Maxillary sinus boundary extraction method based on convolutional neural network
By constructing a convolutional neural network and a classification model, combined with transparency analysis, fine segmentation of the maxillary sinus boundary was achieved, solving the problems of inaccurate identification and loss of structural information in three-dimensional data in traditional methods, and improving the identification accuracy of the maxillary sinus boundary and the accuracy of clinical applications.
Patent Information
- Application Number
- CN202510428130.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-07
- Publication Date
- 2025-12-30
AI Technical Summary
Traditional two-dimensional medical imaging methods struggle to accurately identify the maxillary sinus boundary, especially in three-dimensional data where structural information is easily lost. Existing computer algorithms also have limitations in terms of noise sensitivity and complexity.
A convolutional neural network-based approach was adopted to construct a three-dimensional tissue block boundary model through training and mapping. Combined with a classification model and transparency analysis, this approach enabled fine segmentation of the maxillary sinus boundary.
It improves the accuracy and detail of maxillary sinus boundary identification, and can more realistically reflect its three-dimensional spatial morphology, which is helpful for accurate diagnosis and treatment in clinical applications.
Smart Images
Figure CN121235975A_ABST
Abstract
Description
Technical Field
[0001] This invention proposes a method for extracting the maxillary sinus boundary based on a convolutional neural network, which relates to the field of maxillary sinus image processing technology. Background Technology
[0002] In the medical field, the maxillary sinus is part of the paranasal sinuses, and accurate identification of its morphology and boundaries is crucial for the diagnosis and treatment of various diseases. Traditional diagnostic methods mainly rely on visual observation of images such as X-rays, CT scans (computed tomography), or MRI (magnetic resonance imaging). However, the boundaries of the maxillary sinus in these images may be difficult to determine accurately due to obstruction by surrounding tissues, limitations in image resolution, or differences in individual anatomical structures.
[0003] With the development of digital image processing technology, computer algorithms have begun to be used to assist in boundary extraction in medical images. Early methods included thresholding and edge detection operators. Thresholding is based on the differences in pixel gray values in an image, setting one or more thresholds to segment the image into different regions. However, due to the complexity of medical images, the gray values of the maxillary sinus tissue and surrounding tissues may overlap, leading to inaccurate thresholding results. Edge detection operators mainly find edges by calculating the first or second derivative of the image. However, these operators are sensitive to noise, and in complex medical image scenarios, they may detect too many false edges or miss true boundaries.
[0004] Modern medical imaging technologies increasingly provide three-dimensional data, such as 3D CT and 3D MRI. For tissues with three-dimensional spatial structures, like the maxillary sinus, two-dimensional boundary extraction methods are no longer sufficient. Traditional two-dimensional methods require processing each two-dimensional slice separately and then integrating the results when processing three-dimensional data, which easily leads to the loss of structural information in three-dimensional space. Therefore, a method is needed that can effectively utilize three-dimensional image data to accurately extract the maxillary sinus boundary. This provides the technical background for maxillary sinus boundary extraction methods based on convolutional neural networks (CNNs), which can directly process three-dimensional data and better learn and extract its boundary features. Summary of the Invention
[0005] To address the aforementioned technical problems, this invention proposes a method for extracting the maxillary sinus boundary based on a convolutional neural network, comprising the following steps:
[0006] S1. Construct a convolutional neural network and train it using labeled maxillary sinus images;
[0007] S2. Use a trained convolutional neural network to predict the tissue region boundary position of the actual input maxillary sinus image, and map it to the actual position in three-dimensional space to construct a three-dimensional tissue block boundary model.
[0008] S3. Construct a classification model and use the classification model to output the preliminary block boundary surface of the three-dimensional organizational block boundary model;
[0009] S4. Within the block enclosed by the block boundary surface obtained in step S3, fine boundary surface segmentation of the maxillary sinus tissue is achieved by determining key voxels.
[0010] In a preferred embodiment, step S2 includes:
[0011] S21. Use the trained convolutional neural network to predict the tissue region boundary position of the actual input maxillary sinus image, and transform the two-dimensional coordinates of the boundary vertices of the two-dimensional image to two-dimensional actual physical coordinates.
[0012] S22. Combining the image slice thickness information, the two-dimensional actual physical coordinates are extended to three-dimensional actual spatial coordinates.
[0013] In a preferred embodiment, in step S3, a classification model is constructed using the point v(x,y,z) at the upper boundary of the three-dimensional tissue block boundary model as a voxel, and the voxels are divided into two categories: those belonging to the block boundary surface and those not belonging to the block boundary surface.
[0014] The input parameters f(x,y,z) of the classification model include: voxel gradient features The feature vector Ct(x,y,z) of the voxel is used to obtain the classification result C(v). If C(v) = 1, it means that the voxel belongs to the block boundary surface. If C(v) = 0, it means that it does not belong to the block boundary surface.
[0015] Traverse the points at the upper boundary of the 3D organization block boundary model, mark all voxels classified as block boundary surfaces, and connect the voxels of the block boundary surfaces to form a preliminary block boundary surface.
[0016] In a preferred embodiment, during prediction, for the input parameter F = f(v) of the voxel to be classified, Σ is calculated. R∈R μ r y r T(F r The result of F)+n is used to obtain the classification result C(v) according to the following classification rules:
[0017]
[0018] Among them, y r It is the r-th feature label, F r It is the r-th eigenvector, μ ris the coefficient corresponding to the r-th feature label, T is the mapping function, B is the constant term, and R is the feature vector space.
[0019] In a preferred embodiment, in step S4, within the block enclosed by the preliminary block boundary surface obtained in step S3, key voxels are determined based on the transparency analysis of voxels within the block:
[0020] Calculate the transparency H[v(x,y,z)] of voxels v(x,y,z) within the block, and find the transparency peak g. peak Set the transparency threshold range [g] peak -t1,g peak -t2], t1 and t2 are the upper and lower fluctuation points respectively. Within the block, search for voxels whose transparency H[v(x,y,z)] falls within this range as the initial key voxels;
[0021] If: g peak -t1≤H[v(x,y,z)]≤g peak +t2
[0022] Then v(x,y,z) is marked as the key voxel s(x,y,z).
[0023] In a preferred embodiment, in step S4, a fine segmentation criterion is defined based on transparency similarity and spatial adjacency. Let the set of key voxels be S, and the neighboring voxels of key voxel s(x,y,z)∈S be n(x,y,z). If the following conditions are met:
[0024] |H[n(x,y,z)]-H[s(x,y,z)]|≤t3
[0025] Then, n(x,y,z) is merged into the key voxel set, and the key voxel set S'=S∪{n(x,y,z)} is updated; where H[n(x,y,z)] and H[s(x,y,z)] are the transparency of key voxel s(x,y,z) and adjacent voxel n(x,y,z) respectively, t3 is the transparency difference threshold, and n(x,y,z) and s(x,y,z) are spatially adjacent.
[0026] In a preferred embodiment, in step S21, the maxillary sinus image to be segmented is input into a trained convolutional neural network, which outputs a region boundary enclosed by four borders, and extracts the coordinates of the two opposite corner vertices of the region boundary (X). 左下 ,Y 左下 ) and (X 右上 ,Y 右上 Convert it to two-dimensional physical coordinates; the conversion formula is as follows:
[0027]
[0028] Among them, pixel spacing (P) X ,P Y This represents the actual physical size corresponding to each pixel in the X and Y directions. These are the two-dimensional physical coordinates corresponding to the two diagonal points.
[0029] In a preferred embodiment, in step S22, the index of the slice corresponding to the input maxillary sinus image in the z-axis direction is set to z. k The three-dimensional actual spatial coordinates of the two opposite corners of the maxillary sinus tissue region boundary in three-dimensional space are represented as follows:
[0030]
[0031] Among them, s z The thickness of the slice; These are the three-dimensional actual spatial coordinates of the two opposite corner points of the maxillary sinus tissue region boundary.
[0032] In a preferred embodiment, in step S1, let the pixel value of pixel N on each side l1 of the region C enclosed by the four bounding boxes predicted by the convolutional neural network be B. l1 (N), the pixel value of each edge l2 of the labeled target region C0 in the existing image, representing the pixel value of point M. Construct the following difference function:
[0033]
[0034] Where E is the difference value, M0 is the total number of pixels M, N0 is the total number of pixels N; ε is an empirical constant.
[0035] Compared with the prior art, the present invention has the following beneficial technical effects:
[0036] By constructing and training a convolutional neural network, its powerful feature learning capabilities can be leveraged to automatically learn the features of tissue region boundaries in maxillary sinus images. Compared to traditional methods, it can predict boundary locations more accurately. This is because convolutional neural networks can learn complex boundary patterns from a large number of labeled images, avoiding the subjectivity and limitations of manual feature extraction.
[0037] Mapping in three-dimensional space to construct a three-dimensional tissue block boundary model extends boundary extraction from two-dimensional to three-dimensional. This is very important for tissues with complex three-dimensional structures, such as the maxillary sinus, as it can more realistically reflect its boundary conditions and help to understand the spatial morphology and extent of the maxillary sinus more accurately in clinical and other practical applications.
[0038] A classification model is constructed to detect the block boundary surfaces of a 3D tissue block boundary model. This hierarchical processing approach can further improve the accuracy of boundary detection. The classification model can classify and refine the features of the block boundary surfaces, removing noise and erroneous boundary information, thereby obtaining clearer and more accurate block boundary surfaces.
[0039] Fine-grained boundary segmentation of the maxillary sinus tissue is achieved by identifying key voxels within the block enclosed by the boundary surface. More detailed subdivision of the tissue within the boundary facilitates in-depth analysis of the internal structure of the maxillary sinus tissue and allows for more accurate identification of tissue boundaries. Attached Figure Description
[0040] Figure 1 This is a flowchart of the maxillary sinus boundary extraction method based on convolutional neural networks of the present invention;
[0041] Figure 2 Medical images of the maxillary sinus;
[0042] Figure 3 A schematic diagram of the process for constructing a three-dimensional organizational block boundary model. Detailed Implementation
[0043] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0044] In the accompanying drawings of specific embodiments of the present invention, in order to better and more clearly describe the working principle of each component in the system and show the connection relationship of each part in the device, only the relative positional relationship between each component is clearly distinguished. It does not constitute a limitation on the signal transmission direction, connection sequence, or size, dimension, and shape of each part within the component or structure.
[0045] like Figure 1 The flowchart shown is a method for extracting the maxillary sinus boundary based on a convolutional neural network according to the present invention. The method for extracting the maxillary sinus boundary includes the following steps:
[0046] S1. Construct a convolutional neural network and train it using labeled maxillary sinus images.
[0047] Building a convolutional neural network includes an encoder part and a decoder part.
[0048] Encoder section:
[0049] Set the initial convolutional layer, and let the input image I... norm The boundary of the target region after the first convolutional layer is F. L The formula for calculating convolution is:
[0050]
[0051] Where L represents the layer number, (x, y) are the coordinates of the four borders of the target interval, and w ij is the position weight, b is the bias term, i and j are the coordinate values of the increase in the horizontal and vertical coordinates of the pixels on the four borders of the target interval as they move on the four borders, I is the maximum increase in the horizontal coordinate, and J is the maximum increase in the vertical coordinate.
[0052] As the number of network layers increases, high-level features of the image are gradually extracted by continuously stacking convolutional and pooling layers, resulting in target region boundaries F at different scales. L .
[0053] Decoder section:
[0054] The target region boundary F extracted by each layer of the encoder L Upsampling is performed, and the data is then stitched and merged with the region of the corresponding encoder layer.
[0055] Let the upsampled region be U. L The spliced area is C(U) L ,F L ).
[0056] We collected a large amount of normal maxillary sinus image data, manually labeled the boundaries of target regions in the maxillary sinus images, and used the manually labeled target region boundaries of the maxillary sinus images to train a convolutional neural network.
[0057] Specifically, two-dimensional medical images containing clear images of the maxillary sinus are collected, such as... Figure 2 As shown, images can be axial, coronal, or sagittal. The images need to cover the maxillary sinus conditions of different individuals and at different angles.
[0058] Annotation tools were used to label the maxillary sinus boundaries in the images. For each maxillary sinus instance, the coordinates of the upper left and lower right corners of the target region boundary were labeled. The constructed network model was trained using the prepared labeled data. Image data was input into the network to obtain a four-channel predicted boundary map, and then the difference value was calculated based on the difference function.
[0059] Suppose that the pixel value of pixel N on each edge l1 of region C, which is enclosed by four bounding boxes predicted by the convolutional neural network, is B. l1 (N), the pixel value of each edge l2 of the labeled target region C0 in the existing image, representing the pixel value of point M. Construct the following difference function:
[0060]
[0061] Where E is the difference value, M0 is the total number of pixels M, N0 is the total number of pixels N, and ε is an empirical constant. The difference value E itself is used to measure the similarity between two regions. By optimizing the algorithm to minimize the difference value E, the region enclosed by the four bounding boxes predicted by the convolutional neural network is closer to the real region.
[0062] By optimizing the algorithm and updating the parameters of the convolutional neural network, the difference value E is gradually reduced, making the region enclosed by the four borders predicted by the convolutional neural network closer and closer to the actual annotation. After multiple iterations of training, the convolutional neural network can learn the ability to accurately segment the region enclosed by the four borders from the input image.
[0063] S2. Use a trained convolutional neural network to predict the tissue region boundary position of the actual input maxillary sinus image, and map it to the actual position in three-dimensional space to construct a three-dimensional tissue block boundary model.
[0064] Specifically, it includes the following steps, such as Figure 3 As shown:
[0065] S21. Use the trained convolutional neural network to predict the tissue region boundary position of the actual input maxillary sinus image, and transform the two-dimensional coordinates of the boundary vertices of the two-dimensional image to two-dimensional actual physical coordinates.
[0066] In the prediction phase, to predict the actual input maxillary sinus image using a trained convolutional neural network, the maxillary sinus image to be segmented is input into the trained convolutional neural network. The network outputs the region boundary enclosed by the four corresponding bounding boxes, and extracts the coordinates of the two diagonal vertices (X). 左下 ,Y 左下 ) and (X 右上 ,Y 右上 Then convert it into two-dimensional physical coordinates. and (Actual physical size measurement)
[0067] Assuming the top-left corner of the image is the origin, the transformation formula is as follows:
[0068]
[0069] Here the pixel spacing (P) X ,P Y This represents the actual physical size corresponding to each pixel in the X and Y directions. It refers to the two-dimensional physical coordinates corresponding to the diagonal points, with the units converted from pixels to actual physical size units such as millimeters.
[0070] S22. Combining the image slice thickness information, the two-dimensional actual physical coordinates are extended to three-dimensional actual spatial coordinates.
[0071] Since medical images are three-dimensional data volumes composed of a series of two-dimensional slice images, further information from the slice layers is needed to determine the location of the maxillary sinus boundary in three-dimensional space.
[0072] Assume the slice corresponding to the input maxillary sinus image has an index of z along the z-axis. k (Starting from a predetermined initial slice), the coordinates of the two opposite corners of the maxillary sinus tissue region boundary in three-dimensional space are represented as follows:
[0073]
[0074] Among them, s z The thickness of the slice; These are the three-dimensional actual spatial coordinates of the two opposite corner points of the maxillary sinus tissue region boundary. Based on the three-dimensional actual spatial coordinates of the two opposite corner points, a three-dimensional tissue block boundary model corresponding to the maxillary sinus tissue region can be constructed.
[0075] S3. Construct a classification model and use the output three-dimensional organizational block boundary model to obtain the preliminary block boundary surface.
[0076] by arrive Points v(x,y,z) at the upper boundary of the three-dimensional organization block boundary model are used as voxels. A classification model is constructed based on the extracted voxels to classify them into two categories: those belonging to the block boundary surface and those not belonging to the block boundary surface.
[0077] The training samples for the classification model were obtained by manually labeling typical block boundary surface voxels and the surrounding non-block boundary surface voxels.
[0078] For a voxel v(x,y,z) to be classified, f(v) is used as the input parameter of the classification model, where f(x,y,z) includes the gradient features of the voxel. The eigenvector Ct(v) of the voxel will Input into the classification model,
[0079] During prediction, for the input parameters F = f(v) of the voxel to be classified, Σ is calculated. r∈R μ r y r T(F r The result of F)+B is used to obtain the classification result C(v) according to the following classification rules:
[0080]
[0081] yr It is the r-th feature label, F r It is the r-th eigenvector, μ r is the coefficient corresponding to the r-th feature label, T is the mapping function, B is the constant term, and R is the feature vector space.
[0082] The classification result C(v) is obtained. If C(v) = 1, it means that the voxel belongs to the block boundary surface. If C(v) = 0, it means that it does not belong to the block boundary surface.
[0083] Traverse the points at the upper boundary of the 3D organization block boundary model, mark all voxels classified as block boundary surfaces, and connect all voxels classified as block boundary surfaces to form a preliminary block boundary surface.
[0084] S4. Within the block enclosed by the preliminary block boundary surface obtained in step S3, fine boundary surface segmentation of the maxillary sinus tissue is achieved by determining key voxels.
[0085] Within the block enclosed by the preliminary block boundary surface obtained in step S3, key voxels are determined based on the transparency analysis of voxels within the block.
[0086] Calculate the transparency H[v(x,y,z)] of the voxels within the block, and find the transparency peak g. peak These peaks typically correspond to the main tissue types within the maxillary sinus.
[0087] Set a transparency threshold range [g] peak -t1,g peak -t2], t1 and t2 are the upper and lower fluctuation points respectively. Within the block, search for voxels whose transparency H[v(x,y,z)] falls within this range as the initial key voxels;
[0088] If: g peak -t1≤H[v(x,y,z)]≤g peak +t2
[0089] Then v(x,y,z) is marked as the key voxel s(x,y,z).
[0090] Fine-grained boundary segmentation is performed based on key voxels.
[0091] Based on transparency similarity and spatial adjacency, a fine segmentation criterion is defined. Let the set of key voxels be S, and for a key voxel s(x,y,z)∈S, its neighboring voxels be n(x,y,z). If the following conditions are met:
[0092] |H[n(x,y,z)]-H[s(x,y,z)]|≤t3
[0093] Then, n(x,y,z) is merged into the key voxel set, and the key voxel set S'=S∪{n(x,y,z)} is updated.
[0094] Where H[n(x,y,z)] and H[s(x,y,z)] are the transparency of the key voxel s(x,y,z) and the adjacent voxel n(x,y,z), respectively, t3 is the transparency difference threshold, and n(x,y,z) and s(x,y,z) are spatially adjacent.
[0095] Repeat the above process until there are no adjacent voxels that meet the fine segmentation criteria, thus obtaining the fine boundary of the maxillary sinus tissue.
[0096] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.
[0097] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above method embodiments.
[0098] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps in the above method embodiments.
[0099] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. The databases involved in the embodiments provided in this application can include at least one of relational and non-relational databases. Non-relational databases can include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application can be general-purpose processors, central processing units, graphics processors, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.
[0100] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0101] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.
Claims
1. A method for extracting a maxillary sinus boundary based on a convolutional neural network, characterized by, The method comprises the following steps: S1, constructing a convolutional neural network, and training the convolutional neural network by using the labeled maxillary sinus image; S2, predicting the tissue region boundary position of the actual input maxillary sinus image by using the trained convolutional neural network, and mapping the actual position in the three-dimensional space to construct a three-dimensional tissue block boundary model; S3, constructing a classification model, and outputting a preliminary block boundary surface of the three-dimensional tissue block boundary model by using the classification model; S4, in the block surrounded by the block boundary surface obtained in step S3, the fine boundary surface segmentation of the maxillary sinus tissue is realized by determining the key voxels. 2.The maxillary sinus boundary extraction method based on a convolutional neural network according to claim 1, characterized in that, The step S2 comprises: S21, converting the two-dimensional coordinates of the boundary vertex of the two-dimensional image to the two-dimensional actual physical coordinates by using the trained convolutional neural network to predict the tissue region boundary position of the actual input maxillary sinus image; S22, combining the image slice thickness information to expand the two-dimensional actual physical coordinates to the three-dimensional actual space coordinates. 3.The maxillary sinus boundary extraction method based on a convolutional neural network according to claim 1, characterized in that, In the step S3, the point v(x, y, z) at the upper boundary of the three-dimensional tissue block boundary model is taken as a voxel to construct a classification model, and the voxels are divided into two categories, i.e., belonging to the block boundary surface and not belonging to the block boundary surface; The input parameter f(x, y, z) of the classification model includes: gradient features of the voxels and the feature vector Ct(x, y, z) of the voxels, to obtain a classification result C(v), wherein if C(v) = 1, it indicates that the voxel belongs to the block boundary surface, and if C(v) = 0, it indicates that the voxel does not belong to the block boundary surface; All voxels classified as the block boundary surface are marked by traversing the points at the upper boundary of the three-dimensional tissue block boundary model, and the voxels of the block boundary surface are connected to form a preliminary block boundary surface.
4. The maxillary sinus boundary extraction method based on a convolutional neural network according to claim 3, wherein For the input parameter F = f(v) of the voxel v to be classified at the prediction time, the result of the calculation of Σ r∈R μ r y r T(F r ,F) + B is obtained according to the following classification rule to obtain the classification result C(v): where y r is the rth feature label, F r is the rth feature vector, μ r is the coefficient corresponding to the rth feature label, T is a mapping function, B is a constant term, and R is a feature vector space. 5.The maxillary sinus boundary extraction method based on a convolutional neural network according to claim 4, characterized in that, In the step S4, the key voxels are determined based on the transparency analysis of the voxels in the block in the block surrounded by the preliminary block boundary surface obtained in step S3: Calculate the transparency H[v(x, y, z)] of the voxel v(x, y, z) in the block, find the transparency peak value g peak , set the transparency threshold range [g peak -t1, g peak -t2], t1 and t2 are respectively the upper and lower fluctuation points, search the voxel whose transparency H[v(x, y, z)] falls in the range as the initial key voxel in the block; if: g peak -t1≤ H[v(x, y, z)] ≤ g peak +t2 Then, v(x, y, z) is marked as a key voxel s(x, y, z). 6.The maxillary sinus boundary extraction method based on a convolutional neural network according to claim 5, characterized in that, In the step S4, a fine segmentation criterion is defined based on the transparency similarity and the spatial adjacency, the key voxel set is S, the adjacent voxel of the key voxel s(x, y, z) is n(x, y, z), and if the following conditions are met: |H[n(x, y, z)]-H[s(x, y, z)]|≤y3 Then, n(x, y, z) is merged into the key voxel set, and the key voxel set S' is updated as S' = S ∪ {n(x, y, z)}; wherein H[n(x, y, z)] and H[s(x, y, z)] are the transparencies of the key voxel s(x, y, z) and the adjacent voxel n(x, y, z), respectively, t3 is a transparency difference threshold, and n(x, y, z) and s(x, y, z) are adjacent in space. 7.The maxillary sinus boundary extraction method based on a convolutional neural network according to claim 2, characterized in that, In the step S21, the maxillary sinus image to be segmented is input into the trained convolutional neural network, and a region boundary surrounded by four bounding boxes is output. Two pairs of diagonal vertex position coordinates (X 左下 ,Y 左下 ) and (X 右上 ,Y 右上 ) of the region boundary are extracted and converted into two-dimensional actual physical coordinates. The conversion formula is as follows: wherein the pixel pitch (P X ,P Y ) represents the actual physical size corresponding to each pixel in the X and Y directions, are two-dimensional actual physical coordinates corresponding to the two pairs of diagonal points. 8.The maxillary sinus boundary extraction method based on a convolutional neural network according to claim 7, characterized in that, In the step S22, it is assumed that the index of the slice corresponding to the input maxillary sinus image in the z-axis direction is z k Then, the three-dimensional actual space coordinate representation of the two opposite corner points of the maxillary sinus tissue region boundary in the three-dimensional space is: where s z is the slice thickness; are the three-dimensional actual spatial coordinates of the two pairs of corner points of the boundary of the maxillary sinus tissue region. 9.The maxillary sinus boundary extraction method based on a convolutional neural network according to claim 1, wherein, In the step S1, the pixel value of each pixel point N of the edge l1 of the region C surrounded by the four edge boxes predicted by the convolutional neural network is B l1 (N), the pixel value of each pixel point M of the edge l2 of the marked target region C0 of the existing image The difference function is constructed as follows: Wherein, E is the difference value, M0 is the total number of pixel points M, N0 is the total number of pixel points N, and ε is an empirical constant.