Pulmonary artery hemodynamic parameter prediction method and system

Abstract

The invention provides a pulmonary artery hemodynamic parameter prediction method and system, and relates to the technical field of medical image processing. The method comprises the following steps: acquiring an original CXR image of a patient; preprocessing the original CXR image to obtain a target CXR image; constructing an image segmentation model based on a U-Net segmentation network; inputting the target CXR image into the image segmentation model for segmentation, and outputting a right lower pulmonary artery ROI region image; the method comprises the steps that a pulmonary artery hemodynamic parameter prediction model is constructed, and the pulmonary artery hemodynamic parameter prediction model comprises a Swin-Transform branch and a CNN branch; and inputting the target CXR image to the Swindow-Transform branch, inputting the right lower pulmonary artery ROI region image to the CNN branch, and outputting a prediction result of the pulmonary artery hemodynamic parameters.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and particularly to a method and system for predicting pulmonary artery hemodynamic parameters. Background Art

[0002] With the rapid development of medical imaging and artificial intelligence technologies, image-based disease diagnosis has gradually become an indispensable tool in clinical medicine. Especially in the early screening and diagnosis of cardiopulmonary diseases such as pulmonary hypertension (PH), chest X-ray images (CXRs) have become an important means of routine clinical examinations due to their non-invasive nature, wide application, and low cost.

[0003] However, in the prior art, traditional methods for measuring pulmonary artery hemodynamic parameters mainly rely on right heart catheterization. Although this method has high precision, it is highly invasive, costly, and risky.

[0004] In addition, traditional medical image analysis methods have high requirements for image quality and noise, and their segmentation accuracy and feature extraction capabilities are relatively limited. It is difficult to effectively extract details of important regions such as the pulmonary artery, which affects the accuracy and effectiveness of diagnosis. Summary of the Invention

[0005] In order to solve the technical problems in the prior art that traditional methods for measuring pulmonary artery hemodynamic parameters mainly rely on right heart catheterization, which is highly invasive, costly, and risky, and traditional medical image analysis methods have high requirements for image quality and noise, with relatively limited segmentation accuracy and feature extraction capabilities, and it is difficult to effectively extract details of important regions such as the pulmonary artery, affecting the accuracy and effectiveness of diagnosis, the present invention provides a method and system for predicting pulmonary artery hemodynamic parameters.

[0006] The technical solutions provided by the embodiments of the present invention are as follows:

[0007] First aspect:

[0008] A method for predicting pulmonary artery hemodynamic parameters provided by an embodiment of the present invention includes:

[0009] S1: Obtain the original CXR image of the patient;

[0010] S2: Preprocess the original CXR image to obtain the target CXR image;

[0011] S3: Construct an image segmentation model based on the U-Net segmentation network;

[0012] S4: Input the target CXR image into the image segmentation model for segmentation, and output the image of the right lower pulmonary artery ROI region;

[0013] S5: Construct a prediction model for pulmonary artery hemodynamic parameters, where the prediction model for pulmonary artery hemodynamic parameters includes a Swin-Transformer branch and a CNN branch;

[0014] S6: Input the target CXR image into the Swin-Transformer branch, and input the image of the right lower pulmonary artery ROI region into the CNN branch to output the prediction result of the pulmonary artery hemodynamic parameters.

[0015] Second aspect:

[0016] A pulmonary artery hemodynamic parameter prediction system provided by an embodiment of the present invention includes:

[0017] A processor;

[0018] A memory, on which computer-readable instructions are stored. When the computer-readable instructions are executed by the processor, the pulmonary artery hemodynamic parameter prediction method described in the first aspect is implemented.

[0019] Third aspect:

[0020] A computer-readable storage medium provided by an embodiment of the present invention, on which a computer program is stored. When the program is executed by a processor, the pulmonary artery hemodynamic parameter prediction method described in the first aspect is implemented.

[0021] The beneficial effects brought by the technical solutions provided by the embodiments of the present invention at least include:

[0022] In the embodiments of the present invention, by obtaining the original CXR image of the patient, it no longer depends on right heart catheterization, is non-invasive, has low cost and small risk. By preprocessing the original CXR image, the target CXR image is obtained, the image quality is improved and the noise is reduced. By inputting the target CXR image into the image segmentation model for segmentation, the image of the right lower pulmonary artery ROI region is output, and the target CXR image is input into the Swin-Transformer branch, and the image of the right lower pulmonary artery ROI region is input into the CNN branch to output the prediction result of the pulmonary artery hemodynamic parameters. The segmentation accuracy and feature extraction ability are strong, and the details of important regions such as the pulmonary artery can be effectively extracted, improving the accuracy and effect of diagnosis. Description of the Drawings

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0024] Figure 1 It is a schematic flowchart of a method for predicting pulmonary artery hemodynamic parameters provided by an embodiment of the present invention;

[0025] Figure 2 It is a schematic structural diagram of a system for predicting pulmonary artery hemodynamic parameters provided by an embodiment of the present invention. Specific embodiments

[0026] Next, the technical solutions in the present invention will be described with reference to the accompanying drawings.

[0027] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as an "example" in the present invention should not be construed as being more preferred or more advantageous than other embodiments or design solutions. Rather, the use of the word "example" is intended to present concepts in a specific manner. In addition, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one of the two.

[0028] In the embodiments of the present invention, "image" and "picture" can sometimes be used interchangeably. It should be noted that when their differences are not emphasized, the meanings they express are the same. "(of)", "corresponding", and "corresponding" can sometimes be used interchangeably. It should be noted that when their differences are not emphasized, the meanings they express are the same.

[0029] In the embodiments of the present invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When their differences are not emphasized, the meanings they express are the same.

[0030] To make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the accompanying drawings and specific embodiments.

[0031] Refer to the attached instructions Figure 1 , which shows a schematic flowchart of a method for predicting pulmonary artery hemodynamic parameters provided by an embodiment of the present invention.

[0032] The embodiments of the present invention provide a method for predicting pulmonary artery hemodynamic parameters. This method can be implemented by a device for predicting pulmonary artery hemodynamic parameters, and this device for predicting pulmonary artery hemodynamic parameters can be a terminal or a server. The processing flow of the method for predicting pulmonary artery hemodynamic parameters can include the following steps:

[0033] S1: Obtain the original CXR image of the patient.

[0034] S2: Preprocess the original CXR image to obtain the target CXR image.

[0035] In the embodiments of the present invention, preprocessing the original CXR image can effectively improve the image quality, reduce noise, enhance contrast, and further improve the prediction accuracy of the model for pulmonary artery hemodynamic parameters, ensuring more accurate image segmentation and feature extraction.

[0036] In a possible implementation manner, S2 specifically includes sub-steps S201 to S203:

[0037] S201: Align the original CXR image spatially through an elastic registration algorithm.

[0038] It should be noted that the Elastic Registration Algorithm is an image registration technique aimed at aligning two images through a non-rigid transformation so that they match as closely as possible in shape or position. This algorithm is commonly used in medical image processing, computer vision, etc., and performs particularly well when dealing with images with deformed or flexible shapes. It defines an elastic deformation model that allows corresponding regions of the image to undergo local deformation in space rather than simple translation, rotation, or scaling. In this way, the algorithm can handle image differences caused by different shooting angles, lesions, or object deformations, achieving high-precision image alignment.

[0039] Specifically, first, select a reference image as the alignment benchmark and perform a preliminary registration of the original CXR image with the reference image. Then, apply the elastic registration algorithm to obtain the non-rigid transformation relationship between the original image and the reference image by calculating the deformation field of the image. Next, use an elastic deformation model (such as B-spline transformation) to perform a spatial transformation on the original image to achieve fine alignment with the reference image in the local area. Finally, optimize the deformation field to achieve precise spatial alignment by minimizing the difference between the deformed image and the reference image.

[0040] In the embodiments of the present invention, precise spatial alignment of the image through the elastic registration algorithm can effectively handle image differences caused by different shooting angles, lesions, or object deformations, improve the registration accuracy of the image, and ensure the accuracy of subsequent analysis.

[0041] S202: Perform intensity normalization processing on the spatially aligned original CXR image through the CLAHE algorithm.

[0042] It should be noted that the CLAHE (Contrast Limited Adaptive Histogram Equalization) algorithm is an image enhancement technique used to improve the contrast of images, especially in local regions of the image. Different from traditional histogram equalization, CLAHE enhances the contrast by performing histogram equalization within local regions of the image (such as small blocks or windows), thus avoiding the problems of over-enhancement or noise amplification that may be caused by global equalization. By restricting the amount of contrast enhancement in each small region, this algorithm ensures that there will be no phenomenon of over-lightening or over-darkening, improving the visual quality of the image, and is especially suitable for processing low-contrast images.

[0043] Specifically, first, the original CXR image after spatial alignment is divided into multiple small local regions (i.e., image blocks). Then, for each local region, its local contrast is calculated and adaptive histogram equalization is performed to enhance the details of the local region. Next, the CLAHE algorithm is applied to prevent noise caused by over-enhancement by restricting the amplitude of contrast enhancement. Finally, the processed local image blocks are merged to obtain a CXR image after intensity normalization, making its overall brightness and contrast uniform and improving the visual quality of the image.

[0044] In the embodiment of the present invention, by adaptively enhancing the local contrast of the image through the CLAHE algorithm, the details of the low-contrast image can be improved, the brightness and contrast of the image can be made uniform, and the visual quality of the image can be improved, providing a clearer image for subsequent segmentation and feature extraction.

[0045] S203: Perform data augmentation on the original CXR image after intensity normalization through random rotation, translation, and gamma correction operations to obtain the target CXR image.

[0046] It should be noted that random rotation is a data augmentation technique mainly used in image processing. By randomly rotating the image by a certain angle, it simulates different perspectives. This method can increase the generalization ability of the model, enabling it to better adapt to images in different directions or angles, and is often used in training deep learning models to enhance their robustness to object pose changes.

[0047] It should be noted that translation is an image transformation technique, which refers to moving the image a certain distance in the horizontal or vertical direction. This operation can help the model better learn the local features of the image and improve its adaptability to the position changes of objects in the image. It is a common image data augmentation method and is especially effective in object detection and image classification.

[0048] It should be noted that gamma correction is an image adjustment technique used to change the brightness and contrast of an image. By performing a power transformation on the pixel values of the image, the brightness distribution of the image can be made more uniform, improving the visual effect of overly dark or overly bright areas. Gamma correction has an important impact on the visual perception of images. Especially when dealing with low-contrast images, it can effectively enhance details and clarity.

[0049] Specifically, first, the original CXR image after intensity normalization is randomly rotated by selecting a random angle for rotation transformation, thereby increasing the rotational invariance of the image. Then, a random translation operation is applied to randomly translate the image in the horizontal and vertical directions to simulate changes in different positions. Next, gamma correction is performed by adjusting the gamma value of the image to enhance or suppress the brightness of the image, thereby enhancing the details and contrast of the image to obtain the target CXR image.

[0050] In the embodiment of the present invention, through random rotation, translation, and gamma correction operations, the diversity and robustness of the image are increased, helping the model to better adapt to changes in different perspectives, positions, and brightness, enhancing the generalization ability of the model, and improving its performance in different environments.

[0051] S3: Construct an image segmentation model based on the U-Net segmentation network.

[0052] It should be noted that the U-Net segmentation network is a deep learning model commonly used in image segmentation tasks, especially widely used in medical image processing. It consists of an encoder (downsampling part) and a decoder (upsampling part), adopting a symmetric "U" shape structure. The encoder gradually extracts high-level features of the image through a series of convolutional and pooling operations, while the decoder gradually restores the spatial resolution of the image through upsampling and convolutional operations. At each layer of the U-Net, there are skip connections between the encoder and the decoder, which can retain the low-level features extracted by the encoder and help the decoder make more accurate pixel-level predictions. This structure of the U-Net enables it to achieve high accuracy when segmenting images with a lot of details (such as organs or tumors in medical images).

[0053] Optionally, the image segmentation model includes an encoder and a decoder. Among them, the encoder includes a DCD module, a TripletAttention module, and a CBAM module.

[0054] It should be noted that the DCD module (Dynamic Convolutional Decoder Module) is a module for image feature extraction, which enhances the model's ability to express image features through dynamic convolution operations. In the DCD module, the weights of the convolutional kernels are dynamically adjusted according to the features of the input image, so as to more effectively capture image information of different scales and different types. This module can process multi-scale feature information, improve the expression ability and accuracy of image features, and has good performance especially when dealing with complex images.

[0055] It should be noted that the Triplet Attention module is a module based on the attention mechanism, aiming to enhance the important information in image features. In this module, through cross-dimensional enhancement processing of multi-scale feature maps of the image, a composite feature map containing channel, spatial and position information is generated. By calculating the relationships between different scales, positions and channels, the Triplet Attention module can effectively improve the model's attention to key information, thereby improving the accuracy of segmentation or classification.

[0056] It should be noted that the CBAM module (Convolutional Block Attention Module) is a module that combines channel attention and spatial attention, mainly used to improve the feature expression ability of convolutional neural networks. CBAM first calculates channel attention to select important channel features, and then calculates spatial attention to select important spatial regions. Through this two-level attention mechanism, CBAM can effectively strengthen the network's attention to useful features and suppress irrelevant background information, thereby improving the performance of the network, especially in tasks such as image segmentation and object detection.

[0057] In the embodiments of the present invention, using the U-Net segmentation network combined with the DCD, Triplet Attention and CBAM modules can effectively improve the image feature extraction ability, enhance the model's attention to complex image details, thereby improving the accuracy of the pulmonary artery ROI region segmentation and ensuring more accurate image analysis.

[0058] S4: Input the target CXR image into the image segmentation model for segmentation, and output the image of the right lower pulmonary artery ROI region.

[0059] In the embodiments of the present invention, inputting the target CXR image into the image segmentation model for segmentation can accurately extract the right lower pulmonary artery ROI region, ensure the accurate prediction of pulmonary artery hemodynamic parameters, and improve the accuracy of the segmentation result and the reliability of the model.

[0060] In a possible implementation manner, S4 specifically includes sub-steps S401 to S406:

[0061] S401: In the DCD module, perform dynamic convolution operations on the target CXR image to generate multi-scale feature maps.

[0062] Optionally, perform dynamic convolution operations on the target CXR image according to the following formula to generate multi-scale feature maps:

[0063] F dcd = W dcd * I target

[0064]

[0065] where F dcd represents the multi-scale feature map, W dcd represents the dynamic convolution kernel, * represents the dynamic convolution operation, I target represents the target CXR image, α i represents the weight coefficient of the i-th basic convolution kernel, W i represents the i-th basic convolution kernel, where i = 1, 2,... k, and k represents the total number of basic convolution kernels.

[0066] In the embodiments of the present invention, through dynamic convolution operations, the convolution kernel can be adaptively adjusted according to the features of the input image, effectively capturing multi-scale features and enhancing the expression ability of image features, especially for complex medical images.

[0067] S402: In the Triplet Attention module, perform cross-dimensional enhancement processing on the multi-scale feature map to generate a cross-dimensionally enhanced feature map.

[0068] Optionally, perform cross-dimensional enhancement processing on the multi-scale feature map according to the following formula to generate a cross-dimensionally enhanced feature map:

[0069]

[0070] W c = Sigmoid(MLP(AvgPool(F dcd )) + MLP(MaxPool(F dcd )))

[0071] W s = Sigmoid(f 3×3 ([AvgPool(F dcd );Maxpool(F dcd )))

[0072] W p = Sigmoid(P 1×1(CoordConv(F dcd )))

[0073] Among them, F triplet represents the cross-dimensional enhanced feature map, σ() represents the activation function, W c represents the first-channel attention weight, represents element-wise addition, W s represents the first-spatial attention weight, W p represents the position attention weight, Sigmoid() represents the Sigmoid activation function, MLP() represents the multi-layer perceptron, AvgPool() represents the average pooling operation, MaxPool() represents the max pooling operation, f 3×3 () represents the 3×3 convolution operation, f 1×1 () represents the 1×1 convolution operation, CoordConv represents the coordinate convolution operation.

[0074] In the embodiment of the present invention, the Triplet Attention module is used to enhance the information of different scales, channels and spaces, enabling the model to better focus on key information, improving the expression of image features and enhancing the ability of feature learning, thereby improving the segmentation accuracy.

[0075] S403: In the CBAM module, perform channel-spatial attention fusion operation on the cross-dimensional enhanced feature map to generate a fused feature map.

[0076] Optionally, perform channel-spatial attention fusion operation on the cross-dimensional enhanced feature map to generate a fused feature map:

[0077]

[0078] M c (F triplet ) = Sigmoid(MLP(AvgPool(F triplet )) + MLP(MaxPool(F triplet )))

[0079] M s (F triplet ) = Sigmoid(f 7×7 ([AvgPool(F triplet );MaxPool(F triplet )]))

[0080] Among them, F cbam represents the fused feature map, M s (F triplet ) represents the cross-dimensional enhanced feature map after spatial attention fusion, M c (F triplet) represents the cross-dimensional enhanced feature map after channel attention fusion. represents element-wise multiplication, f 7×7 () represents a 7×7 convolution operation.

[0081] In the embodiment of the present invention, the CBAM module combines channel and spatial attention mechanisms, and further improves the expression effect of image features by focusing on different regions and channels of the image, helping the model to better focus on important regions and reducing the interference of irrelevant information.

[0082] S404: In the decoder, perform upsampling operation, convolution operation, and activation operation on the fused feature map in sequence to generate the probability map of the right lower pulmonary artery ROI region.

[0083] Optionally, perform upsampling operation, convolution operation, and activation operation on the fused feature map in sequence to generate the probability map of the right lower pulmonary artery ROI region:

[0084] P vessel = Sigmoid(Conv(Decode(F cbam )))

[0085] where P vessel represents the probability map of the right lower pulmonary artery ROI region, Conv() represents the convolution operation, and Decode() represents the decoding operation, i.e., the upsampling operation.

[0086] In the embodiment of the present invention, the spatial resolution of the image is restored through the upsampling operation of the decoder, and the probability map of the right lower pulmonary artery ROI region is generated, ensuring an accurate segmentation region.

[0087] S405: Perform binarization processing on the probability map of the right lower pulmonary artery ROI region to generate the image of the right lower pulmonary artery ROI region.

[0088] Optionally, perform binarization processing on the probability map of the right lower pulmonary artery ROI region to generate the image of the right lower pulmonary artery ROI region:

[0089]

[0090] where I ROI represents the image of the right lower pulmonary artery ROI region. When I ROI = 1, the pixel in the probability map of the right lower pulmonary artery ROI region belongs to the right lower pulmonary artery ROI region. When I ROI = 0, the pixel in the probability map of the right lower pulmonary artery ROI region does not belong to the right lower pulmonary artery ROI region. represents the value of the j-th pixel in the probability map of the right lower pulmonary artery ROI region, and τ represents the binarization threshold.

[0091] It should be noted that those skilled in the art can set the size of the binarization threshold according to actual needs, and the present invention does not limit it here.

[0092] In the embodiment of the present invention, through binarization processing, the pulmonary artery ROI region is clearly distinguished, further improving the usability of the image and ensuring the accuracy and clarity of the segmented region.

[0093] S406: Output the image of the right lower pulmonary artery ROI region.

[0094] S5: Construct a pulmonary artery hemodynamic parameter prediction model, where the pulmonary artery hemodynamic parameter prediction model includes a Swin-Transformer branch and a CNN branch.

[0095] It should be noted that the Swin-Transformer branch is a feature extraction network based on the Transformers architecture, especially designed for processing image data in computer vision tasks. Different from traditional Transformers, the Swin-Transformer effectively captures local and global information by introducing a window-based self-attention mechanism, while optimizing computational efficiency and performance. It alternately uses window self-attention (W-MSA) and shifted window self-attention (SW-MSA) in multiple stages, thereby breaking boundary limitations, enhancing feature interaction, and gradually constructing hierarchical image feature representations. When dealing with global morphological features in images, the Swin-Transformer branch can efficiently capture long-range dependencies and is widely used in image classification and segmentation tasks.

[0096] It should be noted that the CNN branch (Convolutional Neural Network branch) is a classic deep learning model that uses convolutional operations to automatically extract local features of images. The CNN branch gradually learns the detailed information in the image through multiple convolutional layers, pooling layers, and activation functions, and converts it into features that can be used for classification or regression through fully connected layers. The CNN branch is very effective in dealing with local texture features, especially in capturing local patterns and details of images. By combining convolutional features and pooling operations, the CNN branch can extract spatial information crucial for image recognition tasks.

[0097] Optionally, the CNN branch includes: multiple convolutional layers, multiple pooling layers, multiple flattening layers, a fully connected layer, and a feature fusion layer. The convolutional layers include a first convolutional layer, a second convolutional layer, and a third convolutional layer. The pooling layers include a first pooling layer, a second pooling layer, a third pooling layer, a fourth pooling layer, and an adaptive pooling layer. The flattening layers include a first flattening layer, a second flattening layer, and a third flattening layer.

[0098] Optionally, the pulmonary artery hemodynamic parameter prediction model further includes: a cross-attention mechanism and multiple prediction heads, where the prediction heads include a pulmonary artery systolic pressure prediction head and a pulmonary to systemic blood flow ratio prediction head.

[0099] It should be noted that each prediction head receives the fused feature vector and maps it to the corresponding parameter space through one or more fully connected layers, and finally outputs the predicted value of the corresponding pulmonary artery hemodynamic parameter. For example, the pulmonary artery systolic pressure prediction head outputs the predicted value of the pulmonary artery systolic pressure, the pulmonary vascular resistance prediction head outputs the predicted value of the pulmonary vascular resistance, and the pulmonary to systemic blood flow ratio prediction head outputs the predicted value of the pulmonary to systemic blood flow ratio.

[0100] In the embodiment of the present invention, combining the Swin-Transformer branch and the CNN branch can make full use of the global morphological features and local texture features, thereby improving the prediction accuracy of the pulmonary artery hemodynamic parameters. Through the cross-attention mechanism and multiple prediction heads, the prediction ability of the model is further enhanced.

[0101] S6: Input the target CXR image into the Swin-Transformer branch, input the image of the right lower pulmonary artery ROI region into the CNN branch, and output the prediction result of the pulmonary artery hemodynamic parameters.

[0102] Among them, the pulmonary artery hemodynamic parameters specifically include: pulmonary artery systolic pressure, pulmonary vascular resistance, and pulmonary to systemic blood flow ratio.

[0103] In the embodiment of the present invention, inputting the target CXR image and the image of the right lower pulmonary artery ROI region into the Swin-Transformer and CNN branches respectively can effectively combine the global morphological features and local texture features, improve the prediction accuracy of the pulmonary artery hemodynamic parameters, and enhance the robustness and accuracy of the model.

[0104] In a possible implementation manner, S6 specifically includes sub-steps S601 to S604:

[0105] S601: According to the target CXR image, output the global morphological feature vector through the Swin-Transformer branch.

[0106] Specifically, first, the target CXR image is segmented into multiple local image patches according to a fixed size. Then, convolution operations are used to linearly embed these local patches to generate a high-dimensional Token sequence. Next, in multiple stages of the Swin-Transformer branch, window multi-head self-attention (W-MSA) and shifted window multi-head self-attention (SW-MSA) are alternately executed. W-MSA captures the morphological correlations between local regions within a fixed-size window, while SW-MSA breaks the boundary limitations by shifting the window to achieve cross-region feature interaction. At the end of each stage, the resolution of the feature map is reduced and the channels are expanded by merging adjacent blocks, thereby gradually constructing a hierarchical feature pyramid. In the last stage, the deep feature map is processed by global average pooling, compressed into a global morphological feature vector of a fixed dimension, and output.

[0107] In the embodiment of the present invention, through the window multi-head self-attention and shifted window multi-head self-attention mechanisms in the Swin-Transformer branch, the global morphological features of the image can be captured, and the understanding of long-range dependencies can be improved through the hierarchical feature pyramid, thereby enhancing the ability to identify key features in the image.

[0108] S602: According to the image of the right lower pulmonary artery ROI region, output a local texture feature vector through the CNN branch.

[0109] In the embodiment of the present invention, the CNN branch can effectively extract the local texture features of the right lower pulmonary artery region, finely capture the details within the region, and ensure that local information is fully utilized.

[0110] In a possible implementation manner, S602 specifically includes sub-steps S602A to S602O:

[0111] S602A: In the first convolutional layer, perform feature extraction on the image of the right lower pulmonary artery ROI region to obtain a first local feature map.

[0112] Optionally, the first convolutional layer uses 32 5×5 convolutional kernels, with a stride of 1 and a padding of 2 layers of pixels at the edges, and the ReLU activation function is used as the activation function.

[0113] S602B: In the first pooling layer, perform a pooling operation on the first local feature map to obtain a first pooled feature map.

[0114] Optionally, the first pooling layer uses a max-pooling operation, with a pooling window size of 2×2 and a stride of 2.

[0115] S602C: In the second convolutional layer, perform feature extraction on the first pooled feature map to obtain a second local feature map.

[0116] Optionally, the second convolutional layer uses 64 3×3 convolutional kernels, with a stride of 1 and a padding of 1 pixel at the edges, and the ReLU activation function is adopted.

[0117] S602D: In the second pooling layer, perform a pooling operation on the second local feature map to obtain a second pooled feature map.

[0118] Optionally, the second pooling layer adopts a max pooling operation, with a pooling window size of 2×2 and a stride of 2.

[0119] S602E: In the third convolutional layer, perform feature extraction on the second pooled feature map to obtain a third local feature map.

[0120] Optionally, the third convolutional layer uses 128 3×3 convolutional kernels, with a stride of 1 and a padding of 1 pixel at the edges, and the ReLU activation function is adopted.

[0121] S602F: In the third pooling layer, perform a pooling operation on the third local feature map to obtain a third pooled feature map.

[0122] Optionally, the third pooling layer adopts a max pooling operation, with a pooling window size of 2×2 and a stride of 2.

[0123] S602G: In the first flattening layer, flatten the third pooled feature map into a one-dimensional vector to obtain a first flattened vector.

[0124] S602H: In the fourth pooling layer, perform a pooling operation on the first pooled feature map to obtain a fourth pooled feature map.

[0125] Optionally, the fourth pooling layer adopts a max pooling operation, with a pooling window size of 4×4 and a stride of 4.

[0126] S602I: In the second flattening layer, flatten the fourth pooled feature map into a one-dimensional vector to obtain a second flattened vector.

[0127] S602J: In the adaptive pooling layer, perform a pooling operation on the second pooled feature map to obtain a fifth pooled feature map.

[0128] Optionally, the adaptive pooling layer adopts an adaptive average pooling operation (i.e., adaptively pools the size of the input second pooled feature map to a preset size).

[0129] It should be noted that those skilled in the art can set the size of the preset size according to actual needs, and the present invention does not limit it here.

[0130] For example, when the preset size is 14×14, if the size of the input second pooling feature map is 56×56, the second pooling feature map with a size of 56×56 is adaptively pooled into a size of 14×14, the pooling window is automatically adjusted to 4×4 (because 56÷14 = 4), and the stride is 4.

[0131] S602K: In the third flattening layer, flatten the fifth pooling feature map into a one-dimensional vector to obtain a third flattened vector.

[0132] S602L: In the fully connected layer, compress the first flattened vector to obtain a first compressed flattened vector.

[0133] Optionally, the number of neurons in the fully connected layer is 512 or 256.

[0134] S602M: In the feature fusion layer, concatenate the first compressed flattened vector, the second flattened vector, and the third flattened vector.

[0135] S602N: In the fully connected layer, compress the concatenated flattened vector to obtain a second compressed flattened vector.

[0136] S602O: Output the second compressed flattened vector as the local texture feature vector.

[0137] In the embodiment of the present invention, the detailed convolution, pooling, flattening, and fully connected operations enable the model to gradually extract and fuse the local texture features of the ROI region of the right lower pulmonary artery, ensuring accurate capture of details. The multi-level convolutional layers (convolution kernels of different sizes) and pooling layers (max pooling and adaptive pooling) can effectively process multi-scale image features, while the flattening and feature fusion operations combine these features into a one-dimensional vector for subsequent classification and prediction.

[0138] S603: Through the cross-attention mechanism, fuse the global morphological feature vector and the local texture feature vector to obtain a fused feature vector.

[0139] Optionally, according to the following formula, fuse the global morphological feature vector and the local texture feature vector to obtain a fused feature vector:

[0140] F fusion = θ·F trans +(1 - θ)F CNN

[0141] θ = σ(W a [F trans ,F CNN )

[0142] Among them, F fusiondenotes the fused feature vector, θ denotes the correlation weight between the global morphological feature vector and the local texture feature vector, and F trans denotes the global morphological feature vector, and F CNN denotes the local texture feature vector, and W a denotes the learnable weight matrix.

[0143] In the embodiment of the present invention, by fusing the global morphological features and the local texture features, the model can more comprehensively understand the spatial and structural information of the image, thereby improving the prediction accuracy of the pulmonary artery hemodynamic parameters.

[0144] S604: According to the fused feature vector, through each prediction head, output the prediction results of the pulmonary artery hemodynamic parameters.

[0145] In the embodiment of the present invention, by processing the fused feature vector through each prediction head, multiple pulmonary artery hemodynamic parameters can be output simultaneously, improving the multi-task learning ability of the model and ensuring the accurate prediction of each parameter.

[0146] In a possible implementation manner, after S6, it further includes:

[0147] S7: Design a hybrid loss function.

[0148] Optionally, design the hybrid loss function according to the following formula:

[0149] L = λ1L MAE + λ2L GradMatch + λ3L Phy

[0150]

[0151] L Phy = ||PVP pred - PVP||

[0152]

[0153] where L denotes the hybrid loss function, λ1 denotes the weight coefficient of the MAE loss function, and L MAE denotes the MAE loss function, λ2 denotes the weight coefficient of the gradient matching loss function, and L GradMatch denotes the gradient matching loss function, λ3 denotes the weight coefficient of the physiological constraint loss function, and L Phy denotes the physiological constraint loss function, y n denotes the true value of the nth sample, Denote the predicted value of the nth sample, where n = 1, 2, …, N and N represents the total number of samples, ΔPVR represents the change in pulmonary vascular resistance (PVR), ΔsPAP represents the change in systolic pulmonary artery pressure (sPAP), and Z represents the measured data obtained from right heart catheterization (RHC). of the mean value, PVP pred Denote the predicted value of pulmonary vascular pressure (PVP), PVP represents the true value of pulmonary vascular pressure (PVP), mPAP represents the mean pulmonary artery pressure, B represents an empirical coefficient (used to adjust the relationship between mean pulmonary artery pressure (mPAP) and pulmonary artery wedge pressure (PAWP)), PAWP represents the pulmonary artery wedge pressure, and CO represents cardiac output, i.e., the amount of blood pumped by the heart per minute.

[0154] In the embodiment of the present invention, by combining the MAE loss, gradient matching loss, and physiological constraint loss, the accuracy and physiological rationality of the prediction results are ensured. The MAE loss ensures the minimization of the difference between the predicted value and the actual value, the gradient matching loss ensures that the prediction results conform to the physical laws of pulmonary artery hemodynamics, and the physiological constraint loss ensures the physiological consistency of the model, making the prediction results more clinically reliable.

[0155] S8: Aiming at minimizing the function value of the mixed loss function, optimize the pulmonary artery hemodynamic parameter prediction model through the whale optimization algorithm.

[0156] It should be noted that the Whale Optimization Algorithm (WOA) is a heuristic optimization algorithm that simulates the hunting behavior of whale groups. The algorithm is inspired by the hunting method of humpback whales, especially their strategy of using "bubble nets" to catch prey. Whales search for food by forming spiral and spiral contraction search paths while swimming in the ocean. WOA simulates this process and uses spiral motion and position update mechanisms for global search and local search to find the optimal solution. The algorithm adjusts the positions of whale individuals, balances the global exploration and local exploitation capabilities, and is suitable for a variety of optimization problems, especially performing well in high-dimensional complex problems and non-linear problems.

[0157] In the embodiment of the present invention, the whale optimization algorithm globally searches for the optimal solution by simulating the hunting behavior of whales, while balancing the local and global exploration capabilities. The algorithm is suitable for complex and high-dimensional optimization problems, does not require a large amount of prior knowledge, can efficiently adjust the model parameters, thereby improving the performance of the model and accelerating the convergence process. Combining this optimization algorithm, the model can achieve higher accuracy in multi-task learning.

[0158] In a possible implementation manner, S8 specifically includes sub-steps S801 to S808:

[0159] S801: Initialize the population to obtain the initial population, which includes multiple whale individuals, and each whale individual represents a set of parameters of a feasible pulmonary artery hemodynamic parameter prediction model.

[0160] S802: Use the reciprocal of the mixed loss function as the fitness function of the whale optimization algorithm, calculate the fitness values of the whale individuals in the initial population, and use the position of the whale individual with the maximum fitness value as the position of the current whale individual.

[0161] S803: Update the control factors:

[0162] A = 2a·r1 - a

[0163] E = 2r2

[0164]

[0165] where A, a, and E all represent control factors, r1 and r2 both represent random numbers between [0, 1], l represents a random number between [-1, 1], T max represents the maximum number of iterations, t represents the current iteration number, and T0 represents the normalized time scale factor.

[0166] S804: Determine whether the predation mechanism probability is less than 0.5. If so, enter S6025. Otherwise, enter the hunting stage.

[0167] Optionally, the calculation formula in the hunting stage is specifically:

[0168]

[0169] where D represents the distance between the position of the current whale individual and the position of the optimal whale individual in the hunting stage, X * (t) represents the position of the optimal whale individual at the t-th iteration, X(t) represents the position of the current whale individual at the t-th iteration, X(t + 1) represents the position of the whale individual at the (t + 1)-th iteration, b represents the logarithmic spiral shape constant, and ω represents the adaptive inertia weight.

[0170] It should be noted that by introducing the adaptive inertia weight ω into the whale optimization algorithm, the flexibility and efficiency of the search process can be significantly improved, and the balance between global exploration and local optimization can be enhanced. The dynamic adjustment of ω ensures that the algorithm intelligently adjusts the search strategy at different stages, avoiding over-reliance on local search or global search, thereby improving the convergence efficiency, calculation accuracy, and quality of the final result. This optimization strategy makes the optimization process of the controller more efficient, stable, and able to adapt to complex dynamic systems.

[0171] S805: Determine whether the absolute value of control factor A is less than 1. If so, enter the encirclement phase. Otherwise, enter the search phase.

[0172] Optionally, the calculation formula in the encirclement phase is specifically:

[0173]

[0174] where D ′ represents the distance between the position of the current whale individual and the position of the optimal whale individual in the encirclement phase.

[0175] Optionally, the calculation formula in the search phase is specifically:

[0176]

[0177] where D″ represents the distance between the position of the current whale individual and the position of a randomly selected whale individual in the search phase, X rand (t + 1) represents the position of a randomly selected whale individual at the (t + 1)-th iteration, p1 represents a random number with a value range of [0, 1], and F0 represents a constant with a value of 0.4.

[0178] S806: Update the position of the current whale individual according to the calculation results in the hunting phase, encirclement phase, and search phase.

[0179] S807: Calculate the fitness value of the updated whale individual. When the fitness value of the updated whale individual is greater than or equal to the fitness value of the current whale individual, update the position of the current whale individual. When the fitness value of the updated whale individual is less than the fitness value of the current whale individual, keep the position of the current whale individual unchanged.

[0180] S808: Determine whether the maximum number of iterations is reached. If so, output the optimal parameter set of the pulmonary artery hemodynamic parameter prediction model. Otherwise, return to step S802.

[0181] In the embodiment of the present invention, model optimization is performed through the whale optimization algorithm (WOA), which can achieve a balance between global and local searches, avoid over-reliance on local or global searches, and thus improve the optimization efficiency. WOA ensures that the optimization process finds the global optimal solution in a complex and high-dimensional parameter space by dynamically adjusting the control factor and search mechanism, significantly improving the accuracy of the prediction model. Its adaptive and dynamic optimization capabilities enable the model to converge efficiently and ensure calculation accuracy, thereby improving the performance and stability of the pulmonary artery hemodynamic parameter prediction model.

[0182] Refer to the attached instructions Figure 2, showing a schematic structural diagram of a pulmonary artery hemodynamic parameter prediction system provided by the present invention.

[0183] The present invention also provides a pulmonary artery hemodynamic parameter prediction system 20, which is applied to the above-mentioned pulmonary artery hemodynamic parameter prediction method, and includes:

[0184] A processor 201;

[0185] A memory 202, on which computer-readable instructions are stored. When the computer-readable instructions are executed by the processor 201, the pulmonary artery hemodynamic parameter prediction method described in the method embodiment is implemented.

[0186] The pulmonary artery hemodynamic parameter prediction system 20 provided by the present invention can execute the above-mentioned pulmonary artery hemodynamic parameter prediction method and achieve the same or similar technical effects. To avoid repetition, the present invention will not be elaborated herein.

[0187] The beneficial effects brought by the technical solution provided by the embodiment of the present invention at least include:

[0188] In the embodiment of the present invention, by obtaining the original CXR image of the patient, it no longer depends on right heart catheterization, is non-invasive, has low cost and small risk. By preprocessing the original CXR image, a target CXR image is obtained, which improves the image quality and reduces noise. By inputting the target CXR image into an image segmentation model for segmentation, an image of the right lower pulmonary artery ROI region is output, and the target CXR image is input into the Swin-Transformer branch, and the image of the right lower pulmonary artery ROI region is input into the CNN branch to output the prediction result of the pulmonary artery hemodynamic parameter. The segmentation accuracy and feature extraction ability are strong, and it can effectively extract the details of important regions such as the pulmonary artery, improving the accuracy and effect of diagnosis.

[0189] It should be understood that the processor in the embodiment of the present invention may be a central processing unit (CPU), and this processor may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or this processor may also be any conventional processor, etc.

[0190] It should also be understood that the memory in the embodiments of the present invention may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchlink DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0191] The above embodiments can be implemented in whole or in part by software, hardware (such as circuits), firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present invention are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (such as infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or a data center that contains one or more collections of available media. The available medium can be a magnetic medium (such as a floppy disk, hard disk, or magnetic tape), an optical medium (such as a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0192] It should be understood that the term "and / or" in this document is merely a description of the association relationship between associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. Here, A and B can be singular or plural. Additionally, the character " / " in this document generally represents an "or" relationship between the associated objects before and after, but it may also represent an "and / or" relationship, which can be specifically understood by referring to the context before and after.

[0193] In the present invention, "at least one" means one or more, and "a plurality" means two or more. "At least one of the following" or its similar expressions refer to any combination of these items, including any combination of single-item or plural-item. For example, at least one of a, b, or c can represent: a, b, c, a - b, a - c, b - c, or a - b - c, where a, b, and c can be single or multiple.

[0194] It should be understood that in various embodiments of the present invention, the magnitudes of the sequence numbers of the above processes do not imply the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

[0195] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in connection with the embodiments disclosed herein can be implemented in electronic hardware, or in a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Skilled professionals can use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of the present invention.

[0196] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the devices, apparatuses, and units described above can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0197] In several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division, and there can be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces, and the indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0198] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they can be located in one place, or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0199] In addition, the functional units in each embodiment of the present invention can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit.

[0200] When the above-mentioned function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0201] An embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored, characterized in that when the program is executed by a processor, it implements the pulmonary artery hemodynamic parameter prediction method as described in the method embodiment.

[0202] The computer-readable storage medium provided by the present invention can implement the steps and effects of the pulmonary artery hemodynamic parameter prediction method in the above method embodiment. To avoid repetition, the present invention will not elaborate further.

[0203] The beneficial effects brought by the technical solution provided by the embodiments of the present invention at least include:

[0204] In the embodiments of the present invention, by obtaining the original CXR image of the patient, it no longer depends on right heart catheterization, is non-invasive, has low cost and small risk. By preprocessing the original CXR image, a target CXR image is obtained, which improves the image quality and reduces noise. By inputting the target CXR image into an image segmentation model for segmentation, an image of the right lower pulmonary artery ROI region is output, and the target CXR image is input into the Swin-Transformer branch, and the image of the right lower pulmonary artery ROI region is input into the CNN branch, and a prediction result of the pulmonary artery hemodynamic parameters is output. The segmentation accuracy and feature extraction ability are strong, and it can effectively extract the details of important regions such as the pulmonary artery, improving the accuracy and effect of diagnosis.

[0205] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

[0206] The following points need to be explained:

[0207] (1) The accompanying drawings of the embodiments of the present invention only relate to the structures involved in the embodiments of the present invention, and other structures can refer to the general design.

[0208] (2) For clarity, in the accompanying drawings used to describe the embodiments of the present invention, the thickness of layers or regions is enlarged or reduced, that is, these drawings are not drawn to actual scale. It can be understood that when an element such as a layer, film, region, or substrate is referred to as being "on" or "under" another element, the element can be "directly" on or under the other element or there can be intervening elements.

[0209] (3) Without conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other to obtain new embodiments.

[0210] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. The protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for predicting pulmonary artery hemodynamic parameters, characterized in that, Including: S1: Obtain the original chest X-ray (CXR) image of the patient; S2: Preprocess the original CXR image to obtain the target CXR image; S3: Construct an image segmentation model based on the U-Net segmentation network; S4: Input the target CXR image into the image segmentation model for segmentation, and output the right lower pulmonary artery ROI region image; S5: Construct a pulmonary artery hemodynamic parameter prediction model, where the pulmonary artery hemodynamic parameter prediction model includes a Swin-Transformer branch and a CNN branch; S6: Input the target CXR image into the Swin-Transformer branch, input the right lower pulmonary artery ROI region image into the CNN branch, and fuse the results of the above two branches to output the prediction result of the pulmonary artery hemodynamic parameters.

2. The pulmonary artery hemodynamic parameter prediction method according to claim 1, wherein The pulmonary artery hemodynamic parameters specifically include: pulmonary artery systolic pressure, pulmonary vascular resistance, and pulmonary systemic blood flow ratio.

3. The method for predicting pulmonary artery hemodynamic parameters according to claim 1, wherein The S2 specifically includes: S201: Perform spatial alignment on the original CXR image through an elastic registration algorithm; S202: Perform intensity normalization on the spatially aligned original CXR image through the CLAHE algorithm; S203: Perform data augmentation on the intensity-normalized original CXR image through random rotation, translation, and gamma correction operations to obtain the target CXR image.

4. The method for predicting pulmonary artery hemodynamic parameters according to claim 1, wherein The image segmentation model includes an encoder and a decoder, where the encoder includes a DCD module, a Triplet Attention module, and a CBAM module.

5. The method for predicting pulmonary artery hemodynamic parameters according to claim 4, wherein The S4 specifically includes: S401: Perform dynamic convolution operations on the target CXR image in the DCD module to generate multi-scale feature maps; S402: Perform cross-dimensional enhancement processing on the multi-scale feature maps in the Triplet Attention module to generate cross-dimensionally enhanced feature maps; S403: Perform channel-spatial attention fusion operations on the cross-dimensionally enhanced feature maps in the CBAM module to generate fused feature maps; S404: In the decoder, perform upsampling operations, convolution operations, and activation operations on the fused feature maps in sequence to generate a right lower pulmonary artery ROI region probability map; S405: Perform binarization on the right lower pulmonary artery ROI region probability map to generate the right lower pulmonary artery ROI region image; S406: Output the right lower pulmonary artery ROI region image.

6. The method for predicting pulmonary artery hemodynamic parameters according to claim 1, wherein The CNN branch includes: multiple convolutional layers, multiple pooling layers, multiple flattening layers, a fully connected layer, and a feature fusion layer. The convolutional layers include a first convolutional layer, a second convolutional layer, and a third convolutional layer. The pooling layers include a first pooling layer, a second pooling layer, a third pooling layer, a fourth pooling layer, and an adaptive pooling layer. The flattening layers include a first flattening layer, a second flattening layer, and a third flattening layer.

7. The method for predicting pulmonary artery hemodynamic parameters according to claim 1, wherein The pulmonary artery hemodynamic parameter prediction model further includes: a cross-attention mechanism and a plurality of prediction heads, wherein the prediction heads include a pulmonary artery systolic pressure prediction head and a pulmonary to systemic blood flow ratio prediction head.

8. The pulmonary artery hemodynamic parameter prediction method according to claim 7, characterized in that, The specific steps of S6 include: S601: According to the target CXR image, output a global morphological feature vector through the Swin-Transformer branch; S602: According to the image of the right lower pulmonary artery ROI region, output a local texture feature vector through the CNN branch; S603: Through the cross-attention mechanism, fuse the global morphological feature vector and the local texture feature vector to obtain a fused feature vector; S604: According to the fused feature vector, output the prediction results of the pulmonary artery hemodynamic parameters through each of the prediction heads.

9. The method for predicting pulmonary artery hemodynamic parameters according to claim 1, wherein After S6, it further includes: S7: Design a hybrid loss function; S8: With the goal of minimizing the function value of the hybrid loss function, optimize the pulmonary artery hemodynamic parameter prediction model through the whale optimization algorithm.

10. A pulmonary artery hemodynamic parameter prediction system, characterized in that, It includes: A processor; A memory, on which computer-readable instructions are stored. When the computer-readable instructions are executed by the processor, the pulmonary artery hemodynamic parameter prediction method according to any one of claims 1 to 9 is implemented.

Images