Spleen and liver trauma model establishing method based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence, and discloses an artificial intelligence-based spleen and liver trauma model establishment method, which comprises the following steps of: S01, acquiring target data; step S02, obtaining research data; s03, inputting research data into the improved U-Net network model, and performing organ segmentation to obtain a target organ contour; s04, performing three-dimensional reconstruction to obtain a trauma model corresponding to the target organ, and performing hemodynamic analogue simulation; through the step S03, construction of the improved U-Net network model is facilitated, compared with a traditional U-Net network model, the improved U-Net network model can significantly improve the accuracy and robustness of organ segmentation, the effectiveness and accuracy of the spleen and liver contour are enhanced, and an effective foundation is laid for subsequent three-dimensional modeling based on the spleen and liver contour.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and more particularly to a method for establishing a spleen and liver trauma model based on artificial intelligence. Background Art

[0002] The spleen and liver are important solid organs in the human body, and trauma may lead to serious complications; in clinical treatment, timely and accurate assessment of the degree of spleen and liver trauma is of great significance for the formulation of treatment plans and the improvement of patient prognosis.

[0003] A method for establishing an artificial intelligence-assisted ultrasonic diagnosis spleen and liver trauma model is disclosed in the published document with the publication number of CN116188424A. By using the method of data enhancement and generation, the ultrasonic images of animal spleen and liver trauma are enhanced into a strong data set closer to the clinical data of human spleen and liver trauma; then, a large number of enhanced ultrasonic images of spleen and liver trauma are used to create a source model, and the source model is fine-tuned with a small amount of human spleen and liver trauma data; finally, an automatic segmentation model for ultrasonic images of spleen and liver trauma and a grade classification model for ultrasonic images of spleen and liver trauma are established. This model analyzes the images generated during the ultrasonic diagnosis of spleen and liver trauma, identifies the anatomical structure contours of the spleen and liver and the positions of spleen and liver trauma in the segmented images, and obtains the grade classification results, showing good performance on the clinical ultrasonic image data set of spleen and liver trauma.

[0004] However, in the process of obtaining the strong data set, a large number of ultrasonic images of animal spleen and liver trauma are required. And in the process of obtaining the ultrasonic images of animal spleen and liver trauma, a large amount of human resources are consumed, and the spleen and liver trauma conditions vary. Therefore, when using a large number of ultrasonic images of animal spleen and liver trauma to fine-tune the trauma clinical images with a small amount of human spleen and liver trauma data, the similarity of trauma conditions cannot be guaranteed, and thus the fine-tuned trauma ultrasonic images cannot represent the actual trauma conditions that will occur. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method for establishing a spleen and liver trauma model based on artificial intelligence to solve the problems existing in the above background art.

[0006] The present invention provides the following technical solutions: A method for establishing a spleen and liver trauma model based on artificial intelligence, comprising the following steps: Step S01: Obtain target data; the target data is divided into spleen target data and liver target data; the spleen target data is used to construct a spleen trauma model, and the liver target data is used to construct a liver trauma model; Step S02: Obtain research data. Use the image data after multi-modal image fusion of CT images and ultrasound images as research data, and the research data is divided into spleen research data and liver research data; Step S03: Input the research data in Step S02 into the improved U-Net network model for organ segmentation to obtain the outline of the target organ; Step S04: Based on the obtained outline of the target organ, perform three-dimensional reconstruction to obtain the trauma model corresponding to the target organ and conduct hemodynamic simulation; when the target organ is the spleen, the corresponding trauma model is the spleen trauma model, and when the target organ is the liver, the corresponding trauma model is the liver trauma model.

[0007] Preferably, the spleen target data is specifically the CT images and ultrasound images of spleen trauma patients; the liver target data is specifically the CT images and ultrasound images of liver trauma patients; the research data is the input data of the spleen and liver trauma model for training the spleen and liver trauma model; the image data after fusion of the spleen target data is the spleen research data and serves as the training data of the spleen trauma model, and the image data after fusion of the liver target data is the liver research data and serves as the training data of the liver trauma model.

[0008] Preferably, the specific method of multi-modal image fusion is as follows: Preprocess the CT images, including bias field correction and image enhancement; Preprocess the ultrasound images, including elastic parameter normalization and motion artifact suppression; Perform rigid registration and elastic registration on the preprocessed CT images and ultrasound images; Perform time synchronization to align the three-phase data of the CT images with the imaging time axis of the ultrasound images.

[0009] Preferably, the improved U-Net network model mainly includes an input layer, an encoder, a decoder, an attention mechanism, a comprehensive loss function, and an output layer; The encoder is responsible for extracting features from the input image, gradually compressing the spatial dimension of the image through multiple convolutional layers and pooling layers to extract high-level features, and introducing residual connections in the encoder; The decoder restores the spatial dimension of the image through deconvolution to reconstruct the segmented image, and the skip connection in the decoder transfers the low-level features of the encoder to the decoder; The attention mechanism includes channel attention and spatial attention; The output layer uses the Sigmoid activation function to generate a binary mask.

[0010] Preferably, the residual connection is expressed as: ; where Represents the input feature map of the residual connection, represents the output, represents the residual feature; If the input feature map of the channel attention is , after global average pooling and global max pooling, a channel descriptor is generated, and the channel weight is generated through a shared MLP , and the formula is expressed as: ; Among them, represents the generated channel weight, represents the activation function, is the representation of the fully connected layer, represents global average pooling, represents global max pooling; After passing through the channel attention, the output is ; As the input feature map of the spatial attention, the global average pooling mean and the global max pooling maximum are calculated along the channel dimension, and after concatenation, a spatial weight is generated by convolution , and the formula is expressed as: ; Among them, represents the generated spatial weight, represents a 7×7 convolution kernel; global average pooling, represents global max pooling; represents concatenation, that is, concatenating the global average pooling result and the global max pooling result; After passing through the spatial attention, the output is .

[0011] Preferably, the comprehensive loss function is expressed as: ; Among them, represents the comprehensive loss, represents the loss, represents the cross-entropy loss, and are the corresponding weight coefficients respectively, and both satisfy being greater than 0 and less than 1; ; Among them, represents the predicted probability of the th pixel point, represents the The true label of each pixel point, i.e., the pre-annotated value; represents the smoothing coefficient, and the value range is ; is the total number of pixels, ; ; among them, represents the base of the logarithm.

[0012] Preferably, the specific method for training the improved U-Net network model is as follows: When constructing the spleen trauma model, the spleen research data is used as the input data of the improved U-Net network model. The spleen research data is divided into a training set, a validation set, and a test set. 80% of the spleen research data is divided into the training set. Among the remaining 20%, 10% of the spleen research data is divided into the validation set, and 10% of the spleen data is divided into the test set; When constructing the liver trauma model, the liver research data is used as the input data of the improved U-Net network model. The liver research data is divided into a training set, a validation set, and a test set. 80% of the liver research data is divided into the training set. Among the remaining 20%, 10% of the liver research data is divided into the validation set, and 10% of the liver data is divided into the test set; The training set is input into the improved U-Net network model. The model learns from the training set and adjusts its internal parameters to minimize the training error. During the model training process, the validation set is regularly input into the trained model to evaluate the model performance, and the hyperparameters of the model are adjusted in a timely manner. The test set is used to evaluate the generalization ability of the trained model to verify the performance of the model on unseen data.

[0013] Preferably, the specific method for performing 3D reconstruction in step S04 to obtain the trauma model corresponding to the target organ is as follows: The surface of the target organ contour is converted into a 3D triangular mesh surface by using a surface rendering algorithm; The 3D triangular mesh is optimized to construct an initial mesh model; Locate the trauma position, mark the trauma position on the surface of the initial mesh model to obtain the trauma area; Perform hemodynamic simulation on the trauma area.

[0014] Preferably, the main steps for obtaining the 3D triangular mesh surface are as follows: After the research data is output by the improved U-Net network model, it can be regarded as a binary mask sequence. The binary mask of the target organ contour is 1, and the binary mask of the non-target organ contour is 0. The binary mask sequence is stacked into a cubic grid, and the voxel value of the target organ is 1, and the voxel value of the non-target organ, i.e., the background, is 0; Preset isosurface threshold , traverse each voxel cube, and generate triangular patches according to the relationship between the vertex values of the cube and ; Connect the triangular patches generated by adjacent voxels to form a continuous surface, and obtain a three-dimensional triangular mesh surface.

[0015] Preferably, the hemodynamic simulation can be expressed by the Navier-Stokes equation formula as: ; Among them, represents the fluid density, represents the velocity vector field of the fluid, indicating the velocity direction and magnitude of the fluid at each point in space; represents time, represents the pressure field, represents the dynamic viscosity, which characterizes the viscosity of the fluid, that is, the internal friction force between adjacent fluid layers; represents the body force per unit mass, is the gradient operator, indicating the directional derivative in space.

[0016] Technical effects and advantages of the present invention: By setting step S03, the present invention is beneficial to constructing an improved U-Net network model. By introducing residual connections in the encoder and embedding an attention mechanism at the skip connection, first dynamically allocate channel weights through channel attention to enhance the target feature response, and then focus on the key spatial regions through spatial attention, such as the boundaries of the spleen and liver; it can fuse multi-scale features, improve the segmentation accuracy, and lay a foundation for the subsequent construction of the three-dimensional model of the spleen and liver; and combine cross-entropy loss with loss to construct a comprehensive loss function. Compared with the traditional U-Net network model, the improved U-Net network model can significantly improve the accuracy and robustness of organ segmentation, enhance the effectiveness and accuracy of the spleen and liver contours, and lay an effective foundation for subsequent three-dimensional modeling based on the spleen and liver contours.

[0017] By setting step S04, the present invention is beneficial to performing three-dimensional reconstruction based on the obtained target organ contour to obtain a trauma model corresponding to the target organ, and performing hemodynamic simulation. The CT blood perfusion parameters in the CT image can be integrated into the Navier-Stokes equation, which can better realize the modeling from image data to hemodynamic simulation, so as to improve the accuracy of hemodynamic simulation modeling; use a large amount of known data of trauma patients as training data samples, without using animals as samples, reducing the human resource cost, and improving the accuracy of the training data samples, avoiding the differences between animals and human trauma situations, and improving the accuracy of trauma model establishment. Brief description of the drawings

[0018] Figure 1 This is a flowchart of the method for establishing a spleen and liver trauma model based on artificial intelligence according to the present invention. Specific embodiments

[0019] Next, the technical solutions in the present invention will be clearly and completely described in conjunction with the accompanying drawings in the present invention. In addition, the forms of the various structures described in the following embodiments are merely illustrative. A method for establishing a spleen and liver trauma model based on artificial intelligence according to the present invention is not limited to the various structures described in the following embodiments. All other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0020] As Figure 1 shown, the present invention provides a method for establishing a spleen and liver trauma model based on artificial intelligence, including the following steps: Step S01: Obtain target data; the target data is divided into spleen target data and liver target data. The spleen target data is used to construct a spleen trauma model, and the liver target data is used to construct a liver trauma model; the spleen target data is specifically the CT images and ultrasound images of spleen trauma patients; the liver target data is specifically the CT images and ultrasound images of liver trauma patients; trauma causes and descriptions of the patients' trauma symptoms can also be additionally collected. The trauma cause is the cause of spleen or liver trauma, and the trauma symptom description is the description of the symptoms of spleen and liver patients after spleen and liver trauma, including but not limited to persistent dull pain, persistent sharp pain, obvious tenderness accompanied by rebound pain, dizziness, fatigue, and cold sweats, etc.; the target data can be collected using big data technology; Step S02: Obtain research data. The image data after multi-modal image fusion of CT images and ultrasound images is used as the research data. The research data is divided into spleen research data and liver research data. The research data is the input data of the spleen and liver trauma model and is used to train the spleen and liver trauma model. The image data after fusion of spleen target data is used as the spleen research data and serves as the training data for the spleen trauma model. The image data after fusion of liver target data is used as the liver research data and serves as the training data for the liver trauma model. Multi-modal image fusion of CT images and ultrasound images can effectively improve the accuracy of the image data and provide reliable basic data for generating the trauma model. Since CT images are usually high-resolution tomographic scans that can clearly display the anatomical structure of organs, especially under enhanced scans, the contrast between blood vessels and tissues is very good. However, the disadvantage of CT is that the radiation dose is relatively high and real-time imaging is not possible. Ultrasound images, on the other hand, are real-time and non-invasive, compensating for the deficiencies of CT images. However, ultrasound images are not clear enough for deep tissues such as the spleen, but CT images can make up for this deficiency. Therefore, multi-modal image fusion of CT images and ultrasound images can obtain more accurate and effective spleen and liver trauma conditions, making the trauma model more accurate and effective during the establishment process. Step S03: Input the fused image data in Step S02 into the improved U-Net network model for organ segmentation to obtain the contour of the target organ. The fused image data is pre-annotated as the training label. The pixel points of the target organ are pre-annotated as 1, and the background pixel points are pre-annotated as 0. The target organ is the spleen or the liver. When constructing the spleen trauma model, the target organ is the spleen, and when constructing the liver trauma model, the target organ is the liver. Step S04: Based on the obtained contour of the target organ, perform three-dimensional reconstruction to obtain the corresponding trauma model of the target organ. When the target organ is the spleen, the corresponding trauma model is the spleen trauma model, and when the target organ is the liver, the corresponding trauma model is the liver trauma model.

[0021] In this embodiment, it should be specifically noted that the specific method of multi-modal image fusion is as follows: Preprocess the CT image, including bias field correction and image enhancement. The bias field correction uses the N4 bias field correction algorithm to eliminate the gray level deviation caused by magnetic field inhomogeneity in the CT image and enhance the contrast between the spleen and liver parenchyma and the bleeding area. The N4 bias field correction algorithm is used to correct the brightness or contrast inhomogeneity (i.e., the bias field) caused by equipment or scanning conditions in the image. The image enhancement is specifically the enhancement of blood vessels and bleeding areas. Based on the Hessian matrix filtering, the spleen and liver blood vessel network is extracted to enhance the linear low-density shadow characteristics in the laceration area. Preprocess the ultrasound image, including elastic parameter normalization and motion artifact suppression; the elastic parameter normalization linearly maps the ultrasound strain ratio to the range of 0 to 1 to generate an elastic feature map matching the CT spatial resolution; the motion artifact suppression uses median filtering to eliminate the blurring of elastography caused by respiratory motion and retains the serrated edge features of the capsule tear; the ultrasound strain ratio is expressed as: ; where represents the ultrasound strain ratio, represents the tissue deformation amount in the trauma area, which can be measured by the method of gently placing the ultrasound probe; represents the deformation amount in the normal area of the same organ, reflecting the tissue hardness difference; when it is severe injury; Perform rigid registration and elastic registration on the preprocessed CT image and ultrasound image. The rigid registration selects a rigid registration reference and realizes the global spatial alignment of the CT image and the ultrasound image through affine transformation; the specific rigid registration reference is: if it is a spleen CT image and an ultrasound image, select the splenic portal vein bifurcation point and the highest point on the outer edge of the spleen as the rigid registration reference; if it is a liver CT image and an ultrasound image, select the hepatic portal vein bifurcation point and the main trunk of the right hepatic vein as the rigid registration reference; the elastic registration uses the SyN (Symmetric Normalization) algorithm, optimizes the spatial deformation field of the CT image and the ultrasound image with mutual information as the similarity measure, and realizes the pixel-level alignment of the spleen parenchyma or liver parenchyma and blood vessels; Perform time synchronization to align the three-phase data of the CT image with the imaging time axis of the ultrasound image, and use linear interpolation to compensate for the acquisition time difference of different modalities , ; the three-phase CT data are arterial phase data, portal phase data, and delayed phase data.

[0022] In this embodiment, it should be specifically noted that the N4 bias field correction algorithm decomposes the image into the product of the real signal and the bias field through iterative optimization. The real signal is the tissue contrast, and the bias field is the low-frequency band interference; the formula is expressed as: ; where represents the CT image to be preprocessed, represents the bias field, represents the real signal; estimate the bias field by maximizing the image entropy or minimizing the local histogram difference, and use B-spline basis functions to smooth the bias field to avoid overfitting; For a certain pixel point in the two-dimensional CT image , the Hessian matrix is composed of second-order partial derivatives and is expressed as: ; where represents the two-dimensional CT image The Hessian matrix of a certain pixel in denotes the second-order partial derivative of a certain pixel in the x direction in , reflecting the curvature in the horizontal direction; denotes the second-order partial derivative of a certain pixel in the y direction in , reflecting the curvature in the vertical direction; denotes the mixed partial derivative of a certain pixel in the x direction and the y direction in , reflecting the correlation between directions; if directly calculating the second-order derivative is sensitive to noise, the two-dimensional CT image can be first subjected to multi-scale Gaussian smoothing using a Gaussian kernel, and each scale corresponds to blood vessels of different diameters, and then the Hessian matrix is calculated for the smoothed image; The value of is usually set to 1 - 2 times the blood vessel radius, and the specific set value can be specifically set by those skilled in the art according to the actual situation; Calculate the eigenvalues of the Hessian matrix and , ; in the CT image, blood vessels are tubular structures. For tubular structures, for bright blood vessels in a dark background, the eigenvalues of the Hessian matrix satisfy: , ; for dark blood vessels in a bright background, the eigenvalues of the Hessian matrix satisfy: , ; enhance the tubular structure through a response function, and the response function is expressed as: ; wherein, denotes the response function, denotes the suppression parameter, controlling the sensitivity to , suppressing non-tubular structures, such as plaques and noise; denotes the enhancement parameter, controlling the sensitivity to , enhancing the axial flatness of the tubular structure; , , in the actual numerical settings of the suppression parameter and the enhancement parameter, it can be set according to the specific situation of the image, adjusted according to the image contrast. For example, a high-noise image may require a larger enhancement parameter; if the eigenvalues of the Hessian matrix and are the eigenvalues of the r-th pixel, then the response function is the response function of the r-th pixel; Take the maximum response value among all scales, set a threshold , and retain the region where the maximum response value is greater than the set threshold as the blood vessel candidate region, suppress non-local maxima along the blood vessel direction, and refine the blood vessel centerline; Effectively enhance the tubular structures in medical images, providing a basis for subsequent analysis.

[0023] In this embodiment, it should be specifically noted that the improved U-Net network model mainly includes an input layer, an encoder, a decoder, an attention mechanism, a comprehensive loss function, and an output layer; The encoder is responsible for extracting features from the input image, gradually compressing the spatial dimension of the image through multiple convolutional layers and pooling layers, and extracting high-level features. Residual connections are introduced in the encoder to alleviate the problem of gradient disappearance and enhance the feature reuse ability; the decoder restores the spatial dimension of the image through deconvolution and reconstructs the segmented image. The skip connections in the decoder transfer the low-level features of the encoder to the decoder to help restore finer segmentation boundaries; the attention mechanism includes channel attention and spatial attention. The channel attention dynamically adjusts the channel weights through the SE module to enhance the response of the lesion area, and the spatial attention aggregates the boundary-blurred areas to suppress background interference; the output layer uses the Sigmoid activation function to generate a binary mask, with the binary mask of the background assigned a value of 0 and the binary mask of the target organ assigned a value of 1; an attention mechanism is embedded at the skip connection, first dynamically allocating channel weights through channel attention to enhance the response of target features, and then focusing on key spatial regions through spatial attention, such as the boundaries of the spleen and liver; it can fuse multi-scale features, improve the segmentation accuracy, and lay a foundation for the subsequent construction of the three-dimensional model of the spleen and liver; If the input feature map of the channel attention is , after global average pooling and global max pooling, a channel descriptor is generated, and the channel weights are generated through a shared MLP , and the formula is expressed as: ; Among them, represents the generated channel weights, represents the activation function, is the representation of the fully connected layer, represents global average pooling, calculating the global average value of each channel along the spatial dimension, represents global max pooling, calculating the global maximum value of each channel along the spatial dimension; After passing through the channel attention, the output is : ; is the input feature map of the spatial attention. Calculate the global average pooling mean and global max pooling maximum along the channel dimension, and splice them and then convolve to generate the spatial weights , and the formula is expressed as: ; Among them, Represents The generated spatial weight Represents a 7×7 convolutional kernel used to locate key regions of the target organ, such as the spleen edge and liver blood vessels Global average pooling, taking the mean along the channel dimension Represents global max pooling, taking the maximum along the channel dimension Represents concatenation, that is, concatenating the global average pooling result and the global max pooling result The output after spatial attention is : ; The residual connection is represented as: ; where Represents the input feature map of the residual connection Represents the output Represents the residual feature; the residual connection directly passes the input to the deep network, avoiding the gradual attenuation of the gradient during backpropagation, improving the training stability of the deep network, alleviating the vanishing gradient, adding the shallow features such as organ edge texture and the deep features through residual addition to achieve multi-scale feature fusion, enhancing the model's expression ability for complex structures, introducing the residual connection allows for the construction of deeper encoders, extracting more abstract features by stacking residual blocks to improve the segmentation accuracy The comprehensive loss function is represented as: ; where Represents the comprehensive loss Represents Loss Represents the cross-entropy loss And Are the corresponding weight coefficients respectively And Both take values that satisfy being greater than 0 and less than 1. The specific values can be set or modified by those skilled in the art according to actual needs. In this embodiment, take , ; to balance the region overlap and classification accuracy; the cross-entropy loss can optimize the local pixel classification The loss can optimize the global region coverage. Combining the two to obtain the comprehensive loss function can improve the sharpness of the segmentation boundary while ensuring the integrity of the overall organ, and the cross-entropy is sensitive to noise, while The loss is sensitive to the region shape. Jointly training the two to obtain the comprehensive loss function can balance the noise interference and structural error, enhancing the model's robustness ; where Represents the predicted probability of the th pixel point Represents the The true label of each pixel, i.e., the pre-annotated value; Denotes the smoothing coefficient, which is a small constant to prevent the denominator from being zero, and the value is ; Is the total number of pixels, ; The loss directly optimizes the overlap ratio between the predicted region and the true region, alleviating the problem of missed segmentation caused by the small volume of organs. Since the spleen and liver only occupy a small part of the region in some medical images, the background pixels are far more than the target pixels. Through The loss can be normalized to prevent the model from biasing towards the majority class; ; Among them, Denotes the base of the logarithm, which can take , or , Is the natural constant, The value of Depends on the specific application scenario and requirements, and can be set by those skilled in the art. However, The value of Has little impact on the actual meaning of the cross-entropy loss function. When using the cross-entropy loss function for training, the goal is to minimize the loss function, The value of

[0024] In this embodiment, it should be specifically noted that the specific training method of the improved U-Net network model is as follows: When constructing the spleen trauma model, the spleen research data is used as the input data of the improved U-Net network model. The spleen research data is divided into a training set, a validation set, and a test set. 80% of the spleen research data is divided into the training set, and among the remaining 20%, 10% of the spleen research data is divided into the validation set, and 10% of the spleen data is divided into the test set; When constructing a liver trauma model, liver research data is used as the input data for the improved U-Net network model. The liver research data is divided into a training set, a validation set, and a test set. 80% of the liver research data is divided into the training set. Among the remaining 20%, 10% of the liver research data is divided into the validation set, and 10% of the liver data is divided into the test set; The training set is input into the improved U-Net network model. The model learns from the training set and adjusts its internal parameters to minimize the training error. During the model training process, the validation set is regularly input into the trained model to evaluate the model performance, and the hyperparameters of the model are adjusted in a timely manner to improve the generalization ability of the model. The model performance can be evaluated through performance indicators such as accuracy, precision, and recall. The test set is used to evaluate the generalization ability of the trained model to verify the performance of the model on unseen data to ensure the generalization ability of the model.

[0025] In this embodiment, it should be specifically noted that the specific method for performing 3D reconstruction to obtain the trauma model corresponding to the target organ in step S04 is as follows: The surface rendering algorithm is used to convert the target organ contour into a 3D triangular mesh surface; the surface rendering algorithm uses the Marching Cubes algorithm; The 3D triangular mesh is optimized to construct an initial mesh model; the purpose is to eliminate surface jaggedness and staircase artifacts caused by voxelization through optimization, improve the visual quality of the 3D model, reduce the number of mesh patches through mesh simplification, reduce the subsequent computational amount, and at the same time retain key shape features; The trauma location is located, and the trauma location is marked on the surface of the initial mesh model to obtain the trauma area; the purpose is to determine the deformation range, simulate the trauma morphology, and lay a foundation for subsequent hemodynamic simulation; Hemodynamic simulation is performed on the trauma area; the purpose is to evaluate the impact of the trauma on the function of the target organ, such as the bleeding risk, through the simulation situation, and provide a visual basis for medical staff to optimize the surgical plan.

[0026] In this embodiment, it should be specifically noted that the main steps for obtaining the 3D triangular mesh surface are as follows: After the research data is output by the improved U-Net network model, it can be regarded as a binary mask sequence. The binary mask of the target organ contour is 1, and the binary mask of the non-target organ contour is 0. The binary mask sequence is stacked into a cubic grid, and the voxel value of the target organ is 1, and the voxel value of the non-target organ, that is, the background, is 0; Preset an isosurface threshold , traverse each voxel cube, and generate triangular patches according to the relationship between the cube vertex values and ; the The value can be specifically set by those skilled in the art according to the actual situation to meet , and in this embodiment, is selected; the isosurface threshold determines the position where the surface passes through the voxel, indicates the intermediate position where the triangular mesh surface is located inside and outside the voxel, When it increases, the surface biases towards the inside of the organ, When it decreases, the surface biases towards the outside of the organ; Connect the triangular patches generated by adjacent voxels to form a continuous surface, and obtain a three-dimensional triangular mesh surface; The optimization process of the three-dimensional triangular mesh includes: denoising and smoothing the mesh, mesh subdivision, and mesh simplification; the denoising and smoothing of the mesh can adopt Laplacian smoothing, and move each vertex of the three-dimensional triangular mesh towards the average value of its neighboring vertices. The formula is expressed as: ; where, represents the coordinate of the b-th vertex after smoothing; represents the original coordinate of the b-th vertex, that is, the coordinate before the smoothing operation; represents the smoothing coefficient, satisfying , and is used to control the movement amplitude; represents the set of neighboring vertices of the b-th vertex, represents the coordinate of the d-th vertex in the set of neighboring vertices; in this embodiment, is taken, which can balance denoising and shape preservation; The Loop subdivision algorithm can be used for mesh subdivision to encrypt local areas with high curvature such as the edge of the trauma to improve the resolution; the edge collapse mesh simplification algorithm can be used for mesh simplification to reduce the number of patches in non-critical areas; both the Loop subdivision algorithm and the edge collapse mesh simplification algorithm are existing technologies, and no technical improvement is made in this embodiment, so they will not be elaborated in this embodiment; The trauma position can be detected and marked by obtaining the surface Gaussian curvature and mean curvature. If the Gaussian curvature or mean curvature is greater than the determination threshold, then this area is the trauma area; the determination threshold can be set or modified by those skilled in the art according to the actual situation. In this embodiment, the value of the determination threshold is selected as 0.1; The formula for the Gaussian curvature is expressed as: ; ; where, represents the Gaussian curvature, represents the mean curvature, and are respectively the radii of the two principal curvatures; The hemodynamic simulation can be expressed by the Navier - Stokes equation formula as: ; Among them, represents the fluid density, represents the flow velocity vector field, indicating the velocity direction and magnitude of the fluid at each point in space; represents time, represents the pressure field, represents the dynamic viscosity, characterizing the viscosity of the fluid, that is, the internal frictional force between adjacent fluid layers; represents the body force per unit mass, is the gradient operator, representing the directional derivative in space.

[0027] Embodiment 2 When performing three-dimensional reconstruction in step S04 to obtain the trauma model corresponding to the target organ, performing hemodynamic simulation on the trauma area can integrate the CT perfusion parameters in the CT image into the Navier-Stokes equation, better realizing the modeling from image data to hemodynamic simulation, so as to improve the accuracy of hemodynamic simulation modeling; Introduce a source term Y in the continuity equation to characterize the volume change rate caused by perfusion; the continuity equation is: ; After introducing the source term, the continuity equation is expressed as: ; Among them, ; Among them, Q represents the blood perfusion volume, represents the tissue volume of the trauma area; ; Among them, represents the change amount of CT value in the trauma area, represents the time interval corresponding to the change amount of CT value, represents the attenuation coefficient, which takes the value of ; When modeling the spleen and liver trauma model, if the established model is a spleen model, spleen-related data is used for modeling, and if the established model is a liver model, liver-related data is used for modeling; during the modeling process of the spleen and liver trauma model, except for the different data used, the modeling process and the technical means used are the same.

[0028] Finally: The above are only the preferred embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0029] As described above, it is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the said claims.

Claims

1. A method for establishing a spleen and liver trauma model based on artificial intelligence, characterized in that: It includes the following steps: Step S01: Obtain target data; the target data is divided into spleen target data and liver target data; The spleen target data is used to construct a spleen trauma model, and the liver target data is used to construct a liver trauma model; Step S02: Obtain research data, and use the image data after multi-modal image fusion of CT images and ultrasound images as research data. The research data is divided into spleen research data and liver research data; Step S03: Input the research data in Step S02 into the improved U-Net network model for organ segmentation to obtain the contour of the target organ; Step S04: Based on the obtained contour of the target organ, perform three-dimensional reconstruction to obtain the trauma model corresponding to the target organ, and perform hemodynamic simulation; when the target organ is the spleen, the corresponding trauma model is the spleen trauma model, and when the target organ is the liver, the corresponding trauma model is the liver trauma model.

2. The method for establishing a spleen and liver trauma model based on artificial intelligence according to claim 1, wherein: The spleen target data is specifically the CT images and ultrasound images of spleen trauma patients; the liver target data is specifically the CT images and ultrasound images of liver trauma patients; the research data is the input data of the spleen and liver trauma models for training the spleen and liver trauma models; the image data after fusing the spleen target data is the spleen research data and serves as the training data of the spleen trauma model, and the image data after fusing the liver target data is the liver research data and serves as the training data of the liver trauma model.

3. The method for establishing a spleen and liver trauma model based on artificial intelligence according to claim 2, characterized in that: The specific method of the multi-modal image fusion is as follows: Perform preprocessing on the CT images, including bias field correction and image enhancement; Perform preprocessing on the ultrasound images, including elastic parameter normalization processing and motion artifact suppression processing; Perform rigid registration and elastic registration on the preprocessed CT images and ultrasound images; Perform time synchronization to align the three-phase data of the CT images with the imaging time axis of the ultrasound images.

4. A method for establishing a spleen and liver trauma model based on artificial intelligence according to claim 3, characterized in that: The improved U-Net network model mainly includes an input layer, an encoder, a decoder, an attention mechanism, a comprehensive loss function, and an output layer; The encoder is responsible for extracting features from the input image, gradually compressing the spatial dimension of the image through multiple convolutional layers and pooling layers to extract high-level features, and introducing residual connections in the encoder; The decoder restores the spatial dimension of the image through deconvolution to reconstruct the segmented image, and the skip connection in the decoder passes the low-level features of the encoder to the decoder; The attention mechanism includes channel attention and spatial attention; The output layer uses the Sigmoid activation function to generate a binary mask.

5. A method for establishing a spleen and liver trauma model based on artificial intelligence according to claim 4, characterized in that: The residual connection is expressed as: ; where represents the input feature map of the residual connection, represents the output, represents the residual feature; If the input feature map of the channel attention is , after global average pooling and global max pooling, a channel descriptor is generated, and the channel weights are generated through a shared MLP , which is expressed by the formula: ； Among them, represents the generated channel weights, represents the activation function, is the representation of the fully connected layer, represents global average pooling, represents global max pooling; The output after channel attention is ; As the input feature map of spatial attention, calculate the mean value of global average pooling and the maximum value of global maximum pooling along the channel dimension, and generate spatial weights after concatenation and convolution , which is expressed by the formula: ； Among them, denotes the generated spatial weight, denotes a 7×7 convolutional kernel; global average pooling, denotes global max pooling; denotes concatenation, that is, concatenating the global average pooling result and the global max pooling result; the output after spatial attention is .

6. The method for establishing a spleen and liver trauma model based on artificial intelligence according to claim 5, wherein: The comprehensive loss function is expressed as: ; where represents the comprehensive loss, represents the loss, represents the cross-entropy loss, and are the corresponding weight coefficients respectively, and both satisfy the condition of being greater than 0 and less than 1. ; among them, represents the predicted probability of the th pixel point, represents the true label of the th pixel point, that is, the pre-annotated value; represents the smoothing coefficient, and the value is ; is the total number of pixels, ; ; wherein, represents the base of the logarithm.

7. A method for establishing a spleen and liver trauma model based on artificial intelligence according to claim 6, characterized in that: The specific method for training the improved U-Net network model is as follows: When constructing the spleen trauma model, use the spleen research data as the input data of the improved U-Net network model. Divide the spleen research data into a training set, a validation set, and a test set. Divide 80% of the spleen research data into the training set, and among the remaining 20%, 10% of the spleen research data is divided into the validation set, and 10% of the spleen data is divided into the test set; When constructing a liver trauma model, the liver research data is used as the input data for the improved U-Net network model. The liver research data is divided into a training set, a validation set, and a test set. 80% of the liver research data is divided into the training set. Among the remaining 20%, 10% of the liver research data is divided into the validation set, and 10% of the liver data is divided into the test set; The training set is input into the improved U-Net network model. The model learns from the training set and adjusts its internal parameters to minimize the training error. During the model training process, the validation set is regularly input into the trained model to evaluate the model performance, and the hyperparameters of the model are adjusted in a timely manner. The test set is used to evaluate the generalization ability of the trained model to verify the performance of the model on unseen data.

8. A method for establishing a spleen and liver trauma model based on artificial intelligence according to claim 7, characterized in that: The specific method for performing 3D reconstruction to obtain the trauma model corresponding to the target organ in step S04 is as follows: The surface of the target organ contour is converted into a 3D triangular mesh surface using a surface rendering algorithm; The 3D triangular mesh is optimized to construct an initial mesh model; The trauma location is located, and the trauma location is marked on the surface of the initial mesh model to obtain the trauma area; Hemodynamic simulation is performed on the trauma area.

9. A method for establishing a spleen and liver trauma model based on artificial intelligence according to claim 8, characterized in that: The main steps for obtaining the 3D triangular mesh surface are as follows: After the research data is output by the improved U-Net network model, it can be regarded as a binary mask sequence. The binary mask of the target organ contour is 1, and the binary mask of the non-target organ contour is 0. The binary mask sequence is stacked into a cubic grid, and the voxel value of the target organ is 1, and the voxel value of the non-target organ, that is, the background, is 0; Preset isosurface threshold , traverse each voxel cube, and generate triangular patches according to the relationship between the vertex values of the cube and . The triangular patches generated by adjacent voxels are connected to form a continuous surface to obtain the 3D triangular mesh surface.

10. The method for establishing a spleen and liver trauma model based on artificial intelligence according to claim 9, characterized in that: The hemodynamic simulation can be expressed by the Navier-Stokes equation formula as: ； Among them, represents the fluid density, represents the velocity vector field, indicating the velocity direction and magnitude of the fluid at each point in space; represents time, represents the pressure field, represents the dynamic viscosity, characterizing the viscosity of the fluid, that is, the internal frictional force between adjacent fluid layers; represents the body force per unit mass, is the gradient operator, representing the directional derivative in space.

Images