A prediction method and system for severe postpartum hemorrhage

Abstract

The method of the embodiment of the present invention proposes a prediction method and system for severe postpartum hemorrhage. The method first obtains multimodal medical imaging data and then preprocesses and aligns it; then the image is segmented by a sliding window strategy, and structural features, texture features, hemodynamic features and vascular distribution features are extracted from the segmentation results; then the features are used as nodes to construct a graph structure, and the relationship between the nodes is encoded using a graph convolutional network, and the correlation of the features at the topological level is calculated; finally, the features are fused according to the correlation, input into the prediction model and the prediction results are output. The method of the present invention improves the prediction ability and accuracy of the prediction through multi-feature fusion and multimodal data integration, enhances the robustness of the model in different data environments, can timely identify the risk of bleeding in pregnant women in the middle and late stages of pregnancy, provide early intervention and treatment recommendations, and significantly reduce the incidence and lethality of severe postpartum hemorrhage.

Description

Technical field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a method and system for predicting severe postpartum hemorrhage. [Background Technology]

[0002] Severe postpartum hemorrhage (SPPH) is a major maternal complication after childbirth and a leading cause of maternal mortality worldwide. Traditionally, the prediction of SPPH relies primarily on clinical risk assessment, which includes factors such as the mother's medical history, labor progress, and mode of delivery. However, the assessment of these factors is subjective and, due to significant individual variability, has limited accuracy and sensitivity. Furthermore, because the onset of severe postpartum hemorrhage is often sudden and unpredictable, it is difficult to effectively screen all high-risk women based solely on existing risk factors. By integrating maternal clinical information, biochemical markers, and imaging data, and establishing a predictive model based on big data and artificial intelligence technologies, the accuracy of SPPH prediction can be effectively improved. This will help improve the early identification of high-risk women, reduce the incidence of severe postpartum hemorrhage and its associated complications, and thus improve the overall health of pregnant women. [Summary of the invention]

[0003] In view of this, an embodiment of the present invention provides a method and system for predicting severe postpartum hemorrhage.

[0004] In a first aspect, an embodiment of the present invention provides a method for predicting severe postpartum hemorrhage, the method comprising:

[0005] S1. Preprocessing and aligning multimodal medical imaging data after acquiring the data, wherein the multimodal medical imaging data includes MRI image data and ultrasound Doppler image data of a pregnant woman in the middle and late stages of pregnancy;

[0006] S2. Segment the image using a sliding window strategy, and extract structural features, texture features, hemodynamic features, and vascular distribution features from the segmentation results;

[0007] S3. Build a graph structure using features as nodes, use a graph convolutional network to encode the relationships between nodes, and calculate the topological correlation between structural features, texture features, hemodynamic features, and vascular distribution features;

[0008] S4. After fusing the structural features, texture features, hemodynamic features, and vascular distribution features according to their correlation, the features are input into the prediction model and the prediction results are output.

[0009] According to the above aspects and any possible implementation, an implementation is further provided, wherein S1 includes:

[0010] S11. After denoising and suppressing artifacts on multimodal medical imaging data, normalize the resolution and grayscale;

[0011] S12. Build a preliminary deep learning model to perform contour detection, locate the uterus and placenta areas, and generate regions of interest;

[0012] S13. Select the uterine horn, placental edge, and vascular entrance of the MRI image region of interest and the ultrasound Doppler image region of interest respectively. If the thin plate spline deformation alignment can be performed, the two modalities are aligned in a unified coordinate system; if the thin plate spline deformation alignment cannot be performed, the two modalities are aligned in an adversarial network.

[0013] According to the above aspects and any possible implementation, an implementation is further provided, wherein S12 specifically includes:

[0014] S121. Build a preliminary deep learning model:

[0015]

[0016] Where X represents the normalized multimodal medical imaging data, Θ represents all trainable parameters of the preliminary deep learning model, and W out and b out represents the weight and bias of the output layer, F e (1) and Represents the feature map of the penultimate layer of the encoder and decoder, CC is the abbreviation of Concat, which means the concatenation of feature maps, US is the abbreviation of UpSample, which means the upsampling operation, σ is the activation function, and Represents the weights of the two convolution kernels of the last decoding layer, and Represents the bias of the two convolution kernels of the last decoding layer;

[0017] S122. Construct the loss function of the preliminary learning model:

[0018] Among them, L Dice is the Dice loss, L CE is the cross entropy loss, p i is the predicted value, g i is the predicted value, ε is the true value, y i is the true label, is the prediction probability, N is the number of training samples, α and β are weight parameters;

[0019] S123, minimizing the loss function L(Θ) by gradient descent based on multiple training samples, and substituting the trainable parameter Θ at this time into the preliminary deep learning model;

[0020] S124. Input the multimodal medical imaging data to be processed into the trained preliminary deep learning model to generate a predicted probability map. Convert the probability map into a binary mask and smooth it. Use the Canny edge detection algorithm to extract the edge contour of the mask. After cropping, generate a region of interest based on the uterus and placenta area.

[0021] According to the above aspects and any possible implementation, an implementation is further provided, wherein S13 specifically includes:

[0022] S131, respectively select the uterine horn, placental edge and blood vessel entrance of the MRI image region of interest and the ultrasound Doppler image region of interest, extract the precise coordinates and form the ultrasound image landmark point set and MRI image landmark point set

[0023] S132. Construct thin plate spline mapping function Where A is the translation vector and B is the affine transformation matrix and ω i is the non-rigid deformation weight and U(r)=r 2 log r 2 is the radial basis function, ||·|| is the Euclidean distance; by minimizing the deformation energy, a linear equation system is established to solve A, B and ω i Solve to get the mapping function f(x);

[0024] S133, through the formula Calculate the error MSE value of the landmark points after registration. If the MSE value is within the preset error range, it is determined that the MRI image region of interest and the ultrasound Doppler image region of interest can be deformed and registered using thin plate spline. i The coordinate system of the MRI image is transformed into the coordinate system through the thin plate spline mapping function for alignment;

[0025] S134. If the MSE value is not within the preset error range, it is determined that the MRI image region of interest and the ultrasound Doppler image region of interest cannot be deformably registered using thin plate splines, and an adversarial network is used for registration and alignment.

[0026] According to the above aspects and any possible implementation, an implementation is further provided, wherein S134 specifically includes:

[0027] Construct a model including generator G, generator F, and discriminator D MRI and the discriminator D US The adversarial network model, in which the generator G is used to transform the ultrasound image X US Converted to MRI style image G(X US ), the generator F is used to transform the MRI image X MRI Converted to ultrasound style image F(X MRI ), discriminator D MRI Used to determine whether the image is a real MRI image, the discriminator D US Used to determine whether the image is a real ultrasound image;

[0028] Construct a total loss function that includes adversarial loss and cycle consistency loss:

[0029] L total (G,F,D MRI ,D US )=L GAN (G,D MRI ,X US ,X MRI )+L GAN (F,D US ,X MRI ,X US )+λL cycle (G,F); where L GAN (G,D MRI ,X US ,X MRI ) represents the generator G and the discriminator D MRI The adversarial loss, L GAN (F,D US ,X MRI ,X US ) represents the generators F and D US The adversarial loss, L cycle (G,F) represents the cycle consistency loss, λ is the weight parameter, is the expectation operator; it is expressed in the following formula:

[0030]

[0031] Initialize the training parameters and iterate the training by continuously updating the discriminator and generator until the total loss function reaches the preset number of training rounds to obtain the final adversarial network model;

[0032] Use the generator G to transform the ultrasound image X US Converted to MRI style image G(X US ), and compare it with the MRI image X MRI to align.

[0033] According to the above aspects and any possible implementation, an implementation is further provided, wherein S2 specifically includes:

[0034] S21, preset window size and sliding step, window area W (x,y) =I[:,y:y+H,x:x+W], where x and y are the horizontal and vertical coordinates of the upper left corner of the window, H and W are the height and width of the input image, and I is the input image;

[0035] S22, W (x,y) Input into the trained segmentation model f to generate the corresponding segmentation mask M (x,y) =f(W (x,y) ; Θ), the segmentation masks of all windows are stitched back to the original image size, where K is the number of categories, Θ is the trainable parameter of the segmentation model;

[0036] S23. Extracting the boundary of the target area using an edge detection algorithm, calculating the number of boundary pixels, thereby obtaining the perimeter of the target area, and obtaining the area, shape factor, and aspect ratio of the target area based on the perimeter, wherein the area, perimeter, shape factor, and aspect ratio constitute structural features;

[0037] S24. Calculate the contrast, correlation, uniformity, and angular second moment of the target area based on the gray level co-occurrence matrix GLCM(i,j), where GLCM(i,j) represents the gray level at a given direction θ and distance d, and the number of times a pixel with gray level i and a pixel with gray level j co-occur in a specific spatial relationship;

[0038] The contrast ratio is given by the formula Perform calculations;

[0039] The correlation is expressed by the formula Perform calculations,

[0040]

[0041] Uniformity is determined by the formula Perform calculations;

[0042] The angular second moment is given by the formula Perform calculations;

[0043] S25. The extracted hemodynamic features include blood flow velocity, blood volume and pulse wave propagation velocity; the vascular distribution features include vascular density, number of vascular branches and vascular curvature.

[0044] According to the above aspects and any possible implementation, an implementation is further provided, wherein S3 specifically includes:

[0045] S31. Construct a graph structure G = (V, E), where the nodes in the node set V represent features, and the edges in the edge set E represent the associations between the nodes;

[0046] S32. Update the feature vector of each node through the graph convolution network on the graph structure G = (V, E), where the node feature H of the lth layer (l+1) Expressed as; in, for The degree matrix of is the adjacency matrix with self-loops added, W (l) is the learnable weight matrix of layer l, and σ is the activation function;

[0047] S33, after the encoding output of the graph convolutional network, the topological correlation between nodes is calculated based on the embedded vector, and the correlation is measured using cosine similarity. Expressed as follows; where z i For node v i The corresponding embedding vector, z i =f(h i ,{h j |j∈N(i)}), h i For node v i The corresponding eigenvector, N(i) node v i The corresponding neighbor node set.

[0048] According to the above aspects and any possible implementation, an implementation is further provided, wherein S4 specifically includes:

[0049] S41. Generate high-order feature representation based on the correlation between features, through the formula Assign dynamic weights to features of different categories and obtain the comprehensive feature vector f fused ; Among them, α k is the dynamic weight, through the formula Calculate, e k =W T tanh(Wf k +b), W is the weight matrix, b is the bias vector, f k is the feature vector of the kth category, K is the total number of feature categories;

[0050] S42, perform nonlinear dimensionality reduction on the comprehensive feature vector through the trained encoder to generate a low-dimensional feature representation z=f(x)=σ(W e x+b e ); where W e is the weight matrix of the encoder, σ is the activation function, b e is the bias vector of the encoder;

[0051] The encoder is trained as follows:

[0052] Construct an encoder with an input layer size of N, a hidden layer size of M, and an output layer size of N; initialize the weight matrix W of the encoder and decoder e and W d , and the bias vectors b of the encoder and decoder e and b d ; Input f fused The low-dimensional feature z is obtained through the encoder; the low-dimensional feature z is reconstructed through the decoder Calculate the reconstruction error loss function and calculate the loss function relative to W e 、W d 、b e and b d The gradient of θ is calculated and the parameters are updated; the iterative calculation is performed until the loss function converges, and an encoder that can be used independently is obtained.

[0053] S43. Input the low-dimensional features into the trained integrated prediction learning model to obtain the prediction results.

[0054] According to the above aspects and any possible implementation, an implementation is further provided, wherein the integrated prediction learning model structure includes:

[0055] Input layer, used to input low-dimensional features z;

[0056] M independent prediction models predict the low-dimensional feature z and generate their own prediction results Independent prediction models include at least: fully connected neural networks, such as support vector machines, random forests, such as gradient boosting machines;

[0057] Ensemble layer, used to determine the weight ω based on the validation set performance of each model m , perform weighted integration of the prediction results of all independent models

[0058] Output layer, used to output the final prediction results

[0059] In a second aspect, an embodiment of the present invention provides a system for predicting severe postpartum hemorrhage, the system comprising:

[0060] a preprocessing module for preprocessing and aligning multimodal medical imaging data after acquiring the data, wherein the multimodal medical imaging data includes MRI image data and ultrasound Doppler image data of pregnant women in the second and third trimesters;

[0061] The segmentation and extraction module is used to segment the image using a sliding window strategy and extract structural features, texture features, hemodynamic features, and vascular distribution features from the segmentation results;

[0062] The computational encoding module is used to construct a graph structure using features as nodes, encode the relationships between nodes using a graph convolutional network, and calculate the topological correlation between structural features, texture features, hemodynamic features, and vascular distribution features;

[0063] The fusion prediction module is used to fuse the structural features, texture features, hemodynamic features and vascular distribution features according to the correlation, input them into the prediction model and output the prediction results.

[0064] One of the above technical solutions has the following beneficial effects:

[0065] The method of the embodiment of the present invention proposes a method and system for predicting severe postpartum hemorrhage. The method first obtains multimodal medical imaging data and then preprocesses and aligns it. The multimodal medical imaging data includes MRI image data and ultrasound Doppler image data of pregnant women in the middle and late stages of pregnancy; then the image is segmented using a sliding window strategy, and structural features, texture features, hemodynamic features, and vascular distribution features are extracted from the segmentation results; then the features are used as nodes to construct a graph structure, and the relationships between the nodes are encoded using a graph convolutional network, and the correlation between the structural features, texture features, hemodynamic features, and vascular distribution features at the topological level is calculated; finally, the structural features, texture features, hemodynamic features, and vascular distribution features are fused according to the correlation, input into a prediction model, and the prediction results are output. The method of the present invention improves the prediction ability and accuracy through multi-feature fusion and multimodal data integration. The application of the graph convolutional network improves the model's understanding and processing capabilities of complex data relationships. The method of the present invention has a flexible architecture and can adjust the feature extraction method, the number of graph neural network layers, and the type of prediction model according to specific application requirements, with good scalability and adaptability. The present invention uses a sliding window strategy to segment the image and extract multiple features, thereby enhancing the robustness of the model in different data environments and reducing the prediction error caused by data noise or missing data; calculating the correlation of features at the topological level helps to identify key features and improve the stability and reliability of the model. In the future, the method of the present invention will not only be limited to MRI and ultrasound Doppler image data. The system will be compatible with more types of medical imaging data and adapt to the development and application needs of future medical imaging technology. Through accurate prediction models, the present invention can timely identify the risk of bleeding in pregnant women in the middle and late stages of pregnancy, provide early intervention and treatment recommendations, and significantly reduce the incidence and lethality of severe postpartum hemorrhage; the integrated prediction system can serve as an auxiliary decision-making tool for clinicians to improve the quality and efficiency of medical services.

Brief Description of the Drawings

[0066] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0067] Figure 1 A schematic flow chart of a method for predicting severe postpartum hemorrhage provided by an embodiment of the present invention;

[0068] Figure 2 A schematic block diagram of a system for predicting severe postpartum hemorrhage provided by an embodiment of the present invention;

[0069] Figure 3 A schematic diagram of the hardware structure of a severe postpartum hemorrhage prediction system provided by an embodiment of the present invention. [Specific implementation method]

[0070] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0071] Please refer to Figure 1 , which is a flow chart of a method for predicting severe postpartum hemorrhage provided by an embodiment of the present invention. As shown in the figure, the method includes the following steps:

[0072] S1. Preprocessing and aligning multimodal medical imaging data after acquiring the data, wherein the multimodal medical imaging data includes MRI image data and ultrasound Doppler image data of a pregnant woman in the middle and late stages of pregnancy;

[0073] S2. Segment the image using a sliding window strategy, and extract structural features, texture features, hemodynamic features, and vascular distribution features from the segmentation results;

[0074] S3. Build a graph structure using features as nodes, use a graph convolutional network to encode the relationships between nodes, and calculate the topological correlation between structural features, texture features, hemodynamic features, and vascular distribution features;

[0075] S4. After fusing the structural features, texture features, hemodynamic features, and vascular distribution features according to their correlation, the features are input into the prediction model and the prediction results are output.

[0076] Specifically, the S1 includes:

[0077] S11. After denoising and suppressing artifacts on multimodal medical imaging data, normalize the resolution and grayscale;

[0078] S12. Build a preliminary deep learning model to perform contour detection, locate the uterus and placenta areas, and generate regions of interest;

[0079] S13. Select the uterine horn, placental edge, and vascular entrance of the MRI image region of interest and the ultrasound Doppler image region of interest respectively. If the thin plate spline deformation alignment can be performed, the two modalities are aligned in a unified coordinate system; if the thin plate spline deformation alignment cannot be performed, the two modalities are aligned in an adversarial network.

[0080] Specifically, the S12 includes:

[0081] S121. Build a preliminary deep learning model:

[0082]

[0083] Where X represents the normalized multimodal medical imaging data, Θ represents all trainable parameters of the preliminary deep learning model, and W out and b out represents the weight and bias of the output layer, F e (1) and Represents the feature map of the penultimate layer of the encoder and decoder, CC is the abbreviation of Concat, which means the concatenation of feature maps, US is the abbreviation of UpSample, which means the upsampling operation, σ is the activation function, and Represents the weights of the two convolution kernels of the last decoding layer, and Represents the bias of the two convolution kernels of the last decoding layer;

[0084] S122. Construct the loss function of the preliminary learning model:

[0085] Among them, L Dice is the Dice loss, L CE is the cross entropy loss, p i is the predicted value, g i is the predicted value, ε is the true value, y i is the true label, is the prediction probability, N is the number of training samples, α and β are weight parameters;

[0086] S123, minimizing the loss function L(Θ) by gradient descent based on multiple training samples, and substituting the trainable parameter Θ at this time into the preliminary deep learning model;

[0087] S124. Input the multimodal medical imaging data to be processed into the trained preliminary deep learning model to generate a predicted probability map. Convert the probability map into a binary mask and smooth it. Use the Canny edge detection algorithm to extract the edge contour of the mask. After cropping, generate a region of interest based on the uterus and placenta area.

[0088] Specifically, the S13 includes:

[0089] S131, respectively select the uterine horn, placental edge and blood vessel entrance of the MRI image region of interest and the ultrasound Doppler image region of interest, extract the precise coordinates and form the ultrasound image landmark point set and MRI image landmark point set

[0090] S132. Construct thin plate spline mapping function Where A is the translation vector and B is the affine transformation matrix and ω i is the non-rigid deformation weight and U(r)=r 2 log r 2 is the radial basis function, ||·|| is the Euclidean distance; by minimizing the deformation energy, a linear equation system is established to solve A, B and ω i Solve to get the mapping function f(x);

[0091] S133, through the formula Calculate the error MSE value of the landmark points after registration. If the MSE value is within the preset error range, it is determined that the MRI image region of interest and the ultrasound Doppler image region of interest can be deformed and registered using thin plate spline. i The coordinate system of the MRI image is transformed into the coordinate system through the thin plate spline mapping function for alignment;

[0092] S134. If the MSE value is not within the preset error range, it is determined that the MRI image region of interest and the ultrasound Doppler image region of interest cannot be deformably registered using thin plate splines, and an adversarial network is used for registration and alignment.

[0093] Specifically, the S134 includes:

[0094] Construct a model including generator G, generator F, and discriminator D MRI and the discriminator DUS The adversarial network model, in which the generator G is used to transform the ultrasound image X US Converted to MRI style image G(X US ), the generator F is used to transform the MRI image X MRI Converted to ultrasound style image F(X MRI ), discriminator D MRI Used to determine whether the image is a real MRI image, the discriminator D US Used to determine whether the image is a real ultrasound image;

[0095] Construct a total loss function that includes adversarial loss and cycle consistency loss:

[0096] L total (G,F,D MRI ,D US )=L GAN (G,D MRI ,X US ,X MRI )+L GAN (F,D US ,X MRI ,X US )+λL cycle (G,F); where L GAN (G,D MRI ,X US ,X MRI ) represents the generator G and the discriminator D MRI The adversarial loss, L GAN (F,D US ,X MRI ,X US ) represents the generators F and D US The adversarial loss, L cycle (G,F) represents the cycle consistency loss, λ is the weight parameter, is the expectation operator; it is expressed in the following formula:

[0097]

[0098] Initialize the training parameters and iterate the training by continuously updating the discriminator and generator until the total loss function reaches the preset number of training rounds to obtain the final adversarial network model;

[0099] Use the generator G to transform the ultrasound image X US Converted to MRI style image G(X US ), and compare it with the MRI image X MRI to align.

[0100] Specifically, the S2 specifically includes:

[0101] S21, preset window size and sliding step, window area W (x,y) =I[:,y:y+H,x:x+W], where x and y are the horizontal and vertical coordinates of the upper left corner of the window, H and W are the height and width of the input image, and I is the input image;

[0102] S22, W (x,y) Input into the trained segmentation model f to generate the corresponding segmentation mask M (x,y) =f(W (x,y) ; Θ), the segmentation masks of all windows are stitched back to the original image size, where K is the number of categories, Θ is the trainable parameter of the segmentation model;

[0103] S23. Extracting the boundary of the target area using an edge detection algorithm, calculating the number of boundary pixels, thereby obtaining the perimeter of the target area, and obtaining the area, shape factor, and aspect ratio of the target area based on the perimeter, wherein the area, perimeter, shape factor, and aspect ratio constitute structural features;

[0104] S24. Calculate the contrast, correlation, uniformity, and angular second moment of the target area based on the gray level co-occurrence matrix GLCM(i,j), where GLCM(i,j) represents the gray level at a given direction θ and distance d, and the number of times a pixel with gray level i and a pixel with gray level j co-occur in a specific spatial relationship;

[0105] The contrast ratio is given by the formula Perform calculations;

[0106] The correlation is expressed by the formula Perform calculations,

[0107]

[0108] Uniformity is determined by the formula Perform calculations;

[0109] The angular second moment is given by the formula Perform calculations;

[0110] S25. The extracted hemodynamic features include blood flow velocity, blood volume and pulse wave propagation velocity; the vascular distribution features include vascular density, number of vascular branches and vascular curvature.

[0111] Specifically, the S3 includes:

[0112] S31. Construct a graph structure G = (V, E), where the nodes in the node set V represent features, and the edges in the edge set E represent the associations between the nodes;

[0113] S32. Update the feature vector of each node through the graph convolution network on the graph structure G = (V, E), where the node feature H of the lth layer (l+1) Expressed as; in, for The degree matrix of is the adjacency matrix with self-loops added, W (l) is the learnable weight matrix of layer l, and σ is the activation function;

[0114] S33, after the encoding output of the graph convolutional network, the topological correlation between nodes is calculated based on the embedded vector, and the correlation is measured using cosine similarity. Expressed as follows; where z i For node v i The corresponding embedding vector, z i =f(h i ,{h j |j∈N(i)}), h i For node v i The corresponding eigenvector, N(i) node v i The corresponding neighbor node set.

[0115] Specifically, S4 includes:

[0116] S41. Generate high-order feature representation based on the correlation between features, through the formula Assign dynamic weights to features of different categories and obtain the comprehensive feature vector f fused ; Among them, α k is the dynamic weight, through the formula Calculate, e k =W T tanh(Wf k +b), W is the weight matrix, b is the bias vector, f k is the feature vector of the kth category, K is the total number of feature categories;

[0117] S42, perform nonlinear dimensionality reduction on the comprehensive feature vector through the trained encoder to generate a low-dimensional feature representation z=f(x)=σ(W e x+b e ); where W e is the weight matrix of the encoder, σ is the activation function, b e is the bias vector of the encoder;

[0118] The encoder is trained as follows:

[0119] Construct an encoder with an input layer size of N, a hidden layer size of M, and an output layer size of N; initialize the weight matrix W of the encoder and decodere and W d , and the bias vectors b of the encoder and decoder e and b d ; Input f fused The low-dimensional feature z is obtained through the encoder; the low-dimensional feature z is reconstructed through the decoder Calculate the reconstruction error loss function and calculate the loss function relative to W e 、W d 、b e and b d The gradient of θ is calculated and the parameters are updated; the iterative calculation is performed until the loss function converges, and an encoder that can be used independently is obtained.

[0120] S43. Input the low-dimensional features into the trained integrated prediction learning model to obtain the prediction results.

[0121] Specifically, the integrated prediction learning model structure includes:

[0122] Input layer, used to input low-dimensional features z;

[0123] M independent prediction models predict the low-dimensional feature z and generate their own prediction results Independent prediction models include at least: fully connected neural networks, such as support vector machines, random forests, such as gradient boosting machines;

[0124] Ensemble layer, used to determine the weight ω based on the validation set performance of each model m , perform weighted integration of the prediction results of all independent models

[0125] Output layer, used to output the final prediction results

[0126] Through the above steps, the present invention achieves the following technical effects:

[0127] The method of an embodiment of the present invention proposes a method and system for predicting severe postpartum hemorrhage. The method first obtains multimodal medical imaging data and then preprocesses and aligns it. The multimodal medical imaging data includes MRI image data and ultrasound Doppler image data of pregnant women in the middle and late stages of pregnancy; then, the image is segmented using a sliding window strategy, and structural features, texture features, hemodynamic features, and vascular distribution features are extracted from the segmentation results; then, the features are used as nodes to construct a graph structure, and the relationships between the nodes are encoded using a graph convolutional network, and the correlation between the structural features, texture features, hemodynamic features, and vascular distribution features at the topological level is calculated; finally, the structural features, texture features, hemodynamic features, and vascular distribution features are fused according to the correlation, input into a prediction model, and the prediction results are output.

[0128] The method of the present invention improves predictive power and accuracy through multi-feature fusion and multimodal data integration. The application of graph convolutional networks enhances the model's understanding and processing capabilities of complex data relationships. The method of the present invention has a flexible architecture and can adjust the feature extraction method, number of graph neural network layers, and predictive model type according to specific application requirements, with good scalability and adaptability. The present invention uses a sliding window strategy to segment images and extract multiple features, enhancing the robustness of the model in different data environments and reducing prediction errors caused by data noise or missing data. Calculating the correlation of features at the topological level helps identify key features and improves the stability and reliability of the model. In the future, the method of the present invention will not be limited to MRI and ultrasound Doppler image data. The system will be compatible with more types of medical imaging data and adapt to the future development and application needs of medical imaging technology. Through accurate predictive models, the present invention can promptly identify bleeding risks in pregnant women in the second and third trimesters, provide early intervention and treatment recommendations, and significantly reduce the incidence and lethality of severe postpartum hemorrhage. The integrated prediction system can serve as a decision-making aid for clinicians, improving the quality and efficiency of medical services.

[0129] The embodiments of the present invention further provide device embodiments for implementing the steps and methods in the above method embodiments.

[0130] Please refer to Figure 2 , which is Figure 2 This is a schematic block diagram of a system for predicting severe postpartum hemorrhage provided by an embodiment of the present invention. The system includes:

[0131] A preprocessing module 210 is configured to acquire multimodal medical imaging data and perform preprocessing and alignment on the data, wherein the multimodal medical imaging data includes MRI image data and ultrasound Doppler image data of pregnant women in the second and third trimesters;

[0132] The segmentation and extraction module 220 is used to segment the image using a sliding window strategy and extract structural features, texture features, hemodynamic features, and vascular distribution features from the segmentation results;

[0133] A computational encoding module 230 is configured to construct a graph structure using features as nodes, encode the relationships between nodes using a graph convolutional network, and calculate the topological correlation between structural features, texture features, hemodynamic features, and vascular distribution features;

[0134] The fusion prediction module 240 is used to fuse the structural features, texture features, hemodynamic features and vascular distribution features according to the correlation, input them into the prediction model and output the prediction results.

[0135] Since each unit module in this embodiment can execute Figure 1 For the method shown in the embodiment, the part not described in detail in this embodiment can be referred to Figure 1 Related instructions.

[0136] Please refer to Figure 3 , which is a schematic diagram of the hardware structure of a severe postpartum hemorrhage prediction system provided by an embodiment of the present invention. The data prediction device includes at least one processor and a memory. The at least one processor is coupled to the memory and is configured to read and execute instructions in the memory to perform the data prediction method provided by an embodiment of the present invention.

[0137] In a third aspect, an embodiment of the present invention provides a computer-readable medium having a program code stored therein, which, when executed on a computer, causes the computer to execute the method for predicting severe postpartum hemorrhage provided by an embodiment of the present invention.

[0138] At the hardware level, the device may include a processor and, optionally, an internal bus, a network interface, and memory. The memory may include internal memory, such as high-speed random-access memory (RAM), and may also include non-volatile memory, such as at least one disk drive. Of course, the device may also include other hardware required for the service.

[0139] The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus. The bus can be divided into an address bus, a data bus, a control bus, and the like.

[0140] The memory is used to store programs. Specifically, the program may include program code, which includes computer operating instructions. The memory may include internal memory and non-volatile memory, and provides instructions and data to the processor.

[0141] The steps of the method disclosed in conjunction with the embodiments of the present invention can be directly implemented and executed by a hardware decoding processor, or by a combination of hardware and software modules within the decoding processor. The software modules can be located in a storage medium well-known in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. The storage medium is located in the memory, and the processor reads the information in the memory and, in conjunction with its hardware, completes the steps of the method described above.

[0142] The systems, devices, modules, or units described in the above embodiments may be implemented by computer chips or entities, or by products having certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0143] For the convenience of description, the above device is described as being divided into various units or modules according to their functions. Of course, when implementing the present invention, the functions of each unit or module can be implemented in the same or multiple software and / or hardware.

[0144] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0145] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0146] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0147] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0148] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0149] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0150] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0151] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0152] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0153] The present invention may be described in the general context of computer-executable instructions, such as program modules, executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. The present invention may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media, including storage devices.

[0154] The various embodiments of the present invention are described in a progressive manner. Similar portions between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the system embodiment is generally similar to the method embodiment, so its description is relatively simple. For relevant portions, refer to the description of the method embodiment.

[0155] The foregoing is merely an embodiment of the present invention and is not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.

Claims

1. A method for predicting severe postpartum hemorrhage, characterized in that: The method comprises: S1. Preprocessing and aligning multimodal medical imaging data after acquiring the data, wherein the multimodal medical imaging data includes MRI image data and ultrasound Doppler image data of a pregnant woman in the middle and late stages of pregnancy; S2. Segment the image using a sliding window strategy, and extract structural features, texture features, hemodynamic features, and vascular distribution features from the segmentation results; S3. Build a graph structure using features as nodes, use a graph convolutional network to encode the relationships between nodes, and calculate the topological correlation between structural features, texture features, hemodynamic features, and vascular distribution features; S4. After fusing the structural features, texture features, hemodynamic features, and vascular distribution features based on their correlation, the features are input into the prediction model and the prediction results are output; Said S1 comprises: S11. After denoising and suppressing artifacts on multimodal medical imaging data, normalize the resolution and grayscale; S12. Build a preliminary deep learning model to perform contour detection, locate the uterus and placenta areas, and generate regions of interest; S13, respectively selecting the uterine horn, placental edge, and vascular inlet of the MRI image region of interest and the ultrasound Doppler image region of interest, and if deformable registration using thin plate splines is possible, aligning the two modalities in a unified coordinate system; if deformable registration using thin plate splines is not possible, aligning the two modalities using an adversarial network; The S12 specifically includes: S121. Build a preliminary deep learning model: Where X represents the normalized multimodal medical imaging data, Θ represents all trainable parameters of the preliminary deep learning model, and W out and b out represents the weight and bias of the output layer, F e (1) and Represents the feature map of the penultimate layer of the encoder and decoder, CC is the abbreviation of Concat, which means the concatenation of feature maps, US is the abbreviation of UpSample, which means the upsampling operation, σ is the activation function, and Represents the weights of the two convolution kernels of the last decoding layer, and Represents the bias of the two convolution kernels of the last decoding layer; S122. Construct the loss function of the preliminary learning model: Among them, L Dice is the Dice loss, L CE is the cross entropy loss, p i is the predicted value, g i is the predicted value, ε is the true value, y i is the true label, is the prediction probability, N is the number of training samples, α and β are weight parameters; S123, minimizing the loss function L(Θ) by gradient descent based on multiple training samples, and substituting the trainable parameter Θ at this time into the preliminary deep learning model; S124, inputting the multimodal medical imaging data to be processed into the trained preliminary deep learning model to generate a prediction probability map, converting the probability map into a binary mask and smoothing it, extracting the edge contour of the mask using the Canny edge detection algorithm, and generating a region of interest based on the uterus and placenta region after cropping; The S13 specifically includes: S131, respectively select the uterine horn, placental edge and blood vessel entrance of the MRI image region of interest and the ultrasound Doppler image region of interest, extract the precise coordinates and form the ultrasound image landmark point set and MRI image landmark point set S132. Construct thin plate spline mapping function Where A is the translation vector and B is the affine transformation matrix and ω i is the non-rigid deformation weight and U(r)=r 2 logr 2 is the radial basis function, ||·|| is the Euclidean distance; by minimizing the deformation energy, a linear equation system is established to solve A, B and ω i Solve to get the mapping function f(x); S133, through the formula Calculate the error MSE value of the landmark points after registration. If the MSE value is within the preset error range, it is determined that the MRI image region of interest and the ultrasound Doppler image region of interest can be deformed and registered using thin plate spline. i The coordinate system of the MRI image is transformed into the coordinate system through the thin plate spline mapping function for alignment; S134. If the MSE value is not within the preset error range, it is determined that the MRI image region of interest and the ultrasound Doppler image region of interest cannot be deformably registered using thin plate spline, and an adversarial network is used to perform registration and alignment. The S134 specifically includes: Construct a model including generator G, generator F, and discriminator D MRI and the discriminator D US The adversarial network model, in which the generator G is used to transform the ultrasound image X US Converted to MRI style image G(X US ), the generator F is used to transform the MRI image X MRI Converted to ultrasound style image F(X MRI ), discriminator D MRI Used to determine whether the image is a real MRI image, the discriminator D US Used to determine whether the image is a real ultrasound image; Construct a total loss function that includes adversarial loss and cycle consistency loss: L total (G,F,D MRI ,D US )=L GAN (G,D MRI ,X US ,X MRI )+L GAN (F,D US ,X MRI ,X US )+λL cycle (G,F); where L GAN (G,D MRI ,X US ,X MRI ) represents the generator G and the discriminator D MRI The adversarial loss, L GAN (F,D US ,X MRI ,X US ) represents the generators F and D US The adversarial loss, L cycle (G,F) represents the cycle consistency loss, λ is the weight parameter, is the expectation operator; it is expressed in the following formula: Initialize the training parameters and iterate the training by continuously updating the discriminator and generator until the total loss function reaches the preset number of training rounds to obtain the final adversarial network model; Use the generator G to transform the ultrasound image X US Converted to MRI style image G(X US ), and compare it with the MRI image X MRI to align.

2. The method for predicting severe postpartum hemorrhage according to claim 1, wherein The S2 specifically includes: S21, preset window size and sliding step, window area W (x,y) =I[:,y:y+H,x:x+W], where x and y are the horizontal and vertical coordinates of the upper left corner of the window, H and W are the height and width of the input image, and I is the input image; S22, W (x,y) Input into the trained segmentation model f to generate the corresponding segmentation mask M (x,y) =f(W (x,y) ; Θ), the segmentation masks of all windows are stitched back to the original image size, where K is the number of categories, Θ is the trainable parameter of the segmentation model; S23. Extracting the boundary of the target area using an edge detection algorithm, calculating the number of boundary pixels, thereby obtaining the perimeter of the target area, and obtaining the area, shape factor, and aspect ratio of the target area based on the perimeter, wherein the area, perimeter, shape factor, and aspect ratio constitute structural features; S24. Calculate the contrast, correlation, uniformity, and angular second moment of the target area based on the gray level co-occurrence matrix GLCM(i,j), where GLCM(i,j) represents the gray level at a given direction θ and distance d, and the number of times a pixel with gray level i and a pixel with gray level j co-occur in a specific spatial relationship; The contrast ratio is given by the formula Perform calculations; The correlation is expressed by the formula Perform calculations, Uniformity is determined by the formula Perform calculations; The angular second moment is given by the formula Perform calculations; S25. The extracted hemodynamic features include blood flow velocity, blood volume and pulse wave propagation velocity; the vascular distribution features include vascular density, number of vascular branches and vascular curvature.

3. The method for predicting severe postpartum hemorrhage according to claim 2, wherein: The S3 specifically includes: S31. Construct a graph structure G = (V, E), where the nodes in the node set V represent features, and the edges in the edge set E represent the associations between the nodes; S32. Update the feature vector of each node through the graph convolution network on the graph structure G = (V, E), where the node feature H of the lth layer (l+1) Expressed as; in, for The degree matrix of is the adjacency matrix with self-loops added, W (l) is the learnable weight matrix of layer l, and σ is the activation function; S33, after the encoding output of the graph convolutional network, the topological correlation between nodes is calculated based on the embedded vector, and the correlation is measured using cosine similarity. Expressed as follows; where z i For node v i The corresponding embedding vector, z i =f(h i ,{h j |j∈N(i)}), h i For node v i The corresponding eigenvector, N(i) node v i The corresponding neighbor node set.

4. The method for predicting severe postpartum hemorrhage according to claim 3, wherein: The S4 specifically includes: S41. Generate high-order feature representation based on the correlation between features, through the formula Assign dynamic weights to features of different categories and obtain the comprehensive feature vector f fused ; Among them, α k is the dynamic weight, through the formula Calculate, e k =W T tanh(Wf k +b), W is the weight matrix, b is the bias vector, f k is the feature vector of the kth category, K is the total number of feature categories; S42, perform nonlinear dimensionality reduction on the comprehensive feature vector through the trained encoder to generate a low-dimensional feature representation z=f(x)=σ(W e x+b e ); where W e is the weight matrix of the encoder, σ is the activation function, b e is the bias vector of the encoder; The encoder is trained as follows: Construct an encoder with an input layer size of N, a hidden layer size of M, and an output layer size of N; initialize the weight matrix W of the encoder and decoder e and W d , and the bias vectors b of the encoder and decoder e and b d ; Input f fused The low-dimensional feature z is obtained through the encoder; the low-dimensional feature z is reconstructed through the decoder Calculate the reconstruction error loss function and calculate the loss function relative to W e 、W d 、b e and b d The gradient of the loss function is calculated and the parameters are updated. The loss function is iterated until convergence, and an encoder that can be used independently is obtained. S43. Input the low-dimensional features into the trained integrated prediction learning model to obtain the prediction results.

5. The method for predicting severe postpartum hemorrhage according to claim 4, wherein: The integrated prediction learning model structure includes: Input layer, used to input low-dimensional features z; M independent prediction models predict the low-dimensional feature z and generate their own prediction results Independent prediction models include at least: fully connected neural networks, such as support vector machines, random forests, such as gradient boosting machines; Ensemble layer, used to determine the weight ω based on the validation set performance of each model m , perform weighted integration of the prediction results of all independent models Output layer, used to output the final prediction results 6. A prediction system for severe postpartum hemorrhage using the method according to any one of claims 1 to 5, characterized in that: The system comprises: a preprocessing module for preprocessing and aligning multimodal medical imaging data after acquiring the data, wherein the multimodal medical imaging data includes MRI image data and ultrasound Doppler image data of pregnant women in the second and third trimesters; The segmentation and extraction module is used to segment the image using a sliding window strategy and extract structural features, texture features, hemodynamic features, and vascular distribution features from the segmentation results; The computational encoding module is used to construct a graph structure using features as nodes, encode the relationships between nodes using a graph convolutional network, and calculate the topological correlation between structural features, texture features, hemodynamic features, and vascular distribution features; The fusion prediction module is used to fuse the structural features, texture features, hemodynamic features and vascular distribution features according to the correlation, input them into the prediction model and output the prediction results.

Images