AI-driven oral-maxillofacial war wound virtual simulation treatment training system

Abstract

The invention relates to the technical field of virtual simulation treatment, in particular to an AI-driven virtual simulation treatment training system for oral-maxillofacial war wounds, which comprises an image data processing module for sensing image data, dividing the image data into a plurality of different image groups and preprocessing each image data in the image groups; the tissue segmentation module is used for receiving the preprocessed image group and then identifying all tissues in each piece of image data by adopting a medical image segmentation algorithm, and the medical image segmentation algorithm specifically comprises a coding unit, a dense jump connection unit, a decoding unit and a depth supervision output unit; the war wound model building and calling module is used for converting a two-dimensional skeleton contour into a three-dimensional grid model, namely a war wound model, through voxel division, contour surface extraction and grid generation on the basis of pixel coordinates of the segmentation mask; basic information of the patient is received, and a mapping relation between the war wound model and the image group is established.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of flow meter testing, in particular to an AI-driven virtual simulation treatment training system for oral and maxillofacial war injuries. BACKGROUND

[0002] In modern war injury treatment, the oral and maxillofacial region is highly exposed, has a complex anatomical structure (including key tissues such as bones, nerves, blood vessels and teeth), is easily injured by explosions and impacts, and the injury is often sudden, complex and highly disabling (such as fracture displacement, tissue defect and occlusion dysfunction). Rapid and accurate treatment of oral and maxillofacial war injuries is directly related to the life safety, functional recovery and quality of life of the wounded, and efficient treatment skill training is a core link to improve battlefield treatment capabilities.

[0003] Traditional oral and maxillofacial war injury treatment training relies on cadaver specimens, physical models or two-dimensional images (such as CT and MRI tomographic images), which have obvious limitations: cadaver specimens are scarce and difficult to preserve, and cannot simulate diversified war injury patterns; physical models have fixed structures and cannot dynamically present anatomical changes of different injuries (such as comminuted fractures and soft tissue lacerations); two-dimensional images are highly abstract, and medical personnel need to restore three-dimensional anatomical relationships through spatial imagination, which makes it difficult to intuitively understand the three-dimensional relationship between the injury and the surrounding tissues, resulting in limited training effectiveness.

[0004] With the development of artificial intelligence (AI) and medical imaging technology, virtual simulation training based on three-dimensional reconstruction has gradually become a trend, but existing systems still have shortcomings: first, the accuracy of the injury model construction is insufficient, and the restoration degree of small structures (such as alveolar bone fractures and nerve branches) is low, which makes it difficult to support fine operation training; second, the model is disconnected from the original image data, and the corresponding tomographic image details of the injury area cannot be traced during training, affecting the depth of understanding of the injury; third, there is a lack of dynamic evolution simulation of war injuries, which cannot reproduce the real-time changes of tissue morphology during treatment (such as bone displacement during fracture reduction), making it difficult to simulate real surgical scenarios.

[0005] Therefore, it is urgent to build an AI-driven virtual simulation treatment training system for oral and maxillofacial war injuries, which accurately extracts anatomical structures and injury characteristics from image data through AI algorithms, establishes a high-precision war injury model, and associates the original image and patient information to realize dynamic calling and simulation operation of the injury model, thereby improving the authenticity, accuracy and efficiency of war injury treatment training, providing technical support for cultivating treatment personnel with real combat capabilities. In view of this, we propose an AI-driven virtual simulation treatment training system for oral and maxillofacial war injuries. SUMMARY

[0006] The present application aims to solve the problems of resource scarcity, model fixation, and abstract anatomical relationship in traditional oral and maxillofacial war trauma treatment training, which relies on cadaver specimens, physical models, or two-dimensional images. In addition, the existing virtual simulation training system based on three-dimensional reconstruction has insufficient model construction accuracy (low restoration degree for small structures such as alveolar bone fractures and nerve branches), the model is disconnected from the original image data (unable to trace the details of the corresponding tomographic image in the trauma area), and lacks dynamic evolution simulation of war trauma (difficult to reproduce real-time changes in tissue morphology during treatment). Therefore, an AI-driven oral and maxillofacial war trauma virtual simulation treatment training system is provided, which accurately extracts anatomical structures and trauma characteristics from image data through AI algorithms, establishes a high-precision war trauma model, and associates the original image and patient information. The dynamic calling and simulation operation of the trauma model are realized, which improves the authenticity, accuracy, and efficiency of war trauma treatment training and provides technical support for cultivating rescue personnel with practical ability.

[0007] The present application provides an AI-driven oral and maxillofacial war trauma virtual simulation treatment training system, which includes an image data processing module, an organization segmentation module, and a war trauma model construction and calling module. The image data processing module perceives image data and divides the image data into multiple different image groups, and pre-processes each image data in the image group. The organization segmentation module receives the pre-processed image group, and then uses a medical image segmentation algorithm to identify all the tissues in each image data. The medical image segmentation algorithm specifically includes four parts: encoding unit, dense skip connection, decoding unit, and deep supervision output. The decoding unit restores the feature map size through transpose convolution upsampling and fuses with the dense connection features of the encoding unit to gradually generate a segmentation mask with the same size as the input image. The war trauma model construction and calling module converts the two-dimensional bone contour into a three-dimensional grid model based on the pixel coordinates of the segmentation mask through voxel division, isosurface extraction, and grid generation, which is the war trauma model. The image group division unit receives the patient's basic information of the image group, establishes a mapping relationship between the war trauma model and the image group, receives the input basic information, and then calls out the corresponding war trauma model.

[0008] As a further improvement of the technical solution, the image data processing module includes an image group division unit and an image preprocessing unit. The image group division unit is used to divide a large number of real case image data into multiple different image groups. The image preprocessing unit receives the image group and pre-processes each image data in the image group. The preprocessing includes removing scanning noise from the image data and normalizing the image data by layer thickness.

[0009] As a further improvement to this technical solution, the image group division unit senses a large amount of real case image data labeled with the patient's basic information, and divides the large amount of real case image data into multiple different image groups based on the patient's basic information. Each image group corresponds to one patient's basic information, and the image group includes a large amount of different image data.

[0010] As a further improvement to this technical solution, the image preprocessing unit removes scanning noise from the image data by setting a neighborhood radius of... Take the pixels in the neighborhood and calculate the weight of the pixels in the neighborhood based on the distance of the pixel to the center pixel of the neighborhood. Then, a weighted average is used to calculate the values ​​of all pixels in the neighborhood and their corresponding weights. The sum of the products ,replace It is the pixel value of the center pixel in the neighborhood.

[0011] As a further improvement to this technical solution, the image preprocessing unit can process any pixel in the image data. A point within its neighborhood weight for: ; in: , It represents the horizontal and vertical offset of a neighboring pixel relative to the center pixel; Standard deviation is the core parameter used to control the smoothness of filtering. The smaller the standard deviation, the more concentrated the weights are at the center, resulting in weaker noise reduction but sharper edges; The larger the value, the more the weight diffuses, resulting in strong noise reduction but blurred edges; The normalization coefficient ensures that the sum of all weights is 1, thus avoiding changes in the overall brightness of the filtered image data. The pixel value of the center pixel within the neighborhood is: .

[0012] As a further improvement to this technical solution, the working principle of layer thickness normalization in the image preprocessing unit is as follows: setting a standard layer thickness. The original layer thickness of the received image data ; If the standard layer thickness <Original layer thickness If a new fault is inserted between two adjacent layers, then the standard layer thickness is... >Original layer thickness Then adjacent faults are merged.

[0013] As a further improvement to this technical solution, the image preprocessing unit inserts a new tomographic image between two consecutive tomographic images in the image data: setting the original slice thickness. Adjacent tomographic images of the perceived image data are and Then calculate Layers and Number of tomographic images to be inserted between layers ,in The floor symbol is used for rounding down. The target layer thickness for inserting new tomographic images; The inserted first layer( Pixel values ​​of image data ,Depend on and Linear weighting yields: ; in For the two-dimensional coordinates of pixels, and These are the weighting coefficients.

[0014] As a further improvement to this technical solution, the image preprocessing unit merges adjacent topographies: Calculation Layers and Number of original faults that need to be merged between layers ,in To ensure the total thickness after merging is not less than the standard layer thickness, the rounding sign is used to round up to the nearest whole number. ; Calculate the structural saliency of each fault in the image data Then, the structural salience of each layer is determined. Normalization yields the first... Layer weights ,in The original number of faults to be merged, and the sum of the weights of all layers is 1; For continuous The coordinates in the original fault layer are The pixels, the merged pixel values The weighted average of pixel values ​​from each layer: ,in for In the original fault layer The pixel value at that location.

[0015] As a further improvement to this technical solution, the medical image segmentation algorithm in the tissue segmentation module is as follows: The encoding unit gradually reduces the feature map size through convolutional blocks and max pooling, extracting hierarchical features from low-level features to high-level features; Each layer extracts hierarchical features through two 3×3 convolutions, batch normalization, and ReLU activation. Layer Level feature map for: ; in, The convolution kernel weights and biases; For standardized batch operations, specifically: ( , For batch mean and variance, , (for learnable parameters) To introduce nonlinearity into the ReLU activation function: ; Max pooling: After every two convolutional blocks, 2×2 max pooling is used to reduce the feature map size. ,in This is a 2×2 max pooling operation, which takes the maximum value of the feature map within a 2×2 window, thus reducing the feature map size. This means taking the feature map within a 2×2 window. The maximum value; Dense skip connections: in the coding unit Layer and Decoding Unit Dense connections are established between layers, and the coding unit is... The feature maps for each stage are as follows: ,in Feature maps are stitched together along the channel dimension. For shallower feature maps in the decoding unit; Decoding unit: Recovers the feature map size through transposed convolution and fuses it with the densely connected features of the encoding unit to gradually generate a segmentation mask of the same size as the input image. Specifically: The feature map size is doubled through 2×2 transposed convolution. The output feature map of the dth stage and the dth layer is Where * denotes the transpose convolution operation, , These are the transposed convolution weights and biases, respectively. After transpose convolution, it is concatenated with the densely connected features of the corresponding layer of the encoding unit, and then processed by convolutional blocks; Deeply supervised output: Finally, the number of feature map channels is converted into the number of classes through 1×1 convolution, and then the pixel-level classification probability is output after Softmax activation; Deep supervision involves setting multiple output layers at different depths in the decoding unit. Each output layer is trained using an independent loss function, and the total loss is a weighted sum of the losses of each layer. ; in The depth of the decoding unit (e.g., 4 layers); Weights (usually deeper layers have higher weights, such as...) ); For the first Loss of layer output: ; To predict the mask at pixels The probability of; For real mask; Total number of pixels; Avoid having a denominator of 0.

[0016] As a further improvement to this technical solution, the specific working principle of establishing the combat trauma model is as follows: Coordinate transformation: Transforming the two-dimensional coordinates of bone pixels in the segmentation mask. Combining the slice thickness information from the tomographic images, it is converted into three-dimensional physical coordinates, specifically: The two-dimensional coordinates of a certain bone pixel in the k-th tomographic image are: The fault number is Then the three-dimensional physical coordinates for: ; in, The physical dimensions of pixels along the X and Y axes. For layer thickness, The coordinates are the global origin. Voxel vertex state encoding: The 3D space is divided into 1mm×1mm×1mm voxels, each voxel has 8 vertices, and a CT value threshold is set ( skeleton Determine if a vertex is inside a skeleton: ; in The CT value of the vertex; These are the vertex states; the 8 vertex states constitute an 8-bit binary code: ; Isosurface intersection calculation: For edges crossing the bone boundary in a voxel, the boundary intersection points are calculated using linear interpolation. : ; in These are the interpolation coefficients; The CT values ​​are the endpoints of the edge. The three-dimensional physical coordinates of the two endpoints; Triangle face generation and mesh stitching: based on voxel encoding The war trauma model is constructed by calling the corresponding triangular facet combination from the preset facet table.

[0017] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall module of the present invention.

[0019] The meanings of the labels in the diagram are as follows: 100. Image data processing module; 110. Image group division unit; 120. Image preprocessing unit; 200. Tissue segmentation module; 300. Combat trauma model construction and calling module. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] refer to Figure 1 As shown, an AI-driven virtual simulation treatment and training system for oral and maxillofacial combat trauma includes an image data processing module 100, a tissue segmentation module 200, and a combat trauma model construction and retrieval module 300, wherein: The image data processing module 100 senses the image data and divides it into multiple different image groups, and preprocesses each image data in the image group; The tissue segmentation module 200 receives the preprocessed image group and then uses a medical image segmentation algorithm to identify all tissues in each image data. The medical image segmentation algorithm specifically includes four parts: an encoding unit, dense skip connections, a decoding unit, and a depth supervision output. The decoding unit recovers the feature map size by transposing convolution upsampling and fuses it with the dense connection features of the encoding unit to gradually generate a segmentation mask of the same size as the input image. The combat trauma model construction and calling module 300, based on the pixel coordinates of the segmentation mask, transforms the two-dimensional skeletal contour into a three-dimensional mesh model through voxel division → isosurface extraction → mesh generation, which is the combat trauma model; it receives the patient's basic information from the image group in the image group division unit 110, establishes the mapping relationship between the combat trauma model and the image group, receives the input basic information, and then calls up the corresponding combat trauma model. Specifically: The image data processing module 100 includes an image group division unit 110 and an image preprocessing unit 120, wherein: The image group division unit 110 senses a large amount of real case image data labeled with the patient's basic information, including severe LeFort III maxillary fracture, gunshot wound with a diameter of 5.56 mm, 2 hours after injury, patient age of 25, male. Then, based on the patient's basic information, the large amount of real case image data is divided into multiple different image groups. Each image group corresponds to one patient's basic information. When the patient's basic information is different, the corresponding image data is divided into other image groups. Moreover, the image group includes a large amount of different image data. For example, the severe LeFort III maxillary fracture group includes complete image data generated by scanning through medical imaging equipment such as CT and MRI. The image preprocessing unit 120 receives the image group and preprocesses each image data in the image group. The preprocessing includes: removing scanning noise from the image data and normalizing the image data by layer thickness. The working principle of the image preprocessing unit 120 in removing scanning noise from image data is as follows: The neighborhood radius is set to... Take the pixels in the neighborhood and calculate the weight of the pixels in the neighborhood based on the distance of the pixel to the center pixel of the neighborhood. Points closer to the center pixel have a higher weight, while those farther away have a lower weight. Then, a weighted average is used to calculate the values ​​of all pixels in the neighborhood and their corresponding weights. The sum of the products ,replace The pixel value of the center pixel in the neighborhood is used to smooth out noise in the image data; For any pixel in the image data A point within its neighborhood weight for: ; in: , It represents the horizontal and vertical offset of a neighboring pixel relative to the center pixel; Standard deviation is the core parameter used to control the smoothness of filtering. The smaller the standard deviation, the more concentrated the weights are at the center, resulting in weaker noise reduction but sharper edges; The larger the value, the more the weight diffuses, resulting in strong noise reduction but blurred edges; The normalization coefficient ensures that the sum of all weights is 1, thus avoiding changes in the overall brightness of the filtered image data. The pixel value of the center pixel within the neighborhood is: .

[0022] To avoid inconsistencies in the thickness of adjacent slices in the image data, which could lead to Z-axis scaling errors when constructing a war trauma model, the working principle of slice thickness normalization in image preprocessing unit 120 is as follows: A standard slice thickness is set. The original layer thickness of the received image data ; If the standard layer thickness <Original layer thickness If a new fault is inserted between two adjacent layers to supplement the missing structural information; if the standard layer thickness is... >Original layer thickness Then, adjacent faults are merged to reduce data redundancy while preserving key structural features; The specific working principle of inserting a new fault between two adjacent layers is as follows: setting the original layer thickness. Adjacent tomographic images of the perceived image data are and Then calculate Layers and Number of tomographic images to be inserted between layers ,in The floor symbol is used for rounding down. The target layer thickness for inserting new tomographic images; The inserted first Pixel values ​​of layer image data ,Depend on and Linear weighting yields: ; in For the two-dimensional coordinates of pixels, and The weighting coefficient ensures a continuous grayscale transition between the inserted layer and the original layer, avoiding discontinuous jumps.

[0023] The specific working principle of merging adjacent faults is as follows: calculation Layers and Number of original faults that need to be merged between layers ,in To ensure the total thickness after merging is not less than the standard layer thickness, the rounding sign is used to round up to the nearest whole number. ; Calculate the structural saliency of each fault in the image data Then, the structural salience of each layer is determined. Normalization yields the first... Layer weights ,in The original number of faults to be merged, and the sum of the weights of all layers is 1; For continuous The coordinates in the original fault layer are The pixels, the merged pixel values The weighted average of pixel values ​​from each layer: ,in for In the original fault layer The pixel value at that location.

[0024] Structural salience The quantification is obtained by calculating the grayscale change rate of pixels within the corresponding fault line—the greater the grayscale change rate, the more salient the structure. The higher the value, the more likely the fault has obvious anatomical boundaries, traumatic interfaces, or morphological abrupt changes, for example: At the junction of the jawbone and surrounding muscles, the grayscale value abruptly changes from a high CT value of 200-1000 HU for the bone to a low value of -100-100 HU for the muscle. Due to the large rate of change in grayscale, the structure is highly significant. ; Specific structural salience The specific expression for calculating the gradient magnitude is as follows: ; in and For pixels in , The gradient in a direction can be calculated using operators such as Sobel and Prewitt, reflecting the rate of change of pixel grayscale.

[0025] The tissue segmentation module 200 receives the preprocessed image group and then uses a medical image segmentation algorithm to identify all tissues in each image data, such as bones, soft tissues, and trauma areas. The medical image segmentation algorithm consists of four parts: encoding unit downsampling, dense skip connections, decoding unit upsampling, and depth supervision output. The specific working principle is as follows: The encoding unit gradually reduces the feature map size through convolutional blocks and max pooling, extracting hierarchical features from low-level features (shallow features containing details such as edges and textures, like the grayscale changes of the alveolar bone edge) to high-level features (deep features containing semantic information, like the overall region of the "maxilla"). Each layer extracts hierarchical features through two 3×3 convolutions, batch normalization (BN), and ReLU activation. Layer Level feature map for: ; in, The convolution kernel weights and biases; For standardized batch operations, specifically: , , For batch mean and variance, , These are learnable parameters; To introduce nonlinearity into the ReLU activation function: ; Max pooling: After every two convolutional blocks, a 2×2 max pooling step of 2 is used to reduce the feature map size to 1 / 4. This is used to preserve the strongest features within the region and extract higher-level semantic information, such as the "overall contour of the maxilla." This is a 2×2 max pooling operation, which takes the maximum value of the feature map within a 2×2 window, thus reducing the feature map size. This means taking the feature map within a 2×2 window. The maximum value; Dense skip connections are used to fuse cross-layer features, specifically: in the encoding unit... Layer and Decoding Unit layer ≥ Dense connections are established between them, enabling the decoding unit to simultaneously acquire low-level details and high-level semantics from the encoding unit. The expression is as follows: ,in, Feature maps are stitched together along the channel dimension. This is a shallower feature map in the decoding unit; for example, the 512-channel feature of the encoding unit and the 256-channel feature of the decoding unit are concatenated to form a 768-channel feature map, which is used to address the problem of large semantic differences between cross-layer features and poor fusion effect, especially to improve the segmentation accuracy of small targets such as fracture lines. The decoding unit is used to restore the feature map size and output the segmentation result. Specifically, it restores the feature map size by upsampling through transposed convolution and fuses it with the densely connected features of the encoding unit to gradually generate a segmentation mask of the same size as the input image. Specifically, it enlarges the feature map size by 2 times by using a 2×2 transposed convolution with a stride of 2. Where * denotes the transpose convolution operation, , These are the transposed convolution weights and biases, respectively. After transpose convolution, the features are concatenated with the densely connected features of the corresponding layer of the encoding unit, and then processed by convolutional blocks to refine the features.

[0026] Deeply supervised output: Finally, the feature map channel number is converted into the number of categories (e.g., "background=0, skeleton=1, soft tissue=2, trauma area=3") through a 1×1 convolution, and then the pixel-level classification probability is output after Softmax activation. ; in For the first The class's logit value is output by a 1×1 convolution. The total number of categories is divided into four categories, such as maxillofacial tissues; For pixels The probability of belonging to category c; Deep supervision is used to optimize the training process. Its specific working principle is as follows: Instead of just the last layer, multiple output layers are set at different depths of the decoding unit. Each output layer is trained under supervision using an independent loss function, and the total loss is a weighted sum of the losses of each layer. ; in The decoding unit depth is 4 layers; For example, deeper layers usually have higher weights. ; For the first The layer output loss uses Dice loss, which is suitable for the problem of imbalanced categories in medical images. ; To predict the mask at pixels The probability belongs to the target class; The target is set to 1 and the background is set to 0. Total number of pixels; Avoid having a denominator of 0.

[0027] The combat trauma model construction and calling module 300, based on the pixel coordinates of the segmentation mask, transforms the two-dimensional skeletal contour into a three-dimensional mesh model through voxel division, isosurface extraction, and mesh generation. This model avoids misalignment of the combat trauma model due to differences in coordinate units or scanning positions. Then, it receives the patient's basic information from the image group division unit 110 and establishes a mapping relationship between the combat trauma model and the image group. During subsequent virtual simulation treatment training, the trainee can quickly obtain the combat trauma model corresponding to the target case by inputting basic information such as the tomographic image details corresponding to the fracture area. This facilitates targeted drills, improves the accuracy and efficiency of training, and enables standardized calling and traceability of case data. The specific working principle of building a combat trauma model in the combat trauma model construction and invocation module 300 is as follows: Coordinate transformation: Transforming the two-dimensional coordinates of bone pixels in the segmentation mask. Combining the slice thickness information (z-axis) from tomographic images with spatial location data for constructing war trauma models, this information is converted into three-dimensional physical coordinates. The two-dimensional coordinates of a certain bone pixel in the k-th tomographic image are: The fault number is Starting from the global origin at layer 0, the three-dimensional physical coordinates... for: ; in, The physical dimensions of pixels along the X and Y axes. For layer thickness, The global origin coordinates are such as the center of the skull base plane; Voxel vertex state encoding: The 3D space is divided into 1mm×1mm×1mm voxels, each voxel has 8 vertices, and a CT value threshold is set. (skeleton Determine if a vertex is inside a skeleton: ; in The CT value of the vertex; Vertex states are 1 = inside the skeleton, 0 = outside the skeleton; 8 vertex states constitute an 8-bit binary code: There are 256 topological patterns in total, used to determine whether a voxel contains a skeletal boundary; Calculate the bone boundary using the intersection points of isosurfaces: For one end of the edge spanning the bone boundary in a voxel. The other end Find the boundary intersection points using linear interpolation. : ; in These are the interpolation coefficients; CT values ​​at both ends of the edge ≥CT value threshold , <CT value threshold ; The three-dimensional physical coordinates of the two endpoints; Triangle face generation and mesh stitching: based on voxel encoding It retrieves the corresponding triangular face combination from the preset "face table", for example... The binary code 00000101 corresponds to two triangular facets, which are formed by connecting the intersection points of four isosurfaces. All facets are pieced together to form a combat trauma model of the skeleton, realizing the transformation from discrete coordinates to a continuous three-dimensional model. Finally, a combat trauma model containing key structures such as the maxilla, mandible, and zygomatic bone is constructed, providing an intuitive three-dimensional visualization tool for injury assessment, surgical planning, and other purposes.

[0028] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-driven virtual simulation treatment training system for oral and maxillofacial war injury, characterized in that, It comprises an image data processing module (100), a tissue segmentation module (200) and a war injury model construction and calling module (300), wherein: The image data processing module (100) perceives image data and divides the image data into a plurality of different image groups, and pre-processes each image data in the image groups; The tissue segmentation module (200) receives the pre-processed image groups, and then uses a medical image segmentation algorithm to identify all tissues in each image data, wherein the medical image segmentation algorithm comprises an encoding unit, a dense skip connection, a decoding unit and a deep supervision output, wherein the decoding unit restores the feature map size through transposed convolution upsampling, and fuses with the dense connection features of the encoding unit, and gradually generates a segmentation mask with the same size as the input image; The war injury model construction and calling module (300) converts the two-dimensional bone contour into a three-dimensional mesh model, i.e. a war injury model, based on the pixel coordinates of the segmentation mask through voxel division, isosurface extraction and mesh generation; receives the patient basic information of the image groups in the image data processing module (100), establishes a mapping relationship between the war injury model and the image groups, receives the input basic information, and then calls out the corresponding war injury model.

2. The AI-driven oral and maxillofacial war injury virtual simulation rescue training system according to claim 1, characterized in that: The image data processing module (100) comprises an image group division unit (110) and an image preprocessing unit (120), wherein: The image group division unit (110) is used for dividing a large number of real case image data into a plurality of different image groups; the image preprocessing unit (120) receives the image groups and pre-processes each image data in the image groups, which includes removing scanning noise in the image data and layer thickness normalization of the image data. 3.The AI-driven oromaxillofacial war injury virtual simulation rescue training system according to claim 2, characterized in that: The image group division unit (110) perceives a large number of real case image data with patient basic information, divides the large number of real case image data into a plurality of different image groups according to the patient basic information, each image group corresponds to a patient basic information, and the image group includes a large number of different image data.

4. The AI-driven oromaxillofacial war injury virtual simulation rescue training system according to claim 2, characterized in that: The image preprocessing unit (120) removes scanning noise from the image data: the neighborhood radius is set to... Take the pixels in the neighborhood and calculate the weight of the pixels in the neighborhood based on the distance of the pixel to the center pixel of the neighborhood. Then, a weighted average is used to calculate the values ​​of all pixels in the neighborhood and their corresponding weights. The sum of the products ,replace It is the pixel value of the center pixel in the neighborhood.

5. The AI-driven oromaxillofacial war injury virtual simulation rescue training system according to claim 4, characterized in that: The weight of any pixel point in the image data for the image pre-processing unit (120) is: ​​ ； Wherein: , is the horizontal and vertical offset of the pixel point within the neighborhood relative to the center pixel; For the standard deviation, the core parameter for controlling the filtering smoothing degree, the standard deviation The smaller, the weight is concentrated in the center, the weak denoising but the clear edge; the standard deviation The greater, the weight is diffused, the strong denoising but the fuzzy edge; is a normalization coefficient, ensuring that the sum of all weights is 1, avoiding overall brightness change of the filtered image data; The pixel value of the center pixel point in the neighborhood is: . 6.The AI-driven oromaxillofacial war injury virtual simulation rescue training system according to claim 5, characterized in that: The working principle of the layer thickness normalization in the image preprocessing unit (120) is as follows: setting a standard layer thickness , receiving the original layer thickness of the image data ; If the standard layer thickness < original layer thickness , a new fault is inserted between the two adjacent layers; if the standard layer thickness > original layer thickness , the adjacent faults are merged.

7. The AI-driven oromaxillofacial war injury virtual simulation rescue training system according to claim 3, characterized in that: The image preprocessing unit (120) inserts a new tomographic image between two tomographic images arranged in succession in the image data: setting an original slice thickness , perceiving adjacent tomographic images of the image data as and , and then calculating the number of tomographic images needed to be inserted between the slice and the slice , wherein is a floor symbol, is a target slice thickness for inserting a new tomographic image; inserted first pixel values of the layer image data , by and linearly weighted ； wherein is a two-dimensional coordinate of a pixel, and is a weight coefficient. 8.The AI-driven oromaxillofacial war injury virtual simulation rescue training system according to claim 7, characterized in that: The image preprocessing unit (120) merges adjacent layers: calculates the number of original layers that need to be merged between layers where is the ceiling symbol, ensuring that the total thickness after merging is not less than the standard layer thickness ; Computing structure saliency of each slice in the image data and normalizing the structure saliency of each slice to obtain the first weight of each slice wherein is the number of original slices merged, and the sum of the weights of all slices is 1; for the continuous layer original fault in the pixel, the pixel value of the merged pixel is the weighted average of each layer pixel value: , wherein is the pixel value of layer original fault in place.​ 9.The AI-driven oromaxillofacial war injury virtual simulation rescue training system according to claim 1, characterized in that: The medical image segmentation algorithm in the tissue segmentation module (200) comprises: The encoding unit gradually reduces the feature map size through convolution blocks and maximum pooling to extract hierarchical features from low-level features to high-level features; Each layer extracts hierarchical features by 2 times 3*3 convolution, batch normalization, and ReLU activation. The layer is the first level feature map is: ； wherein, are convolution kernel weights and bias; is a batch normalization operation, in particular: , , are batch mean and variance, is a feature map, , are learnable parameters; is a ReLU activation function, introducing non-linearity: ; Max pooling: reduce feature map size by 2x2 max pooling after every 2 convolutional blocks: where is a 2x2 max pooling operation that takes the maximum value of the feature map in a 2x2 window, reducing the feature map size, denotes taking the maximum value of the feature map in a 2x2 window. dense skip connection: a dense connection is established between the encoding unit stage and the decoding unit stage , the feature map of the encoding unit stage is concatenated with the feature map of the decoding unit stage , wherein the feature map of the decoding unit stage is a feature map of a shallower layer in the decoding unit; The decoding unit recovers the feature map size by transposed convolution and fuses with the dense connection features of the encoding unit to gradually generate a segmentation mask with the same size as the input image. Specifically, the feature map size is expanded by 2 times through 2*2 transposed convolution, and the output feature map of the first stage and the dth layer is wherein * is a transposed convolution operation, , , are the transposed convolution weights and bias, respectively. After transposed convolution, the corresponding hierarchical dense connection features of the encoding unit are spliced, and then processed through convolution blocks; Deep supervision output: finally convert the feature map channel number to the class number through 1x1 convolution, and output the pixel-level classification probability through Softmax activation; The deep supervision is that multiple output layers are set at different depths of the decoding unit, each output layer is supervised by an independent loss function, and the total loss is a weighted sum of the losses of the layers: ; wherein is the decoding unit depth; w is the weight; Loss for the first layer output: ; P is the probability of the pixel being masked out by the predicted mask; M is the true mask; N is the total number of pixels; : Avoid division by zero.

10. The AI-driven oromaxillofacial war injury virtual simulation rescue training system according to claim 1, characterized in that: The specific working principle of establishing a war injury model is: Coordinate transformation: Transforming the two-dimensional coordinates of bone pixels in the segmentation mask. Combining the slice thickness information from the tomographic images, it is converted into three-dimensional physical coordinates, specifically: Receiving the two-dimensional coordinate of the certain bone pixel in the k-th tomographic image as , the tomographic sequence number as , the three-dimensional physical coordinate is: ； wherein, is the X, Y axis pixel physical size, is the layer thickness, is the global origin coordinate; Encoding of the vertex state of the voxel: divide the three-dimensional space into voxels of 1mm x 1mm x 1mm, 8 vertices per voxel, set the CT value threshold Determine whether the vertex is in the bone: ； wherein CT value of the vertex; vertex state; 8 vertex states make up an 8-bit binary code: ; Isosurface intersection point calculation: for edges in a voxel that cross a bone boundary, the boundary intersection point is found by linear interpolation : ； wherein is an interpolation coefficient; is the CT value of the two end points of the edge; is the three-dimensional physical coordinates of the two end points. Triangular patch generation and mesh stitching: according to voxel encoding The combat wound model is constructed by calling corresponding triangular patch combination from the preset patch table.

Images