A liver ultrasound standard section image recognition method and system

Abstract

The present application relates to the technical field of liver ultrasound examination, and more particularly to a liver ultrasound standard section image recognition method and system, which extracts multi-scale blood vessel topological features from a target ultrasound image, obtains node coordinates, connectivity and trend information, and constructs a blood vessel topological feature dataset; divides liver anatomical regions according to the dataset, calculates the recognition contribution of the topological features of each region to different standard sections, and generates a regional feature weight distribution map; constructs a blood vessel topological registration model based on the weight distribution map, implements position mapping and deformation deviation correction on the topological features of the input image, and obtains a corrected topological feature map; combines the corrected feature map and the weight distribution map, and uses a region-adaptive topological matching algorithm to complete standard section classification and recognition, effectively overcoming the interference of respiratory motion and individual differences, and improving the robustness and accuracy of liver ultrasound standard section recognition in complex scenarios.

Description

Technical Field

[0001] This invention relates to the field of liver ultrasound examination technology, and in particular to a method and system for recognizing standard cross-sectional images of liver ultrasound. Background Technology

[0002] Liver ultrasound is one of the most commonly used imaging techniques for the clinical diagnosis of hepatobiliary diseases. It is convenient to operate, provides real-time imaging, and is radiation-free, making it widely used in health checkups and disease screening. In liver ultrasound examinations, obtaining standardized sectional images is a crucial prerequisite for ensuring diagnostic accuracy. Standardized sections clearly display the anatomical structures of the liver and the course of important blood vessels, especially the topological distribution of the portal vein and hepatic veins, which is of great significance for judging liver morphology, assessing hemodynamics, and detecting space-occupying lesions. However, the acquisition of ultrasound images is highly dependent on the operator's experience and skill level. Images acquired by different physicians or by the same physician under different conditions may have deviations in sectional angles and incomplete structural visualization, leading to subjectivity and inconsistency in subsequent diagnoses.

[0003] Currently, some studies have attempted to utilize computer-aided techniques to achieve automatic recognition of standard ultrasound sections. Common methods are mostly based on image texture, organ contours, or local feature points for classification, such as using convolutional neural networks for end-to-end classification of the entire image. However, these methods often neglect the functional anatomical structures within the liver, especially the spatial topology of the intrahepatic vascular system, resulting in poor robustness in low-quality, low-contrast, or partially occluded ultrasound images. Furthermore, traditional methods typically treat the entire image as a homogeneous region, failing to fully consider the differences in contribution of different anatomical regions across different standard sections, and lacking a weighted guidance mechanism for key recognition areas.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide a method and system for recognizing standard ultrasound images of the liver. This invention aims to address the technical problems of existing methods for recognizing standard ultrasound images of the liver, which rely excessively on image texture or contour and neglect the topological features of intrahepatic blood vessels, and lack a differentiated weighting mechanism for the contribution of different anatomical regions, resulting in insufficient robustness and accuracy in low-quality or complex scenarios.

[0006] To achieve the above objectives, the present invention provides a method for recognizing standard ultrasound cross-sectional images of the liver, the method comprising: Multi-scale vascular topology feature extraction was performed on the input target liver ultrasound image to extract the node coordinates, connectivity and orientation information of intrahepatic blood vessels, and a vascular topology feature dataset was constructed. Based on the aforementioned vascular topology feature dataset, the liver interior is divided into anatomical regions, and the contribution of the topology features of each region to the recognition of different standard sections is calculated to obtain a region feature weight distribution map. Based on the region feature weight distribution map, a blood vessel topology registration model is constructed, and the topological features of the input image are mapped to their positions and deformed to correct for distortion, resulting in a corrected topological feature map. Based on the corrected topological feature map and the regional feature weight distribution map, the input image is classified and identified using a regional adaptive topological matching algorithm to obtain a recognition result that includes the cross-section category.

[0007] Optionally, the step of extracting multi-scale vascular topology features from the input target liver ultrasound image, extracting the node coordinates, connectivity, and orientation information of intrahepatic vessels, and constructing a vascular topology feature dataset includes: The input target liver ultrasound image is preprocessed for denoising and enhancement. A segmentation network is used to extract the binary mask of intrahepatic blood vessels. The center line and bifurcation nodes of all blood vessels are extracted by a skeletonization algorithm. Traverse all bifurcation nodes and record the spatial coordinates, vascular branch direction angle, and branch level information of each bifurcation node; Based on the information of all bifurcation nodes and vessel segments, an adjacency matrix describing the connectivity of vessels is constructed. Combining the spatial coordinates and angle information of the bifurcation nodes, a feature vector field of the vessel topology is established. The vascular topology feature dataset is constructed based on the relative positional relationship between the feature vector field and the intrahepatic anatomical regions.

[0008] Optionally, the step of dividing the liver interior into anatomical regions based on the vascular topology feature dataset and calculating the contribution of each region's topology features to the recognition of different standard cross-sections to obtain a region feature weight distribution map includes: The degree of topological difference between each bifurcation node and its adjacent vessel segments is calculated based on the vascular topology feature dataset, and the liver interior is initially divided into anatomical regions based on the degree of topological difference to obtain the initial region division results. The mean and variance of the topological features of all nodes within each initial region in the initial region partitioning result are calculated based on the vascular topological feature dataset. The initial regions are then adaptively merged and split based on the mean and variance of the topological features to obtain the target anatomical region partitioning result, which includes multiple anatomical regions. The contribution coefficient of the topological features of each region to the recognition of different standard sections is calculated based on the vascular topological feature data of each region, and a corresponding feature weight value is assigned to each region based on the recognition contribution coefficient. The target anatomical region division results and the feature weight values ​​corresponding to each region are integrated into a regional feature weight distribution map of the liver.

[0009] Optionally, the step of calculating the contribution coefficient of the topological features of each region to the recognition of different standard cross-sections based on the vascular topological feature data within each region, and assigning a corresponding feature weight value to each region based on the recognition contribution coefficient, includes: All vascular topological nodes in each region are sorted according to spatial coordinates. The topological feature distribution surface function in the region is obtained by fitting the least squares method. The region topological feature data is obtained by calculating the gradient and curvature of the topological feature distribution surface function. The gradient of the topological feature distribution surface function is used to calculate the normal vector of the blood vessel orientation at each point inside the liver. Based on the normal vector and the regional topological feature data, the matching angle between the standard anatomical orientation corresponding to each standard section and the current regional normal vector is calculated. By combining the matching angle with the gradient and curvature of the topological feature distribution surface, the initial recognition contribution coefficient is obtained; The initial identification contribution coefficient is weighted and averaged using the spatial distribution density of vascular nodes within the region, and the weighted average result is normalized to obtain the corrected identification contribution coefficient for each region in each standard section. When the correction recognition contribution coefficient of a region corresponding to a certain standard section is greater than the preset weight threshold, the corresponding feature weight value is set to the logarithm of the correction recognition contribution coefficient. When it is less than the weight threshold, the corresponding feature weight value is set to zero, thus obtaining the feature weight value of each region corresponding to each standard section.

[0010] Optionally, the step of constructing a blood vessel topology registration model based on the region feature weight distribution map, and performing position mapping and deformation deviation correction on the topological features of the input image to obtain a corrected topological feature map includes: Based on the pre-stored vascular topology template parameters of various standard sections and liver anatomy priors, a set of basic transformation equations for topological registration is established for each anatomical region. The feature weight values ​​in the region feature weight distribution map are combined with the basic transformation equations to construct a vascular topology registration model that includes weight factors. Based on the vascular topology registration model, calculate the transformation path and offset of each vascular topology node in the input image to the standard topology template; Based on the transformation path and offset, the spatial position mapping relationship of the topological nodes is calculated, and the topological features of the input image are corrected using the spatial position mapping relationship to obtain a corrected topological feature map.

[0011] Optionally, the step of calculating the spatial location mapping relationship of the topological nodes based on the transformation path and offset, and using the spatial location mapping relationship to correct the topological features of the input image to obtain a corrected topological feature map, includes: Using the transformation path and offset, combined with the overall anatomical contour data of the liver, the local deformation rotation angle of the blood vessel topology of the input image relative to the standard template is calculated, and a topological transformation equation considering respiratory motion deformation is established. Substitute the topological transformation equations into the topological prior parameters of different cross-sections to calculate the topological position deviation and feature weight deviation of each node. Based on the topological position deviation and feature weight deviation, establish the spatial position mapping matrix and feature weight compensation matrix of the topological nodes; For each node of the input topological feature, the spatial location mapping matrix is ​​used for position correction, and the feature weight compensation matrix is ​​used for feature weight compensation to obtain the corrected topological feature map.

[0012] Optionally, the step of using a region adaptive topology matching algorithm to perform standard cross-section classification and recognition on the input image based on the corrected topology feature map and the region feature weight distribution map, to obtain a recognition result containing cross-section categories, includes: Within each anatomical region, the regional topological feature matrix of the corrected topological feature map is extracted, and a feature description vector reflecting the regional topological distribution is constructed based on the regional topological feature matrix. Calculate the adaptive weight coefficient for each region based on the regional feature weight distribution map. The adaptive weight coefficient is proportional to the variance of the feature weight values ​​within the region. A region adaptive fusion function is constructed using the aforementioned weight adaptive coefficients, and the fusion parameters are dynamically adjusted according to the uniformity of the region feature distribution. The region adaptive fusion function is applied to the feature description vectors of all regions to generate global fusion topological features of the input image. The global fusion topological features are then matched and classified with a pre-constructed standard section feature library to obtain the final liver ultrasound standard section recognition result.

[0013] Furthermore, to achieve the above objectives, the present invention also provides a liver ultrasound standard section image recognition system, the system comprising: The topology extraction module is used to extract and acquire multi-scale vascular topology features from the input target liver ultrasound image, extracting the node coordinates, connectivity and orientation information of intrahepatic blood vessels, and constructing a vascular topology feature dataset. The weight mapping module is used to divide the anatomical regions inside the liver according to the vascular topology feature dataset, and calculate the contribution of the topology features of each region to the recognition of different standard sections, so as to obtain a region feature weight distribution map. The registration and correction module is used to construct a blood vessel topology registration model based on the region feature weight distribution map, perform position mapping and deformation deviation correction on the topological features of the input image, and obtain a corrected topological feature map. An adaptive recognition module is used to perform standard cross-section classification and recognition on the input image using a region adaptive topology matching algorithm based on the corrected topology feature map and the region feature weight distribution map, so as to obtain a recognition result containing cross-section categories.

[0014] In addition, to achieve the above objectives, the present invention also provides a liver ultrasound standard section image recognition device, the device comprising: a memory, a processor, and a liver ultrasound standard section image recognition program stored in the memory and executable on the processor, the liver ultrasound standard section image recognition program being configured to implement the steps of the liver ultrasound standard section image recognition method as described above.

[0015] In addition, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a liver ultrasound standard section image recognition program, which, when executed by a processor, implements the steps of the liver ultrasound standard section image recognition method as described above.

[0016] This invention provides a method for recognizing standard ultrasound images of the liver. This method extracts the node coordinates, connectivity, and orientation information of intrahepatic blood vessels to construct a multi-scale vascular topology feature dataset. It fully utilizes the stable and specific spatial distribution pattern of the liver vascular system in different standard sections, significantly improving the biological interpretability and structural sensitivity of image recognition. Furthermore, by combining topological features to divide the liver into anatomical regions and calculating the contribution of each region to the recognition of different standard sections, a regional feature weight distribution map is generated, achieving differentiated weighting of key recognition regions and enhancing the model's responsiveness to important anatomical areas. Based on this, a vascular topology registration model is constructed that can perform position mapping and deformation deviation correction on the topological features in the input image, effectively overcoming structural deformation problems caused by respiratory motion, probe angle deviation, or individual anatomical differences, and improving the consistency and robustness of feature expression. Finally, combining the corrected topology feature map and regional weight distribution, a region-adaptive topology matching algorithm is used to achieve standard section classification. This not only improves the recognition accuracy in complex clinical scenarios such as low quality, low contrast, or partial occlusion, but also enhances the algorithm's adaptability to different patients and operators. In summary, this invention provides a high-precision, robust, and interpretable scheme for identifying standard liver ultrasound sections, offering reliable technical support for automated quality control, intelligent auxiliary diagnosis, and teaching and training systems in ultrasound examinations. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an embodiment of the liver ultrasound standard section image recognition method of the present invention; Figure 2 This is a structural block diagram of an embodiment of the liver ultrasound standard section image recognition system of the present invention.

[0018] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0020] Reference Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the liver ultrasound standard section image recognition method of the present invention, which presents an embodiment of the liver ultrasound standard section image recognition method of the present invention.

[0021] In one embodiment, the liver ultrasound standard section image recognition method includes: Step S100: Multi-scale vascular topology feature extraction is performed on the input target liver ultrasound image to extract the node coordinates, connectivity and orientation information of intrahepatic blood vessels, and a vascular topology feature dataset is constructed.

[0022] The target liver ultrasound image can be an ultrasound examination image of the liver to be identified, containing grayscale or color Doppler imaging data of liver tissue and internal vascular structures. It can be used as the input source for the entire identification process, used for subsequent vascular feature extraction and cross-sectional classification. In an exemplary embodiment, the target liver ultrasound image can be acquired by an ultrasound device during a clinical examination and can be a two-dimensional B-mode image. The node coordinates of intrahepatic vessels can be the spatial location representation of the intersection or endpoints of major intrahepatic vascular branches in the image coordinate system. These can be used to construct spatial anchor points for the vascular topology, characterizing key geometric information about vascular distribution. For example, the node coordinates of intrahepatic vessels can be generated by detecting vascular bifurcation points or termination points after vascular segmentation and skeletonization. Further, the node coordinates of intrahepatic vessels can include, but are not limited to, the bifurcation point of the portal vein, the point where the hepatic vein merges into the inferior vena cava, and the terminal points of tertiary branches. Connectivity relationships can be the connection logic between intrahepatic vascular nodes, describing which nodes are directly connected through vascular segments. This can be used to reflect the network structure of the vascular system and support topological invariance modeling. In a specific embodiment, connectivity relationships can be expressed based on an adjacency matrix or edge list of the vascular skeleton graph.

[0023] Orientation information can be the direction vector or local curvature feature of a blood vessel segment within the image plane. It can be used to characterize the spatial extension trend of blood vessels and help distinguish typical course patterns of blood vessels in different standard cross-sections. Furthermore, orientation information can be obtained by differentiating the blood vessel centerline or fitting spline curves to describe directional changes. The blood vessel topology feature dataset can be a multi-scale structured blood vessel network representation formed by integrating node coordinates, connectivity relationships, and orientation information. It can be used to provide anatomically significant structural features for subsequent region segmentation and registration correction. In this embodiment, the blood vessel topology feature dataset can be constructed by extracting and fusing topological elements at multiple image resolutions or blood vessel scale levels. For example, the blood vessel topology feature dataset can include, but is not limited to, one or more of coarse-grained trunk topology, medium-grained secondary branch topology, and fine-grained terminal topology. Furthermore, the blood vessel topology feature dataset can provide a basis for anatomical region segmentation and also serve as input features for the blood vessel topology registration model.

[0024] Multi-scale vascular topology feature extraction is performed on the input target liver ultrasound image. This can be achieved by performing vascular enhancement, segmentation, and skeletonization on the image at multiple scales, extracting topological elements layer by layer. Further, this operation can be implemented by using Laplacian pyramids for multi-scale decomposition and extracting topology separately, or by using U-Net series networks to segment blood vessels at different coding layers and then uniformly constructing the topology. This captures the complete vascular network structure from the trunk to the tip, enhancing adaptability to images at different resolutions. Extracting the node coordinates, connectivity, and orientation information of intrahepatic vessels can be achieved by performing graph theory analysis on the vascular skeleton map to identify nodes, edges, and their geometric attributes. In a specific embodiment, this operation can be achieved by morphological refinement followed by node extraction using an intersection detection algorithm, predicting the vascular centerline using deep learning, and then constructing the graph. This forms a structured vascular network description that preserves anatomical spatial relationships. The resulting vascular topology feature dataset can be constructed by integrating the extracted node coordinates, connectivity, and orientation information into a unified data structure. For example, this operation can be implemented by storing the adjacency matrix and node feature matrix in the form of graph neural network compatible data, or by recording the edge list with coordinates in JSON format, thereby establishing a standardized topological representation that can be used for subsequent computation.

[0025] Step S200: Based on the vascular topology feature dataset, the liver interior is divided into anatomical regions, and the contribution of the topology features of each region to the recognition of different standard sections is calculated to obtain a region feature weight distribution map.

[0026] The anatomical region can be a functional or structural sub-region of the liver defined by the intrahepatic vascular topology. This sub-region can be used to divide the liver into local units with different diagnostic values, supporting differentiated weight calculation. In one specific embodiment, the anatomical region can be partitioned based on vascular branch attribution or watershed analysis methods to divide the liver parenchyma. For example, the anatomical region may include, but is not limited to, the upper segment of the left lateral lobe, the lower segment of the right anterior lobe, and the caudate lobe. The regional feature weight distribution map can be a two-dimensional weight mapping map that quantifies the contribution of each anatomical region to the recognition of different standard sections. This map can be used to guide the subsequent matching process to focus on high-discriminative regions, achieving a weighted response of key structures. In this embodiment, the regional feature weight distribution map can be used to construct a vascular topology registration model and input together with the corrected topology feature map into a region adaptive topology matching algorithm. Furthermore, the regional feature weight distribution map can be obtained by evaluating the correlation between the topological features of each region and the standard section label through statistical learning or attention mechanisms.

[0027] The liver's internal anatomical regions are divided based on a vascular topology feature dataset. This can be achieved by dividing the liver into several sub-regions based on vascular branch attribution or watershed segmentation methods. For example, this operation can be implemented using the Dijkstra algorithm to expand the region division outward from the main vessels and labeling regions based on the portal vein's blood supply range, thus achieving anatomically structure-driven region definition, distinct from traditional pixel clustering. The contribution of each region's topological features to the recognition of different standard cross-sections is calculated, which can be achieved by evaluating the correlation strength between region features and cross-section labels through statistical analysis or learnable weight mechanisms. In an exemplary embodiment, this operation can use gradient-weighted class activation mapping to infer region contributions and train region weights through a multi-instance learning framework, thereby quantifying the importance of different regions in cross-section discrimination and supporting differentiated processing. The resulting region feature weight distribution map can be obtained by mapping the contribution of each region back to the image space to form a continuous or discrete weight map. Furthermore, this operation can visualize the weights as a heatmap and store them as a single-channel floating-point image of the same size as the original image, thus providing a basis for spatial weighted matching.

[0028] Step S300: Based on the regional feature weight distribution map, construct a blood vessel topology registration model, perform position mapping and deformation deviation correction on the topological features of the input image, and obtain a corrected topological feature map.

[0029] The vascular topology registration model can be a computational model used to spatially align and correct the deformation of vascular topological features in an input image. It can compensate for topological deformations caused by respiratory motion, probe angle deviation, or individual anatomical differences. In one specific embodiment, the vascular topology registration model can be constructed based on a non-rigid registration algorithm guided by a region feature weight distribution map, such as thin-plate spline transformation or graph neural network alignment. For example, the vascular topology registration model can include, but is not limited to, graph matching-based registration models, field deformation-based registration models, and hybrid optimization registration models. The corrected topology feature map can be a representation of vascular topology features after position mapping and deformation correction. It can be used to eliminate structural deviations caused by physiological motion or operational factors, improving feature consistency. In this embodiment, the corrected topology feature map can be used as one of the main inputs to a region adaptive topology matching algorithm, in conjunction with the region feature weight distribution map. Furthermore, the corrected topology feature map can be obtained by spatially transforming the original topological features using the vascular topology registration model.

[0030] Constructing a vascular topology registration model based on the regional feature weight distribution map can be achieved by introducing regional weights as constraints or guiding terms into the registration optimization objective. For example, this operation can be implemented by adding weighted terms to the thin-plate spline energy function and using a graph attention network to achieve weighted node alignment, thereby allowing deformation correction to focus more on high-discriminative regions and improving registration accuracy. Position mapping and deformation deviation correction of the input image's topological features can be performed by transforming the original topological features to a standard space or reference template space through the registration model. In an exemplary embodiment, this operation can be achieved by applying a non-rigid transformation matrix to relocate node coordinates and implicitly correcting the topology structure through a graph convolutional network, thereby reducing structural variations caused by individual differences and operational factors. The corrected topological feature map is obtained, which can be the output of the corrected topological feature representation. Further, this operation can be implemented by reconstructing the graph structure with the corrected node coordinates and connectivity relationships, generating a topological mask map aligned with the original image, thereby providing a more consistent structural input for final matching.

[0031] Step S400: Based on the corrected topological feature map and the regional feature weight distribution map, the region adaptive topological matching algorithm is used to perform standard cross-section classification and recognition on the input image to obtain the recognition result containing the cross-section category.

[0032] The region-adaptive topology matching algorithm can be a matching calculation method that combines a corrected topological feature map with regional weight distribution to classify and identify standard sections. It can be used to accurately determine the category of standard sections, maintaining robustness, especially when image quality is poor. In a specific embodiment, the region-adaptive topology matching algorithm can calculate the similarity score between the input image and various standard section templates in a weighted topological space. For example, the region-adaptive topology matching algorithm may include, but is not limited to, weighted graph edit distance matching, region-aware topological embedding matching, and attention-guided similarity matching. The section category can be a predefined standard section type label in liver ultrasound examination, which can be used as the output category for the recognition task to determine whether the image conforms to the standard. For example, the section category may include, but is not limited to, the subxiphoid transverse section, the right intercostal oblique section, and the right subcostal longitudinal section. The recognition result can be the output of the standard section category to which the input target liver ultrasound image belongs, which can be used to provide a basis for structural compliance judgment for automated quality control or assisted diagnosis.

[0033] Based on the corrected topological feature map and the regional feature weight distribution map, a region-adaptive topological matching algorithm is used to classify and identify standard facets in the input image. This can be achieved by calculating the matching score between the input and various standard facet templates in the weighted topological space and selecting the optimal category. Further, this operation can be implemented by calculating the weighted map edit distance, taking the minimum value corresponding to the category, fusing the corrected topological embedding vector with the weights, and then inputting it into the classifier. This enables highly robust automatic standard facet discrimination. The resulting recognition result, including the facet category, can be the output label of the standard facet with the highest matching score, thus completing the final determination of the entire recognition process.

[0034] Taking the liver ultrasound quality control system in a primary hospital as an example, the liver ultrasound standard section image recognition method in this embodiment can be as follows: ultrasound physicians collect liver images for physical examinations in township health centers. The system receives the target liver ultrasound image in real time, automatically extracts the node coordinates, connectivity, and direction information of intrahepatic blood vessels, and constructs a multi-scale vascular topology feature dataset. Subsequently, based on this dataset, anatomical regions are divided and a regional feature weight distribution map is generated, with a focus on strengthening the weight of key regions such as the portal vein bifurcation area. Then, the vascular topology registration model is used to correct the vascular displacement caused by the patient's breathing, and a corrected topology feature map is obtained. Finally, combining the weight distribution and the corrected features, a regional adaptive topology matching algorithm is used to determine whether the image is a standard right intercostal oblique section and outputs the recognition result. If it does not meet the standard, a re-collection is prompted, thereby ensuring the quality of examinations at the primary level.

[0035] In one embodiment, multi-scale vascular topology feature extraction is performed on the input target liver ultrasound image to extract the node coordinates, connectivity, and orientation information of intrahepatic vessels, thereby constructing a vascular topology feature dataset, including: The input target liver ultrasound image is preprocessed for denoising and enhancement. A segmentation network is used to extract the binary mask of intrahepatic blood vessels. The center line and bifurcation nodes of all blood vessels are extracted by a skeletonization algorithm. The denoising and enhancement preprocessing can be an image preprocessing operation that suppresses noise and enhances structural contrast in the original liver ultrasound image. This can improve the visibility of vascular structures in low-contrast or noisy images, providing clearer input for subsequent segmentation. In this embodiment, the denoising and enhancement preprocessing can employ methods such as nonlocal mean filtering, anisotropic diffusion, or deep learning denoising models to improve image quality. For example, the denoising and enhancement preprocessing can use the BM3D algorithm for denoising followed by CLAHE contrast enhancement, and then use a U-Net denoising network for end-to-end image quality optimization. The binary mask of intrahepatic vessels can be a set of pixels in the image that identify the intrahepatic vascular region, represented in 0-1 form. This can be used to accurately separate vascular structures from background tissue, serving as the input basis for skeletonization processing. Furthermore, the binary mask of intrahepatic vessels can be generated by pixel-level classification of the preprocessed image using a segmentation network. The vascular centerline can be the central axis of the vascular structure in a two-dimensional image, reflecting the vascular orientation and topological skeleton. This can be used to simplify planar vessels into linear structures, facilitating node detection and graph construction. In one specific embodiment, the vessel centerline can be extracted as a single-pixel wide skeleton by applying morphological thinning or distance transformation to a binary mask. Bifurcation nodes can be intersections of three or more connecting directions on the vessel centerline, representing the locations of vessel branches. They can be used as key vertices in the topology graph, carrying spatial and hierarchical semantic information. Furthermore, bifurcation nodes can be identified by detecting pixels on the centerline with a degree greater than 2. Exemplarily, bifurcation nodes can include, but are not limited to, one or more of the following: primary trunk bifurcation nodes, secondary interlobular bifurcation nodes, and tertiary intrasegment bifurcation nodes.

[0036] The input target liver ultrasound image undergoes denoising and enhancement preprocessing, which can involve applying image enhancement algorithms to suppress speckle noise and improve the contrast of blood vessel edges. Further, this denoising and enhancement preprocessing can be achieved by using the BM3D algorithm for denoising followed by CLAHE contrast enhancement, and employing a U-Net denoising network for end-to-end image quality optimization. This improves the input conditions of low-quality images and reduces subsequent segmentation errors. The segmentation network extracts a binary mask of intrahepatic blood vessels. This can be achieved by inputting the preprocessed image into a trained blood vessel segmentation depth network, outputting a pixel-level blood vessel probability map, and then binarizing it. Further, the extraction of the intrahepatic blood vessel binary mask can be achieved using the nnU-Net adaptive segmentation framework and a ResUNet structure enhanced with an attention mechanism. This enables high-precision blood vessel region extraction, overcoming the failure of traditional thresholding methods under low contrast. The skeletonization algorithm extracts the centerline and bifurcation nodes of all blood vessels. This can be achieved by performing morphological refinement of the binary mask to a single pixel width and detecting pixels with a degree ≥3 as bifurcation nodes. Furthermore, the extraction of the centerline and bifurcation nodes of all blood vessels through the skeletonization algorithm can be achieved by using the Zhang-Suen thinning algorithm to extract nodes through graph traversal, and by combining distance transformation-based skeleton extraction with curvature detection to locate bifurcation. This allows the planar blood vessels to be transformed into a linear skeleton structure that can be used for graph modeling.

[0037] Traverse all bifurcation nodes and record the spatial coordinates, vascular branch direction angle, and branch level information of each bifurcation node; The spatial coordinates of the bifurcation node can be its (x, y) position in the image coordinate system, providing geometric anchor points for the topological structure and facilitating vector field and registration mapping. The branch orientation angle can be the angle value of each branch relative to a reference direction, representing the spatial orientation of the branches and enhancing the directional sensitivity of topological features. In one specific embodiment, the branch orientation angle can be calculated by fitting a local centerline segment to generate a direction vector and converting it to a polar angle. Branch hierarchy information can be the depth or generational number of the bifurcation node within the vascular tree structure, reflecting anatomical hierarchical relationships and supporting multi-scale topological modeling. Furthermore, branch hierarchy information can be assigned hierarchical labels through recursive traversal starting from the root node of the main vessel. Traversing all bifurcation nodes and recording the spatial coordinates, branch orientation angle, and branch hierarchy information for each node can be done by performing a depth-first or breadth-first traversal of the skeleton diagram, collecting geometric and hierarchical attributes at each bifurcation node. Furthermore, traversing all bifurcation nodes and recording the spatial coordinates, vascular branch direction angle, and branch level information of each bifurcation node can be achieved by recursively traversing and assigning levels starting from the portal vein entrance and simultaneously calculating the local direction, and implicitly encoding node attributes using a graph convolutional network. This allows the construction of a node feature set with rich semantics, supporting the subsequent establishment of a vector field.

[0038] Based on the information of all bifurcation nodes and vessel segments, an adjacency matrix describing the connectivity of vessels is constructed. Combining the spatial coordinates and angle information of the bifurcation nodes, a feature vector field of the vessel topology is established. Among them, the blood vessel segment can be a centerline segment connecting two adjacent bifurcation nodes or between a bifurcation node and an endpoint, and can be used to form the edge unit in the adjacency matrix to represent the blood vessel connectivity path.

[0039] An adjacency matrix can be a data structure that explicitly expresses the connectivity relationships between all branch nodes in matrix form, and can be used to provide a structured connectivity description for graph matching and topology analysis. Furthermore, the adjacency matrix can be obtained by setting the corresponding matrix element to 1 if there is a direct connection between two nodes via a blood vessel segment, and 0 otherwise. In an exemplary embodiment, the adjacency matrix, together with the spatial coordinates of the branch nodes, can form an input format that can be processed by a graph neural network. The feature vector field can be a topological field representation with directional information formed by fusing the spatial coordinates of the branch nodes and the directional angle of the blood vessel branches, which can be used to give the blood vessel topology both geometric location and local orientation semantics, improving structural discriminative power. Furthermore, the feature vector field can be obtained by embedding a direction vector defined by the directional angle at each node position to form a joint space-direction field.

[0040] Based on information from all bifurcation nodes and vessel segments, an adjacency matrix describing vascular connectivity is constructed, which can be based on the filled adjacency matrix between nodes in the skeleton graph. Furthermore, based on information from all bifurcation nodes and vessel segments, constructing an adjacency matrix describing vascular connectivity can be achieved by constructing a sparse adjacency matrix to save storage and introducing edge weights to represent vessel segment length or curvature, thus explicitly expressing the topological connectivity of the vascular network and facilitating graph matching computation. Combining the spatial coordinates and angle information of bifurcation nodes, a feature vector field of the vascular topology is established. This can be achieved by embedding a direction vector defined by the branch direction angle at each node location, forming a joint spatial-directional representation. Furthermore, combining the spatial coordinates and angle information of bifurcation nodes, establishing a feature vector field of the vascular topology can be achieved by interpolating around the nodes to obtain a continuous vector field and encoding the angles as two-dimensional vectors on the unit circle, thus endowing the topological structure with directional sensitivity and enhancing the ability to capture cross-sectional specific travel patterns.

[0041] Based on the relative positional relationship between the feature vector field and the intrahepatic anatomical regions, a vascular topology feature dataset is constructed.

[0042] The relative positional relationships of intrahepatic anatomical regions can be considered as spatial correspondences between vascular topology and standard liver anatomical divisions (such as Couinaud segments). This can be used to constrain the anatomical rationality of topological features, ensuring the dataset conforms to clinical standards. Furthermore, the relative positional relationships of intrahepatic anatomical regions can be obtained through prior anatomical atlas registration or by inferring region affiliation based on the location of major blood vessels. For example, the relative positional relationships of intrahepatic anatomical regions may include, but are not limited to, their positional relationships relative to the hepatic fissure, relative to the portal vein supply area, and relative to the hepatic vein drainage area. Based on the feature vector field and the relative positional relationships of intrahepatic anatomical regions, a vascular topology feature dataset is constructed. This can be achieved by fusing the feature vector field with anatomical region constraints to form a structured dataset that conforms to clinical anatomical patterns. Furthermore, the vascular topology feature dataset can be constructed by using region relationships as regularization terms to constrain the vector field and by adding region label fields to the dataset. This ensures that the topological features possess both discriminative power and anatomical rationality, providing high-quality input for subsequent steps.

[0043] Taking the compliance assessment of the liver ultrasound section in the teaching and training system as an example, the liver ultrasound standard section image recognition method in this embodiment can be achieved by medical students acquiring liver images during ultrasound simulation training. The system first performs noise reduction and enhancement preprocessing on the target liver ultrasound image to compensate for image blurring caused by unstable operation. Then, a segmentation network is used to extract the binary mask of intrahepatic vessels, which can accurately separate vessels even in low-contrast areas. Next, the skeletonization algorithm is used to obtain the vessel centerline and bifurcation nodes, and the spatial coordinates, branch direction angles and hierarchical information of each node are recorded. On this basis, an adjacency matrix is ​​constructed to express connectivity, and coordinates and angles are fused to establish a feature vector field. Finally, the relative positional relationship of intrahepatic anatomical regions (such as the right branch of the portal vein corresponding to the right anterior lobe) is combined to construct a vascular topology feature dataset, which is used to determine whether the image meets the requirements of the standard right subcostal longitudinal section, thereby providing real-time feedback on the operation quality.

[0044] In one embodiment, based on a vascular topology feature dataset, the liver interior is divided into anatomical regions, and the contribution of the topological features of each region to the recognition of different standard cross-sections is calculated to obtain a region feature weight distribution map, including: The degree of topological difference between each bifurcation node and its adjacent vessel segments is calculated based on the vascular topology feature dataset, and the initial anatomical region of the liver is divided based on the degree of topological difference to obtain the initial region division results. In this context, a bifurcation node can be a vascular node in the intrahepatic vascular network with three or more connecting edges, typically corresponding to an anatomical branch point. It can serve as a key anchor point for region partitioning, reflecting the branching structure of the vascular system. In an exemplary embodiment, bifurcation nodes can be obtained from nodes with a recognition score greater than or equal to 3 in a vascular topology feature dataset. Adjacent vascular segments can be a set of vascular segments directly connected to the same bifurcation node, used to calculate local topological differences and characterize the structural characteristics at the branch. For example, adjacent vascular segments can be obtained by traversing their adjacent edges starting from the bifurcation node through connectivity relationships. In a specific embodiment, adjacent vascular segments may include, but are not limited to, inflow segments, outflowing left branch segments, and outflowing right branch segments. The degree of topological difference can be a measure of the structural inconsistency between a bifurcation node and its adjacent vascular segments in terms of direction, length, or connectivity pattern. It can be used as a criterion for initial region partitioning, with high differences tending to form region boundaries. In this embodiment, the degree of topological difference can be calculated by comparing the angle between the direction vectors of each vascular segment at the node, changes in curvature, or scale differences. Furthermore, the degree of topological difference may include, but is not limited to, directional difference, scale difference, and connectivity complexity.

[0045] Initial anatomical region segmentation can be a preliminary region segmentation operation on the liver interior based on the degree of topological difference, which can be used to generate coarse-grained regions initially aligned with vascular branching structures. In a specific embodiment, the initial anatomical region segmentation can use regions with high topological difference as segmentation boundaries, and employ region growing or graph cut methods to segment the liver. For example, the initial anatomical region segmentation can be achieved by using Voronoi partitioning centered on bifurcation nodes, or by using the Watershed algorithm with the degree of difference as the gradient field, thereby aligning the initial region boundaries with the actual vascular branching structures. The initial region segmentation result can be a set of liver sub-regions obtained after the initial anatomical region segmentation, which can be used to provide the basis for regions to be optimized for subsequent statistical analysis and structural adjustment. In an exemplary embodiment, the initial region segmentation result can include, but is not limited to, one or more of the following: star-shaped radial regions, tree-like nested regions, and isolated clump regions.

[0046] Calculating the topological difference between each bifurcation node and its adjacent vessel segments based on a vascular topology feature dataset can be achieved by traversing all bifurcation nodes, extracting features such as the orientation and length of adjacent vessel segments, and calculating a structural difference metric. Further, this operation can be implemented by calculating the cosine distance between the direction vectors of adjacent vessel segments and measuring local irregularities based on graph Laplacian eigenvalues, thereby capturing local topological abrupt changes at vascular branches and providing a basis for region boundary localization. Initial anatomical region division within the liver based on the degree of topological difference can be performed by setting regions with high topological difference as segmentation boundaries and using region growing or graph cut methods to divide the liver. Further, this operation can be implemented by performing Voronoi partitioning centered on bifurcation nodes and using the Watershed algorithm with the degree of difference as the gradient field, thereby aligning the initial region boundaries with the actual vascular branch structure. The initial region division results can be output as a labeled map or a list of regions, providing a base set for subsequent optimization.

[0047] The mean and variance of the topological features of all nodes within each initial region in the initial region partitioning result are calculated based on the vascular topological feature dataset. The initial regions are then adaptively merged and split based on the mean and variance of the topological features to obtain the target anatomical region partitioning result, which includes multiple anatomical regions. The topological feature mean can be the statistical average of the topological features (such as orientation and connectivity) of all nodes within an initial region. It can be used to characterize the overall topological tendency of the region and is used for regional consistency assessment. In one specific embodiment, the topological feature mean can be obtained by calculating the arithmetic mean of the node feature vectors within the region along their dimensions. The topological feature variance can be a measure of the dispersion of the topological features of all nodes within an initial region. It can be used to reflect structural heterogeneity within the region; high variance suggests the need for splitting. For example, the topological feature variance can be obtained by calculating the variance of the node feature vectors within the region along their dimensions. The target anatomical region partitioning result can be the final liver anatomical region partitioning obtained after adaptive merging and splitting adjustments. It can be used to form functional regional units that conform to anatomical structure and possess internal consistency. In this embodiment, the target anatomical region partitioning result can be obtained by dynamically adjusting the initial region boundaries based on a threshold or clustering criterion set according to the topological feature mean and variance. Furthermore, the target anatomical region partitioning result can include, but is not limited to, merging dominant regions, splitting dominant regions, and stable preservation regions, among others. In an exemplary embodiment, the target anatomical region partitioning result serves as the spatial skeleton of the regional feature weight distribution map, supporting weight allocation.

[0048] The mean and variance of topological features of all nodes within each initial region are calculated based on the vascular topological feature dataset. This can be achieved by calculating the statistical mean and variance of the node features within each initial region separately. Furthermore, this operation can be implemented by independently calculating statistics along the feature dimension and using a robust estimator to reduce the influence of outlier nodes, thereby quantifying the consistency and heterogeneity of the internal structure of the region. Adaptive merging and splitting adjustments are then made to the initial regions based on the mean and variance of the topological features. This can be achieved by setting a mean similarity threshold for merging adjacent regions and setting a variance upper limit to trigger region splitting. Further, this operation can be achieved by using hierarchical clustering to merge low-discrepancy regions and performing bisectioning along the principal component direction in high-variance regions, thereby optimizing the region partitioning to achieve both anatomical rationality and feature homogeneity. The resulting target anatomical region partitioning can be output as the adjusted final region partitioning, thus forming a stable anatomical region structure suitable for weight allocation.

[0049] The contribution coefficient of the topological features of each region to the recognition of different standard sections is calculated based on the vascular topological feature data of each region, and a corresponding feature weight value is assigned to each region based on the recognition contribution coefficient. The identification contribution coefficient can be a numerical indicator that quantifies the ability of the topological features of an anatomical region to discriminate against a specific standard section, and can be used as a direct basis for assigning feature weight values. In a specific embodiment, the identification contribution coefficient can be obtained through correlation analysis between region features and section labels, gradient importance, or learnable attention scores. The feature weight value can be a numerical weight assigned to each anatomical region, reflecting its importance in standard section recognition, and can be used to construct a region feature weight distribution map to achieve differentiated weighted matching. For example, the feature weight value can be obtained by normalizing or nonlinearly mapping the identification contribution coefficient. In an exemplary embodiment, the feature weight value can map the contribution coefficient to a weight value in the interval [0, 1], which can be obtained through softmax or sigmoid transformation.

[0050] Calculating the contribution coefficient of the topological features of each region to the recognition of different standard sections based on the vascular topological feature data can analyze the discriminative association strength between regional features and various standard section labels. Furthermore, this operation can be achieved by using linear discriminant analysis projection scores as contribution coefficients and learning region-section association weights through a multi-head attention mechanism, thereby enabling a quantitative assessment of regional diagnostic value. Assigning corresponding feature weight values ​​to each region based on the recognition contribution coefficients can be done by mapping the contribution coefficients to weight values ​​within the interval [0, 1], which can be achieved through softmax or sigmoid transformation. Further, this operation can be achieved by exponentially scaling and normalizing the contribution coefficients, and generating weight values ​​according to section categories, thereby assigning higher response weights to key regions.

[0051] Integrating the target anatomical region segmentation results and the corresponding feature weight values ​​for each region into a regional feature weight distribution map of the liver can be achieved by fusing discrete region labels with their corresponding weight values ​​to generate a continuous weight map spatially aligned with the original image. Furthermore, this operation can be implemented by multiplying the region mask by the weight values ​​and then overlaying them to generate a heatmap, which is then stored as a region annotation file with weighted attributes. This results in a spatially weighted guiding map that can be used for subsequent registration and matching.

[0052] Taking the standardization assessment of the slicing in an ultrasound teaching and training system as an example, the liver ultrasound standard slicing image recognition method in this embodiment can be implemented by medical students acquiring liver ultrasound images during simulated training. The system first identifies bifurcation nodes and their adjacent vessel segments from the vascular topology feature dataset, calculates the degree of topological difference, and divides the initial anatomical regions. Subsequently, it analyzes the mean and variance of the topological features of nodes within each initial region, automatically merges structurally similar regions, and splits regions with high heterogeneity to form the target anatomical region division result. Then, it calculates the contribution coefficient of each region to the recognition of standard slicing such as the subxiphoid transverse section, and assigns feature weight values ​​accordingly. Finally, it integrates and generates a regional feature weight distribution map. Based on this, the system judges whether the images acquired by the student accurately display high-weight regions such as the bifurcation of the portal vein. If key regions are missing or the weight response is weak, it indicates a slicing angle deviation, assisting in teaching feedback.

[0053] In one embodiment, the contribution coefficient of the topological features of each region to the recognition of different standard cross-sections is calculated based on the vascular topological feature data within each region, and a corresponding feature weight value is assigned to each region based on the recognition contribution coefficient, including: All vascular topological nodes in each region are sorted according to spatial coordinates. The topological feature distribution surface function in the region is obtained by fitting the least squares method. The gradient and curvature of the topological feature distribution surface function are calculated to obtain the region topological feature data. In this context, vascular topology nodes can be discrete points with coordinate positions and topological attributes within the intrahepatic vascular network, including bifurcation points and endpoints. These nodes can serve as the basic data units for fitting a surface function representing the topological feature distribution. In an exemplary embodiment, vascular topology nodes may include, but are not limited to, trunk connection nodes, branch transition nodes, and terminal termination nodes. Spatial coordinates can be the two-dimensional position representation (x, y) of the vascular topology nodes in the image plane, serving as a spatial reference for node sorting and surface fitting. For example, spatial coordinates can be pixel coordinates, normalized coordinates, or anatomically standardized coordinates. Sorting all vascular topology nodes within each region according to their spatial coordinates can be done by ascending or descending order based on x or y coordinates, or by organizing them according to the spatial scanning order. Furthermore, sorting all vascular topology nodes within each region according to their spatial coordinates can be achieved by sorting by row priority or by polar coordinate angle, thereby improving fitting stability.

[0054] Least squares can be a mathematical optimization method that fits function parameters by minimizing the sum of squared errors. It can be used to fit a continuous topological feature distribution surface function from discrete vascular topological nodes. The topological feature distribution surface function within a region can be obtained by fitting the least squares method with node coordinates as input and topological attributes (such as orientation angle) as output, solving for the optimal surface parameters. Furthermore, the topological feature distribution surface function within a region can be obtained by using quadratic polynomial basis functions for global fitting and local weighted least squares for piecewise fitting, thereby establishing a continuous regional topological representation and supporting differential geometric analysis. The topological feature distribution surface function can be a mathematical function describing the continuous change of vascular topological features (such as orientation and density) with spatial location within a region. It can be used to upgrade discrete topological information to a continuous geometric representation, supporting gradient and curvature calculations. In this embodiment, the topological feature distribution surface function is obtained by fitting the topological features of vascular topological nodes with the spatial coordinates as independent variables and their topological attributes as dependent variables using the least squares method. For example, the topological feature distribution surface function can be a quadratic polynomial surface, a radial basis function surface, a spline interpolation surface, etc.

[0055] The gradient can be the rate of change vector of the topological feature distribution surface function at various points in space. It can be used to reflect the direction of the most drastic change in local topological features and to derive the normal vector of blood vessel orientation. In one specific embodiment, the gradient is obtained by taking the first-order partial derivative of the topological feature distribution surface function. Curvature can be a measure of the degree of bending of the topological feature distribution surface function at various points in space. It can be used to characterize the complexity and nonlinearity of the regional structure and to assist in the calculation of contribution coefficients. In an exemplary embodiment, curvature is obtained by further differentiating the gradient or by calculating the Hessian matrix. By calculating the gradient and curvature of the topological feature distribution surface function, regional topological feature data can be obtained. This can be achieved by analytically differentiating or numerically differencing the fitted surface to extract the first and second derivative information. Furthermore, by calculating the gradient and curvature of the topological feature distribution surface function, regional topological feature data can be obtained by calculating the exact gradient through symbolic differentiation and approximating the gradient using the Sobel operator, thereby obtaining structure-sensitive local geometric features. Regional topological feature data can be a set of structure-sensitive features composed of the gradient and curvature of the topological feature distribution surface function, which can be used as input for calculating the matching angle and the initial identification contribution coefficient. In this embodiment, the regional topological feature data originates from the topological feature distribution surface function and serves to calculate the vessel orientation normal vector and the matching angle.

[0056] The gradient of the topological feature distribution surface function is used to calculate the normal vector of the blood vessel orientation at each point inside the liver. Based on the normal vector and the regional topological feature data, the matching angle between the standard anatomical orientation corresponding to each standard section and the current regional normal vector is calculated. Calculating the normal vector of blood vessel orientation at points within the liver using the gradient of the topological feature distribution surface function can be achieved by rotating the gradient vector counterclockwise by 90 degrees and normalizing it to obtain the normal direction. Furthermore, calculating the normal vector of blood vessel orientation at points within the liver using the gradient of the topological feature distribution surface function can be achieved by obtaining the normal through gradient orthogonalization and estimating the local orientation using principal component analysis, thus realizing the conversion from topological change rate to anatomical orientation. The blood vessel orientation normal vector can be a unit vector perpendicular to the local orientation of the blood vessel, derived from the gradient direction, and can be used to characterize the spatial orientation of the blood vessels in the current region for orientation matching with standard anatomical orientations. Standard anatomical orientations can be the ideal blood vessel orientation direction or normal reference corresponding to various standard liver ultrasound sections, and can be used as a benchmark for orientation matching to assess the consistency between the current region and the standard section. In a specific embodiment, standard anatomical orientations may include, but are not limited to, the portal vein horizontal normal corresponding to the subxiphoid transverse section, the hepatic vein oblique normal corresponding to the right intercostal oblique section, and the vertical normal corresponding to the right subcostal longitudinal section, etc. Calculating the matching angle between the standard anatomical orientation corresponding to each standard section and the current region's normal vector based on the normal vector and region topological feature data can be achieved by calculating the cosine and inverse cosine of the angle between the current normal vector and the pre-stored standard anatomical orientation vector. Furthermore, calculating the matching angle between the standard anatomical orientation corresponding to each standard section and the current region's normal vector based on the normal vector and region topological feature data can be achieved by using vector dot product to calculate the angle and measuring the deviation using the absolute value of the angle difference, thereby quantifying the directional consistency between the current region and the standard section.

[0057] By combining the matching angle with the gradient and curvature of the topological feature distribution surface, the initial recognition contribution coefficient is obtained; The matching angle can be the angle between the normal vector of the current region's blood vessel orientation and the corresponding standard anatomical orientation of a standard cross-section. It can be used to quantify directional consistency; the smaller the angle, the better it conforms to the anatomical requirements of the standard cross-section. In an exemplary embodiment, the matching angle is calculated using the vector dot product formula between two normal vectors. Combining the matching angle with the gradient and curvature of the topological feature distribution surface, an initial identification contribution coefficient is obtained. This can be achieved by designing a fusion function that weights directional consistency and structural complexity. Furthermore, the initial identification contribution coefficient, obtained by combining the matching angle with the gradient and curvature of the topological feature distribution surface, can be achieved by mapping the angle to weights using an exponential decay function and then multiplying it by the curvature, and then fusing multidimensional features using a multilayer perceptron, thereby generating a comprehensive discrimination index.

[0058] The initial identification contribution coefficient is weighted and averaged using the spatial distribution density of vascular nodes within the region, and the weighted average result is normalized to obtain the corrected identification contribution coefficient for each region in each standard section. The initial identification contribution coefficient can be the original region discrimination score calculated by fusing matching angle, gradient, and curvature, which can be used to initially reflect the region's identification value to a specific standard cross-section. In a specific embodiment, the initial identification contribution coefficient is generated by weighted combination of directional consistency and structural complexity indicators. For example, the initial identification contribution coefficient can be a linearly weighted initial coefficient, a product modulation initial coefficient, or a multi-channel splicing initial coefficient. The spatial distribution density of vascular nodes can be the number of vascular topological nodes per unit area, reflecting the density of the regional vascular structure, and can be used to reliably weight the initial identification contribution coefficient, suppressing the influence of noise in sparse regions. In an exemplary embodiment, the spatial distribution density of vascular nodes is calculated by kernel density estimation or grid counting. The weighted average of the initial identification contribution coefficient using the spatial distribution density of vascular nodes within the region can be achieved by using density as the weight and calculating the weighted average of the initial contribution coefficients of each point within the region. Furthermore, the weighted average of the initial identification contribution coefficient using the spatial distribution density of vascular nodes within the region can be achieved by using Gaussian kernel density as the weight and the number of nodes in the grid as the discrete weight, thereby reducing the interference of sparse or noisy regions on the overall evaluation.

[0059] The weighted average result can be the intermediate value after averaging the initial identification contribution coefficients with the spatial distribution density of blood vessel nodes as weights, which can be used to improve the statistical robustness of the contribution coefficients. In this embodiment, the weighted average result is used as the input for normalization processing, and the output is the corrected identification contribution coefficient. Normalization processing can be a mathematical transformation process that maps the weighted average result to a uniform numerical range, which can be used to make the contribution coefficients of different regions and different cross-sections comparable. In a specific embodiment, the normalization processing uses min-max scaling, z-score standardization, or softmax normalization. Normalizing the weighted average result to obtain the corrected identification contribution coefficient of each region corresponding to each standard cross-section can be achieved by mapping the weighted average of all regions to the same standard cross-section to the interval [0, 1]. Furthermore, normalizing the weighted average result to obtain the corrected identification contribution coefficient of each region corresponding to each standard cross-section can be achieved by using softmax normalization to highlight high contribution regions and using min-max scaling to preserve relative size, thereby ensuring the comparability of contribution coefficients between different regions. The corrected identification contribution coefficient can be the final region discrimination score after spatial density weighting and normalization, which can be used as a direct basis for setting feature weight values.

[0060] When the correction recognition contribution coefficient of a region corresponding to a certain standard section is greater than the preset weight threshold, the corresponding feature weight value is set to the logarithm of the correction recognition contribution coefficient. When it is less than the weight threshold, the corresponding feature weight value is set to zero, thus obtaining the feature weight value of each region corresponding to each standard section.

[0061] The weight threshold can be a preset numerical limit used to distinguish between valid and invalid region contributions. It can be used for contribution filtering to avoid interference from low-discriminatory regions in the matching process. When the corrected identification contribution coefficient of a region corresponding to a standard section is greater than the preset weight threshold, the corresponding feature weight value is set to the logarithm of the corrected identification contribution coefficient. This can be achieved by determining whether the coefficient exceeds the threshold; if so, the logarithm is used as the weight. Furthermore, setting the corresponding feature weight value to the logarithm of the corrected identification contribution coefficient when the corrected identification contribution coefficient of a region corresponding to a standard section is greater than the preset weight threshold can be achieved by using log(1+x) to avoid negative infinity and employing a logarithmic transformation with offset, thereby enabling nonlinear enhancement of high-contribution regions.

[0062] The logarithmic value can be the value obtained by transforming the corrected identification contribution coefficient using a logarithmic function. It can be used to amplify the weight differences in high-contribution regions and enhance the nonlinearity of the model response. In a specific embodiment, the logarithmic value is usually nonlinearly compressed using the natural logarithm or the logarithm to base 10. When the value is less than the weight threshold, the corresponding feature weight value is set to zero. This can be done by directly setting the low-contribution region to zero, thus shielding its influence in the matching process. Furthermore, when the value is less than the weight threshold, setting the corresponding feature weight value to zero can be achieved by setting a soft threshold for gradual decay or by using a hard threshold for direct truncation, thereby enabling region-selective activation. The feature weight value can be the final numerical weight assigned to each region under a specific standard section. It can be used to construct a regional feature weight distribution map to guide subsequent matching algorithms to focus on key regions. In a specific embodiment, the feature weight value is determined based on the relationship between the corrected identification contribution coefficient and the weight threshold, using a logarithmic mapping or a zeroing strategy. Obtaining the feature weight value of each region corresponding to each standard section can be done by outputting the final weight of each region under all standard sections. Furthermore, obtaining the feature weight value of each region corresponding to each standard section can complete the weight allocation and support the construction of the regional feature weight distribution map.

[0063] Taking the recognition of standard sections in low-quality ultrasound images as an example, the liver ultrasound standard section image recognition method in this embodiment can be used in low-contrast liver ultrasound images acquired in the emergency department, where some blood vessels are blurred. The system first sorts the vascular topological nodes in each anatomical region according to their spatial coordinates, and uses the least squares method to fit a topological feature distribution surface function. The gradient of this surface is used to calculate the normal vector of the blood vessel orientation, and the matching angle is obtained by comparing it with the standard anatomical orientation of the subxiphoid transverse section. An initial recognition contribution coefficient is generated by combining the angle, gradient amplitude, and curvature. Then, a weighted average is performed based on the spatial distribution density of the vascular nodes within the region, and normalized to obtain a corrected recognition contribution coefficient. For the portal vein bifurcation area, its correction coefficient is higher than the weight threshold, so a logarithmic value is assigned as the feature weight value, while the coefficient in the surrounding sparse areas is set to zero because it is lower than the threshold. The resulting regional feature weight distribution map significantly enhances the response of key structures, enabling the system to accurately identify standard sections even when the image quality is poor.

[0064] In one embodiment, a blood vessel topology registration model is constructed based on the region feature weight distribution map. The topological features of the input image are then mapped to their positions and deformed to correct for distortion, resulting in a corrected topological feature map, including: Based on the pre-stored vascular topology template parameters of various standard sections and liver anatomy priors, a set of basic transformation equations for topological registration is established for each anatomical region. The vascular topology template parameters for various standard sections can be pre-stored parametric descriptions representing typical topological structures of the liver vascular system under various standard sections. These parameters can provide anatomically standardized reference benchmarks for the registration process, guiding the input image to align with the standard structures. In this embodiment, the vascular topology template parameters for various standard sections are obtained through vascular topology extraction and statistical modeling of a large number of high-quality standard section images, including the mean of node distribution, connectivity patterns, and orientation constraints. Furthermore, the vascular topology template parameters for various standard sections may include, but are not limited to, the portal vein trifurcation template parameters corresponding to the subxiphoid transverse section, the radial distribution template parameters of the hepatic veins corresponding to the right intercostal oblique section, and the parallel orientation template parameters of the main trunk corresponding to the right subcostal longitudinal section. Liver anatomical priors can be medical knowledge or statistical models regarding the spatial layout, branching patterns, and regional affiliation of the liver's internal vascular system. These can be used to constrain the physiological rationality of the transformation equations and prevent the registration results from violating known anatomical rules. For example, liver anatomical priors may originate from anatomical atlases, clinical imaging databases, or expert-annotated structured knowledge bases. In one exemplary embodiment, liver anatomical priors may include, but are not limited to, Couinaud segmentation rules, portal vein blood supply basin models, and hepatic vein drainage region mappings. The basic transformation equation set may be a set of mathematical equations established for each anatomical region to describe the spatial transformation required from its input state to its standard template state. This can be used to provide an initial transformation framework for subsequent weighted registration, ensuring that region-level deformation conforms to anatomical logic. In a specific embodiment, the basic transformation equation set combines vascular topology template parameters with liver anatomical priors to set allowed degrees of freedom and constraints for deformation at the region level. Furthermore, the basic transformation equation set may include, but is not limited to, affine transformation equation sets, thin-plate spline non-rigid transformation equation sets, and graph Laplacian smoothing constraint equation sets.

[0065] Based on pre-stored vascular topology template parameters for various standard cross-sections and liver anatomical priors, a set of fundamental transformation equations for topological registration is established for each anatomical region. This can be achieved by combining the target structure defined by the template parameters with the constraint rules provided by the anatomical priors, setting the allowed transformation forms and degrees of freedom at the region level. Furthermore, this operation can be achieved by independently fitting local affine transformation matrices to each anatomical region and constructing regional smoothing constraint equations based on graph Laplacian regularization. This ensures that the deformation correction of each region is anatomically reasonable and avoids violating physiological structural laws.

[0066] By combining the feature weight values ​​in the regional feature weight distribution map with the basic transformation equations, a vascular topology registration model containing weight factors is constructed. The weighting factor can be a coefficient derived from the weight values ​​of the corresponding regions in the region feature weight distribution map, used to adjust the constraint strength of the transformation equation. It can be used to enable high-discriminative regions to obtain stronger position preservation or more precise alignment priority during registration. In this embodiment, the weighting factor directly references or functions the values ​​in the region feature weight distribution map and embeds them into the objective function or regularization term of the basic transformation equation. Furthermore, the weighting factor acts as a bridge connecting the region feature weight distribution map and the basic transformation equation set, jointly constituting a vascular topology registration model containing the weighting factor. The vascular topology registration model containing the weighting factor can be an optimized registration model that integrates region weight information and anatomical constraints, and can be used to guide the spatial correction of the topology. For example, the vascular topology registration model containing the weighting factor is constructed by introducing the weighting factor as a weighting term into the energy minimization objective function of the basic transformation equation set. In an exemplary embodiment, the vascular topology registration model containing the weighting factor receives the topological features of the input image as the object to be corrected, and outputs the spatial position mapping relationship to generate a corrected topological feature map.

[0067] By combining the feature weight values ​​in the regional feature weight distribution map with the basic transformation equations, a vascular topology registration model incorporating weight factors can be constructed. This can be achieved by introducing the weight values ​​as coefficients into the objective function or regularization term of the transformation equations, forming a weighted optimization problem. Furthermore, this operation can be implemented by using the squared weights as regularization coefficients in the energy function and by using the weights as loss weights to modulate the node matching error. This allows high-contribution regions to receive stronger constraints during correction, achieving differentiated registration.

[0068] Based on the vascular topology registration model, calculate the transformation path and offset of each vascular topology node in the input image to the standard topology template; In this context, vascular topology nodes can be key points of the intrahepatic vascular network extracted from the input image, such as bifurcation points or endpoints. These nodes can serve as basic units for registration, and their positional correction forms the basis for overall topological alignment. In a specific embodiment, vascular topology nodes may include, but are not limited to, trunk anchor points, branch connection points, and terminal termination points. The standard topology template can be a graphical model representing the ideal standard cross-sectional vascular structure, generated by instantiating vascular topology template parameters from various standard cross-sections. It can be used as a target reference for registration, defining the spatial configuration of vessels under each cross-section. Furthermore, the standard topology template and the topological features of the input image establish a correspondence through transformation paths and offsets. The transformation path can be the spatial transformation trajectory or mapping sequence experienced by a single vascular topology node from its original position to its corresponding position in the standard template. This can be used to describe the dynamic process of node correction and support nonlinear deformation modeling. For example, the transformation path may include, but is not limited to, direct linear interpolation paths, piecewise smooth transition paths, and continuous paths based on flow field integrals. The offset can be the vector difference between the positions of the vascular topology node before and after correction. This offset can be used to quantify the magnitude and direction of deformation correction and to construct spatial positional mapping relationships. In one exemplary embodiment, the offset may include, but is not limited to, a two-dimensional offset in the image pixel coordinate system, a relative offset in the normalized anatomical coordinate system, and radial and angular offsets in the polar coordinate system.

[0069] Based on the vascular topology registration model, the transformation path and offset of each vascular topology node in the input image to the standard topology template are calculated. This can be achieved by solving the optimal solution of the weighted registration model, obtaining the mapping parameters of each node from its original position to its target position. Furthermore, this operation can be implemented by matching nodes and calculating offsets using an iterative nearest-point algorithm, and transforming the path using implicit regression of a graph neural network. This allows for fine-grained, individualized node-level correction, preserving topological connectivity while aligning the structure.

[0070] Based on the transformation path and offset, the spatial position mapping relationship of the topological nodes is calculated, and the topological features of the input image are corrected using the spatial position mapping relationship to obtain the corrected topological feature map.

[0071] The spatial location mapping relationship can be a one-to-one correspondence function between all vascular topological nodes in the input image and their corrected positions. This can be used as the basis for the final correction operation, ensuring overall consistency of the topological structure. In this embodiment, the spatial location mapping relationship integrates the transformation paths and offsets of each node to construct a global or local spatial transformation field. Furthermore, the spatial location mapping relationship is applied to the topological features of the input image to generate a corrected topological feature map. Calculating the spatial location mapping relationship of the topological nodes based on the transformation path and offset can be achieved by integrating the correction parameters of all nodes to construct a unified spatial transformation field or mapping function. Further, this operation can be implemented by interpolating to obtain a dense displacement field and constructing a thin-plate spline transformation driven by sparse control points, thereby forming a global mapping rule that can be used for overall feature correction. Correcting the topological features of the input image using the spatial location mapping relationship to obtain a corrected topological feature map can be achieved by repositioning the node coordinates in the original topological features according to the mapping relationship while maintaining connectivity. Furthermore, this operation can be achieved by reconstructing edge connections after correcting only the node coordinates and simultaneously correcting the orientation information to maintain local geometric properties, thereby outputting a topological representation aligned with standard cross-sectional anatomy specifications and improving the consistency of subsequent identification.

[0072] For example, in the scenario of a multi-center ultrasound image standardization processing platform, the liver ultrasound standard section image recognition method in this embodiment can be used to upload liver ultrasound images from different hospitals to a cloud platform. The system first extracts the vascular topology features; then it calls the pre-stored right intercostal oblique section vascular topology template parameters and Couinaud segmentation prior knowledge to establish basic transformation equations for anatomical regions such as the left lateral lobe and right anterior lobe; next, it incorporates the previously generated regional feature weight distribution map (showing that the portal vein bifurcation area has a higher weight) into the equations to construct a vascular topology registration model containing weight factors; the model calculates the offset and transformation path of each vascular node from the standard template. For example, the high-weight portal vein trunk node is strictly aligned, while the low-weight terminal region is allowed to undergo moderate deformation; finally, a corrected topology feature map is generated, so that the image that originally caused the vascular tilt due to the difference in probe angle restores the standard radial distribution, providing structurally consistent input for subsequent cross-institutional unified classification and recognition.

[0073] In one embodiment, based on the transformation path and offset, the spatial location mapping relationship of the topological nodes is calculated, and the topological features of the input image are corrected using the spatial location mapping relationship to obtain a corrected topological feature map, including: By utilizing the transformation path and offset, combined with the overall anatomical contour data of the liver, the local deformation rotation angle of the blood vessel topology of the input image relative to the standard template is calculated, and a topological transformation equation considering respiratory motion deformation is established. The overall anatomical contour data of the liver can be geometric or semantic contour information describing the shape of the liver's outer boundary and its relative position with surrounding structures. This can be used to provide global spatial constraints for local deformation modeling and help distinguish between overall displacement and local rotational deformation. In an exemplary embodiment, the overall anatomical contour data of the liver can be obtained by segmenting ultrasound images to obtain the liver's outer edge, or by extracting a typical contour template from a standard atlas. The input image vascular topology can be a representation of the original vascular network structure extracted from the ultrasound image of the target liver, and can be used as the original object for deformation analysis and correction. For example, the input image vascular topology can include trunk-dominated topology, sparsely branched topology, partially occluded topology, etc. The local deformation rotation angle can be the amount of rotational deformation of the input image vascular topology relative to a standard template within a specific anatomical region. This can be used to characterize non-translational deformation caused by breathing or probe pressure, improving the accuracy of transformation modeling. In a specific embodiment, the local deformation rotation angle can be jointly estimated based on the transformation path, offset, and overall anatomical contour data of the liver, and calculated through local coordinate system alignment. Furthermore, local deformation rotation angles can include rotation around the Z-axis within the image plane, tilting rotation approximately around the portal vein trunk axis, and regional vortex deformation. Respiratory motion deformation can be a physiological phenomenon caused by the periodic movement of the liver position and relative displacement of its internal structures due to the patient's breathing. It can be used as a key interference factor in topological transformation equation modeling to enhance registration robustness. For example, respiratory motion deformation can include overall vertical displacement accompanied by slight rotation, interlobar shear deformation, and local compression deformation caused by diaphragmatic compression.

[0074] By utilizing the transformation path and offset, combined with overall liver anatomical contour data, the local deformation rotation angle of the vascular topology in the input image relative to the standard template is calculated. This can be achieved by estimating the rotation component within the local anatomical region based on the relationship between the node displacement vector and the contour normal vector. Further, this operation can be implemented by fitting the local vascular orientation using principal component analysis and calculating the angle difference, and by estimating the rotation using contour-guided polar coordinate transformation. This allows for the separation of translational and rotational deformation components, improving the accuracy of deformation modeling. A topological transformation equation considering respiratory motion deformation is established by embedding the local deformation rotation angle into the transformation model and introducing respiratory phase or contour constraint terms. In a specific embodiment, this operation can be achieved by constructing a periodic transformation equation with a time phase parameter and using the contour distance field as a regularization term to constrain the deformation range, making the transformation equation physiologically reasonable and adaptable to dynamic examination scenarios. The topological transformation equation can be a mathematical expression that integrates the local deformation rotation angle and the respiratory motion model, used to describe the mapping relationship of the vascular topology from the input state to the standard state. This can be used to achieve explicit modeling of complex physiological deformations, improving the anatomical rationality of the transformation. Furthermore, the topological transformation equations can include affine transformation equations with rotational compensation, contour-guided non-rigid field equations, and temporal breathing phase interpolation equations. These topological transformation equations serve as topological prior parameters for bridge connection transformation paths, offsets, and different cross-sections, used to calculate deviations.

[0075] Substitute the topological transformation equations into the topological prior parameters of different cross-sections to calculate the topological position deviation and feature weight deviation of each node. The topological prior parameters for different cross-sections can be pre-defined statistical parameters of vascular topology for various standard cross-sections, including node distribution, orientation range, and regional weight benchmarks. These parameters can be used to provide cross-section-specific reference standards for deviation calculation. For example, the topological prior parameters for different cross-sections may include the trichiasis symmetry parameter of the subxiphoid transverse section, the radial divergence parameter of the right intercostal oblique section, and the parallel trunk parameter of the right subcostal longitudinal section. The topological position deviation can be the geometric difference between the input topological nodes and their corresponding positions in the standard template before correction. It can be used to quantify the degree of spatial misalignment to be corrected and drive the construction of the spatial position mapping matrix. In an exemplary embodiment, the topological position deviation can be obtained by substituting the topological transformation equations into the topological prior parameters of different cross-sections and solving for the residuals.

[0076] Feature weight deviation can be the difference between the current regional weight of an input topological node and the weight it should have at that position under the standard section. It can be used to reflect the degree of semantic importance distortion and guide feature weight compensation. In a specific embodiment, feature weight deviation can be obtained by comparing the theoretical weight of the expected position of the node after correction in the regional feature weight distribution map with the current actual weight. Substituting the topological transformation equation into the topological prior parameters of different sections, the topological position deviation and feature weight deviation of each node can be calculated by comparing the transformed node position with the prior parameters and calculating the geometric and semantic residuals respectively. Furthermore, this operation can be achieved by obtaining the standard position through nearest neighbor matching, calculating the Euclidean distance and weight difference, and using the vector difference in the graph embedding space to simultaneously represent the two types of deviations, thereby realizing the quantitative evaluation of both structural and semantic deviations.

[0077] Based on the topological position deviation and feature weight deviation, establish the spatial position mapping matrix and feature weight compensation matrix of the topological nodes; The spatial location mapping matrix can be a linear or nonlinear transformation matrix used to correct the geometric positions of input topological nodes. It can be used to achieve precise relocation of vascular node coordinates and restore the standard anatomical layout. In one specific embodiment, the spatial location mapping matrix can be constructed based on topological position deviations and can be a globally unified matrix or a node-local matrix. The spatial location mapping matrix and the feature weight compensation matrix work together to complete the two-dimensional correction. The feature weight compensation matrix can be an adjustment matrix used to semantically correct the regional weights of input topological nodes. It can be used to restore the discriminative importance of nodes in a specific standard section and avoid semantic mismatch caused by geometric correction. In an exemplary embodiment, the feature weight compensation matrix can be constructed based on feature weight deviations and applied to the original weights in an additive or multiplicative manner. The feature weight compensation matrix and the spatial location mapping matrix are used in parallel to ensure that the structure and semantics are synchronously aligned.

[0078] Based on the topological position deviation and feature weight deviation, a spatial position mapping matrix and a feature weight compensation matrix for the topological nodes are established. These can be achieved by constructing a transformation matrix and a weight adjustment matrix for correction, with the deviation as the optimization objective. Furthermore, this operation can be implemented by using thin-plate spline interpolation to obtain a dense transformation field for spatial mapping and by using local Gaussian kernel smoothing for weight compensation, or by uniformly learning both matrices end-to-end through a graph neural network. This forms a dual-channel correction mechanism that simultaneously repairs geometric and semantic distortions.

[0079] For each node of the input topological feature, position correction is performed using a spatial location mapping matrix, and feature weight compensation is performed using a feature weight compensation matrix to obtain a corrected topological feature map.

[0080] Each node in the input topological feature can be a basic unit in the vascular topology map to be corrected, possessing coordinates, connectivity, and initial weight attributes, and can be used as the object of the dual-matrix correction operation. For example, each node in the input topological feature can include high-weight key nodes, low-weight auxiliary nodes, boundary transition nodes, etc. For each node in the input topological feature, position correction is performed using a spatial position mapping matrix. This can be achieved by multiplying or substituting the original node coordinates into the spatial position mapping matrix, outputting the corrected coordinates. Furthermore, this operation can be implemented by applying direct matrix multiplication to the node coordinates or by interpolating and querying the dense displacement field to obtain a new position, thereby achieving spatial alignment of anatomical structures.

[0081] For each node in the input topological features, feature weight compensation is performed using a feature weight compensation matrix. This can be achieved by operating on the original node weights and the corresponding elements of the compensation matrix to update its region discrimination weights. In a specific embodiment, this operation can be implemented using additive compensation (original weights + bias correction value) or multiplicative compensation (original weights × confidence gain factor), thereby restoring the semantic importance of the node in the target section. The corrected topological feature map is obtained by integrating all nodes and their connectivity relationships after double-matrix correction to form the final topological representation. Furthermore, this operation can be implemented by storing the corrected node coordinates and updated weights in a graph structure and generating a weighted topological mask map aligned with the original graph, thereby outputting structured features that simultaneously possess geometric consistency and semantic sensitivity.

[0082] For example, in a scenario of real-time ultrasound quality control under dynamic breathing conditions, the liver ultrasound standard section image recognition method in this embodiment can be as follows: The patient undergoes a liver ultrasound examination while breathing shallowly and rapidly. The images acquired by the system show that the main portal vein is significantly deflected clockwise. The system first calculates a local rotation of approximately 12 degrees in the left medial lobe region based on the transformation path and offset obtained in the previous steps, combined with the real-time segmented liver contour. Then, a topological transformation equation containing this rotation term is established, and the topological prior parameters of the right subcostal longitudinal section are substituted. It is found that the main portal vein node has a significant positional deviation and its regional weight is underestimated. Based on this, a spatial position mapping matrix is ​​constructed to correct the node's inverse rotation, and its weight is increased from 0.4 to 0.75 through a feature weight compensation matrix. The final output corrected topological feature map not only restores the parallel orientation of the main vein but also correctly reflects its high discriminative value in the longitudinal section, enabling the subsequent classification module to accurately identify it as a qualified standard section.

[0083] In one embodiment, based on the corrected topological feature map and the region feature weight distribution map, a region adaptive topological matching algorithm is used to perform standard cross-section classification and recognition on the input image, obtaining a recognition result containing cross-section categories, including: Within each anatomical region, the regional topological feature matrix is ​​extracted from the corrected topological feature map, and a feature description vector reflecting the regional topological distribution is constructed based on the regional topological feature matrix. The regional topological feature matrix can be a structured vascular topological representation matrix extracted from the corrected topological feature map within a single anatomical region. It can be used to preserve the spatial distribution details of blood vessels within a local anatomical region, serving as the basis for constructing a regional semantic description. In an exemplary embodiment, the regional topological feature matrix can crop and encode the corrected node coordinates, connectivity relationships, and orientation information into a matrix form according to the region. The feature description vector can be a low-dimensional vector generated from the regional topological feature matrix, used to characterize the spatial distribution pattern of vascular topology within the region. For example, the feature description vector can be compressed from the regional topological feature matrix into a fixed-length vector using graph embedding, principal component analysis, or pooling operations. In a specific embodiment, the feature description vector may include, but is not limited to, graph neural network embedding vectors, statistical moment feature vectors, and topological invariant combination vectors.

[0084] Extracting the regional topological feature matrix from the corrected topological feature map within each anatomical region can be achieved by cropping and structuring the corresponding subgraph from the corrected topological feature map based on the pre-defined anatomical region boundaries. Furthermore, extracting the regional topological feature matrix from the corrected topological feature map within each anatomical region can be accomplished by constructing a subgraph matrix based on the region mask index nodes and edges, and by using a graph convolutional network to aggregate features within the region's neighborhood. This allows for the spatial local decomposition of topological features while preserving region-specific structural information. Constructing a feature description vector reflecting the regional topological distribution based on the regional topological feature matrix can be achieved by performing dimensionality reduction or embedding transformations on the regional topological feature matrix. Further, constructing a feature description vector reflecting the regional topological distribution based on the regional topological feature matrix can be achieved by applying graph pooling operations to generate vectors, and by combining topological statistics such as the number of branches, total length, and average curvature into vectors. This transforms the local topological structure into a computable numerical vector.

[0085] Calculate the adaptive weight coefficient for each region based on the regional feature weight distribution map. The adaptive weight coefficient is proportional to the variance of the feature weight values ​​within the region. The adaptive weight coefficient can be a scalar parameter reflecting the dispersion of feature weight values ​​within a single anatomical region. It can be used to quantify the consistency of discriminative power within a region; a larger variance indicates that the region is more sensitive to section recognition and should be assigned a higher fusion weight. In a specific embodiment, the adaptive weight coefficient can be calculated based on the variance of pixel weight values ​​within the corresponding region in the regional feature weight distribution map, and used as the coefficient. Furthermore, the adaptive weight coefficient can be used to construct a regional adaptive fusion function, directly affecting the dynamic adjustment of the fusion parameters. The variance of feature weight values ​​within a region can be the statistical variance of weight values ​​at all locations within a certain anatomical region in the regional feature weight distribution map. It can be used as a direct basis for the adaptive weight coefficient, characterizing the degree of heterogeneity in the response within the region. For example, the variance of feature weight values ​​within a region can be calculated by performing standard deviation calculations on the weight values ​​within the region. In an exemplary embodiment, the variance of feature weight values ​​within a region can include, but is not limited to, unbiased sample variance, weighted local variance, and sliding window variance.

[0086] The adaptive weight coefficient for each region can be calculated based on the regional feature weight distribution map. This can be achieved by calculating the variance of the weight values ​​within each anatomical region and using this variance as the coefficient for the corresponding region. Furthermore, the adaptive weight coefficient for each region can be calculated using the unbiased sample variance formula and local variance estimation via a sliding window, thereby establishing a quantitative indicator of discriminative consistency within the region.

[0087] A region-adaptive fusion function is constructed using weight adaptive coefficients, and the fusion parameters are dynamically adjusted according to the uniformity of the region feature distribution. The region-adaptive fusion function can be a nonlinear fusion mapping function designed based on adaptive weight coefficients. It can be used to achieve differentiated fusion of feature description vectors from different regions, enhancing the contribution of highly discriminative regions. In one specific embodiment, the region-adaptive fusion function can use the adaptive weight coefficients as input parameters to construct an adjustable weighted average or attention fusion mechanism. For example, the region-adaptive fusion function can include, but is not limited to, Softmax weighted fusion functions, gated attention fusion functions, and learnable affine fusion functions. The fusion parameters can be adjustable variables in the region-adaptive fusion function that control the fusion strength or method. They can be used to enable the fusion process to adapt to changes in regional information density, avoiding information dilution caused by uniform weighting. Furthermore, the fusion parameters can be dynamically set according to the uniformity of the regional feature distribution (i.e., the adaptive weight coefficients). In an exemplary embodiment, the fusion parameters can include, but are not limited to, relative scaling factors between regions, nonlinear activation thresholds, and attention temperature coefficients.

[0088] Constructing a region-adaptive fusion function using adaptive weight coefficients can be achieved by designing a fusion weight allocation mechanism with these coefficients as parameters. Furthermore, this function can be implemented by constructing a Softmax normalized weight function and designing a gating mechanism to control the contribution ratio of each region's vector, thus enabling the fusion process to respond to the characteristics of the region's internal feature distribution. Dynamically adjusting the fusion parameters based on the uniformity of the region's feature distribution can be achieved by mapping the adaptive weight coefficients to specific parameter values ​​in the fusion function. Further, this dynamic adjustment can be achieved by setting a temperature coefficient through linear mapping and generating a scaling factor using a nonlinear function, thereby enabling adaptive adjustment of the fusion intensity and providing stronger representation of high-variance regions.

[0089] The region adaptive fusion function is applied to the feature description vectors of all regions to generate global fusion topological features of the input image. The global fusion topological features are then matched and classified with a pre-built standard section feature library to obtain the final liver ultrasound standard section recognition result.

[0090] The global fusion topological feature can be a unified topological feature representation of the entire image generated after processing by a region adaptive fusion function. It can be used as input features for final matching and classification, and includes structural topological information and regional importance differences. In one specific embodiment, the global fusion topological feature can be obtained by weighting and combining the feature description vectors of all regions according to the fusion function. Furthermore, the global fusion topological feature can be directly used for matching and classification with a standard section feature library. The standard section feature library can be a pre-constructed set of global fusion topological feature templates corresponding to various liver ultrasound standard sections, which can be used to provide a matching reference benchmark and support the category determination of the input image. For example, the standard section feature library can be constructed by performing the same feature extraction and fusion process on a large number of labeled standard section images. In an exemplary embodiment, the standard section feature library may include, but is not limited to, a subxiphoid transverse section template library, a right intercostal oblique section template library, and a right subcostal longitudinal section template library. The liver ultrasound standard section recognition result can be the final output of the standard section category to which the input image belongs, obtained through matching and classification. It can be used to complete the judgment loop of the entire recognition task for quality control or auxiliary diagnostic decision-making.

[0091] Applying the region-adaptive fusion function to the feature description vectors of all regions can be achieved by calculating a weighted combination of the feature description vectors of all regions according to the fusion function. Furthermore, applying the region-adaptive fusion function to the feature description vectors of all regions can be achieved by performing weighted summation or weighted aggregation through an attention mechanism, thereby generating a global representation that balances structure and importance. Generating the global fusion topological features of the input image can be achieved by outputting a unified fused feature vector. Furthermore, generating the global fusion topological features of the input image can be achieved by concatenating the weighted vectors of each region, reducing dimensionality, and directly outputting the fusion result as the final feature, thus providing highly discriminative features for final matching. Matching and classifying the global fusion topological features with a pre-built standard cross-section feature library can be achieved by calculating the similarity between the input features and various templates in the library, and selecting the category corresponding to the highest score. Furthermore, matching and classifying the global fusion topological features with the pre-built standard cross-section feature library can be achieved by using cosine similarity matching and template retrieval using a nearest neighbor classifier, thereby achieving automatic discrimination of standard cross-sections. The final liver ultrasound standard section recognition result can be obtained by outputting the section category label with the highest matching score, thus completing the final output of the recognition task.

[0092] For example, in the scenario of a real-time feedback module in an ultrasound teaching and training system, the liver ultrasound standard section image recognition method of this embodiment can be used when medical students collect liver ultrasound images during simulated training. The system first extracts the regional topological feature matrix in each anatomical region of the corrected topological feature map and converts it into a feature description vector. Then, it calculates the weight adaptive coefficient of each region based on the regional feature weight distribution map and finds that the weight variance of the portal vein bifurcation area is significantly higher than that of other regions, indicating that this region is highly sensitive to section discrimination. Based on this, the system constructs a regional adaptive fusion function to dynamically improve the fusion parameters of the region. The global fusion topological feature generated after fusion is compared with the standard section feature library to accurately identify that the current image is close to but not completely aligned with the right intercostal oblique section. The system then prompts that the probe angle needs to be finely adjusted to clearly show that the right branch of the portal vein runs parallel to the right hepatic vein, thereby helping students understand the importance of key anatomical landmarks.

[0093] In addition, refer to Figure 2 To achieve the above objectives, the present invention also provides a liver ultrasound standard section image recognition system, the system comprising: The topology extraction module 10 is used to extract and acquire multi-scale vascular topology features from the input target liver ultrasound image, extract the node coordinates, connectivity and orientation information of intrahepatic blood vessels, and construct a vascular topology feature dataset. The weight mapping module 20 is used to divide the anatomical region inside the liver according to the vascular topology feature dataset, and calculate the contribution of the topology features of each region to the recognition of different standard sections, so as to obtain a region feature weight distribution map. The registration and correction module 30 is used to construct a blood vessel topology registration model based on the region feature weight distribution map, perform position mapping and deformation deviation correction on the topological features of the input image, and obtain a corrected topological feature map. The adaptive recognition module 40 is used to perform standard cross-section classification and recognition on the input image using a regional adaptive topology matching algorithm based on the corrected topology feature map and the regional feature weight distribution map, so as to obtain a recognition result containing cross-section categories.

[0094] Other embodiments or specific implementations of the liver ultrasound standard section image recognition system described in this invention can be referred to the above-described method embodiments, and will not be repeated here.

[0095] In addition, to achieve the above objectives, the present invention also provides a liver ultrasound standard section image recognition device, the device comprising: a memory, a processor, and a liver ultrasound standard section image recognition program stored in the memory and executable on the processor, the liver ultrasound standard section image recognition program being configured to implement the steps of the liver ultrasound standard section image recognition method as described above.

[0096] In addition, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a liver ultrasound standard section image recognition program, which, when executed by a processor, implements the steps of the liver ultrasound standard section image recognition method as described above.

[0097] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for recognizing standard ultrasound cross-sectional images of the liver, characterized in that, The method includes: Multi-scale vascular topology feature extraction was performed on the input target liver ultrasound image to extract the node coordinates, connectivity and orientation information of intrahepatic blood vessels, and a vascular topology feature dataset was constructed. Based on the aforementioned vascular topology feature dataset, the liver interior is divided into anatomical regions, and the contribution of the topology features of each region to the recognition of different standard sections is calculated to obtain a region feature weight distribution map. Based on the region feature weight distribution map, a blood vessel topology registration model is constructed, and the topological features of the input image are mapped to their positions and deformed to correct for distortion, resulting in a corrected topological feature map. Based on the corrected topological feature map and the regional feature weight distribution map, the input image is classified and identified using a regional adaptive topological matching algorithm to obtain a recognition result that includes the cross-section category.

2. The liver ultrasound standard section image recognition method as described in claim 1, characterized in that, The process involves multi-scale vascular topology feature extraction from the input target liver ultrasound image, extracting the node coordinates, connectivity, and orientation information of intrahepatic blood vessels to construct a vascular topology feature dataset, including: The input target liver ultrasound image is preprocessed for denoising and enhancement. A segmentation network is used to extract the binary mask of intrahepatic blood vessels. The center line and bifurcation nodes of all blood vessels are extracted by a skeletonization algorithm. Traverse all bifurcation nodes and record the spatial coordinates, vascular branch direction angle, and branch level information of each bifurcation node; Based on the information of all bifurcation nodes and vessel segments, an adjacency matrix describing the connectivity of vessels is constructed. Combined with the spatial coordinates and angle information of the bifurcation nodes, a feature vector field of the vessel topology is established. The vascular topology feature dataset is constructed based on the relative positional relationship between the feature vector field and the intrahepatic anatomical regions.

3. The liver ultrasound standard section image recognition method as described in claim 1, characterized in that, The process involves dividing the liver interior into anatomical regions based on the vascular topology feature dataset, calculating the contribution of each region's topology features to the recognition of different standard cross-sections, and obtaining a region feature weight distribution map, including: The degree of topological difference between each bifurcation node and its adjacent vessel segments is calculated based on the vascular topology feature dataset, and the liver interior is initially divided into anatomical regions based on the degree of topological difference to obtain the initial region division results. The mean and variance of the topological features of all nodes within each initial region in the initial region partitioning result are calculated based on the vascular topological feature dataset. The initial regions are then adaptively merged and split based on the mean and variance of the topological features to obtain the target anatomical region partitioning result, which includes multiple anatomical regions. The contribution coefficient of the topological features of each region to the recognition of different standard sections is calculated based on the vascular topological feature data of each region, and a corresponding feature weight value is assigned to each region based on the recognition contribution coefficient. The target anatomical region division results and the feature weight values ​​corresponding to each region are integrated into a regional feature weight distribution map of the liver.

4. The liver ultrasound standard section image recognition method as described in claim 3, characterized in that, The step of calculating the contribution coefficient of the topological features of each region to the recognition of different standard cross-sections based on the vascular topological feature data within each region, and assigning a corresponding feature weight value to each region based on the recognition contribution coefficient, includes: All vascular topological nodes in each region are sorted according to spatial coordinates. The topological feature distribution surface function in the region is obtained by fitting the least squares method. The region topological feature data is obtained by calculating the gradient and curvature of the topological feature distribution surface function. The gradient of the topological feature distribution surface function is used to calculate the normal vector of the blood vessel orientation at each point inside the liver. Based on the normal vector and the regional topological feature data, the matching angle between the standard anatomical orientation corresponding to each standard section and the current regional normal vector is calculated. By combining the matching angle with the gradient and curvature of the topological feature distribution surface, the initial recognition contribution coefficient is obtained; The initial identification contribution coefficient is weighted and averaged using the spatial distribution density of vascular nodes within the region, and the weighted average result is normalized to obtain the corrected identification contribution coefficient for each region in each standard section. When the correction recognition contribution coefficient of a region corresponding to a certain standard section is greater than the preset weight threshold, the corresponding feature weight value is set to the logarithm of the correction recognition contribution coefficient. When it is less than the weight threshold, the corresponding feature weight value is set to zero, thus obtaining the feature weight value of each region corresponding to each standard section.

5. The liver ultrasound standard section image recognition method as described in claim 1, characterized in that, The step involves constructing a vascular topology registration model based on the region feature weight distribution map, performing position mapping and deformation deviation correction on the topological features of the input image, and obtaining a corrected topological feature map, including: Based on the pre-stored vascular topology template parameters of various standard sections and liver anatomy priors, a set of basic transformation equations for topological registration is established for each anatomical region. The feature weight values ​​in the region feature weight distribution map are combined with the basic transformation equations to construct a vascular topology registration model that includes weight factors. Based on the vascular topology registration model, calculate the transformation path and offset of each vascular topology node in the input image to the standard topology template; Based on the transformation path and offset, the spatial position mapping relationship of the topological nodes is calculated, and the topological features of the input image are corrected using the spatial position mapping relationship to obtain the corrected topological feature map.

6. The liver ultrasound standard section image recognition method as described in claim 5, characterized in that, The step of calculating the spatial location mapping relationship of topological nodes based on the transformation path and offset, and using the spatial location mapping relationship to correct the topological features of the input image to obtain a corrected topological feature map includes: Using the transformation path and offset, combined with the overall anatomical contour data of the liver, the local deformation rotation angle of the blood vessel topology of the input image relative to the standard template is calculated, and a topological transformation equation considering respiratory motion deformation is established. Substitute the topological transformation equations into the topological prior parameters of different cross-sections to calculate the topological position deviation and feature weight deviation of each node. Based on the topological position deviation and feature weight deviation, establish the spatial position mapping matrix and feature weight compensation matrix of the topological nodes; For each node of the input topological feature, the spatial location mapping matrix is ​​used for position correction, and the feature weight compensation matrix is ​​used for feature weight compensation to obtain the corrected topological feature map.

7. The liver ultrasound standard section image recognition method as described in claim 1, characterized in that, The step involves using a region adaptive topology matching algorithm to perform standard cross-section classification and recognition on the input image based on the corrected topology feature map and the region feature weight distribution map, to obtain recognition results containing cross-section categories, including: Within each anatomical region, the regional topological feature matrix of the corrected topological feature map is extracted, and a feature description vector reflecting the regional topological distribution is constructed based on the regional topological feature matrix. Calculate the adaptive weight coefficient for each region based on the regional feature weight distribution map. The adaptive weight coefficient is proportional to the variance of the feature weight values ​​within the region. A region adaptive fusion function is constructed using the aforementioned weight adaptive coefficients, and the fusion parameters are dynamically adjusted according to the uniformity of the region feature distribution. The region adaptive fusion function is applied to the feature description vectors of all regions to generate global fusion topological features of the input image. The global fusion topological features are then matched and classified with a pre-constructed standard section feature library to obtain the final liver ultrasound standard section recognition result.

8. A liver ultrasound standard section image recognition system, characterized in that, The system includes: The topology extraction module is used to extract and acquire multi-scale vascular topology features from the input target liver ultrasound image, extracting the node coordinates, connectivity and orientation information of intrahepatic blood vessels, and constructing a vascular topology feature dataset. The weight mapping module is used to divide the anatomical regions inside the liver according to the vascular topology feature dataset, and calculate the contribution of the topology features of each region to the recognition of different standard sections, so as to obtain a region feature weight distribution map. The registration and correction module is used to construct a blood vessel topology registration model based on the region feature weight distribution map, perform position mapping and deformation deviation correction on the topological features of the input image, and obtain a corrected topological feature map. An adaptive recognition module is used to perform standard cross-section classification and recognition on the input image using a region adaptive topology matching algorithm based on the corrected topology feature map and the region feature weight distribution map, so as to obtain a recognition result containing cross-section categories.

9. A liver ultrasound standard section image recognition device, characterized in that, The device includes: a memory, a processor, and a liver ultrasound standard section image recognition program stored in the memory and executable on the processor, the liver ultrasound standard section image recognition program being configured to implement the steps of the liver ultrasound standard section image recognition method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a liver ultrasound standard section image recognition program, which, when executed by a processor, implements the steps of the liver ultrasound standard section image recognition method as described in any one of claims 1 to 7.

Images