Upper airway segmentation and airflow simulation method based on CBCT image

Abstract

The invention provides an upper airway segmentation and airflow simulation method based on a CBCT image. The method comprises the following steps: constructing an initial three-dimensional model; obtaining original image data and initial three-dimensional model data of the CBCT to be detected; processing the initial three-dimensional model data to obtain model parameter data; performing parameter setting on the initial three-dimensional model according to the model parameter data to obtain an airway three-dimensional model; and inputting the original image data of the CBCT to be detected into the airway three-dimensional model to obtain a speed cloud picture and a stress cloud picture. The problems that in the prior art, the imaging accuracy of the upper airway is low, and the overall change of the airway in the full-breathing process cannot be simulated are solved.

Description

Technical Field

[0001] The invention relates to the technical field of upper airway dynamic airflow field simulation, and in particular to an upper airway segmentation and airflow simulation method based on CBCT images. Background Art

[0002] Orthodontic treatment is often accompanied by changes in the airway, especially airway stenosis in children and adolescents. Orthodontic treatment can timely block subsequent adverse progression. Adenoids are lymphatic tissue located in the upper back of the nasopharynx. According to the development of the human immune system, adenoids proliferate rapidly when children are 2 to 6 years old.

[0003] After the age of 10, they gradually shrink and basically disappear in adulthood. Due to repeated stimulation from environmental factors, adenoids are prone to hyperplasia in childhood, which in turn causes diseases such as adenoid hypertrophy and obstructive sleep apnea hypopnea syndrome (OASHS). In my country, the prevalence of OSAHS in people under 14 years old is as high as 4%, and the incidence of adenoid hypertrophy in children and adolescents is as high as 49.7%. In children and adolescents, these two diseases have a great impact on physical development. If not treated in time, they may cause a series of problems such as adenoid facies, slow development of the nervous system, abnormal metabolic capacity, and poor sleep quality. Adenoid hypertrophy and OSAHS and other related diseases are currently mainly treated by stomatology and otolaryngology. In the clinical practice of stomatology, lateral cephalograms and CBCT images are often used to obtain airway images of patients to assist in diagnosis. Compared with laryngoscopy in otolaryngology, imaging methods are more comfortable and easier for patients to accept.

[0004] However, the lateral cephalogram is a two-dimensional plane image that cannot accurately obtain airway information. Therefore, the three-dimensional image of CBCT is more important. Due to the fact that CBCT still has large image noise, low contrast, large amount of CBCT image data, complex airway anatomical structure variation, unclear boundary between airway soft tissue and airway cavity, and differences in the level of physician reading, it takes a long time to read CBCT images. Deep learning technology has been widely used in the field of image segmentation. At present, deep learning methods used for CBCT airway segmentation include CNN, 3DU-Net and RegressionNeuralNetwork. It is reported that the fastest segmentation speed can reach 6 minutes, and the highest DICE coefficient (DSC) reaches 0.92, which still has a lot of room for improvement.

[0005] With the rapid development of science and technology, especially the advancement of technology in the engineering field, CFD (Computational FluidDynamics, CFD) numerical simulation technology has been applied in scenarios such as blood flow and heart beating. Conventional medical examinations can only analyze the results of a certain moment, and the impact on the inside of the airway in different parts of the airway, in different environments, and under different breathing and intake conditions cannot be obtained.

[0006] In the existing airway flow simulation research, most of them use no-slip wall boundary conditions for the upper respiratory tract wall, and assume that the wall is a rigid wall, without considering the influence of wall elasticity. Only a few studies use flexible walls in the airway to study the local changes of the airway during breathing. However, none of them can simulate the overall changes of the airway during the whole breathing process. Summary of the invention

[0007] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide an upper airway segmentation and airflow simulation method based on CBCT images, which solves the problems in the prior art that the upper airway imaging has low accuracy and cannot simulate the overall changes of the airway during the whole breathing process.

[0008] To achieve the above object, the present invention provides the following solutions:

[0009] An upper airway segmentation and airflow simulation method based on CBCT images, comprising:

[0010] Constructing an initial 3D model;

[0011] Acquire the CBCT original image data and initial three-dimensional model data to be tested;

[0012] Processing the initial three-dimensional model data to obtain model parameter data;

[0013] Setting parameters of the initial three-dimensional model according to the model parameter data to obtain an airway three-dimensional model;

[0014] Inputting the CBCT original image data to be tested into the airway three-dimensional model to obtain a velocity cloud map and a stress cloud map;

[0015] The method for constructing the initial three-dimensional model is:

[0016] Obtaining CBCT raw image data;

[0017] Preprocessing the CBCT raw image data to obtain image input data;

[0018] According to the image input data, a first sub-model is constructed based on the 3DUX-Net network;

[0019] Optimizing the first sub-model to obtain a second sub-model;

[0020] Performing inverse modeling and smoothing processing on the second sub-model to obtain a third sub-model;

[0021] The sub-model is meshed to obtain an initial three-dimensional model.

[0022] Preferably, the initial three-dimensional model data includes:

[0023] The flow parameters and temperature parameters of airway intake, the distance parameters of airway wall movement during breathing, and the temperature parameters and flow parameters during exhalation.

[0024] Preferably, the preprocessing of the CBCT raw image data to obtain image input data includes:

[0025] Manually segmenting the airway on the CBCT original image data to obtain first image data;

[0026] Performing bilateral filtering on the first image data to obtain second image data;

[0027] Performing weighted summation on eight regions of each point in the second image to obtain third image data;

[0028] The third image data is resized and data set enhanced to obtain image input data.

[0029] Preferably, the learning rate of the 3DUX-Net network is 0.0001, the batch size is set to 1, and the training step size is set to 60,000 steps.

[0030] Preferably, optimizing the first sub-model to obtain the second sub-model includes:

[0031] Obtain a first segmented image output by the first sub-model;

[0032] Determine whether the first segmented image has over-division or omission. If so, determine whether the pixel threshold of the first segmented image is greater than -550. If so, remove the corresponding part to obtain a second segmented image. If omission exists, determine whether the pixel position of the omission part is within the eight neighborhoods of the segmented foreground pixels. If so, classify the corresponding part as the foreground area to obtain a third segmented image.

[0033] Based on the first sub-model, a second sub-model is obtained according to the second segmented image and the third segmented image.

[0034] Preferably, performing inverse modeling and smoothing on the second sub-model to obtain a third sub-model includes:

[0035] Importing the second sub-model file into Geomagicwrap, using a mesh doctor to remove particles and non-structures on the surface of the second sub-model, and smoothing or deleting and reconstructing the irregular plane of the second sub-model after removing particles and non-structures, to obtain a smoothed second sub-model;

[0036] Dividing the obtained smoothed second sub-model into curved surface patches to obtain a regular curved surface patch distribution model, converting the regular curved surface patch distribution model into a solid model and exporting it;

[0037] Import the solid model into SpacesClaim, fit or reconstruct the surface patches at the air inlet and exhaust port to obtain the third sub-model.

[0038] Preferably, the gridding of the sub-model to obtain an initial three-dimensional model comprises:

[0039] The third sub-model is divided into boundaries, named, and the inlet and outlet walls are pointed out. The corresponding fluid domain is set, and a boundary layer grid is set at the boundary with a growth rate of 1.2 and a minimum unit size of 2 mm to obtain an initial three-dimensional model.

[0040] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0041] The present invention provides an upper airway segmentation and airflow simulation method based on CBCT images, including: constructing an initial three-dimensional model; obtaining CBCT original image data to be tested and initial three-dimensional model data; processing the initial three-dimensional model data to obtain model parameter data; setting parameters of the initial three-dimensional model according to the model parameter data to obtain an airway three-dimensional model; inputting the CBCT original image data to be tested into the airway three-dimensional model to obtain a velocity cloud map and a stress cloud map; the construction method of the initial three-dimensional model is: obtaining CBCT original image data; preprocessing the CBCT original image data to obtain image input data; constructing a first sub-model based on the 3DUX-Net network according to the image input data; optimizing the first sub-model to obtain a second sub-model; performing reverse modeling and smoothing processing on the second sub-model to obtain a third sub-model; and meshing the sub-model to obtain an initial three-dimensional model. The present invention provides that CBCT is a rapid imaging technology for displaying anatomical structures such as jaws, teeth, and temporomandibular joints in the oral and maxillofacial region. In recent years, it has been widely used in oral implantation, root canal treatment, orthodontic diagnosis, and other aspects. However, the data contained in CBCT images is huge, and making a clear diagnosis based on imaging manifestations requires the doctor's technical sensitivity. The present invention uses a deep learning algorithm to automatically segment the upper airway part in CBCT images based on the UX-Net neural network for three-dimensional medical image segmentation, and simulates the dynamic respiratory airflow field, constructing an efficient, convenient and accurate upper airway three-dimensional model imaging method, which provides a new method for orthodontists to make accurate and comprehensive preoperative diagnosis, carry out orthodontic treatment and orthodontic prognosis evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0043] Figure 1 A flow chart of an upper airway segmentation and airflow simulation method based on CBCT images provided in an embodiment of the present invention;

[0044] Figure 2 A schematic diagram of the 3DUX-Net network structure provided by an embodiment of the present invention;

[0045] Figure 3 A gas flow rate diagram of a breathing intake cycle provided by an embodiment of the present invention;

[0046] Figure 4A velocity vector diagram of an airway internal cross section provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] The purpose of the present invention is to provide an upper airway segmentation and airflow simulation method based on CBCT images, which solves the problems in the prior art that upper airway imaging has low accuracy and cannot simulate the overall changes of the airway during the whole breathing process.

[0049] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0050] like Figure 1 As shown, the present invention provides an upper airway segmentation and airflow simulation method based on CBCT images, comprising:

[0051] Step 100: constructing an initial three-dimensional model;

[0052] Step 200: Acquire the CBCT original image data and initial three-dimensional model data to be tested;

[0053] Step 300: Processing the initial three-dimensional model data to obtain model parameter data;

[0054] Step 400: setting parameters of the initial three-dimensional model according to the model parameter data to obtain an airway three-dimensional model;

[0055] Step 500: inputting the CBCT original image data to be tested into the airway three-dimensional model to obtain a velocity cloud map and a stress cloud map;

[0056] The method for constructing the initial three-dimensional model is:

[0057] Step 101: Acquire CBCT original image data;

[0058] Specifically, images from a certain image database were selected. Based on the medical digital imaging data acquisition and transmission protocol, CBCT data files of the upper airway in the DICOM3.0 standard format of different patients visiting the dental department were obtained. The CS9300C3D intelligent CBCT machine (Carestream Dental, USA) was used, with a tube voltage of 60-90 kV, a tube current of 2-15 mA, a frequency of 140 kHz, a tube focal spot of 0.7 mm (with X-ray tube OPX110) or 0.6 mm (with X-ray tube D-067), a total filter of >2.5 mm aluminum equivalent and 0.1 mm copper equivalent, a TFT sensor with a resolution of 64×1536 pixels, an image exposure time of 4-15 seconds, and a voxel size of 90-500 μm. Patients were required to take a standardized upright position and occlude their anterior teeth, maintain calm breathing, ensure that the orbitoauricular plane is parallel to the ground plane, and the midsagittal plane of the head is perpendicular to the ground plane, and keep the tongue and lips still when exhaling. The scanning site was determined, and the patient was instructed not to swallow during the scan, and a 360° full-scale scan was performed. The inclusion criteria for image acquisition in this study were: CBCT scan images had no obvious motion or metal artifacts, and contained a complete upper airway (the plane from the top of the nasopharynx to the lower edge of the fourth cervical vertebra). CBCT images with obvious skeletal craniofacial deformities, metal dental fillings or implants, and metal orthodontic devices were excluded.

[0059] Step 102: preprocessing the CBCT raw image data to obtain image input data;

[0060] Specifically, the CBCT images were imported into ITK-SNAP3.2 (Pennsylvania, USA) software, the upper airway boundary was identified, and the upper airway nasopharynx, velopharynx, oropharynx, and laryngopharynx were manually segmented and marked with different colors; after all regions were segmented, the obtained data were exported as a whole into an independent Neuroimaging Informatics Technology Initiative (NIFTI) format file.

[0061] Specifically, the nasopharyngeal region (NPV): the section where the roof of the nasopharyngeal cavity is located is the upper boundary, where the posterior pharyngeal wall and the posterior end of the nasal septum fuse; the anterior boundary of the nasopharynx is the section through the PNS perpendicular to the orbitoauricular plane; the posterior and lateral boundaries are the pharyngeal wall; the lower boundary is the HP plane. The velopharyngeal region (VPV): the HP plane is the upper boundary, the SP plane is the lower boundary, the soft palate is the anterior boundary, and the pharyngeal wall is the posterior and lateral boundaries. The glossopharyngeal region (GPV): the SP plane to the Et plane are the upper and lower boundaries respectively, the anterior boundary is the base of the tongue, and the posterior and lateral boundaries are the pharyngeal wall. The hypopharyngeal region (HPV): the area below the Et plane to the lower edge of the fourth cervical vertebra (C4), the anterior boundary is the epiglottis and the anterior wall of the hypoglottic pharynx, and the posterior and lateral boundaries are the pharyngeal wall.

[0062] The original CBCT image is bilaterally filtered to remove high-frequency noise. By processing the spatial position and pixel value of the regional pixel, the category of the pixel is given a relatively high weight. Finally, a weighted sum of the eight domains of the point is performed, and this operation is performed on the entire image. All data sets are uniformly scaled to a size of 256*256 and data enhancement is performed on the existing data sets.

[0063] Step 103: constructing a first sub-model based on the 3DUX-Net network according to the image input data;

[0064] Specifically, the preprocessed data set is put into the 3DUX-Net network for training, the learning rate is set to 0.0001, and the batch size is set to 1. The training step length is set to 60,000 steps, and a verification is performed every 300 steps to see whether the learning weight of the current network has reached the historical optimal. Figure 2 shown.

[0065] Specifically, an image of size H*W*D*C is extracted as the encoder input, and features are calculated using a convolution kernel of size 7*7*7. The encoding layer consists of four modules, each of which contains two convolution kernels of size 7*7*7. Downsampling is performed between layers using a standard convolution kernel of size 2*2*2 and stride 2. The decoding layer consists of four modules, each of which contains two convolution kernels of size 3*3*3. The processed features of each stage are upsampled through the transposed convolution layer and merged with the features of the previous stage. Among them, Layer norm: layer normalization; DWC: deep convolution layer; DCS: deep compression sampling layer; Large Kernel Projection: large convolution kernel layer; 3D UX-Netblock: 3D UX-Net block; Downsamplingblock: downsampling block; Res-Block: residual block.

[0066] Step 104: Optimize the first sub-model to obtain a second sub-model;

[0067] Specifically, the image obtained by neural network segmentation is subjected to threshold segmentation judgment to remove the segmentation that is obviously not part of the pharynx. The segmented foreground part is subjected to eight-neighborhood threshold judgment to segment the missed foreground. For the segmentation of multiple pharyngeal categories at the interface, the number of pixels in each category is calculated and compared, and the largest number is taken as the interface, such as Figure 3 shown.

[0068] Step 105: performing inverse modeling and smoothing processing on the second sub-model to obtain a third sub-model;

[0069] Step 106: Meshing the sub-model to obtain an initial three-dimensional model.

[0070] Specifically, the method for simulating the internal gas of the airway based on ANSYS software has the following steps: collecting airway model data, i.e. extracting the geometric shape of the airway, inlet and outlet setting parameters, and basic dimensions of the airway through CBCT;

[0071] Airway data processing, collecting the gas flow, gas temperature, migration distance of the airway wall and the range of change of the airway's internal ambient temperature during the intake and exhalation stages; airway geometry model, converting the STL type airway initial model into an operational solid model; airway meshing, meshing the airway solid model to improve the mesh quality.

[0072] Numerical calculation, set the inlet and outlet gas parameters, wall boundary parameters, operating conditions, coupling method, solution control parameters, start solving; confirm convergence; result analysis, that is, import post-processing software CFD-POST, obtain the velocity change and stress change at the wall at different positions such as Figure 4 shown.

[0073] Furthermore, the initial three-dimensional model data includes:

[0074] The flow parameters and temperature parameters of airway intake, the distance parameters of airway wall movement during breathing, and the temperature parameters and flow parameters during exhalation.

[0075] Furthermore, the preprocessing of the CBCT raw image data to obtain image input data includes:

[0076] Manually segmenting the airway on the CBCT original image data to obtain first image data;

[0077] Performing bilateral filtering on the first image data to obtain second image data;

[0078] Performing weighted summation on eight regions of each point in the second image to obtain third image data;

[0079] The third image data is resized and data set enhanced to obtain image input data.

[0080] Specifically, the CBCT image data files of the upper airway in the DICOM3.0 standard format were imported into the ITK-SNAP3.2 (Pennsylvania, USA) tool in the format of an image series (DICOM Image Series). The three-dimensional coordinate system was adjusted to the mid-sagittal plane of the skull with reference to the cross section of the first quadrant in the software, and the reference plane for airway measurement was determined in the sagittal plane image of the second quadrant. The lowest point of the infraorbital margin and the upper edge of the external auditory foramen were connected to form the orbitoacoustic plane (Frankfurt horizontal plane, FH plane); the line connecting the anterior nasal spine (ANS) and the posterior nasalspine (PNS) determined the hard palate (HP) plane; the section through the lowest point of the uvula (SP) (i.e., the lowest point of the soft palate) parallel to the FH plane was the SP plane; the section through the apex of the epiglottis (Et) and parallel to the FH plane was the Et plane. The present invention divides the upper airway into four parts according to the reference plane. Nasopharyngeal area: The section where the roof of the nasopharyngeal cavity is located is the upper boundary, where the posterior pharyngeal wall and the posterior end of the nasal septum merge; the anterior boundary of the nasopharynx is the section through the PNS perpendicular to the orbitoauricular plane; the posterior and lateral boundaries are the pharyngeal wall; the lower boundary is the HP plane. Velopharyngeal area: The HP plane is the upper boundary, the SP plane is the lower boundary, the soft palate is the anterior boundary, and the pharyngeal wall is the posterior and lateral boundaries. Glossopharyngeal area: The SP plane to the Et plane are the upper and lower boundaries respectively, the anterior boundary is the base of the tongue, and the posterior and lateral boundaries are the pharyngeal wall. Laryngopharyngeal area: The area below the Et plane to the lower edge of the fourth cervical vertebra (the fourth Cervical vertebra, C4), the anterior boundary is the epiglottis and the anterior wall of the subepiglottic pharynx, and the posterior and lateral boundaries are the pharyngeal wall. The CBCT images in the DICOM3.0 standard format obtained by the present invention consist of 441 slices along the Z axis (axial direction). Due to the different heights of the upper airway, the number of airway slices contained in the upper airway images of each patient varies, ranging from 179 to 346. The staff uses the ITK-SNAP3.2 (Pennsylvania, USA) tool to mark the airway areas layer by layer along the axial direction, namely the nasopharyngeal, palatopharyngeal, glossopharyngeal, and laryngopharyngeal areas specified above. After the slices of all areas are marked, the obtained data is exported as a whole into an independent NIFTI format file, which can truly reflect multidimensional metadata, contains spatial direction information, and is convenient for storing 3D data.

[0081] There is a lot of noise in the original CBCT image, which will affect the network learning performance. Therefore, bilateral filtering is performed on the existing data set. Considering the spatial position and pixel value of the pixel point in the processing area, a relatively high weight is given to the category divided by the pixel point. Finally, a weighted sum is performed on the eight fields of the point, and this operation is performed on the entire image. The image is filtered by spatial information, and the edge features of the image are retained by the information of the pixel value. The cross-sectional image of the upper airway three-dimensional model after CBCT image preprocessing has been shown above. The image resolution of the original data set is between 500-600. Most neural networks require that the image size input to the network be consistent. The image is uniformly resized to 256*256 to facilitate the downsampling of the neural network while reducing the resolution and reducing the training time. The data set is expanded to 20 times the original size by rotating, randomly flipping, randomly cropping, grayscale transformation, linear transformation, nonlinear transformation, adding Gaussian noise, etc. This makes the network have better generalization performance.

[0082] Furthermore, the learning rate of the 3DUX-Net network is 0.0001, the batch size is set to 1, and the training step size is set to 60,000 steps.

[0083] Specifically, in the present invention, a 3DUX-Net network is used for airway segmentation learning. 3DUXNet is a lightweight 3D convolutional neural network that uses a ConvNet module to adjust a hierarchical Transformer to achieve robust voxel segmentation. The network uses a large-scale 7*7 convolution kernel, which can provide a large receptive field when extracting features similar to the MSA (Multi-head Self-Attention) block. DCS (Deep Compressive Sensing) is introduced to linearly scale each channel feature independently, minimize the redundant information generated by cross-channel context, and reduce model complexity. BN (Batch Normalization) will have an adverse effect on the generalization ability of the model. BN is replaced by LN (Layer Normalization), which does not rely on mini-batch. Different features of the same sample have the same mean and variance. GELU (Gaussian Error Linear Unit) is a linear transformation unit based on Gaussian error, which is smoother than ReLU (Rectified Linear Unit) and solves the problem of neuron inactivation caused by hard truncation of negative gradients in ReLU. The input part uses a large convolution kernel projection layer to map the original image to a low-dimensional spatial resolution latent spatial feature representation, reducing the amount of memory parameter calculation and increasing the network receptive field. The network was trained on NVIDIA GeForce RTX 3080 Ti, with a training learning rate of 0.0001 and a batch size of 1. The training step size was set to 60,000 steps, and verification was performed every 300 steps. The segmentation capability of the network has reached the top level in China. Table 1 is a table of 3DUX-Net network segmentation capability parameters, as shown below:

[0084] Table 13 DUX-Net network segmentation capability parameter table

[0085]

[0086] Furthermore, the optimizing the first sub-model to obtain the second sub-model includes:

[0087] Obtain a first segmented image output by the first sub-model;

[0088] Determine whether the first segmented image has over-division or omission. If so, determine whether the pixel threshold of the first segmented image is greater than -550. If so, remove the corresponding part to obtain a second segmented image. If omission exists, determine whether the pixel position of the omission part is within the eight neighborhoods of the segmented foreground pixels. If so, classify the corresponding part as the foreground area to obtain a third segmented image.

[0089] Based on the first sub-model, a second sub-model is obtained according to the second segmented image and the third segmented image.

[0090] Specifically, the images predicted by the network will have obvious over-segmentation and partial omission, and there will be misclassification at the interfaces of different sub-pharyngeal parts. For over-segmentation, it is necessary to determine whether the pixel threshold of the corresponding CT image is greater than -550. If it is greater, it means that this pixel should not be the target area and should be removed. For unsegmented pixels, if they are within the eight neighborhoods of the segmented foreground pixels, the pixel is classified as the foreground area. For the interface part, the method of pixel number statistics is used to determine the attribution of the interface. By counting the number of pixels of different categories of pharyngeal parts of this surface and comparing the sizes. The network predicts that this interface is more inclined to the category that occupies more.

[0091] Furthermore, the second sub-model is subjected to inverse modeling and smoothing processing to obtain a third sub-model, including:

[0092] Importing the second sub-model file into Geomagicwrap, using a mesh doctor to remove particles and non-structures on the surface of the second sub-model, and smoothing or deleting and reconstructing the irregular plane of the second sub-model after removing particles and non-structures, to obtain a smoothed second sub-model;

[0093] Dividing the obtained smoothed second sub-model into curved surface patches to obtain a regular curved surface patch distribution model, converting the regular curved surface patch distribution model into a solid model and exporting it;

[0094] Import the solid model into SpacesClaim, fit or reconstruct the surface patches at the air inlet and exhaust port to obtain the third sub-model.

[0095] Specifically, the present invention uses GeomagicWrap and SpaceClaim for reverse modeling and smoothing. The first step is to import the obtained STL model file into GeomagicWrap, use the mesh doctor to remove the particles and non-structures on the surface, smooth or delete and reconstruct the irregular plane, divide the obtained smoothed model into surface patches to obtain a regular distribution of surface patches, convert it into a solid model and export it; the second step is to import the solid model into SpaceClaim, fit or reconstruct the surface patches at the air inlet and exhaust port to obtain a complete boundary of the air inlet and exhaust port.

[0096] Furthermore, the gridding of the sub-model to obtain an initial three-dimensional model includes:

[0097] The third sub-model is divided into boundaries, named, and the inlet and outlet walls are pointed out. The corresponding fluid domain is set, and a boundary layer grid is set at the boundary with a growth rate of 1.2 and a minimum unit size of 2 mm to obtain an initial three-dimensional model.

[0098] Specifically, the boundaries of the solid model are divided and named, the inlet and outlet walls are clearly pointed out, the corresponding fluid domain is set, the dynamic mesh technology of the boundary wall is involved, the boundary layer mesh is set at the boundary, the growth rate is 1.2, the minimum unit size is 2mm, the calculation domain is mainly hexahedral elements, and the total number of meshes ranges from 700,000 to 900,000.

[0099] Furthermore, data collection mainly involves the retrieval of literature and field experimental results. For airway intake simulation, relevant Chinese and English literature is searched on platforms such as CNKI and Web of Science. Parameters such as airway intake flow, temperature, and airway wall movement distance during breathing are collected, and parameters such as temperature and flow in the exhalation state are obtained through field experimental data. After the airway intake and exhalation data are collected, the intake and exhalation flow data are fitted into sinusoidal parameters to be close to breathing in the real state. As time changes, the maximum peak values ​​of intake and exhalation flow are 200ml / s, 400ml / s, and 600ml / s, respectively. The equation is expressed as, k = 200ml / s, 400ml / s, 600ml / s corresponding to the speed under the flow, t represents time; according to the airway in vivo ambient temperature, the airway wall is set to 33℃, and the airway wall is 0.005mm in the normal direction of the plane during breathing, and UDF is used to write displacement parameters.

[0100] Furthermore, the mesh file output by mesh is read in FLUENT software, and the mesh is checked to confirm that there is no negative volume. In FLUENT, the unit and size of the fluid domain are handled to be consistent with the set parameter units. The airflow and temperature in the airway are described by the laws of conservation of mass, momentum and energy, where the flow is assumed to be a three-dimensional steady incompressible turbulent flow. Turbulence is modeled using a turbulence model. In order to solve the equations, the following boundary conditions are specified: at the inlet end of the airway, a uniform velocity and temperature field is provided, and a turbulence intensity of 10% is used to determine the turbulent kinetic energy k, the turbulent dissipation rate e and the velocity variance. At the outlet end of the airway, the outlet pressure is assumed to be zero gauge pressure. In the exhalation state, the inlet and outlet boundaries are swapped to simulate the process of gas being discharged from the body, and the control equations are solved by coupling pressure and velocity using a separation scheme and the SIMPLE algorithm. Regarding discretization, a second-order upwind scheme is used to calculate nonlinear terms and a second-order central difference scheme is used to calculate viscous terms.

[0101] The simulation results were imported into CFD-POST, and planes were constructed in the direction perpendicular to the intake velocity. The distances from the intake end were 0.1m, 0.12m, 0.16m, 0.19m, and 0.24m, respectively. Velocity cloud maps and stress cloud maps were drawn based on the constructed planes, and the changes under different intake and exhalation states were compared. This was applied to the postoperative airway model, and the plane divisions were made at the same distance. The velocity and stress cloud maps were drawn at the same time. The changes before and after the operation were compared to obtain the postoperative correction results.

[0102] The beneficial effects of the present invention are as follows:

[0103] (1) The present invention establishes a high-precision three-dimensional upper airway model with a respiratory airflow field. The model is completed by automatically extracting CBCT images and gridding the simulated airflow field distribution through deep learning technology.

[0104] (2) The orthodontic process is often accompanied by changes in the upper airway. Patients with upper airway stenosis caused by enlarged adenoids and tonsils have a poor orthodontic prognosis and high requirements for orthodontic technology and plans. The upper airway airflow visualization simulation of the present invention can more comprehensively display the patient's airway problems, making it easier for doctors to conduct comprehensive and accurate pre-orthodontic diagnosis and formulate orthodontic plans.

[0105] (3) The visual simulation of the upper airway and airflow field involved in this method fills the gap in the current digital orthodontic technology in the direction of soft tissue simulation. The automatic extraction of upper airway images using medical image segmentation technology is extremely accurate, solving the problem of unclear soft tissue display in traditional CBCT images.

[0106] (4) This method has good operability. It only needs to upload the original CBCT image data in a specific software, and then the system can automatically extract the airway area and construct a three-dimensional model. This method is easy to operate, simple to implement, efficient and accurate, and repeatable.

[0107] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0108] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for upper airway segmentation and airflow simulation based on CBCT images, characterized in that: include: Constructing an initial 3D model; Acquire the CBCT original image data and initial three-dimensional model data to be tested; Processing the initial three-dimensional model data to obtain model parameter data; Setting parameters of the initial three-dimensional model according to the model parameter data to obtain an airway three-dimensional model; Inputting the CBCT original image data to be tested into the airway three-dimensional model to obtain a velocity cloud map and a stress cloud map; The method for constructing the initial three-dimensional model is: Obtaining CBCT raw image data; Preprocessing the CBCT raw image data to obtain image input data; According to the image input data, a first sub-model is constructed based on the 3DUX-Net network; Optimizing the first sub-model to obtain a second sub-model; Performing inverse modeling and smoothing processing on the second sub-model to obtain a third sub-model; The sub-model is meshed to obtain an initial three-dimensional model.

2. The upper airway segmentation and airflow simulation method based on CBCT images according to claim 1, characterized in that: The initial three-dimensional model data includes: The flow parameters and temperature parameters of airway intake, the distance parameters of airway wall movement during breathing, and the temperature parameters and flow parameters during exhalation.

3. The upper airway segmentation and airflow simulation method based on CBCT images according to claim 1, characterized in that: The preprocessing of the CBCT raw image data to obtain image input data includes: Manually segmenting the airway on the CBCT original image data to obtain first image data; Performing bilateral filtering on the first image data to obtain second image data; Performing weighted summation on eight regions of each point in the second image to obtain third image data; The third image data is resized and data set enhanced to obtain image input data.

4. The upper airway segmentation and airflow simulation method based on CBCT images according to claim 1, characterized in that: The learning rate of the 3DUX-Net network is 0.0001, the batch size is set to 1, and the training step size is set to 60,000 steps.

5. The upper airway segmentation and airflow simulation method based on CBCT images according to claim 1, characterized in that: The step of optimizing the first sub-model to obtain the second sub-model includes: Obtain a first segmented image output by the first sub-model; Determine whether the first segmented image has over-division or omission. If so, determine whether the pixel threshold of the first segmented image is greater than -550. If so, remove the corresponding part to obtain a second segmented image. If omission exists, determine whether the pixel position of the omission part is within the eight neighborhoods of the segmented foreground pixels. If so, classify the corresponding part as the foreground area to obtain a third segmented image. Based on the first sub-model, a second sub-model is obtained according to the second segmented image and the third segmented image.

6. The upper airway segmentation and airflow simulation method based on CBCT images according to claim 1, characterized in that: The performing inverse modeling and smoothing processing on the second sub-model to obtain a third sub-model includes: Importing the second sub-model file into Geomagicwrap, using a mesh doctor to remove particles and non-structures on the surface of the second sub-model, and smoothing or deleting and reconstructing the irregular plane of the second sub-model after removing particles and non-structures, to obtain a smoothed second sub-model; Dividing the obtained smoothed second sub-model into curved surface patches to obtain a regular curved surface patch distribution model, and converting the regular curved surface patch distribution model into a solid model and exporting it; Import the solid model into SpacesClaim, fit or reconstruct the surface patches at the air inlet and exhaust port to obtain the third sub-model.

7. The upper airway segmentation and airflow simulation method based on CBCT images according to claim 1, characterized in that: The meshing process of the sub-model to obtain an initial three-dimensional model includes: The third sub-model is divided into boundaries, named, and the inlet and outlet walls are pointed out. The corresponding fluid domain is set, and a boundary layer grid is set at the boundary with a growth rate of 1.2 and a minimum unit size of 2 mm to obtain an initial three-dimensional model.

Images