxillary sinus based on the labeled data;

[0025] Supervised data is extracted from unlabeled skull CBCT images, and the supervised data is input into the Student training model and the Teacher auxiliary model respectively. The Teacher auxiliary model assists the Student training model in self-training of maxillary sinus labeling and segmentation based on the supervised data.

[0026] Based on the co-training of multiple networks, an automatic segmentation model is established through mutual training between the Student training model and the Teacher auxiliary model.

[0027] Optionally, it also includes a method for mutual training between the Student training model and the Teacher auxiliary model:

[0028] The labeled data are input into the Student training model and the Teacher auxiliary model respectively. The Student training model learns the labeling features and segmentation mode of the maxillary sinus through back propagation. The Teacher auxiliary model is updated by using the average EMA of the labeled data and the Student model to obtain a pseudo label with a certain resolution by adding a fixed parameter VNet based on the MT architecture that is not updated with back propagation to the Teacher auxiliary model.

[0029] The supervised data are input into the Student training model and the Teacher auxiliary model respectively. The Student training model performs self-training of maxillary sinus annotation and segmentation based on the supervised data. The Teacher auxiliary model imposes conditional constraints on the self-training of the Student training model and performs consistency detection with the Student training model by updating parameters through EMA average.

[0030] The pseudo labels of the Teacher auxiliary model are used as new supervised data and input into the Student training model to guide the Student training model to learn and train.

[0031] Optionally, the method further includes a method for processing the labeled image and the unlabeled skull CBCT image before establishing the automatic segmentation model:

[0032] Integrate the annotated images and the unannotated skull CBCT images into CBCT image data in a data format, and delete the CBCT image data that does not meet a preset data volume size;

[0033] For all CBCT image data after the deletion process, the CBCT image data with a pixel gray value range of -1000-2000 is converted into CBCT image data with a pixel gray value range of 0-255, and all the CBCT image data are normalized to obtain CBCT image data in int8 format;

[0034] After normalization, all CBCT image data are cropped to obtain CBCT image data containing only maxillary sinus images;

[0035] The cropped CBCT image data is saved as an h5 file.

[0036] Optionally, the method of obtaining the linear data and volume of all maxillary sinus segmentation images by measuring includes:

[0037] The skull CBCT images of some selected target samples are annotated one by one, and the range of the maxillary sinus is annotated in each section of the skull CBCT image where the maxillary sinus exists;

[0038] According to the maxillary sinus segmentation image obtained by the automatic segmentation model, in each maxillary sinus segmentation image, the linear length, linear width, and linear distance between the two lateral walls of the maxillary sinus are obtained by measuring the continuous cross-sections of each section, and the linear height of the maxillary sinus is obtained by measuring the continuous coronal planes of each section;

[0039] The data obtained from multiple tomographic measurements in each maxillary sinus segmentation image are merged to obtain the volume data of the corresponding maxillary sinus.

[0040] Optionally, a method for obtaining gender difference parameters and age correlation parameters for each age range is also included:

[0041] The Shapiro-Wilk test was used to test the normal distribution of the corresponding measured data of all target samples in each age range;

[0042] If all the measured data of the target samples obey the normal distribution, set the measured data as statistical variables, draw ridge diagrams for each variable by gender, and obtain ridge diagrams corresponding to each variable in each age range;

[0043] Based on the ridge plot, the responses of each variable on different genders in each age range were visualized, and the independent sample t-test was used to compare the differences of each measurement data between different genders, and the parameter data with gender differences in each age range were obtained;

[0044] The Pearson test was used to evaluate the linear correlation between each measurement data and the corresponding target sample's true age, and the age-related parameter data in each age range were obtained.

[0045] Optionally, the method for establishing a gender inference model and an age inference model is also included:

[0046] All target samples were randomly divided into training group and test group in a ratio of 8:2;

[0047] Select multiple preset regression models, input the gender difference parameters of all target samples in the training group into the multiple regression models for gender inference training, and after the model training, input the gender difference parameters of all target samples in the test group into the multiple regression models for test evaluation, judge the model fitting efficiency according to the evaluation results, and determine the random forest model as the gender inference model;

[0048] Select multiple preset regression models, input the age-related parameters of all target samples in the training group into multiple regression models for age inference training, and after model training, input the age-related parameters of all target samples in the test group into multiple regression models for test evaluation, judge the model fitting efficiency according to the evaluation results, and determine the fitting model of the random forest model and the multivariate linear regression model as the age inference model.

[0049] The present invention also provides a model building method, which comprises:

[0050] A preset number of target samples are selected to obtain skull CBCT images of all target samples;

[0051] Select the skull CBCT images of some target samples to annotate and segment the maxillary sinus to obtain annotated images, and establish an automatic segmentation model based on the annotated images and the unannotated skull CBCT images of some target samples;

[0052] Automatically segment the maxillary sinus of all unlabeled skull CBCT images according to the automatic segmentation model to obtain maxillary sinus segmentation images, and obtain the linear data and volume of all maxillary sinus segmentation images by measurement;

[0053] The selected target samples are grouped according to the preset age ranges, and the difference test between the measured data and the real gender is performed to obtain the gender difference parameters of each age range, and the correlation test between the measured data and the real age is performed to obtain the age correlation parameters of each age range;

[0054] A gender inference model is established based on the gender difference parameters of each age range, and an age inference model is established based on the age correlation parameters of each age range.

[0055] The present invention also provides an inference system, comprising:

[0056] A sample screening module is used to select a preset number of target samples and obtain skull CBCT images of all target samples;

[0057] An automatic segmentation model building module is used to select skull CBCT images of some target samples to mark and segment the maxillary sinus, obtain a marked image, and build an automatic segmentation model based on the marked image and unmarked skull CBCT images of some target samples;

[0058] An image segmentation and measurement module is used to automatically segment the maxillary sinus of all unlabeled skull CBCT images according to the automatic segmentation model to obtain maxillary sinus segmentation images, and obtain various linear data and volumes of all maxillary sinus segmentation images by measurement;

[0059] The data verification module is used to group the selected target samples according to the preset age ranges, perform difference tests between the measured data and the real gender, and obtain the gender difference parameters of each age range, and perform correlation tests between the measured data and the real age, and obtain the age correlation parameters of each age range;

[0060] An inference model building module is used to build a gender inference model based on the gender difference parameters of each age interval, and to build an age inference model based on the age correlation parameters of each age interval;

[0061] The gender and age inference module is used to input the skull CBCT image to be detected into the automatic segmentation model to obtain the linear data and volume of the maxillary sinus of the skull CBCT image to be detected, input the various measurement data of the maxillary sinus into the gender inference model and the age inference model for inference, and output the gender and age corresponding to the skull CBCT image to be detected respectively.

[0062] Compared with the prior art, the method for inferring gender and age based on maxillary sinus provided by the embodiment of the present invention has the following beneficial effects:

[0063] By collecting the skull CBCT images of the target samples and establishing an automatic segmentation model, the training and data collection of the automatic segmentation of the maxillary sinus of the skull CBCT images of the target samples are realized. Then, through data analysis and model fitting, the difference test between the measured data and the real gender is carried out to obtain the gender difference parameters of each age range, and the correlation test between the measured data and the real age is carried out to obtain the age correlation parameters of each age range. The gender inference model is established according to the gender difference parameters of each age range, and the age inference model is established according to the age correlation parameters of each age range. Therefore, only the skull CBCT images need to be input later, and the relevant data of the maxillary sinus can be automatically segmented and measured through the synergy of various models, and the gender and age inference results can be obtained based on this. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments, in which:

[0065] Figure 1 A schematic flowchart of a method for inferring gender and age based on the maxillary sinus provided in an embodiment of the present invention;

[0066] Figure 2 A CBCT image of the skull with the maxillary sinus range marked provided in an embodiment of the present invention;

[0067] Figure 3 The image diagrams of the CBCT image data before and after cropping provided by the embodiment of the present invention;

[0068] Figure 4 An image diagram of various linear data of tomographic measurement in the maxillary sinus segmentation image provided by an embodiment of the present invention;

[0069] Figure 5 An image diagram of maxillary sinus segmentation provided by an embodiment of the present invention;

[0070] Figure 6 A ridge plot of each variable provided in an embodiment of the present invention is plotted separately by gender. DETAILED DESCRIPTION

[0071] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. Now, in conjunction with the accompanying drawings, the preferred embodiments of the present invention are described in detail.

[0072] The embodiment of the present invention provides a method for inferring gender and age based on maxillary sinus. Figure 1 As shown, including:

[0073] S1, selecting a preset number of target samples and obtaining skull CBCT images of all target samples;

[0074] S2, selecting skull CBCT images of some target samples to label and segment the maxillary sinus, obtaining labeled images, and establishing an automatic segmentation model based on the labeled images and unlabeled skull CBCT images of some target samples;

[0075] S3, automatically segmenting the maxillary sinus of all unlabeled skull CBCT images according to the automatic segmentation model to obtain maxillary sinus segmentation images, and obtaining the linear data and volume of all maxillary sinus segmentation images by measurement;

[0076] S4, grouping the selected target samples according to preset age ranges, performing difference tests between the measured data and the real gender, and obtaining gender difference parameters of each age range, and performing correlation tests between the measured data and the real age, and obtaining age correlation parameters of each age range;

[0077] S5. Establishing a gender inference model based on the gender difference parameters of each age interval, and establishing an age inference model based on the age correlation parameters of each age interval;

[0078] S6. Input the CBCT image of the skull to be detected into the automatic segmentation model to obtain the linear data and volume of the maxillary sinus of the CBCT image of the skull to be detected, input the various measurement data of the maxillary sinus into the gender inference model and the age inference model for inference, and output the gender and age corresponding to the CBCT image of the skull to be detected respectively.

[0079] Through the implementation of the above method, the skull CBCT images of the target sample are first collected, and an automatic segmentation model is established to realize the training and data collection of the automatic segmentation of the maxillary sinus of the skull CBCT images of the target sample, and then through data analysis and model fitting, the difference and correlation test of the measured data and the real gender and age are carried out, and the gender inference model and the age inference model are established according to the test parameters. Subsequently, only the skull CBCT image needs to be input, and the relevant data of the maxillary sinus can be automatically segmented and measured through the synergy of each model, and the gender and age inference results can be obtained based on this. Compared with the previous manual data collection, the method of the embodiment of the present invention is faster, more accurate, and more objective, and there is no professional requirement for the user in the relevant field. The established model has good prediction accuracy, low actual application cost, and is not limited by space.

[0080] As described above, the method of selecting a preset number of target samples and obtaining skull CBCT images of all target samples includes:

[0081] According to the preset inclusion criteria, a target sample of more than 200 cases was selected;

[0082] The skull CBCT images of all target samples are acquired in DICOM format by a cone beam X-ray computed tomography system;

[0083] According to the target samples corresponding to each skull CBCT image, the corresponding actual gender and actual age are recorded for all skull CBCT images.

[0084] Through the implementation of the above method, preferably, 515 skull CBCT images were selected, including 257 males and 258 females. All skull CBCT images were scanned by a cone beam X-ray computed tomography system (KaVO 3D eXim I, USA) and exported to DICOM format. The inclusion criteria include: (1) living in the Northwest, no trauma, no surgery, no deformity, no systemic disease, no malnutrition; (2) the patient registered complete personal information; (3) both maxillary sinuses were photographed completely, and the boundaries with the surrounding tissues were clear; (4) there was no severe inflammation of the maxillary sinus.

[0085] Furthermore, the method of selecting a part of the target sample's skull CBCT images for maxillary sinus annotation and segmentation includes:

[0086] The skull CBCT images of all target samples were imported into ITK-SNAP software;

[0087] The skull CBCT images of some selected target samples were retrieved one by one, and the range of the maxillary sinus was manually marked using the ROI tool in each slice where the maxillary sinus existed in the retrieved skull CBCT images;

[0088] According to the maxillary sinus range marked on the skull CBCT image, select the segmentation tool to perform maxillary sinus segmentation on the retrieved skull CBCT image to obtain a marked image containing only the maxillary sinus;

[0089] The annotated images and other unannotated skull CBCT images are stored in nii.gz format for future use.

[0090] Through the implementation of the above method, TK-SNAP is a commercial image stitching software widely used in photogrammetry, medical image processing and other fields. The skull CBCT images were imported into ITK-SNAP 4.0.2 for segmentation and annotation of the maxillary sinus. Among them, based on the above preferred embodiment of selecting 515 skull CBCT images, 40 samples were selected for segmentation, such as Figure 2 As shown, the extent of the maxillary sinus was manually annotated in each slice where the maxillary sinus was present.

[0091] In ITK-SNAP, select the "Open" option in the "File" menu, then find and select the skull CBCT image file to be processed. After importing the image, you can see the tomographic image of the skull in ITK-SNAP. You can use tools such as zoom and rotation to adjust the view of the image. In each section where the maxillary sinus exists, use the annotation tools provided by ITK-SNAP (such as brushes, polygons, etc.) to manually outline the range of the maxillary sinus to ensure that the boundaries of the maxillary sinus are accurately outlined. After completing the annotation of each section, select the "Save" or "Export" option to save the annotation results in a specific format. Generally, the nii.gz format is a common medical image annotation format.

[0092] Furthermore, the method of establishing an automatic segmentation model based on the labeled image and the unlabeled skull CBCT image of some target samples includes:

[0093] The VNet architecture based on 3D convolution and 3D deconvolution is selected as the basic model, and the Student training model and Teacher auxiliary model are established. The UNet architecture with cross-layer links is added between the Encoder part of the Student training model and the Decoder part of the Teacher auxiliary model.

[0094] Extracting labeled data from the labeled images, and using the labeled data to train the Student training model and the Teacher auxiliary model respectively, the Teacher auxiliary model assists the Student training model to learn the labeling features and segmentation patterns of the maxillary sinus based on the labeled data;

[0095] Supervised data is extracted from unlabeled skull CBCT images, and the supervised data is input into the Student training model and the Teacher auxiliary model respectively. The Teacher auxiliary model assists the Student training model in self-training of maxillary sinus labeling and segmentation based on the supervised data.

[0096] Based on the co-training of multiple networks, an automatic segmentation model is established through mutual training between the Student training model and the Teacher auxiliary model.

[0097] Furthermore, it also includes a method for mutual training between the Student training model and the Teacher auxiliary model:

[0098] The labeled data are input into the Student training model and the Teacher auxiliary model respectively. The Student training model learns the labeling features and segmentation mode of the maxillary sinus through back propagation. The Teacher auxiliary model is updated by using the average EMA of the labeled data and the Student model to obtain a pseudo label with a certain resolution by adding a fixed parameter VNet based on the MT architecture that is not updated with back propagation to the Teacher auxiliary model.

[0099] The supervised data are input into the Student training model and the Teacher auxiliary model respectively. The Student training model performs self-training of maxillary sinus annotation and segmentation based on the supervised data. The Teacher auxiliary model imposes conditional constraints on the self-training of the Student training model and performs consistency detection with the Student training model by updating parameters through EMA average.

[0100] The pseudo labels of the Teacher auxiliary model are used as new supervised data and input into the Student training model to guide the Student training model to learn and train.

[0101] Through the implementation of the above method, a semi-supervised training method is adopted, and the training data includes labeled data and supervised data. Although a large amount of data is used, only a small part of the data is labeled images, and good training performance is still achieved. Based on the above preferred embodiment of selecting 515 skull CBCT images, the embodiment of the present invention uses 4 data for testing and 76 data for training, of which 24 are supervised data.

[0102] Traditional semi-supervised learning uses a variety of paradigms. After condensing and refining existing work, the embodiment of the present invention proposes an optimized training paradigm and achieves good results. The basic model in its semi-supervised learning framework consists of two VNets and a Mean Teacher architecture. VNet is a UNet architecture that uses three-dimensional convolution and three-dimensional deconvolution and adds cross-layer links between the Encoder and Decoder parts. It is widely used in semi-supervised medical image segmentation. On this basis, a Mean Teacher architecture (MT) is added to each VNet. The MT architecture uses a VNet that does not update parameters with backpropagation as the Teacher model of the existing model. Its parameter update is updated by the Student model through EMA averaging. The consistency between the Teacher and the Student on unsupervised data is detected, and the model is constrained, so that it has good generalization ability and can enhance the representation ability of the model on unlabeled data. Co-training between multiple networks is also used as a widespread semi-supervised learning training method. Co-training is a multi-view learning method that uses multiple related but not completely consistent data sources to train a model. In the embodiment of the present invention, based on the use of two MT frameworks, the models are mutually trained by using a pseudo-label generation method and utilizing the stable and high-resolution pseudo-labels output by the MT frameworks.

[0103] Furthermore, the method also includes a method for processing the labeled image and the unlabeled skull CBCT image data before establishing the automatic segmentation model:

[0104] Integrate the annotated images and the unannotated skull CBCT images into CBCT image data in a data format, and delete the CBCT image data that does not meet a preset data volume size;

[0105] For all CBCT image data after the deletion process, the CBCT image data with a pixel gray value range of -1000-2000 is converted into CBCT image data with a pixel gray value range of 0-255, and all the CBCT image data are normalized to obtain CBCT image data in int8 format;

[0106] After normalization, all CBCT image data are cropped to obtain CBCT image data containing only maxillary sinus images;

[0107] The cropped CBCT image data is saved as an h5 file.

[0108] Through the implementation of the above method, the CBCT image data is normalized and the image format can be converted to int8 format to speed up the data throughput of the model, and then it is stored together with the annotation data as an h5 file for easy reading. At the same time, the image is cropped so that the irrelevant parts are cropped. Figure 3 As shown, Figure 3 The part marked (a) is the image before cropping. Figure 3 The portion marked (b) in the figure is the cropped image.

[0109] Furthermore, the method of obtaining the linear data and volume of all maxillary sinus segmentation images by measuring includes:

[0110] The skull CBCT images of some selected target samples are annotated one by one, and the range of the maxillary sinus is annotated in each section of the skull CBCT image where the maxillary sinus exists;

[0111] According to the maxillary sinus segmentation image obtained by the automatic segmentation model, in each maxillary sinus segmentation image, the linear length, linear width, and linear distance between the two lateral walls of the maxillary sinus are obtained by measuring the continuous cross-sections of each section, and the linear height of the maxillary sinus is obtained by measuring the continuous coronal planes of each section;

[0112] The data obtained from multiple tomographic measurements in each maxillary sinus segmentation image are merged to obtain the volume data of the corresponding maxillary sinus.

[0113] By implementing the above method, Figure 4 As shown in the figure, the maximum linear length of the maxillary sinus (MSL) (a), which is the distance between the most anterior and posterior points of the maxillary sinus inner wall; the maximum linear width of the maxillary sinus (MSW) (b), which is the distance between the most lateral point and the most medial point of a single maxillary sinus; and the distance between the two lateral walls of the maxillary sinus (distance) (d), which is the distance between the two lateral sides of the maxillary sinus. The maximum linear height of the maxillary sinus (MSH) (c), which is the vertical distance between the highest point and the lowest point of the maxillary sinus inner wall in the vertical direction, was measured on the continuous coronal plane. The maxillary sinus range was marked in each section of the skull CBCT, and the sections were finally merged to measure the maxillary sinus volume.

[0114] The method of the embodiment of the present invention realizes automatic segmentation of the maxillary sinus from the skull CBCT image, and further realizes automatic measurement of various linear data and volumes. Figure 5 As shown, Figure 5 The middle mark (a) is the segmentation effect of the maxillary sinus in each section. Figure 5 The middle mark (b) shows the overall effect of the maxillary sinus segmented in the skull CBCT image.

[0115] Preferably, after the maxillary sinus is segmented, DICE can also be used to evaluate the various indicators of the segmentation effect. DICE is a commonly used image segmentation performance measurement method, which can be used to evaluate the similarity between two sets. Specifically, the DICE indicator quantifies the quality of the segmentation result by calculating the ratio of the intersection and union between the predicted value and the true value. The value of the DICE indicator is usually between 0-1. The higher the value, the better the performance of the segmentation result. The Jaccard indicator measures the similarity of two sets by calculating the ratio between the intersection and the union of the two sets. The value range of the Jaccard indicator is [0,1], where 0 means that the two sets have no intersection and 1 means that the two sets are exactly the same. 95HD is mainly used to measure the accuracy of boundary segmentation. The smaller its value, the higher the segmentation accuracy. ASD refers to the average surface distance of all points in the image. The smaller its value, the higher the segmentation accuracy. Among them, the evaluation results based on the above-mentioned preferred embodiment of 515 skull CBCT images are shown in the segmentation efficiency index table of the maxillary sinus automatic segmentation system:

[0116] Segmentation performance indicators of the maxillary sinus automatic segmentation system

[0117]

[0118] Furthermore, it also includes a method for obtaining gender difference parameters and age correlation parameters for each age range:

[0119] The Shapiro-Wilk test was used to test the normal distribution of the corresponding measured data of all target samples in each age range;

[0120] If all the measured data of the target samples obey the normal distribution, the independent sample t test is used to obtain the gender difference parameters of each age range, and the Pearson test is used to obtain the age correlation parameters of each age range.

[0121] Through the implementation of the above method, the embodiment of the present invention uses R4.3.1 for data analysis. R is a powerful programming language commonly used for data analysis and statistical modeling. For the gender inference part, the samples were divided into two groups according to age: those over 18 years old (162 males and 159 females) and those under 18 years old (75 males and 81 females). The Shapiro-Wilk method was used to test the normal distribution of the data and obtain descriptive indicators. Based on the above preferred embodiment of selecting 515 skull CBCT images, the normality test table of maxillary sinus three-dimensional data is as follows:

[0122] Normality test of maxillary sinus three-dimensional data

[0123]

[0124]

[0125]

[0126] According to the above normality test table of maxillary sinus three-dimensional data, all indicators are normally distributed, and in the two groups of samples under 18 years old and over 18 years old, the average value of each variable in males is greater than that in females.

[0127] Furthermore, the method for obtaining the gender difference parameters and age correlation parameters for each age range also includes:

[0128] Each measurement data is set as a statistical variable, and a ridge map is drawn for each variable by gender, and a ridge map corresponding to each variable is obtained in each age range;

[0129] Based on the ridge plot, the responses of each variable on different genders in each age range were visualized, and the independent sample t-test was used to compare the differences of each measurement data between different genders, and the data with gender differences in each age range were obtained;

[0130] The Pearson test was used to evaluate the linear correlation between each measurement data and the corresponding target sample's true age, and the age-related data in each age range were obtained.

[0131] Through the implementation of the above method, sample grouping and normal distribution test were performed on the basis of the above preferred embodiment of selecting 515 skull CBCT images. The ridge map is as follows: Figure 6 shown. Figure 6 The middle left half of the figure is a ridge plot of each variable in the sample over 18 years old, plotted separately by gender. Figure 6 The middle right half of the figure is a ridge plot of each variable in the sample under 18 years old drawn by gender. The independent sample t-test table between the genders of the maxillary sinus indicators in the two age groups and the Pearson test table between the maxillary sinus indicators and age in the two age groups are as follows:

[0132] Independent sample t-test table of maxillary sinus indicators between genders in two age groups

[0133]

[0134]

[0135] Pearson test table between maxillary sinus indicators and age in two age groups

[0136] From the above ridge plot, independent sample t-test table and Pearson test table, we can see that when a ridge plot is made for each variable to make an intuitive comparison between males and females, for sample data aged 18 and above, there is a significant difference in each variable between males and females, and males are higher than females ( Figure 5 However, for the sample data under 18 years old, this difference cannot be observed ( Figure 5 hn, p in ). The results of the independent sample t test showed that there were significant differences in all data in the group over 18 years old (p<0.01), while there was only a significant difference in the variable distance between the sexes in the group under 18 years old (p<0.01) (Table 3-4-1). There was a linear correlation between each linear data and age (p<0.01) (Table 3-4-2).

[0137] Furthermore, it also includes methods for establishing gender inference models and age inference models:

[0138] All target samples were randomly divided into training group and test group in a ratio of 8:2;

[0139] Select multiple preset regression models, input the gender difference parameters of all target samples in the training group into the multiple regression models for gender inference training, and after the model training, input the gender difference parameters of all target samples in the test group into the multiple regression models for test evaluation, judge the model fitting efficiency according to the evaluation results, and determine the random forest model as the gender inference model;

[0140] Select multiple preset regression models, input the age-related parameters of all target samples in the training group into multiple regression models for age inference training, and after model training, input the age-related parameters of all target samples in the test group into multiple regression models for test evaluation, judge the model fitting efficiency according to the evaluation results, and determine the fitting model of the random forest model and the multivariate linear regression model as the age inference model.

[0141] Through the implementation of the above method, regarding the effectiveness of the gender inference model, the effectiveness index table of different regression models is as follows:

[0142] Different regression model performance indicators

[0143]

[0144]

[0145] As shown in the performance index table of different regression models, by comparing the trained random forest with the other two commonly used classifier models on the test set, it can be seen that the random forest model has a higher accuracy rate, i.e. 83.87% (Table 3-5), so the random forest is selected as the regression model finally used in the gender inference model.

[0146] Regarding the effectiveness of the age inference model, the age inference model was fitted using maxillary sinus related data. According to the developmental characteristics of the maxillary sinus, random forest models and multivariate linear regression models were fitted for people aged 18 and above and under, respectively. For the test group aged 18 and above, the final age inference mean absolute error (MAE) was 3.91 years. For the test group under 18 years old, the inference effect was good, with a MAE of 1.39 years.

[0147] The present invention also provides a model building method, comprising:

[0148] A preset number of target samples are selected to obtain skull CBCT images of all target samples;

[0149] Select the skull CBCT images of some target samples to annotate and segment the maxillary sinus to obtain annotated images, and establish an automatic segmentation model based on the annotated images and the unannotated skull CBCT images of some target samples;

[0150] Automatically segment the maxillary sinus of all unlabeled skull CBCT images according to the automatic segmentation model to obtain maxillary sinus segmentation images, and obtain the linear data and volume of all maxillary sinus segmentation images by measurement;

[0151] The selected target samples are grouped according to the preset age ranges, and the difference test between the measured data and the real gender is performed to obtain the gender difference parameters of each age range, and the correlation test between the measured data and the real age is performed to obtain the age correlation parameters of each age range;

[0152] A gender inference model is established based on the gender difference parameters of each age range, and an age inference model is established based on the age correlation parameters of each age range.

[0153] The present invention also provides an inference system, comprising:

[0154] A sample screening module is used to select a preset number of target samples and obtain skull CBCT images of all target samples;

[0155] An automatic segmentation model building module is used to select skull CBCT images of some target samples to mark and segment the maxillary sinus, obtain a marked image, and build an automatic segmentation model based on the marked image and unmarked skull CBCT images of some target samples;

[0156] An image segmentation and measurement module is used to automatically segment the maxillary sinus of all unlabeled skull CBCT images according to the automatic segmentation model to obtain maxillary sinus segmentation images, and obtain various linear data and volumes of all maxillary sinus segmentation images by measurement;

[0157] The data verification module is used to group the selected target samples according to the preset age ranges, perform difference tests between the measured data and the real gender, and obtain the gender difference parameters of each age range, and perform correlation tests between the measured data and the real age, and obtain the age correlation parameters of each age range;

[0158] An inference model building module is used to build a gender inference model based on the gender difference parameters of each age interval, and to build an age inference model based on the age correlation parameters of each age interval;

[0159] The gender and age inference module is used to input the skull CBCT image to be detected into the automatic segmentation model to obtain the linear data and volume of the maxillary sinus of the skull CBCT image to be detected, input the various measurement data of the maxillary sinus into the gender inference model and the age inference model for inference, and output the gender and age corresponding to the skull CBCT image to be detected respectively.

[0160] As described above, the system established by the present invention can enter the import interface by clicking upload CBCT on the initial interface. After uploading the CBCT file of the corresponding format, the system can automatically perform data measurement and gender and age inference to obtain corresponding results.

Claims

1. A method for inferring gender and age based on maxillary sinus, characterized in that: The method for inferring gender and age based on the maxillary sinus comprises: A preset number of target samples are selected to obtain skull CBCT images of all target samples; Select the skull CBCT images of some target samples to annotate and segment the maxillary sinus to obtain annotated images, and establish an automatic segmentation model based on the annotated images and the unannotated skull CBCT images of some target samples; Automatically segment the maxillary sinus of all unlabeled skull CBCT images according to the automatic segmentation model to obtain maxillary sinus segmentation images, and obtain the linear data and volume of all maxillary sinus segmentation images by measurement; The selected target samples are grouped according to the preset age ranges, and the difference test between the measured data and the real gender is performed to obtain the gender difference parameters of each age range, and the correlation test between the measured data and the real age is performed to obtain the age correlation parameters of each age range; A gender inference model is established based on the gender difference parameters of each age interval, and an age inference model is established based on the age correlation parameters of each age interval; The CBCT image of the skull to be tested is input into the automatic segmentation model to obtain the linear data and volume of the maxillary sinus of the CBCT image of the skull to be tested, and the various measurement data are respectively input into the gender inference model and the age inference model for inference, and the gender and age corresponding to the CBCT image of the skull to be tested are output respectively.

2. The method for inferring gender and age based on maxillary sinus according to claim 1, characterized in that: The method for selecting a portion of the target sample's skull CBCT images to label and segment the maxillary sinus includes: The skull CBCT images of all target samples were imported into ITK-SNAP software; The skull CBCT images of some selected target samples were retrieved one by one, and the range of the maxillary sinus was manually marked using the ROI tool in each slice where the maxillary sinus existed in the retrieved skull CBCT images; According to the maxillary sinus range marked on the skull CBCT image, select the segmentation tool to perform maxillary sinus segmentation on the retrieved skull CBCT image to obtain a marked image containing only the maxillary sinus; The annotated images and other unannotated skull CBCT images are stored in nii.gz format for future use.

3. The method for inferring gender and age based on maxillary sinus according to claim 1, characterized in that: The method for establishing an automatic segmentation model based on the labeled image and the unlabeled skull CBCT image of a part of the target sample comprises: The VNet architecture based on 3D convolution and 3D deconvolution is selected as the basic model, and the Student training model and Teacher auxiliary model are established. The UNet architecture with cross-layer links is added between the Encoder part of the Student training model and the Decoder part of the Teacher auxiliary model. Extracting labeled data from the labeled images, and using the labeled data to train the Student training model and the Teacher auxiliary model respectively, the Teacher auxiliary model assists the Student training model to learn the labeling features and segmentation patterns of the maxillary sinus based on the labeled data; Supervised data is extracted from unlabeled skull CBCT images, and the supervised data is input into the Student training model and the Teacher auxiliary model respectively. The Teacher auxiliary model assists the Student training model in self-training of maxillary sinus labeling and segmentation based on the supervised data. Based on the co-training of multiple networks, an automatic segmentation model is established through mutual training between the Student training model and the Teacher auxiliary model.

4. The method for inferring gender and age based on maxillary sinus according to claim 3, characterized in that: It also includes methods for mutual training between the Student training model and the Teacher auxiliary model: The labeled data are input into the Student training model and the Teacher auxiliary model respectively. The Student training model learns the labeling features and segmentation mode of the maxillary sinus through back propagation. The Teacher auxiliary model is updated by using the average EMA of the labeled data and the Student model to obtain a pseudo label with a certain resolution by adding a fixed parameter VNet based on the MT architecture that is not updated with back propagation to the Teacher auxiliary model. The supervised data are input into the Student training model and the Teacher auxiliary model respectively. The Student training model performs self-training of maxillary sinus annotation and segmentation based on the supervised data. The Teacher auxiliary model imposes conditional constraints on the self-training of the Student training model and performs consistency detection with the Student training model by updating parameters through EMA average. The pseudo labels of the Teacher auxiliary model are used as new supervised data and input into the Student training model to guide the Student training model to learn and train.

5. The method for inferring gender and age based on maxillary sinus according to claim 1, characterized in that: Also included is a method for processing the labeled images and the unlabeled skull CBCT images before establishing the automatic segmentation model: Integrate the annotated images and the unannotated skull CBCT images into CBCT image data in a data format, and delete the CBCT image data that does not meet a preset data volume size; For all CBCT image data after the deletion process, the CBCT image data with a pixel gray value range of -1000-2000 is converted into CBCT image data with a pixel gray value range of 0-255, and all the CBCT image data are normalized to obtain CBCT image data in int8 format; After normalization, all CBCT image data are cropped to obtain CBCT image data containing only maxillary sinus images; The cropped CBCT image data is saved as an h5 file.

6. The method for inferring gender and age based on maxillary sinus according to claim 1, characterized in that: The method for obtaining the linear data and volume of all maxillary sinus segmentation images by measuring includes: The skull CBCT images of some selected target samples are annotated one by one, and the range of the maxillary sinus is annotated in each section of the skull CBCT image where the maxillary sinus exists; According to the maxillary sinus segmentation image obtained by the automatic segmentation model, in each maxillary sinus segmentation image, the linear length, linear width, and linear distance between the two lateral walls of the maxillary sinus are obtained by measuring the continuous cross-sections of each section, and the linear height of the maxillary sinus is obtained by measuring the continuous coronal planes of each section; The data obtained from multiple tomographic measurements in each maxillary sinus segmentation image are merged to obtain the volume data of the corresponding maxillary sinus.

7. The method for inferring gender and age based on maxillary sinus according to claim 1, characterized in that: It also includes methods for obtaining gender difference parameters and age-related parameters for each age range: The Shapiro-Wilk test was used to test the normal distribution of the corresponding measured data of all target samples in each age range; If all the measured data of the target samples obey the normal distribution, set the measured data as statistical variables, draw ridge diagrams for each variable by gender, and obtain ridge diagrams corresponding to each variable in each age range; Based on the ridge plot, the responses of each variable on different genders in each age range were visualized, and the independent sample t-test was used to compare the differences of each measurement data between different genders, and the parameter data with gender differences in each age range were obtained; The Pearson test was used to evaluate the linear correlation between each measurement data and the corresponding target sample's true age, and the age-related parameter data in each age range were obtained.

8. The method for inferring gender and age based on maxillary sinus according to claim 1, characterized in that: It also includes methods for establishing gender inference models and age inference models: All target samples were randomly divided into training group and test group in a ratio of 8:2; Select multiple preset regression models, input the gender difference parameters of all target samples in the training group into the multiple regression models for gender inference training, and after the model training, input the gender difference parameters of all target samples in the test group into the multiple regression models for test evaluation, judge the model fitting efficiency according to the evaluation results, and determine the random forest model as the gender inference model; Select multiple preset regression models, input the age-related parameters of all target samples in the training group into multiple regression models for age inference training, and after model training, input the age-related parameters of all target samples in the test group into multiple regression models for test evaluation, judge the model fitting efficiency according to the evaluation results, and determine the fitting model of the random forest model and the multivariate linear regression model as the age inference model.

9. A model building method, characterized in that: The model building method comprises: A preset number of target samples are selected to obtain skull CBCT images of all target samples; Select the skull CBCT images of some target samples to annotate and segment the maxillary sinus to obtain annotated images, and establish an automatic segmentation model based on the annotated images and the unannotated skull CBCT images of some target samples; Automatically segment the maxillary sinus of all unlabeled skull CBCT images according to the automatic segmentation model to obtain maxillary sinus segmentation images, and obtain the linear data and volume of all maxillary sinus segmentation images by measurement; The selected target samples are grouped according to the preset age ranges, and the difference test between the measured data and the real gender is performed to obtain the gender difference parameters of each age range, and the correlation test between the measured data and the real age is performed to obtain the age correlation parameters of each age range; A gender inference model is established based on the gender difference parameters of each age range, and an age inference model is established based on the age correlation parameters of each age range.

10. An inference system, characterized in that The inference system comprises: A sample screening module is used to select a preset number of target samples and obtain skull CBCT images of all target samples; An automatic segmentation model building module is used to select skull CBCT images of some target samples to mark and segment the maxillary sinus, obtain a marked image, and build an automatic segmentation model based on the marked image and unmarked skull CBCT images of some target samples; An image segmentation and measurement module is used to automatically segment the maxillary sinus of all unlabeled skull CBCT images according to the automatic segmentation model to obtain maxillary sinus segmentation images, and obtain various linear data and volumes of all maxillary sinus segmentation images by measurement; The data verification module is used to group the selected target samples according to the preset age ranges, perform difference tests between the measured data and the real gender, and obtain the gender difference parameters of each age range, and perform correlation tests between the measured data and the real age, and obtain the age correlation parameters of each age range; An inference model building module is used to build a gender inference model based on the gender difference parameters of each age interval, and to build an age inference model based on the age correlation parameters of each age interval; The gender and age inference module is used to input the skull CBCT image to be detected into the automatic segmentation model to obtain the linear data and volume of the maxillary sinus of the skull CBCT image to be detected, input the various measurement data of the maxillary sinus into the gender inference model and the age inference model for inference, and output the gender and age corresponding to the skull CBCT image to be detected respectively.

Images