Liver tumor benign and malignant identification system based on ultrasonic shear wave viscoelasticity imaging

Abstract

The present application relates to the technical field of image enhancement display, in particular to a liver tumor benign and malignant differential system based on ultrasonic shear wave viscoelasticity imaging. The present application firstly locks a reference seed region in a target tissue region; further extracts a feature enhancement weight between adjacent pixel nodes and constructs a feature transfer matrix; further takes the reference seed region as an initial energy source, executes random walk iteration with a source restart mechanism based on the feature transfer matrix, and obtains a steady-state confidence distribution matrix; further extracts a potential infiltration region in the steady-state confidence distribution matrix, and generates a benign growth envelope; finally, visual marking is performed in the image, which reduces the interference of high background hardness and image noise on the identification of the infiltration region, and provides a visual basis for clinical benign and malignant differentiation.

Description

Technical Field

[0001] This invention relates to the field of image enhancement and display technology, specifically to a system for differentiating benign and malignant liver tumors based on ultrasound shear wave viscoelastic imaging. Background Technology

[0002] Hepatocellular carcinoma is a common malignant liver tumor, and its early differential diagnosis is crucial for improving patient prognosis. Currently, ultrasound shear wave elastography (SWE) is widely used to differentiate between benign and malignant liver tumors because it can non-invasively measure tissue stiffness (Young's modulus). Clinically, doctors typically diagnose based on the characteristic that the tumor's stiffness value is higher than that of normal liver tissue.

[0003] In clinical practice, existing ultrasound shear wave elastography (SWE) techniques for differentiating benign from malignant liver tumors primarily rely on absolute stiffness thresholds or local image gradients for analysis. However, these methods exhibit significantly reduced performance in complex scenarios involving high background stiffness interference from concomitant liver cirrhosis and inherent speckle noise in ultrasound images. Specifically, the general increase in background liver parenchyma stiffness leads to overlap in stiffness ranges between benign and malignant nodules, rendering absolute thresholds ineffective. Simultaneously, local noise can mask the continuous characteristics of subtle infiltration of malignant tumors into surrounding tissues, causing gradient-based edge detection algorithms to produce misjudgments. Summary of the Invention

[0004] To address the technical problem of accurately visually marking potential infiltrating areas under the influence of high background hardness and image noise, the present invention aims to provide a liver tumor benign / malignant differentiation system based on ultrasound shear wave viscoelastic imaging. The specific technical solution adopted is as follows:

[0005] A system for differentiating benign and malignant liver tumors based on ultrasound shear wave viscoelastic imaging, the system comprising:

[0006] A preparation module is used to obtain the hardness numerical matrix and viscosity numerical matrix within the target tissue region; the hardness numerical matrix is ​​also preprocessed with a confidence mask to extract the effective computational domain; based on the local relative extreme value distribution of the hardness numerical matrix within the effective computational domain, a reference seed region is locked within the target tissue region;

[0007] The analysis module is used to extract feature enhancement weights and construct a feature transfer matrix based on the difference in hardness values ​​between adjacent pixel nodes and the deviation of the mean viscosity values ​​of adjacent pixel nodes from a preset viscosity reference value; using the reference seed region as the initial energy source, it performs a random walk iteration with a source restart mechanism based on the feature transfer matrix to obtain a steady-state confidence distribution matrix;

[0008] The labeling module is used to extract potential infiltration regions based on the numerical distribution characteristics of the steady-state confidence distribution matrix; generate a benign growth envelope based on the contour expansion of the reference seed region at a preset anatomically safe physical distance; and visually label the image by comparing the distribution range of the potential infiltration regions with the benign growth envelope.

[0009] Furthermore, before obtaining the hardness and viscosity numerical matrices within the target tissue region, the original tissue hardness and viscosity numerical matrices at the same acquisition time are resampled and aligned using the row and column dimensions of the anatomical grayscale matrix as a unified pixel coordinate system to obtain the hardness and viscosity numerical matrices.

[0010] Furthermore, the method for obtaining the reference seed region includes:

[0011] Based on the hardness values ​​of all pixels in the effective computation domain, a candidate set is constructed by extracting pixels whose hardness values ​​are within a preset proportion; the largest connected region constructed by the pixels in the candidate set is selected as the reference seed region.

[0012] Furthermore, the method for obtaining the feature enhancement weights includes:

[0013] The region located within the effective computational domain but excluding the reference seed region is designated as the background region.

[0014] Based on the distribution of viscosity values ​​of pixels in the background area, a preset viscosity reference value is obtained, and the standard deviation of hardness values ​​and standard deviation of viscosity values ​​in the background area are extracted.

[0015] For any pair of adjacent pixel nodes within the target tissue region, based on the standard deviation of the hardness values, a hardness difference hindrance term is obtained according to the absolute value of the difference between the hardness values ​​of adjacent pixel nodes; based on the standard deviation of the viscosity values, a viscosity anomaly gain term is obtained according to the absolute value of the difference between the mean viscosity values ​​of adjacent pixel nodes and a preset viscosity reference value; the hardness difference hindrance term and the viscosity anomaly gain term are fused to obtain the feature enhancement weights corresponding to a pair of adjacent pixel nodes.

[0016] Furthermore, the method for obtaining the feature transition matrix includes:

[0017] The feature enhancement weights of each pixel and all its neighboring pixel nodes are summed and normalized to obtain the transition connection strength between neighboring pixel nodes, and a feature transition matrix is ​​constructed.

[0018] Furthermore, the method for obtaining the steady-state confidence distribution matrix includes:

[0019] The reference seed region is mapped to an energy-normalized seed indicator vector, and the transient confidence distribution vector is initialized.

[0020] The diffusion component is obtained by fusing the feature transition matrix with the transient confidence distribution vector; the source maintenance component is obtained according to the seed indicator vector and the preset restart probability; the diffusion component and the source maintenance component are added to update the transient confidence distribution vector until the vector difference between two iterations is less than the preset convergence threshold, and the converged distribution vector is reconstructed into the steady-state confidence distribution matrix.

[0021] Furthermore, the method for obtaining the potential infiltration area includes:

[0022] Obtain the probability distribution histogram of non-zero values ​​within the steady-state confidence distribution matrix, and extract pixel nodes within a preset distribution interval as connected components, serving as potential infiltration regions.

[0023] Furthermore, the method for obtaining the benign growth envelope includes:

[0024] The spatial scale factor is obtained by parsing the underlying ultrasound image metadata; the preset anatomically safe physical distance is converted into pixel expansion steps using the spatial scale factor; and a morphological uniform expansion operation is performed on the boundary of the reference seed region based on the pixel expansion steps to generate the closed benign growth envelope.

[0025] Furthermore, the method for visual labeling in an image includes:

[0026] Extract the set of nodes that belong to the potential infiltration region and whose spatial coordinates are located outside the benign growth envelope to generate the deviation infiltration region; map the deviation infiltration region to the original anatomical display interface, and use a preset warning color thermal layer to overlay and highlight the deviation infiltration region.

[0027] Furthermore, the original tissue hardness matrix and the original tissue viscosity matrix are resampled and aligned using a bicubic interpolation algorithm to generate the hardness matrix and the viscosity matrix.

[0028] The present invention has the following beneficial effects:

[0029] This invention first locks a reference seed region within the target tissue area as a computational source, providing a high-confidence, robust energy starting point for subsequent feature propagation. It then extracts feature enhancement weights between adjacent pixel nodes and constructs a feature transfer matrix, achieving robust extraction of continuous infiltration paths under low signal-to-noise ratio conditions. Next, using the reference seed region as the initial energy source, it performs a random walk iteration with a source restart mechanism based on the feature transfer matrix to obtain a steady-state confidence distribution matrix, directly representing the biomechanical connectivity strength between the corresponding pixel and the reference seed region. Furthermore, it extracts potential infiltration regions from the steady-state confidence distribution matrix, generating a benign growth envelope that represents the theoretical boundary range when the tumor is assumed to be benign. Finally, it performs visual labeling in the image, transforming abstract infiltration features into visually observable spatial distribution differences. This invention, by locking a seed, fusing hardness differences and viscosity anomalies to construct a transfer matrix, obtaining a steady-state confidence distribution through random walk iterations, and finally performing visual labeling, reduces the interference of high background hardness and image noise on infiltration region identification, providing a visual basis for clinical benign-malignant differentiation. Attached Figure Description

[0030] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a system block diagram of a liver tumor benign and malignant differentiation system based on ultrasound shear wave viscoelastic imaging, provided as an embodiment of the present invention. Detailed Implementation

[0032] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a liver tumor benign / malignant differentiation system based on ultrasound shear wave viscoelastic imaging proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0033] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0034] The following description, in conjunction with the accompanying drawings, details a specific scheme for a liver tumor benign / malignant differentiation system based on ultrasound shear wave viscoelastic imaging provided by the present invention.

[0035] Please see Figure 1 The diagram illustrates a system block diagram of a liver tumor benign and malignant differentiation system based on ultrasound shear wave viscoelastic imaging according to an embodiment of the present invention. The system includes: a preparation module 101, an analysis module 102, and a labeling module 103.

[0036] The preparation module 101 is used to obtain the hardness numerical matrix and viscosity numerical matrix within the target tissue region; the hardness numerical matrix is ​​also preprocessed with a confidence mask to extract the effective computational domain; based on the local relative extreme value distribution of the hardness numerical matrix within the effective computational domain, a reference seed region is locked within the target tissue region.

[0037] In one embodiment of the present invention, the system responds to the operator's instructions, locks onto the target tissue region, and extracts multiple sets of raw data matrices from the same acquisition time to form an ultrasound multi-parameter image set, specifically including:

[0038] Dissecting the grayscale matrix Record the differences in acoustic impedance of tissues to reveal their morphological structure;

[0039] Original tissue hardness numerical matrix: Records the Young's modulus readings of the tissue (unit: kPa), which characterizes the tissue's ability to resist deformation. The larger the value, the harder the tissue.

[0040] Original tissue viscosity numerical matrix: Records the shear viscosity reading of the shear wave (unit: Pa·s), characterizing the rheological properties of the tissue. The larger the value, the higher the tissue viscosity. If the measured value is the dispersion slope, then the corresponding shear wave dispersion slope reading is recorded (unit: (m / s) / kHz).

[0041] The original signal confidence value matrix records the quality score of the shear wave signal (usually normalized to between 0 and 1; this normalized confidence value is directly output from the original data interface of the ultrasound front end and can be directly called by the system without secondary extreme value calibration), characterizing the reliability of the measurement data. The larger the value, the more reliable the measurement result.

[0042] In response to the operator's command to select or outline on the ultrasound image display interface, the system defines the image area specified by the operator as the target tissue area.

[0043] Because ultrasound B-mode imaging (used for anatomical structure observation) and shear wave elastography (used for tissue stiffness and viscosity measurement) differ in their imaging principles, their data acquisition frame rates and spatial pixel densities are usually inconsistent. To ensure that each matrix coordinate point in the subsequent algorithm... To accurately correspond to the tissue characteristics of the same anatomical location within the patient's body, a unified coordinate mapping relationship must be established.

[0044] In a preferred embodiment of the present invention, considering that the anatomical grayscale matrix has the highest spatial resolution and the clearest anatomical structure boundaries, and can provide a precise spatial positioning reference for multimodal data fusion, while ensuring that the final generated visual markers can be accurately superimposed on the corresponding positions of the grayscale image observed by the doctor, the row and column dimensions of the anatomical grayscale matrix are used as a unified pixel coordinate system. The original signal confidence value matrix, the original tissue stiffness value matrix, and the original tissue viscosity value matrix at the same acquisition time are resampled and aligned to obtain the corresponding signal confidence value matrix. Hardness numerical matrix With viscous numerical matrix .

[0045] The original signal confidence matrix, the original tissue hardness matrix, and the original tissue viscosity matrix were resampled and aligned using the bicubic interpolation algorithm to generate the signal confidence matrix, the hardness matrix, and the viscosity matrix.

[0046] Among them, bicubic interpolation, compared with simple linear interpolation, can better maintain the smoothness and continuity of numerical values ​​in spatial distribution, reduce jagged effects, and all matrices in the image set have completely consistent row and column dimensions, and the correspondence is fixed.

[0047] Furthermore, due to rib obstruction, gas interference, or patient respiratory movements, ultrasound shear waves may not propagate effectively in some areas, leading to random jumps or gaps (NaN) in the measured values. To prevent these invalid noise data from interfering with subsequent statistical analysis and path calculation, the system utilizes a signal confidence matrix. Pre-screening for data validity, specifically:

[0048] System traversal For each coordinate in the dataset, read its confidence score. Call the system's preset validity threshold In this example, ,like Greater than or equal to If so, the corresponding position in the confidence mask is marked as 1, indicating that the data at that point is available. Less than If the confidence level is 0, then the corresponding position in the confidence mask is marked as 0, indicating that the data is unreliable; all mask regions marked as 1 are taken as the effective computation domain.

[0049] The confidence mask is simply a conventional binary masking process, with a validity threshold. 0.5 represents the median critical point for the signal-to-noise ratio. A value below 0.5 indicates that the minute displacement signal of the shear wave has been masked by probe compression, patient breathing displacement, or deep acoustic shadowing noise. Forced adoption would lead to physically meaningless extreme jumps in hardness and viscosity features. Using 0.5 as the threshold allows for the removal of severely artifact-laden data while retaining a sufficient number of effective pixel nodes. This prevents excessive mask shrinkage that could cause subsequent feature transfer networks to break into isolated segments, ensuring the connectivity of feature confidence along the spatial propagation path.

[0050] Thus, the numerical matrices of hardness and viscosity within the target tissue region were obtained. The hardness numerical matrix was preprocessed using a confidence mask to extract the effective computational domain, providing a basis for subsequent analysis.

[0051] It should be noted that the specific implementation of the system's response to the operator's instructions to extract the image region and the bicubic interpolation algorithm itself are already known prior art in this field, and will not be elaborated upon here.

[0052] In other embodiments of the present invention, when capturing the target tissue region, the area selected by the operator can be expanded by a predetermined number of pixels (e.g., 20 pixels outward) and the expanded peripheral area can be used together with the area originally selected by the operator as the target tissue region. This expansion operation can introduce sufficient background reference information for subsequent neighborhood relationship establishment and confidence propagation calculations, avoiding insufficient background samples or boundary truncation effects caused by the selection boundary being too close to the lesion edge. This ensures the statistical stability of the background viscosity reference value in the feature transfer matrix and the completeness of the hardness difference hindrance term calculation. However, it increases peak memory usage and prolongs algorithm convergence time, placing higher demands on the computing power of the medical terminal hardware. The implementer can choose this option independently.

[0053] Considering that a higher hardness value indicates a stronger ability of tissue to resist deformation, and that the density of the core region of a solid tumor is usually higher than that of the surrounding normal tissue and sclerotic nodules, a reference seed region is locked in the target tissue region based on the local relative extreme value distribution of the hardness value matrix in the effective computational domain. This serves as the computational source and provides a high-confidence, robust energy starting point for subsequent feature propagation.

[0054] Preferably, in one embodiment of the present invention, since the relative hardness ranking of the tumor core is still among the top in the region even when the overall background hardness increases due to cirrhosis, in order to adaptively locate the most reliable computational source without relying on the absolute hardness threshold, a candidate set is constructed based on the hardness values ​​of all pixels in the effective computational domain, pixels with hardness values ​​in the top preset proportion are extracted; the largest connected region constructed by the pixels in the candidate set is selected as the reference seed region.

[0055] As an example, quantile statistics are used to represent the local relative extreme value distribution characteristics within the hardness numerical matrix. In this example, the preset ratio is set to 10%, which corresponds to the interval between the 90th and 100th percentiles, selecting the top 10% of the numerical points. This ratio is determined based on the anatomical fact that the core area of ​​solid tumors usually accounts for 10%-20% of the total lesion volume in clinical experience. This ensures that the seed area has sufficient spatial continuity while avoiding interference from surrounding background tissues due to an excessively large ratio.

[0056] To eliminate discrete noise (such as reflections from blood vessel walls or calcifications), the system performs a morphological opening operation on the extracted pixel set. This operation first performs erosion to remove isolated, small noise points, and then performs dilation to restore the shape of the main region.

[0057] Next, the system uses the Connected Component Labeling algorithm to identify all independent connected regions and calculates the number of pixels (area) of each connected region. The connected region with the largest area is selected as the reference seed region.

[0058] It should be noted that if a connected region has extremely high hardness but its area is smaller than a preset small threshold (such as 5 pixels), the system will determine it to be a calcification point or artifact and remove it.

[0059] If the reference seed region cannot be extracted effectively If all regions are eliminated or the candidate set is abnormally empty, a fallback correction mechanism is triggered: the system prompts the operator to click on the center of the solid part of the tumor on the image interface. In response to the user's click, the system generates a circular region with a preset radius (e.g., 3mm) centered on the click point and forcibly designates it as the seed region for calculation. .

[0060] It should be noted that percentile method, morphological opening operation, and connected component labeling algorithm are all known basic image processing technologies in this field, and will not be elaborated upon in this invention.

[0061] Analysis module 102 is used to extract feature enhancement weights and construct a feature transfer matrix based on the difference in hardness values ​​between adjacent pixel nodes and the deviation of the mean viscosity values ​​of adjacent pixel nodes from the preset viscosity reference value; using the reference seed region as the initial energy source, it performs random walk iterations with a source restart mechanism based on the feature transfer matrix to obtain a steady-state confidence distribution matrix.

[0062] Considering that malignant tumor invasion often manifests as continuous spread to surrounding tissues, and that ultrasound spot noise can disrupt local edge gradients, making it difficult for traditional boundary detection algorithms to identify the true invasion channels, and that hardness differences may not be obvious in the context of cirrhosis, while abnormal viscosity can reflect the activity of the malignant microenvironment, we use the hardness difference between adjacent pixel nodes to characterize the degree of obstruction of feature propagation by anatomical structures (such as capsules), and combine the deviation of the mean viscosity value of adjacent pixel nodes from the preset viscosity reference value to characterize the enhancement effect of pathological active areas on feature propagation. We fuse these two values ​​to extract feature enhancement weights and construct a feature transfer matrix accordingly, thereby achieving robust extraction of continuous invasion paths under low signal-to-noise ratio conditions.

[0063] Using the reference seed region as the initial energy source, a random walk iteration with a source restart mechanism is executed on the propagation rules defined by the feature transfer matrix. This allows the feature confidence to preferentially spread along continuous paths with weak hardness barriers or abnormally high viscosity. At the same time, each iteration forces a re-injection of energy from the seed region to suppress background noise that is not connected to the seed. Finally, a steady-state confidence distribution matrix is ​​obtained, which directly represents the connectivity strength between the corresponding pixel and the reference seed region in terms of biomechanical features.

[0064] Preferably, in one embodiment of the present invention, in order to analyze the connectivity of an image using graph theory algorithms, the system needs to map image pixels to network nodes, that is, define each pixel as a network node (pixel node).

[0065] Methods for obtaining feature enhancement weights include:

[0066] First, the region located within the effective computational domain but excluding the reference seed region is designated as the background region. The reference seed region represents the solid core of the tumor, and excluding it from the background computation ensures that the background reference value is not contaminated by the characteristics of the tumor itself.

[0067] Secondly, based on the distribution of viscosity values ​​of pixels in the background area, a preset viscosity reference value is obtained, which represents the baseline value of viscosity level in the background of normal liver tissue. The standard deviation of hardness values ​​and viscosity values ​​in the background area are extracted. The former represents the dispersion of hardness values ​​in the background area and is used to standardize and adjust the hardness difference, while the latter represents the dispersion of viscosity values ​​in the background area and is used to standardize and adjust the viscosity deviation.

[0068] Next, for any pair of adjacent pixel nodes within the target tissue region, the larger the absolute value of the difference in hardness values, the more significant the tissue hardness difference between the two adjacent nodes is from a data perspective, and the higher the possibility of the existence of anatomical boundaries (such as fibrous capsules). At the same time, the hardness difference at different scales is standardized and adjusted by using the standard deviation of hardness values. Therefore, based on the standard deviation of hardness values, the hardness difference hindrance term is obtained according to the absolute value of the difference in hardness values ​​between adjacent pixel nodes.

[0069] As an example, the arithmetic mean of the viscosity values ​​of all pixels in the background area is calculated as the preset viscosity reference value; the absolute value of the difference in hardness values ​​between adjacent pixel nodes is used as the numerator, the standard deviation of the hardness values ​​is used as the denominator, and the ratio of the fractions is used as the hardness difference retardation term; the larger the ratio, the more significant the hardness gradient and the stronger the corresponding retardation effect.

[0070] Similarly, considering that the larger the absolute value of the difference between the local viscosity mean and the background viscosity reference value, the more significant the abnormality of the tissue rheological properties and the higher the activity of the pathological microenvironment in the local area (common in the matrix degradation area of ​​malignant tumors), and the standard deviation of viscosity values ​​is used to standardize and adjust the viscosity deviation at different scales, the viscosity abnormality gain term is obtained based on the standard deviation of viscosity values ​​and the absolute value of the difference between the viscosity mean of adjacent pixel nodes and the preset viscosity reference value.

[0071] As an example, the average viscosity value of adjacent pixel nodes is the average value obtained by adding the viscosity values ​​of the two nodes and dividing by 2.

[0072] The absolute value of the difference between the mean viscosity value and the preset viscosity reference value is used as the numerator, and the standard deviation of the viscosity value is used as the denominator. The ratio of this fraction is used as the viscosity anomaly gain term. The larger the ratio, the more significant the viscosity anomaly, and the stronger the corresponding gain effect.

[0073] By fusing the hardness difference hindrance term and the viscosity anomaly gain term, feature enhancement weights corresponding to a pair of adjacent pixel nodes are obtained.

[0074] As an example, the viscous anomalous gain term is coupled with a preset coupling coefficient. The product of these terms, minus the difference in the hardness differential retardation term, is used as an exponential function. The exponential part of the independent variable The mapping is performed using an exponential function with the natural constant e as the base, and the mapping result serves as the feature enhancement weight for a pair of adjacent pixel nodes. .

[0075] The role of the exponential function is to map the real number range of the independent variable from negative infinity to positive infinity to the positive real number range of 0 to positive infinity. This makes the weight approach 0 when the independent variable is a large negative number (i.e., when the hardness difference hindrance term is significantly greater than the viscosity anomaly gain term) (characterizing a strong hindrance effect), and the weight greater than 1 when the independent variable is a positive number (i.e., when the viscosity anomaly gain term is significantly greater than the hardness difference hindrance term) (characterizing a strong gain effect).

[0076] Preset coupling coefficient Used to adjust the relative importance of the viscosity anomaly gain term in fusion. The larger the value, the more significant the impact of viscosity anomalies on the weights, and the higher the system's sensitivity to viscous signals. In this example... The value of 0.8 is an empirically optimized value derived from the ratio of the contribution of hardness and viscosity to the identification of malignant tumor infiltration in clinical trials. This value can effectively "illuminate" potential malignant pathways with abnormal viscosity signals in scenarios where hardness differences are not significant, while avoiding excessive amplification of viscosity noise.

[0077] It should be noted that when the standard deviation of the hardness value is zero, it means that all pixels in the background area have the same hardness value. Therefore, the hardness difference hindrance term is directly set to zero, meaning that hardness differences in this local area do not provide any hindrance information. Similarly, when the standard deviation of the viscosity value is zero, the viscosity anomaly gain term is directly set to zero.

[0078] It should be noted that establishing a conventional 8-neighborhood connection relationship means that for any pixel node, it establishes a connection edge with its eight surrounding pixel nodes (top, bottom, left, right, top-left, top-right, bottom-left, bottom-right). For pixels located at the image edge, connections are only established with their existing neighbors. This topological structure ensures that features can only be transmitted between spatially consecutive adjacent pixels. The analysis process is consistent for each pair of adjacent pixel nodes, and only one example is described here.

[0079] Furthermore, after calculating the weights of all adjacent node pairs... Subsequently, the system needs to transform it into a probability matrix describing the transmission rules of the entire graph. In order to satisfy probability conservation (i.e., the total probability of flowing out from a node is 1), the feature enhancement weights of each pixel and all its neighboring pixel nodes are summed and normalized to obtain the transfer connection strength between neighboring pixel nodes and construct the feature transfer matrix.

[0080] The summation and normalization process involves: adjusting the feature enhancement weights from source node j to target node i. As a molecule, the feature enhancement weights from source node j to all its neighboring nodes k are... The sum is used as the denominator, and the ratio of the fractions is used as the single-step transition probability from source node j to target node i, representing the transition connection strength. ;

[0081] Feature transition matrix The construction method is as follows: Assuming there are N pixel nodes in the target tissue region, the feature transition matrix is... It is an N x N square matrix. The element in the i-th row and j-th column This represents the single-step transition probability of the feature confidence score from source node j to target node i. For any pair of nodes, if node i is not a neighbor of node j (i.e., the two pixels are not spatially adjacent), then... Set it directly to 0.

[0082] Since each pixel node is only connected to a few neighboring nodes (e.g., in 8-neighborhood connectivity, it is adjacent to at most 8 nodes), the feature transition matrix... The number of non-zero elements in the array is linearly related to the total number of nodes N (approximately 8N), but not... In order to efficiently store the matrix and support subsequent sparse matrix-vector multiplication operations, the system uses the existing Compressed Sparse Row (CSR) or Compressed Sparse Column (CSC) format for storage. For example, when N=65536, the memory usage is reduced from about 16GB (double-precision floating-point N×N matrix) to about 4MB, ensuring that the algorithm can run in real time on conventional medical device hardware.

[0083] Preferably, in one embodiment of the present invention, the feature transition matrix defines the local rules for each propagation step, but performing only single-step propagation cannot extract global connectivity information. An iterative calculation method is needed so that the feature confidence can gradually diffuse from the reference seed region along the path defined by the feature transition matrix, while maintaining its association with the seed region. If only pure diffusion is performed without maintaining the source energy, the confidence will eventually be uniformly flattened across the entire graph, making it impossible to distinguish between connected and disconnected regions. Therefore, it is necessary to forcibly inject a portion of seed energy in each iteration to maintain the anchoring relationship between the propagation process and the seed region. The method for obtaining the steady-state confidence distribution matrix includes:

[0084] The reference seed region is mapped to an energy-normalized seed indicator vector, and the transient confidence distribution vector is initialized.

[0085] Specifically, the system maps the reference seed region to a seed indicator vector of dimension N×1. For each element in the vector, if the corresponding pixel node belongs to the reference seed region, the value of the element is set to the reciprocal of the total number of pixels in the reference seed region, making the sum of all elements in the vector equal to 1, i.e., normalizing the total energy; if the corresponding pixel node does not belong to the reference seed region, the value of the element is set to 0. Simultaneously, the system initializes a transient confidence distribution vector. Initially, this vector is set to be the same as the seed indicator vector.

[0086] Then the iterative calculation process begins. In each iteration, the system performs the following update operations:

[0087] The diffusion component is obtained by fusing the feature transition matrix and the transient confidence distribution vector;

[0088] The source maintenance component is obtained based on the seed indicator vector and the preset restart probability;

[0089] The transient confidence distribution vector is updated by adding the diffusion component to the source sustaining component;

[0090] Convergence determination: When the difference between the vectors of the two consecutive iterations is less than the preset convergence threshold, the converged distribution vector is reconstructed into a steady-state confidence distribution matrix.

[0091] As an example, in the first In the next iteration ( In the calculation of the diffusion component, the characteristic transfer matrix is ​​transformed into a single matrix. Transient confidence distribution vector at the previous iteration time Multiply by the product, and then multiply by 1-c to obtain the diffusion component. This step simulates the process of feature confidence propagating to neighboring nodes according to transition probabilities.

[0092] Calculate the source sustaining component: use the seed indicator vector With preset restart probability Multiply to obtain the source-maintaining component. .

[0093] The preset restart probability is a constant between 0 and 1, controlling the proportion of total energy re-injected into the seed region in each iteration, and determining the scope of feature exploration. In this example, the R&D and debugging personnel comprehensively considered the typical size and infiltration range of liver tumors and adopted an empirical value. .

[0094] Update the transient confidence distribution vector: Add the diffusion component to the source sustaining component to obtain the transient confidence distribution vector at the current time step. .

[0095] calculate and The Euclidean distance represents the vector difference between the two iterations, and a preset convergence threshold is set to... When the value is less than a preset convergence threshold, the iteration stops, and the resulting vector... It is The long vector. To facilitate subsequent image analysis and display, the system assigns it to the length and width dimensions of the target tissue region (...). The matrix is ​​reshaped to generate a two-dimensional steady-state confidence distribution matrix. .

[0096] Specifically, initialize a size of The zero matrix is ​​generated by mapping the values ​​in the converged distribution vector to the original coordinate indices of the N pixel nodes in the unified pixel coordinate system, thus generating a two-dimensional steady-state confidence distribution matrix.

[0097] Each element in the matrix corresponds one-to-one with a pixel node within the target tissue region. A larger value in the matrix indicates a higher risk that the point belongs to the tumor infiltration area; a lower value indicates that the point belongs to irrelevant background. This matrix naturally achieves denoising and feature enhancement, providing a high-quality data foundation for the quantitative analysis in the next module.

[0098] The labeling module 103 is used to extract potential infiltration regions based on the numerical distribution features of the steady-state confidence distribution matrix; generate benign growth envelopes based on the contour expansion of the reference seed region at a preset anatomical safety physical distance; and visually label the image by comparing the distribution range of the potential infiltration regions with the benign growth envelopes.

[0099] Considering the need for clinical differentiation to transform abstract algorithm analysis results into intuitive and objective visual markers to assist doctors in decision-making, and that benign tumors usually have a complete capsule and exhibit expansive growth, while malignant tumors are typically characterized by border infiltration, a potential infiltration region is extracted based on the numerical distribution characteristics of the steady-state confidence distribution matrix. The potential infiltration region is characterized by a spatial range that is highly connected to the core pathology of the tumor. At the same time, a benign growth envelope is generated based on the contour expansion of the reference seed region at a preset anatomical safety physical distance. This envelope represents the theoretical boundary range when the tumor is assumed to be benign.

[0100] Then, by comparing the distribution range of potential infiltration areas with the benign growth envelope, visual markings are made in the images. The system transforms abstract infiltration features into spatial distribution differences that can be observed intuitively, thereby providing doctors with a quantitative auxiliary decision-making basis based on the fusion calculation results of hardness and viscosity, independent of subjective experience.

[0101] Preferably, in one embodiment of the present invention, the values ​​in the steady-state confidence distribution matrix represent the connectivity strength between each pixel node within the target tissue region and the reference seed region along the propagation path defined by the dual-parameter coupling of hardness and viscosity. Pixel nodes with higher connectivity strength are more likely to belong to the tumor infiltration region; pixel nodes with lower connectivity strength are more likely to belong to the irrelevant background or benign tissue.

[0102] Considering that the infiltration channels of malignant tumors are represented by high-value branches extending outward from the reference seed region in the steady-state confidence distribution matrix, while the background tissue is represented by low-value regions, there is a natural numerical difference between the two. Using quantile thresholds can adaptively preserve high-confidence regions without relying on artificially set absolute thresholds.

[0103] Based on this, the probability distribution histogram of non-zero values ​​in the steady-state confidence distribution matrix is ​​obtained, and the pixel nodes in the preset distribution interval are extracted and mapped as connected components as potential infiltration regions.

[0104] As an example, the horizontal axis of the histogram represents the range of confidence scores, and the vertical axis represents the frequency of occurrence of pixel nodes within each range. The preset distribution range corresponds to P95-P100, which means retaining the top 5% of the regions with the highest scores.

[0105] The system extracts all pixel nodes whose confidence values ​​fall within a preset distribution range from the steady-state confidence distribution matrix, forming a pixel set. It then executes a connected component labeling algorithm to identify all independent connected regions and designates these regions as potential infiltration regions. For isolated connected regions with excessively small areas (e.g., tiny regions with fewer than 5 pixels), the system can classify them as residual noise and remove them.

[0106] Preferably, in one embodiment of the present invention, benign tumors typically exhibit expansive growth pathologically, meaning that tumor cells expand into surrounding tissues in a pushing rather than invasive manner, with their boundaries surrounded by a complete fibrous capsule, forming a relatively regular, closed outline. Based on this pathological fact, an ideal benign morphological model can be constructed by uniformly expanding outward from a reference seed region (representing the solid core of the tumor) by a predetermined anatomical safety distance. This model characterizes the theoretical boundary of the pathological influence range when the tumor is assumed to be benign.

[0107] However, the reference seed region is defined in pixel coordinates (in pixels), while the anatomical safety distance is defined in physical length (such as millimeters). There is a difference in dimensions between the two, which needs to be converted using the spatial scale factor in the ultrasound image metadata.

[0108] Therefore, the underlying ultrasound image metadata is analyzed to obtain the spatial scale factor; the spatial scale factor is used to convert the preset anatomically safe physical distance into the number of pixel expansion steps; based on the number of pixel expansion steps, a morphological uniform expansion operation is performed on the boundary of the reference seed region to generate a closed benign growth envelope.

[0109] As an example, the spatial scale factor can be read from the DICOM (Digital Imaging and Communications in Medicine) standard file format of the ultrasound equipment, or calculated based on imaging depth and image height, typically in millimeters per pixel. The preset anatomically safe physical distance is 3 millimeters, a distance based on clinical experience regarding safe resection margins in liver surgery, representing the upper limit of a reasonable range for the expansive growth of benign tumors.

[0110] Pixel expansion steps = Anatomical safe physical distance ÷ Spatial scale factor, and the result is rounded down;

[0111] Next, the system performs a morphological dilation operation, starting from the boundary of the reference seed region. The structuring element of the dilation operation is a circle or a square, and the number of dilation steps is equal to the number of pixel expansion steps. This operation uniformly advances the boundary of the reference seed region outward by the number of pixel expansion steps, generating a closed contour line as a benign growth envelope.

[0112] It should be noted that morphological dilation is a well-known fundamental image processing technique in this field, and will not be elaborated upon further.

[0113] Finally, the set of nodes belonging to the potential infiltration region and whose spatial coordinates are located outside the benign growth envelope is extracted to generate the deviation infiltration region;

[0114] The potential infiltration region characterizes the spatial extent with extremely high connectivity to the tumor core, while the benign growth envelope characterizes the theoretical boundary under the benign assumption. Each pixel node in the potential infiltration region is traversed, and it is determined whether the node's spatial coordinates lie outside the benign growth envelope (i.e., outside the range allowed by the benign theory). All potential infiltration pixel nodes located outside the benign growth envelope are collected to generate the deviation infiltration region. The deviation infiltration region represents the portion of the actual high connectivity region that breaks through the benign theoretical boundary.

[0115] The deviated infiltration area is mapped to the original anatomical display interface, and a preset warning color thermal layer is used to overlay and highlight the deviated infiltration area.

[0116] Finally, the system uses a heatmap layer with a preset warning color (e.g., red) to overlay and highlight the deviated infiltration area. The transparency of the heatmap layer can be set to semi-transparent (e.g., 50% transparency), allowing doctors to see both the spatial location and shape of the deviated infiltration area, as well as the underlying anatomical structure.

[0117] It should be noted that if no infiltrative area is detected, no warning color highlighting area will be generated. In this case, the system can optionally output the text message "No significant infiltrative features detected." During actual clinical scanning, this system acquires multimodal data (including anatomical grayscale matrix, hardness matrix, viscosity matrix, and signal confidence matrix) in real time according to the image acquisition frame rate of the ultrasound equipment. For each data acquired at the same time, the system automatically completes the analysis and annotation results, updates them in real time, and overlays them on the current anatomical image interface. The analysis and annotation frequency of the system is jointly determined by the shear wave imaging acquisition frame rate of the ultrasound equipment and the real-time interaction needs of the clinical operator.

[0118] In summary, to address the technical challenge of accurately visually marking potential infiltrating regions under the influence of high background hardness and image noise, this invention aims to provide a system for differentiating benign and malignant liver tumors based on ultrasound shear wave viscoelastic imaging. This invention first locks a reference seed region within the target tissue area; then extracts feature enhancement weights between adjacent pixel nodes and constructs a feature transfer matrix; further, using the reference seed region as the initial energy source, it performs a random walk iteration with a source restart mechanism based on the feature transfer matrix to obtain a steady-state confidence distribution matrix; next, it extracts potential infiltrating regions from the steady-state confidence distribution matrix, generating benign growth envelopes; finally, it performs visual marking in the image. This invention, by locking the seed region, fusing hardness differences and viscosity anomalies to construct a transfer matrix, obtaining a steady-state confidence distribution through random walk iterations, and finally performing visual marking, reduces the interference of high background hardness and image noise on infiltrative region identification, providing a visual basis for clinical benign and malignant tumor differentiation.

[0119] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0120] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A system for differentiating benign and malignant liver tumors based on ultrasound shear wave viscoelastic imaging, characterized in that, The system includes: A preparation module is used to obtain the hardness numerical matrix and viscosity numerical matrix within the target tissue region; the hardness numerical matrix is ​​also preprocessed with a confidence mask to extract the effective computational domain; based on the local relative extreme value distribution of the hardness numerical matrix within the effective computational domain, a reference seed region is locked within the target tissue region; The analysis module is used to extract feature enhancement weights and construct a feature transfer matrix based on the difference in hardness values ​​between adjacent pixel nodes and the deviation of the mean viscosity values ​​of adjacent pixel nodes from a preset viscosity reference value; using the reference seed region as the initial energy source, it performs a random walk iteration with a source restart mechanism based on the feature transfer matrix to obtain a steady-state confidence distribution matrix; The labeling module is used to extract potential infiltration regions based on the numerical distribution characteristics of the steady-state confidence distribution matrix; generate a benign growth envelope based on the contour expansion of the reference seed region at a preset anatomically safe physical distance; and visually label the image by comparing the distribution range of the potential infiltration regions with the benign growth envelope.

2. The liver tumor benign / malignant differentiation system based on ultrasound shear wave viscoelastic imaging according to claim 1, characterized in that, Before obtaining the hardness and viscosity numerical matrices within the target tissue region, the original tissue hardness and viscosity numerical matrices at the same acquisition time are resampled and aligned using the row and column dimensions of the anatomical grayscale matrix as a unified pixel coordinate system to obtain the hardness and viscosity numerical matrices.

3. The liver tumor benign / malignant differentiation system based on ultrasound shear wave viscoelastic imaging according to claim 1, characterized in that, The method for obtaining the reference seed region includes: Based on the hardness values ​​of all pixels in the effective computation domain, a candidate set is constructed by extracting pixels whose hardness values ​​are within a preset proportion; the largest connected region constructed by the pixels in the candidate set is selected as the reference seed region.

4. The liver tumor benign / malignant differentiation system based on ultrasound shear wave viscoelastic imaging according to claim 1, characterized in that, The method for obtaining the feature enhancement weights includes: The region located within the effective computational domain but excluding the reference seed region is designated as the background region. Based on the distribution of viscosity values ​​of pixels in the background area, a preset viscosity reference value is obtained, and the standard deviation of hardness values ​​and standard deviation of viscosity values ​​in the background area are extracted. For any pair of adjacent pixel nodes within the target tissue region, based on the standard deviation of the hardness values, a hardness difference hindrance term is obtained according to the absolute value of the difference between the hardness values ​​of adjacent pixel nodes; based on the standard deviation of the viscosity values, a viscosity anomaly gain term is obtained according to the absolute value of the difference between the mean viscosity values ​​of adjacent pixel nodes and a preset viscosity reference value; the hardness difference hindrance term and the viscosity anomaly gain term are fused to obtain the feature enhancement weights corresponding to a pair of adjacent pixel nodes.

5. A system for differentiating benign and malignant liver tumors based on ultrasound shear wave viscoelastic imaging according to claim 1, characterized in that, The method for obtaining the feature transition matrix includes: The feature enhancement weights of each pixel and all its neighboring pixel nodes are summed and normalized to obtain the transition connection strength between neighboring pixel nodes, and a feature transition matrix is ​​constructed.

6. The liver tumor benign / malignant differentiation system based on ultrasound shear wave viscoelastic imaging according to claim 1, characterized in that, The method for obtaining the steady-state confidence distribution matrix includes: The reference seed region is mapped to an energy-normalized seed indicator vector, and the transient confidence distribution vector is initialized. The diffusion component is obtained by fusing the feature transition matrix with the transient confidence distribution vector; the source maintenance component is obtained according to the seed indicator vector and the preset restart probability; the diffusion component and the source maintenance component are added to update the transient confidence distribution vector until the vector difference between two iterations is less than the preset convergence threshold, and the converged distribution vector is reconstructed into the steady-state confidence distribution matrix.

7. The liver tumor benign / malignant differentiation system based on ultrasound shear wave viscoelastic imaging according to claim 1, characterized in that, The method for obtaining the potential infiltration area includes: Obtain the probability distribution histogram of non-zero values ​​within the steady-state confidence distribution matrix, and extract pixel nodes within a preset distribution interval as connected components, serving as potential infiltration regions.

8. A system for differentiating benign and malignant liver tumors based on ultrasound shear wave viscoelastic imaging according to claim 1, characterized in that, The method for obtaining the benign growth envelope includes: The spatial scale factor is obtained by parsing the underlying ultrasound image metadata; the preset anatomically safe physical distance is converted into pixel expansion steps using the spatial scale factor; and a morphological uniform expansion operation is performed on the boundary of the reference seed region based on the pixel expansion steps to generate the closed benign growth envelope.

9. A system for differentiating benign and malignant liver tumors based on ultrasound shear wave viscoelastic imaging according to claim 1, characterized in that, The method for visual labeling in an image includes: Extract the set of nodes that belong to the potential infiltration region and whose spatial coordinates are located outside the benign growth envelope to generate the deviation infiltration region; map the deviation infiltration region to the original anatomical display interface, and use a preset warning color thermal layer to overlay and highlight the deviation infiltration region.

10. A system for differentiating benign and malignant liver tumors based on ultrasound shear wave viscoelastic imaging according to claim 2, characterized in that, The original tissue hardness matrix and the original tissue viscosity matrix are resampled and aligned using a bicubic interpolation algorithm to generate the hardness matrix and the viscosity matrix.

Images