A lung injury recognition system based on critical ultrasound multi-feature fusion

Abstract

This invention discloses a lung injury recognition system based on multi-feature fusion of critical care ultrasound, relating to the field of critical care ultrasound technology. The system includes acquiring multi-organ ultrasound images and measurement parameters of the lungs, heart, inferior vena cava, and kidneys from critically ill patients, obtaining a raw multi-organ ultrasound dataset. From the raw dataset, features such as lung slippage, pleural line and lung parenchymal changes, cardiac function and pulmonary circulation features, volume status and microcirculation features are extracted and standardized to form a multi-organ ultrasound feature set. This feature set is then used to construct a multi-organ ultrasound feature atlas and input into a pre-trained graph network model for inference, obtaining node-level and global-level abnormal indicators. Based on these abnormal indicators, the lung injury status category is determined. This invention, by constructing a multi-organ ultrasound feature atlas and introducing a graph network for feature inference, achieves a shift from judging based on single lung information to collaborative judging based on multi-organ related information.

Description

Technical Field

[0001] This invention relates to the field of critical care ultrasound technology, and in particular to a lung injury identification system based on multi-feature fusion of critical care ultrasound. Background Technology

[0002] With the development of critical care medicine and bedside ultrasound technology, bedside assessment of lung injury has gradually shifted from relying on offline imaging such as chest X-rays and CT scans to a dynamic monitoring model centered on lung ultrasound. In recent years, lung ultrasound has been widely used in diseases such as acute respiratory distress syndrome, mechanical ventilation-related lung injury, infectious pneumonia, and volume overload-related edema, and has gradually formed a series of qualitative and semi-quantitative assessment indicators represented by B-lines, pleural lines, and lung consolidation. At the same time, cardiac ultrasound, inferior vena cava ultrasound, and renal Doppler ultrasound are also used to reflect circulatory status and microcirculation perfusion, providing important evidence for the overall assessment of critically ill patients. The combined application of multi-organ ultrasound has become an important development direction in critical care monitoring.

[0003] However, existing methods for identifying lung injury primarily rely on single lung imaging features or simple multi-indicator parallel analysis, lacking structured modeling of the interactions between lung regions, between the lung and heart, between volume status, and between renal microcirculation. This makes it difficult for current methods to depict the true formation mechanism of lung injury under multi-organ coupling conditions, especially in the early and subclinical stages where typical consolidation has not yet appeared on imaging, resulting in significantly insufficient identification sensitivity. On the other hand, existing multi-organ information fusion methods mostly use feature splicing or simple weighting for statistical analysis, failing to utilize the spatial adjacency relationships of lung regions and the physiological coupling relationships of multiple organs for information propagation and correlation reasoning, making it difficult to simultaneously depict the spatial distribution characteristics of lung injury and the overall degree of involvement. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a lung injury identification system based on critical care ultrasound multi-feature fusion to solve the problems of existing technologies, such as the lack of multi-organ coupling relationship modeling and the difficulty in using spatial adjacency relationships of the lung region and physiological coupling relationships of multiple organs for correlation reasoning.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a lung injury identification system based on critical care ultrasound multi-feature fusion, which includes a data acquisition module that divides critically ill patients according to lung regions, acquires multi-organ ultrasound images and measurement parameters, and obtains a multi-organ ultrasound raw dataset.

[0008] The feature processing module extracts lung slippage, pleural line and lung parenchymal changes features from lung images, cardiac function and pulmonary circulation features from heart images, and volume status and microcirculation features from inferior vena cava and kidney images from the original multi-organ ultrasound dataset. It then performs standardization processing to form a multi-organ ultrasound feature set.

[0009] The graph inference module uses the feature vectors in the multi-organ ultrasound feature set as graph nodes, constructs node connections based on anatomical proximity and physiological coupling, and obtains multi-organ ultrasound feature graphs. The multi-organ ultrasound feature graphs are then input into a pre-trained graph network model for inference to obtain node-level and overall-level abnormality indicators.

[0010] The status determination module aggregates and categorizes the degree of abnormality in the lung region through node-level and overall-level abnormal indicators, and determines the lung injury status to obtain the lung injury status category.

[0011] As a preferred embodiment of the lung injury recognition system based on multi-feature fusion of critical care ultrasound described in this invention, the specific steps for dividing critically ill patients according to lung regions, acquiring multi-organ ultrasound images and measurement parameters, and obtaining a raw multi-organ ultrasound dataset are as follows.

[0012] Read the basic information of critically ill patients to generate patient identifiers and set up lung zone division schemes. Assign a number and probe placement range to each standard lung zone and generate standard lung zone information corresponding to the patient identifier.

[0013] By acquiring multiple consecutive frames of B-mode, M-mode and Doppler images in each standard lung region using standard lung region information, and reading lung region measurement parameters, heart region measurement parameters, inferior vena cava measurement parameters and kidney measurement parameters, multi-organ ultrasound images and measurement parameters are obtained and the acquisition time is recorded.

[0014] The acquisition time, patient identification, and examination round are combined into examination time information. The examination time information is then associated with and stored with multi-organ ultrasound images and measurement parameters to generate a raw multi-organ ultrasound dataset.

[0015] As a preferred embodiment of the lung injury recognition system based on multi-feature fusion of critical care ultrasound described in this invention, the specific steps for extracting lung slippage, pleural line, and lung parenchymal changes features from the original multi-organ ultrasound dataset are as follows:

[0016] The lung ultrasound segmentation model based on convolutional neural network is called to locate the pleural line in the image and segment the pleural line and the region of interest below the pleural line to obtain the segmentation region corresponding to each lung area.

[0017] By segmenting the region, the number and density of linear hyperechoic B-line in the area near the pleural line are counted, and the morphology of B-line is divided into isolated strips, bundled morphology and fused morphology, so as to obtain the number and density characteristics and morphological characteristics of B-line.

[0018] Optical flow tracking was performed on the pleural line position of the same lung region in multiple time frames to obtain lung feature vectors.

[0019] As a preferred embodiment of the lung injury recognition system based on multi-feature fusion of critical care ultrasound described in this invention, the specific steps for extracting cardiac function and pulmonary circulation features from cardiac images are as follows:

[0020] The curves of cardiac chamber area change during systole and diastole are obtained from cardiac images using inter-frame contour matching and area calculation methods.

[0021] Peak velocity and time integral features of the corresponding waveforms were extracted from the Doppler spectra of the pulmonary outflow tract and tricuspid valve region, and together with the cardiac chamber area change curve, they formed a feature vector of cardiac function and pulmonary circulation.

[0022] As a preferred embodiment of the lung injury recognition system based on multi-feature fusion of critical care ultrasound described in this invention, the steps for extracting volume status and microcirculation features from inferior vena cava and kidney images, and then standardizing these features to form a multi-organ ultrasound feature set, are as follows:

[0023] Edge detection was used to extract the contour of the inferior vena cava lumen from the inferior vena cava image, and the lumen diameter at different stages of respiration was extracted to obtain the maximum and minimum diameters. The degree of collapse was calculated by using the maximum and minimum diameters, and the maximum diameter, minimum diameter and degree of collapse were combined to obtain the volume state feature vector.

[0024] The thickness and echo homogeneity of the renal cortex were calculated from the kidney images. The systolic peak and end-diastolic velocity were extracted from the Doppler spectrum of the interrenal artery. The resistance index was calculated using the systolic peak and end-diastolic velocity. The thickness of the renal cortex, echo homogeneity, and resistance index were used as the renal feature vector.

[0025] The lung feature vector, cardiac function and pulmonary circulation feature vector, volume status feature vector, and kidney feature vector are converted into standardized values ​​with zero mean and unit variance and combined to obtain a multi-organ ultrasound feature set.

[0026] As a preferred embodiment of the lung injury identification system based on critical care ultrasound multi-feature fusion described in this invention, the steps of using feature vectors from the multi-organ ultrasound feature set as atlas nodes, and constructing node connections based on anatomical proximity and physiological coupling relationships to obtain a multi-organ ultrasound feature atlas are as follows:

[0027] Write the feature vectors from the multi-organ ultrasound feature set into the node attributes to generate a node representation of the multi-organ ultrasound feature set with a node list and feature matrix.

[0028] Different edge weights are set according to the connection relationships between different organ nodes;

[0029] By using the node representation of the multi-organ ultrasound feature set, and by establishing connections and assigning edge weights between nodes using the lung region adjacency matrix and multi-organ coupling rules, a multi-organ ultrasound feature map is obtained.

[0030] As a preferred embodiment of the lung injury recognition system based on multi-feature fusion of critical care ultrasound described in this invention, the specific steps for inputting multi-organ ultrasound feature maps into a pre-trained graph network model for inference to obtain node-level and global-level abnormality indicators are as follows.

[0031] Based on multi-organ ultrasound feature maps, the node feature matrix and connection relationship are input into a pre-trained graph network model, and multi-layer graph convolution operation is performed to obtain the node-level embedding vector corresponding to each node.

[0032] By using node-level embedding vectors, node-level anomaly indicators for each node are calculated at the output layer. Global pooling is then performed on the node-level embedding vectors to generate graph-level embedding vectors, and overall anomaly indicators are calculated based on the graph-level embedding vectors.

[0033] As a preferred embodiment of the lung injury identification system based on multi-feature fusion of critical care ultrasound described in this invention, the steps of aggregating and classifying the degree of abnormality in the lung region through node-level and global-level abnormal indicators, and determining the lung injury state to obtain the lung injury state category are as follows.

[0034] The lung region is divided into anterior, lateral and posterior regions. The node-level abnormal indicators in each region are compared with the preset abnormal threshold. Lung region nodes with node-level abnormal indicators greater than the abnormal threshold are recorded as abnormal lung region nodes.

[0035] The proportion of abnormal lung nodes to the total number of lung nodes in the region is calculated to obtain the proportion of abnormal lung nodes. The proportion of abnormal lung nodes is then compared with a preset regional proportion threshold to obtain the degree of regional abnormality.

[0036] The overall abnormality index is compared with the preset overall abnormality threshold to obtain a multidimensional abnormality description vector. The lung injury status is determined by the multidimensional abnormality description vector to obtain the lung injury status category.

[0037] As a preferred embodiment of the lung injury recognition system based on multi-feature fusion of critical care ultrasound according to the present invention, the specific steps for comparing the proportion of abnormal lung region nodes with a preset region proportion threshold to obtain the degree of regional abnormality are as follows:

[0038] When the proportion of abnormal lung region nodes is equal to zero, the region is marked as low-level abnormal;

[0039] When the proportion of abnormal lung region nodes is greater than zero and less than the preset region proportion threshold, the region is marked as medium-level abnormal.

[0040] When the proportion of anomalies in a region exceeds a preset threshold, the region will be marked as a high-level anomaly.

[0041] As a preferred embodiment of the lung injury recognition system based on multi-feature fusion of critical care ultrasound according to the present invention, the specific steps for comparing the overall abnormality index with a preset overall abnormality threshold to obtain a multi-dimensional abnormality description vector are as follows.

[0042] When the overall abnormality index is lower than the overall abnormality threshold, the current inspection is initially classified as a state of low overall impact.

[0043] When the overall abnormality index is greater than or equal to the overall abnormality threshold, the current inspection is initially classified as a state of high overall impact.

[0044] By combining the overall affected group data corresponding to the overall-level anomaly index with the regional-level anomaly degree, a multi-dimensional anomaly description vector is obtained.

[0045] The beneficial effects of this invention are as follows: by constructing a multi-organ ultrasound feature atlas and introducing a graph network for feature reasoning, the transformation from judging single lung information to collaborative judging of multi-organ related information is realized; by establishing node connections and information propagation based on the lung region adjacency matrix and multi-organ coupling rules, the risk of lung injury can be detected in advance by utilizing the correlation changes of adjacent lung regions and cardiac function, volume status, and kidney status when no typical abnormalities have appeared in a single lung region; by combining node-level abnormal indicators and overall-level abnormal indicators to determine the state of lung injury, the spatial distribution characteristics of lung injury and the overall degree of involvement are simultaneously characterized. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a schematic diagram of a lung injury identification system based on multi-feature fusion of critical care ultrasound.

[0048] Figure 2 This is a flowchart for acquiring ultrasound data from multiple organs in critically ill patients.

[0049] Figure 3 This is a flowchart for the extraction and standardization of ultrasound features from multiple organs.

[0050] Figure 4 A flowchart for constructing multi-organ ultrasound feature maps and identifying lung injuries. Detailed Implementation

[0051] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0052] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0053] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0054] Reference Figures 1-4 This is one embodiment of the present invention, which provides a lung injury identification system based on critical care ultrasound multi-feature fusion, comprising the following steps:

[0055] The data acquisition module divides critically ill patients into lung regions, acquires multi-organ ultrasound images and measurement parameters, and obtains raw multi-organ ultrasound datasets.

[0056] The system reads the basic information of critically ill patients, generates patient identifiers, sets up a lung region division scheme, and divides both lungs into several standard lung regions according to fixed anatomical landmarks. Each lung region has a number and a corresponding probe placement range. An ultrasound probe is placed in each corresponding lung region. High-frequency linear array or convex array probes are used for lung examination, phased array probes are used for cardiac examination, and abdominal probes are used for inferior vena cava and kidney examination.

[0057] During lung ultrasound acquisition, each lung region is scanned. The ultrasound equipment acquires multiple consecutive B-mode images and necessary M-mode and Doppler images to obtain image data for the corresponding lung region. Measurement parameters representing lung slippage, lung parenchymal echo, and pleural line display patterns are read as lung region measurement parameters. Left ventricular systolic function indices, right ventricular cavity diameter, pulmonary artery outflow tract Doppler waveform, and tricuspid regurgitation-related parameters are acquired as cardiac region measurement parameters. For the inferior vena cava, after the abdominal probe locates the inferior vena cava, multiple frames of images changing with respiration are continuously acquired, and an edge detection algorithm is used to identify the contour of the inferior vena cava lumen. The diameters at different respiratory phases were calculated and saved as measurement parameters along with corresponding image frames to obtain inferior vena cava measurement parameters and image data. For the kidneys, B-mode images and Doppler spectra of the interrenal or interlobar arteries were acquired on standard kidney sections. Peak systolic velocity, end-diastolic velocity, and resistance-related parameters were obtained through spectral analysis to acquire kidney measurement parameters and image data. The image data and measurement parameters of the lungs, heart, inferior vena cava, and kidneys were combined to obtain multi-organ ultrasound images and measurement parameters.

[0058] Record the acquisition time of each image and measurement parameter, combine the acquisition time, patient identification and examination round to form the examination time information, and store the examination time information in association with the multi-organ ultrasound images and measurement parameters to obtain the raw multi-organ ultrasound dataset.

[0059] The feature processing module extracts lung slippage, pleural line and lung parenchymal changes features from the original multi-organ ultrasound dataset, extracts cardiac function and pulmonary circulation features from the heart images, and extracts volume status and microcirculation features from the inferior vena cava and kidney images. These features are then standardized to form a multi-organ ultrasound feature set.

[0060] For lung images in the multi-organ ultrasound raw dataset, preprocessing was performed on the dynamic B-mode images corresponding to each lung region, including grayscale normalization, noise filtering, and contrast enhancement. A spatial filtering-based denoising algorithm and a histogram equalization-based contrast adjustment algorithm were employed to enhance the visibility of the pleural line and lung parenchymal structures. A lung ultrasound segmentation model based on a convolutional neural network was used to locate the pleural line in the image and segment the region of interest, obtaining the segmented region. Based on the segmented region, the number and density of B-line hyperechoic lines near the pleural line were statistically analyzed using a vertical line detection algorithm and spatiotemporal texture analysis. The morphology of the B-lines was classified into different categories such as isolated strips, bundled forms, and fused forms, obtaining the quantity and density features and morphological features of the B-lines. For lung sliding features, optical flow tracking is performed on the position of the pleural line in the same lung region across multiple time frames. The displacement trajectory of the pleural line during the respiratory cycle is calculated using a motion estimation algorithm based on optical flow. Specifically, multiple consecutive frames of images are input into the motion estimation algorithm based on optical flow. The displacement vector field of the gray block of the segmented region is calculated between adjacent frames, and the displacement vector is projected along the normal direction of the pleural line to obtain a one-dimensional displacement sequence of the pleural line on the time axis. The ratio of the one-dimensional displacement sequence to the time interval between adjacent frames is taken as the instantaneous displacement change rate. The statistic of the displacement change rate within the complete respiratory cycle is taken as the lung sliding velocity. The difference between the maximum and minimum values ​​of the displacement sequence within the same respiratory cycle is taken as the lung sliding amplitude. Lung sliding features containing lung sliding velocity and lung sliding amplitude are obtained.

[0061] It should be noted that the lung ultrasound segmentation model based on convolutional neural networks first collects lung ultrasound images with different lung regions, body positions, and image qualities from multiple centers. In the early stages of training, the pleural line contour is drawn and the subpleural region of interest is marked on each image using a manual annotation tool, forming a pixel-level or contour-level labeled dataset. The labeled dataset is divided into training and validation sets. Data augmentation processing such as random cropping, horizontal flipping, and brightness and contrast perturbation is performed on the training set images. The augmented images and corresponding annotations are input into a preset convolutional neural network segmentation structure. A loss function based on pixel-level cross-entropy is used, and the network parameters are iteratively updated through stochastic gradient descent. During training, the segmentation accuracy and loss changes are continuously calculated on the validation set. The learning rate, training epochs, and early stopping conditions are adjusted based on the performance on the validation set. Finally, a set of model parameters that performs stably in segmenting the pleural line and region of interest on the validation set is fixed as the lung ultrasound segmentation model for deployment.

[0062] To address the pleural ridge features, an edge detection algorithm is used to extract the pleural ridge contour. Specifically, within the longitudinal strip region containing the pleural ridge, a gradient operator-based edge detection algorithm is applied to extract continuous high-echo edges with significant gradient magnitude and direction. Edge pixels are sorted by row and column position, and the central contour curve of the pleural ridge is fitted. Texture analysis methods are used to calculate the continuity index and roughness of the pleural ridge. Specifically, the central contour curve of the pleural ridge is discretized into several sampling points at fixed intervals. For each sampling point, the presence of pleural ridge pixels in the neighborhood that satisfy the consistency of edge strength and direction is detected. The ratio of the number of effective sampling points to the total number of sampling points is used as the pleural ridge continuity index. The width and grayscale changes of the pleural ridge band structure along the normal direction near each sampling point are statistically analyzed. The roughness of the pleural ridge is obtained by calculating the bandwidth and the variance of the grayscale gradient. When interruptions or significantly irregular regions of the pleural ridge are found, region growing and connected component analysis algorithms are used. The process involves marking broken segments and thickened regions. Specifically, when a segment with edge breaks appears during continuous detection, the edges of the pleural line at both ends of the broken segment are used as seed points. A region growing algorithm is used to expand the abnormal region within the local area based on gray-level similarity and spatial proximity. Connected component analysis is then performed on the expanded pixel set to identify connected components representing pleural line interruptions, which are then marked as broken segments. In areas where the pleural line width is significantly greater than the surrounding average level, width anomalous points are used as seeds, and region growing and connected component analysis are used to mark the widened strip regions as pleural line thickening regions. For lung parenchymal changes, homogeneous echo areas, consolidated echo areas, and mixed echo areas are classified in the region below the pleural line. Subpleural consolidation is identified, and the area and shape parameters of the consolidated echo area are measured. The number and density features of B-lines, the morphological features of B-lines, lung sliding features, pleural line features, and the area and shape parameters of the consolidated area are combined to obtain a lung feature vector.

[0063] It should be noted that the roughness of the pleural ridge is obtained by calculating the variance of bandwidth and gray-level gradient because the pleural ridge is considered as a bright band extending longitudinally. If the pleural ridge is smooth and regular, then along this line, the bandwidth and gray-level gradient (edge ​​strength and variation) near each sampling point change very little, and the statistically obtained bandwidth and gray-level gradient sequences are relatively stable with small corresponding variances. Once the pleural ridge shows jaggedness, unevenness, local erosion, or fibrosis, it will manifest as sudden changes in bandwidth at certain locations. When the width or narrowing of the grayscale edge changes, and the intensity of the grayscale edge changes, the fluctuation of the bandwidth sequence and the grayscale gradient sequence will increase significantly, and the corresponding variance will increase. Based on this, the roughness of the pleural line can be obtained. The homogeneous echo zone is the lung parenchyma area without obvious consolidation or obvious turbidity. The consolidation echo zone refers to the continuous dense hyperechoic or relatively solid echo-like area. The internal texture of the consolidation echo zone is similar to that of the liver. The mixed echo zone is the area where there are obvious dense hyperechoic foci and surrounding relatively uniform punctate echoes, or areas with interspersed high and low echoes.

[0064] For cardiac images, the boundaries of the left and right ventricles are segmented. The systolic and diastolic chamber area change curves are obtained through inter-frame contour matching and area calculation methods. Specifically, a simple edge tracking method is used to delineate the chamber contours on the cardiac image as an initial reference contour. Intima boundaries with similar gray levels and significant edge gradients are searched near the initial reference contour. Inter-frame contour matching is used to fine-tune the initial reference contour frame by frame towards the true boundary position, obtaining the chamber closure contour for each frame. Pixels within the contour are counted based on the spatial calibration information of the image and converted into the corresponding chamber area value for the frame. The cardiac cycle position of each frame is marked according to the image time axis. The area values ​​of each frame are arranged in chronological order to obtain the systolic and diastolic chamber area change curves. Peak velocity and time integral features of the corresponding waveforms are extracted from the Doppler spectra of the pulmonary outflow tract and tricuspid valve region. Combined with right ventricular size and tricuspid regurgitation parameters, cardiac function and pulmonary circulation feature vectors of the cardiac image are constructed.

[0065] For inferior vena cava images, edge detection methods are used to extract the contour of the inferior vena cava lumen. The lumen diameter is extracted at different stages of respiration in a series of consecutive frames, and the change in diameter with the respiratory cycle is detected. The degree of collapse is calculated by the ratio of the difference between the maximum and minimum diameters to the maximum diameter. The maximum diameter, minimum diameter, and degree of collapse are combined to obtain a volume state feature vector. For kidney images, the renal cortex and medulla are segmented in the B-mode image, and the thickness and echo uniformity of the renal cortex are calculated. Specifically, the renal capsule boundary line and the corticomedullary boundary line are drawn along the contour of the renal capsule and the corticomedullary junction in the B-mode image, respectively. The volume state feature vector is obtained between these two boundary lines. Multiple vertical measurement lines are generated at fixed intervals. The distance between the renal capsule boundary and the corticomedullary junction is calculated along each measurement line. The distances of all measurement lines are statistically analyzed and used as the renal cortical thickness. Several small windows of interest are generated within the renal cortex region. The average gray value and gray value dispersion within each window are statistically analyzed. The gray value dispersion of all windows is summarized as the echo uniformity. The systolic peak value and end-diastolic velocity are extracted from the Doppler spectrum of the interrenal artery. The ratio of the difference between the systolic peak value and end-diastolic velocity to the systolic peak value is used as the resistance index. The renal cortical thickness, echo uniformity, and resistance index are used as the renal feature vector.

[0066] The lung feature vector, cardiac function and pulmonary circulation feature vector, volume status feature vector, and kidney feature vector are converted into standardized values ​​with zero mean and unit variance and combined to obtain a multi-organ ultrasound feature set.

[0067] The graph inference module uses the feature vectors in the multi-organ ultrasound feature set as graph nodes. It constructs node connections based on anatomical proximity and physiological coupling to obtain multi-organ ultrasound feature graphs. The multi-organ ultrasound feature graphs are then input into a pre-trained graph network model for inference to obtain node-level and overall-level abnormality indicators.

[0068] The feature vectors in the multi-organ ultrasound feature set are used as atlas nodes.

[0069] Furthermore, a unique identifier is assigned to each lung node, and the corresponding lung feature vector is written into the node attribute; for the heart node, the cardiac function and pulmonary circulation feature vectors are used as the attributes of the heart node; for the volume status node, the volume status feature vector is used as the attribute of the volume status node; for the kidney node, the kidney feature vector is used as the attribute of the kidney node; all nodes and node attributes are organized into a node list and a feature matrix, the node list records the node identifier and type label, and the feature matrix stores the feature vector of each node row by row.

[0070] Node connections are constructed using anatomical proximity and physiological coupling relationships.

[0071] Furthermore, an anatomical proximity relationship is described using a pre-defined lung region adjacency matrix. This matrix records the adjacency of different lung regions on the chest wall. Connection edges are created between adjacent lung region nodes based on the lung region adjacency matrix, representing the continuity of adjacent lung regions in anatomical structure and ventilation distribution. Physiological coupling relationships are represented using pre-defined multi-organ coupling rules. Functional connections are established between all lung region nodes and cardiac nodes to represent the impact of pulmonary circulatory load on blood perfusion and interstitial status in each lung region. Connections are established between lung region nodes and volume status nodes to express the impact of volume status on pulmonary venous pressure and pulmonary interstitial water content. Connections are established between lung region nodes and kidney nodes to express the impact of systemic inflammation and microcirculatory impairment on pulmonary capillary permeability. Interconnection edges are established between cardiac nodes, volume status nodes, and kidney nodes to reflect the relationship between cardiac output, venous return, and distal organ perfusion.

[0072] It should be noted that the preset lung region adjacency matrix is ​​generated after the division of the left and right anterior, lateral and posterior lung regions into upper and lower lung regions. Based on the relative position of each lung region in the body representation, two lung regions that share a boundary or are adjacent only by an anatomical boundary line are marked as adjacent lung regions. An adjacency table is generated internally, and the matrix element corresponding to each pair of adjacent lung regions in the adjacency table is set as adjacent, while other lung regions are set as non-adjacent.

[0073] Assign edge weights to each edge, organize the information of all nodes and edges, and obtain multi-organ ultrasound feature maps.

[0074] It should be noted that the edge weights are set at the beginning of the construction of the multi-organ ultrasound feature map. The connections between the same lung lobe or adjacent lung regions are set to higher initial weights, the connections between the lung region and the heart node and the lung region and the volume status node are set to medium initial weights, and the connections between the lung region and the kidney node are set to relatively low initial weights. The edge weights are iteratively updated during the training of the multi-organ ultrasound feature map.

[0075] Multi-organ ultrasound feature maps are input into a pre-trained graph network model for inference. The node feature matrix and connection relationship are subjected to multi-layer graph network operations. The pre-trained graph network model adopts a graph convolution structure. In each layer operation, the features of the current node are weighted, aggregated, and nonlinearly transformed according to the feature vectors of neighboring nodes and edge weights. Through multi-layer stacking, the graph network model generates an updated node embedding representation, forming a node-level embedding vector that can comprehensively reflect local features and neighborhood structure information.

[0076] It should be noted that the training process of the pre-trained graph network model involves collecting samples from historical severe cases that have undergone multi-organ ultrasound examinations and whose lung injury status has been determined by comprehensive clinical diagnosis. A corresponding multi-organ ultrasound feature atlas is constructed for each patient, and the lung injury labels (e.g., presence and severity of lung injury) or node-level labels associated with each atlas are compiled into a training labeled dataset. The dataset is divided into training and validation sets based on patients. The atlases in the training set are sequentially input into a graph convolutional network. In the forward computation, feature aggregation and nonlinear transformation are performed layer by layer based on the node feature matrix and adjacency relationships to obtain node-level outputs. The loss function is calculated together with the corresponding labels. A parameter update algorithm based on backpropagation is used to iteratively update the output layer bias term, node transformation weights, output layer weight vector, and edge weight parameters in the pre-trained graph network model. Multiple rounds of training are performed on the training set, and the model parameter combination that performs stably on the validation set is selected as the pre-trained graph network model parameters.

[0077] The mapping function of the output layer of the pre-trained graph network model maps the embedding vectors of each lung node and other organ nodes to node-level anomaly indicators, as expressed in the following expression:

[0078] ;

[0079] ;

[0080] in, Indicates the first The embedding vector of each node. This represents the weight vector corresponding to the output layer. This represents the bias term of the output layer. Indicates the first The intermediate quantity of each node after linear transformation at the output layer. Indicates the first Node-level anomaly indicators for each node.

[0081] Global aggregation is performed on all node embedding vectors, and global pooling operation is performed on all node embedding vectors by averaging across the feature dimensions. Multiple node embedding vectors are compressed into a single graph-level embedding vector, and an overall anomaly index is calculated based on the graph-level embedding vector.

[0082] The status determination module aggregates and categorizes the degree of abnormality in the lung region through node-level and overall-level abnormal indicators, and determines the lung injury status to obtain the lung injury status category.

[0083] The lung region is divided into multiple areas, including the anterior, lateral, and posterior regions. Lung nodes are grouped according to these regions. Node-level abnormality indicators are read for each node within each region and compared to a preset abnormality threshold. Nodes with abnormality indicators exceeding the threshold are designated as abnormal lung nodes. The number of abnormal lung nodes and their proportion to the total number of nodes in each region are calculated to obtain the abnormal lung node ratio. This ratio is then compared to a preset regional ratio threshold to classify regions into different levels of regional abnormality. Specifically, a region is marked as low-level abnormal when the abnormal lung node ratio is zero; a region is marked as medium-level abnormal when the ratio is greater than zero and less than the preset regional ratio threshold; and a region is marked as high-level abnormal when the ratio is greater than the preset regional ratio threshold.

[0084] It should be noted that the abnormal threshold is determined by collecting a large number of historical cases where clinical conclusions of lung injury / non-lung injury have been reached. The node-level abnormal indicators of the corresponding lung regions are compiled together with the true categories. Then, within the numerical range of the node-level abnormal indicators, several candidate thresholds are generated at fixed step sizes. For each candidate threshold, all training samples are judged according to the rule that an abnormal indicator greater than the candidate threshold indicates lung injury, otherwise it is judged as no lung injury. The results are compared with the true categories, and the number of samples incorrectly judged at the candidate threshold is counted. Finally, the value that minimizes the number of incorrectly judged samples is selected as the abnormal threshold, limiting the value range of the abnormal threshold to [0,1]. The preset region proportion threshold is determined by using historical cases to calculate the proportions of the anterior and lateral lung regions. The proportion of abnormal lung nodes in each region is statistically analyzed in the front and back regions. Within the range of abnormal lung node proportions, several candidate region proportion thresholds are generated with a fixed step size. For each candidate region proportion threshold, all regions of all samples are divided according to the following rules: abnormal lung node proportion greater than or equal to the candidate region proportion threshold is judged as high-level abnormality, abnormal lung node proportion greater than zero and less than the candidate region proportion threshold is judged as medium-level abnormality, and abnormal lung node proportion equal to zero is judged as low-level abnormality. The division is compared with manual judgment, and the number of incorrect divisions of each region under the candidate region proportion thresholds is counted. The value that minimizes the number of incorrect divisions is selected from all candidate region proportion thresholds as the preset region proportion threshold, and the value range of the preset region proportion threshold is limited to the range of [0,1].

[0085] The overall-level anomaly index is compared with the preset overall anomaly threshold. When the overall-level anomaly index is lower than the overall anomaly threshold, the current inspection is initially classified as a state of low overall impact. When the overall-level anomaly index is greater than or equal to the overall anomaly threshold, the current inspection is initially classified as a state of high overall impact. The overall impact group data corresponding to the overall-level anomaly index is combined with the low, medium and high regional anomaly levels of each region to obtain a multi-dimensional anomaly description vector.

[0086] It should be noted that the preset overall abnormal threshold is obtained by collecting a batch of historical cases that are judged to have a low overall level of involvement and a high overall level of involvement. The overall level abnormal indicators corresponding to the two types of cases are plotted as distribution histograms. The concentration intervals and overlap intervals of the two types of indicators on the numerical axis are compared. A valley point located between the main concentration interval of cases with low overall involvement and the main concentration interval of cases with high overall involvement is selected at the boundary between the two distribution curves as the overall abnormal threshold. The value range of the overall abnormal threshold is limited to the range of [0,1].

[0087] Lung injury status is determined using a multidimensional anomaly description vector. Specifically, when the overall severity of involvement is low and the regional anomalies in the anterior, lateral, and posterior regions are all low-grade, the lung injury status is classified as "no obvious lung injury." When the overall severity of involvement is high and at least two of the anterior, lateral, and posterior regions are high-grade, the lung injury status is classified as "definite lung injury." When the overall severity of involvement is low but there are intermediate or high-grade abnormal regions, or when the overall severity of involvement is high but only some regions are intermediate-grade abnormal, the lung injury status is classified as "early or subclinical lung injury."

[0088] In summary, this invention achieves a shift from judging single lung information to collaboratively judging multi-organ related information by constructing a multi-organ ultrasound feature atlas and introducing graph networks for feature reasoning; by establishing node connections and information propagation based on lung region adjacency matrices and multi-organ coupling rules, it enables the early detection of lung injury risk by utilizing the correlation changes of adjacent lung regions, cardiac function, volume status, and kidney status, even before typical abnormalities appear in a single lung region; and by combining node-level and overall-level abnormal indicators to determine the state of lung injury, it achieves simultaneous characterization of the spatial distribution characteristics of lung injury and the overall degree of involvement.

[0089] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A lung injury identification system based on multi-feature fusion of critical care ultrasound, characterized in that: include, The data acquisition module divides critically ill patients into lung regions, acquires multi-organ ultrasound images and measurement parameters, and obtains raw multi-organ ultrasound datasets. The feature processing module extracts lung slippage, pleural line and lung parenchymal changes features from lung images, cardiac function and pulmonary circulation features from heart images, and volume status and microcirculation features from inferior vena cava and kidney images from the original multi-organ ultrasound dataset. It then performs standardization processing to form a multi-organ ultrasound feature set. The graph inference module uses the feature vectors in the multi-organ ultrasound feature set as graph nodes, constructs node connections based on anatomical proximity and physiological coupling, and obtains multi-organ ultrasound feature graphs. The multi-organ ultrasound feature graphs are then input into a pre-trained graph network model for inference to obtain node-level and overall-level abnormality indicators. The status determination module aggregates and categorizes the degree of abnormality in the lung region through node-level and overall-level abnormal indicators, and determines the lung injury status to obtain the lung injury status category.

2. The lung injury identification system based on multi-feature fusion of critical care ultrasound as described in claim 1, characterized in that: The specific steps for classifying critically ill patients by lung region, acquiring multi-organ ultrasound images and measurement parameters, and obtaining raw multi-organ ultrasound datasets are as follows: Read the basic information of critically ill patients to generate patient identifiers and set up lung zone division schemes. Assign a number and probe placement range to each standard lung zone and generate standard lung zone information corresponding to the patient identifier. By acquiring multiple consecutive frames of B-mode, M-mode and Doppler images in each standard lung region using standard lung region information, and reading lung region measurement parameters, heart region measurement parameters, inferior vena cava measurement parameters and kidney measurement parameters, multi-organ ultrasound images and measurement parameters are obtained and the acquisition time is recorded. The acquisition time, patient identification, and examination round are combined into examination time information. The examination time information is then associated with and stored with multi-organ ultrasound images and measurement parameters to generate a raw multi-organ ultrasound dataset.

3. The lung injury identification system based on multi-feature fusion of critical care ultrasound as described in claim 2, characterized in that: The specific steps for extracting lung slippage, pleural line, and lung parenchymal changes from lung images in the multi-organ ultrasound raw dataset are as follows: The lung ultrasound segmentation model based on convolutional neural network is called to locate the pleural line in the image and segment the pleural line and the region of interest below the pleural line to obtain the segmentation region corresponding to each lung area. By segmenting the region, the number and density of linear hyperechoic B-line in the area near the pleural line are counted, and the morphology of B-line is divided into isolated strips, bundled morphology and fused morphology, so as to obtain the number and density characteristics and morphological characteristics of B-line. Optical flow tracking was performed on the pleural line position of the same lung region in multiple time frames to obtain lung feature vectors.

4. The lung injury identification system based on multi-feature fusion of critical care ultrasound as described in claim 3, characterized in that: The specific steps for extracting cardiac function and pulmonary circulation features from cardiac images are as follows: The curves of cardiac chamber area change during systole and diastole are obtained from cardiac images using inter-frame contour matching and area calculation methods. Peak velocity and time integral features of the corresponding waveforms were extracted from the Doppler spectra of the pulmonary outflow tract and tricuspid valve region, and together with the cardiac chamber area change curve, they formed a feature vector of cardiac function and pulmonary circulation.

5. The lung injury identification system based on multi-feature fusion of critical care ultrasound as described in claim 4, characterized in that: The extraction of volume status and microcirculation features from images of the inferior vena cava and kidneys is standardized to form a multi-organ ultrasound feature set. The specific steps are as follows: Edge detection was used to extract the contour of the inferior vena cava lumen from the inferior vena cava image, and the lumen diameter at different stages of respiration was extracted to obtain the maximum and minimum diameters. The degree of collapse was calculated by using the maximum and minimum diameters, and the maximum diameter, minimum diameter and degree of collapse were combined to obtain the volume state feature vector. The thickness and echo homogeneity of the renal cortex were calculated from the kidney images. The systolic peak and end-diastolic velocity were extracted from the Doppler spectrum of the interrenal artery. The resistance index was calculated using the systolic peak and end-diastolic velocity. The thickness of the renal cortex, echo homogeneity, and resistance index were used as the renal feature vector. The lung feature vector, cardiac function and pulmonary circulation feature vector, volume status feature vector, and kidney feature vector are converted into standardized values ​​with zero mean and unit variance and combined to obtain a multi-organ ultrasound feature set.

6. The lung injury identification system based on multi-feature fusion of critical care ultrasound as described in claim 5, characterized in that: The process involves using feature vectors from a multi-organ ultrasound feature set as atlas nodes, and constructing node connections based on anatomical proximity and physiological coupling relationships to obtain a multi-organ ultrasound feature atlas. The specific steps are as follows: Write the feature vectors from the multi-organ ultrasound feature set into the node attributes to generate a node representation of the multi-organ ultrasound feature set with a node list and feature matrix. Different edge weights are set according to the connection relationships between different organ nodes; By using the node representation of the multi-organ ultrasound feature set, and by establishing connections and assigning edge weights between nodes using the lung region adjacency matrix and multi-organ coupling rules, a multi-organ ultrasound feature map is obtained.

7. The lung injury identification system based on multi-feature fusion of critical care ultrasound as described in claim 6, characterized in that: The specific steps for inputting multi-organ ultrasound feature maps into a pre-trained graph network model for inference to obtain node-level and system-level abnormality indicators are as follows: Based on multi-organ ultrasound feature maps, the node feature matrix and connection relationship are input into a pre-trained graph network model, and multi-layer graph convolution operation is performed to obtain the node-level embedding vector corresponding to each node. By using node-level embedding vectors, node-level anomaly indicators for each node are calculated at the output layer. Global pooling is then performed on the node-level embedding vectors to generate graph-level embedding vectors, and overall anomaly indicators are calculated based on the graph-level embedding vectors.

8. The lung injury identification system based on multi-feature fusion of critical care ultrasound as described in claim 7, characterized in that: The process involves aggregating and categorizing lung region abnormalities using node-level and global-level abnormal indicators, determining the lung injury status, and obtaining the lung injury status category. The specific steps are as follows: The lung region is divided into anterior, lateral and posterior regions. The node-level abnormal indicators in each region are compared with the preset abnormal threshold. Lung region nodes with node-level abnormal indicators greater than the abnormal threshold are recorded as abnormal lung region nodes. The proportion of abnormal lung nodes to the total number of lung nodes in the region is calculated to obtain the proportion of abnormal lung nodes. The proportion of abnormal lung nodes is then compared with a preset regional proportion threshold to obtain the degree of regional abnormality. The overall abnormality index is compared with the preset overall abnormality threshold to obtain a multidimensional abnormality description vector. The lung injury status is determined by the multidimensional abnormality description vector to obtain the lung injury status category.

9. The lung injury identification system based on multi-feature fusion of critical care ultrasound as described in claim 8, characterized in that: The step of comparing the proportion of abnormal lung region nodes with a preset region proportion threshold to obtain the degree of regional abnormality is as follows: When the proportion of abnormal lung region nodes is equal to zero, the region is marked as low-level abnormal; When the proportion of abnormal lung region nodes is greater than zero and less than the preset region proportion threshold, the region is marked as medium-level abnormal. When the proportion of anomalies in a region exceeds a preset threshold, the region will be marked as a high-level anomaly.

10. The lung injury identification system based on multi-feature fusion of critical care ultrasound as described in claim 9, characterized in that: The step of comparing the overall anomaly index with a preset overall anomaly threshold to obtain a multi-dimensional anomaly description vector is as follows: When the overall abnormality index is lower than the overall abnormality threshold, the current inspection is initially classified as a state of low overall impact. When the overall abnormality index is greater than or equal to the overall abnormality threshold, the current inspection is initially classified as a state of high overall impact. By combining the overall affected group data corresponding to the overall-level anomaly index with the regional-level anomaly degree, a multi-dimensional anomaly description vector is obtained.

Images