Automated analysis method and system for ultrasound shear wave elastography

Abstract

The application relates to an automatic analysis method and system of ultrasonic shear wave elastography, which acquires a standard anatomical section image of a target tissue, performs pre-quality control, performs uniformity evaluation on a preset measurement region, and determines a target region to be measured; the ultrasonic imaging equipment is automatically controlled to switch to a shear wave elastography mode through a software control interface, color-coded elastic images of the target region to be measured are collected, a multi-modal guided cascade optimization segmentation network is used to perform automatic quality evaluation and region segmentation on the collected color-coded elastic images, and an effective elastic signal region is obtained; from the effective elastic signal region, multi-dimensional pixel value distribution features which are the same as those in a training stage of an elastic value prediction model are extracted; the extracted multi-dimensional pixel value distribution features are input into the pre-trained elastic value prediction model, and an elastic value prediction result corresponding to the target tissue is output, so that full-process automation and standardization from image collection to result analysis are realized, and the analysis efficiency is improved.

Description

Technical Field

[0001] This application relates to the field of ultrasonic image processing technology, and in particular to an automated analysis method and system for ultrasonic shear wave elastography. Background Technology

[0002] Shear wave elastography (SWE) quantifies tissue stiffness (Young's modulus) by measuring the propagation speed of shear waves in tissues, and has become an important non-invasive clinical method for assessing tissue elasticity (such as in liver fibrosis). However, the current clinical application of SWE technology heavily relies on manual operation and analysis. Operators need to manually switch imaging modes, locate standard sections, and freeze images, which is cumbersome and influenced by personal experience, resulting in uncertain examination procedures and time consumption.

[0003] Therefore, how to achieve fully automated and efficient shear wave elastic imaging is an urgent problem to be solved. Summary of the Invention

[0004] Therefore, it is necessary to provide an efficient automated analysis method and system for ultrasonic shear wave elastography to address the above problems.

[0005] The first aspect of this application provides an automated analysis method for ultrasonic shear wave elastography, comprising:

[0006] Obtain standard anatomical cross-sectional images of the target tissue and perform pre-quality control; evaluate the uniformity of the preset measurement area to determine the target area to be measured.

[0007] The ultrasonic imaging equipment is automatically switched to shear wave elastography mode via the software control interface to acquire color-coded elastography images of the target area.

[0008] A multimodal guided cascaded optimization segmentation network is used to automatically assess the quality and segment the regions of the acquired color-coded elastic images to obtain effective elastic signal regions.

[0009] From the effective elastic signal region, extract the same multi-dimensional pixel value distribution features as those used in the training phase of the elastic value prediction model;

[0010] The extracted multi-dimensional pixel value distribution features are input into a pre-trained elasticity value prediction model, which outputs the elasticity value prediction results of the target tissue. The elasticity value prediction model adopts an attention-guided cross-scale feature fusion network structure to represent the mapping relationship between pixel features and elasticity values.

[0011] In one embodiment, a standard anatomical section image of the target tissue is acquired and pre-quality controlled. The uniformity of a preset measurement region is evaluated to determine the target region to be tested. This includes: using a composite evaluation strategy that integrates global gray-level statistical distribution and multi-scale frequency domain texture features to pre-quality control the preset measurement region of the standard anatomical section image. When the histogram statistical index and wavelet high-frequency energy of the preset measurement region are both lower than the corresponding thresholds, it is determined to be the target region to be tested. The histogram statistical index includes gray-level entropy and histogram skewness.

[0012] In one embodiment, the composite evaluation strategy that integrates global gray-level statistical distribution and multi-scale frequency domain texture features includes:

[0013] A grayscale image of the preset measurement area is obtained to generate a grayscale histogram. The grayscale entropy and histogram skewness are analyzed based on the grayscale histogram. If the grayscale entropy is less than or equal to a preset threshold and the absolute value of the histogram skewness is less than or equal to a set threshold, the histogram statistical index is qualified, and the preset measurement area is determined to be a region that passes the initial screening.

[0014] Discrete wavelet transform is performed on the regions that pass the initial screening to extract high-frequency sub-band coefficients containing horizontal, vertical, and diagonal details. The average energy of each high-frequency sub-band is calculated based on the high-frequency sub-band coefficients. If the average energy is less than or equal to the texture energy threshold, the wavelet high-frequency energy is qualified, and the regions that pass the initial screening are determined as the target regions to be tested.

[0015] In one embodiment, a multimodal guided cascaded optimization segmentation network is used to automatically assess the quality and segment regions of the acquired color-coded elastic images, including:

[0016] The color elastic image and the hardness confidence map generated based on the color-hardness mapping are used as bimodal inputs. The network adopts a cascaded encoder-decoder structure. The first-level encoder learns edge and texture features from the color-coded elastic image, and the first-level decoder outputs a preliminary segmentation probability map. The second-level network is guided by the preliminary segmentation probability map and the hardness confidence map to perform fine segmentation. The second-level encoder is responsible for fusing spatial confidence information, and the second-level decoder is responsible for finely correcting the boundaries of the first-level result and filling internal holes. The segmentation result output by the second-level decoder is subjected to boundary adjustment by an adaptive iterative optimization module, and finally outputs a binary segmentation mask.

[0017] In one embodiment, after using a multimodal guided cascaded optimization segmentation network to automatically assess the quality and segment the acquired color-coded elastic image, the method further includes: if the percentage of the binary segmentation mask output by the region segmentation occupies the target area to be tested is greater than or equal to a preset quality control threshold, then the elastic image acquisition is qualified and it is determined to be a valid elastic signal area.

[0018] In one embodiment, the elasticity value prediction model includes:

[0019] Parallel branches are used to generate feature maps at local, intermediate, and global scales.

[0020] A scale attention gating module is used for dynamically weighted fusion of the three scale feature maps;

[0021] The main prediction stream and the auxiliary regularization stream are parallel. The main prediction stream outputs the elasticity value, while the auxiliary regularization stream outputs a score to constrain the physical rationality of the prediction results.

[0022] In one embodiment, the method further includes:

[0023] Learning is performed using a historical dataset, which contains multiple sets of associated data pairs. Each set of associated data pairs includes a quality-controlled effective elastic signal region and a reference elastic value obtained by measuring the effective elastic signal region using the gold standard method.

[0024] Multidimensional pixel value distribution features of the effective elastic signal region are extracted. Using the multidimensional pixel value distribution features as input and the corresponding reference elastic value as the target, a machine learning regression algorithm is used for training to construct an elastic value prediction model that can characterize the mapping relationship from pixel features to elastic values.

[0025] In one embodiment, the training process of the elasticity value prediction model uses a loss function that includes an error loss term between the predicted value and the gold standard, and a constraint loss term guided by an auxiliary regularization flow to ensure that the predicted values ​​conform to physical laws.

[0026] In one embodiment, the multi-dimensional pixel value distribution features include color statistical features, texture features, and spatial distribution features; the elasticity value prediction results include the average Young's modulus value, the maximum Young's modulus value, and the minimum Young's modulus value.

[0027] A second aspect of this application provides an automated analysis system for ultrasonic shear wave elastography, comprising:

[0028] The first module is used to acquire standard anatomical cross-sectional images of the target tissue and perform pre-quality control, evaluate the uniformity of the preset measurement area, and determine the target area to be measured.

[0029] The second module is used to automatically control the ultrasound imaging equipment to switch to shear wave elastography mode through the software control interface, and to acquire color-coded elastography images of the target area to be tested.

[0030] The third module is used to automatically assess the quality and segment the acquired color-coded elastic images using a multimodal guided cascaded optimization segmentation network to obtain effective elastic signal regions.

[0031] The fourth module is used to extract the same multi-dimensional pixel value distribution features from the effective elastic signal region as those in the training phase of the elastic value prediction model.

[0032] The fifth module is used to input the extracted multi-dimensional pixel value distribution features into the pre-trained elasticity value prediction model and output the elasticity value prediction result of the corresponding target organization. The elasticity value prediction model adopts an attention-guided cross-scale feature fusion network structure to represent the mapping relationship between pixel features and elasticity values.

[0033] The aforementioned automated analysis method and system for ultrasound shear wave elastography acquires standard anatomical cross-sectional images of the target tissue and performs pre-quality control, evaluates the uniformity of the preset measurement area, and determines the target area to be measured. It automatically controls the ultrasound imaging device to switch to shear wave elastography mode via a software control interface, acquiring color-coded elastic images of the target area. A multimodal guided cascaded optimization segmentation network is used to automatically assess the quality and segment the acquired color-coded elastic images to obtain effective elastic signal regions. From these effective elastic signal regions, multi-dimensional pixel value distribution features identical to those used in the elasticity value prediction model training phase are extracted. These extracted multi-dimensional pixel value distribution features are input into the pre-trained elasticity value prediction model, outputting the corresponding elasticity value prediction results for the target tissue. This system achieves full automation and standardization from image acquisition to result analysis, improving analysis efficiency and accuracy. Attached Figure Description

[0034] Figure 1 This is a flowchart of an automated analysis method for ultrasonic shear wave elastography in one embodiment;

[0035] Figure 2 This is a schematic diagram of the elastic pixel distribution area in one embodiment;

[0036] Figure 3 Example diagram of a dataset used for model training in one embodiment;

[0037] Figure 4 This is a flowchart illustrating the construction and application principles of an elasticity value prediction model in one embodiment.

[0038] Figure 5 Here is a diagram of the feature fusion network architecture in one embodiment;

[0039] Figure 6 This is a pre-quality control flowchart for a standard intercostal anatomical section image of the liver in one embodiment;

[0040] Figure 7 This is a diagram of a multimodal segmentation network architecture in one embodiment;

[0041] Figure 8This is a block diagram of an automated analysis system for ultrasonic shear wave elastography in one embodiment;

[0042] Figure 9 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0044] 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 application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0045] When used herein, the singular forms of “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising / including” or “having,” etc., specify the presence of the stated features, wholes, steps, operations, components, parts, or combinations thereof, but do not preclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof. Meanwhile, the term “and / or” as used in this specification includes any and all combinations of the associated listed items.

[0046] Shear wave imaging technology generates transverse shear waves in the tissue particles of the focused area by emitting continuously focused acoustic radiation force pulses. These shear waves are then acquired by an ultra-high-speed imaging system, and real-time elasticity images and Young's modulus values ​​are obtained using color-coded technology. In the color-coded image, red represents harder tissue and blue represents softer tissue. The Young's modulus value is calculated using the formula E=3ρc. 2 (E is Young's modulus, c represents shear wave propagation velocity, and ρ represents tissue density) It is calculated and the unit is kPa, reflecting tissue stiffness.

[0047] However, the current clinical application of SWE technology heavily relies on manual operation and analysis, which has the following inherent drawbacks:

[0048] First, the measurement process lacks standardization. Operators need to manually switch imaging modes, locate standard sections, and freeze images, which is cumbersome and influenced by personal experience, resulting in uncertain inspection procedures and time consumption.

[0049] Secondly, the results analysis is highly subjective and lacks repeatability. Measurements rely on the operator manually delineating the region of interest (ROI). The choice of ROI location, size, and shape is highly subjective, and results from different operators or even multiple measurements by the same operator can vary significantly. To ensure stable results, operators often tend to choose smaller ROIs, but this sacrifices a large amount of usable flexibility information in the image, resulting in low information utilization.

[0050] Therefore, there is an urgent need for a method that can automate and standardize the entire process from image acquisition to result analysis, and utilize richer image information for intelligent calculation, in order to overcome the limitations of manual operation and improve the objectivity, repeatability and efficiency of SWE measurement.

[0051] In one embodiment, such as Figure 1 As shown, an automated analysis method for ultrasonic shear wave elastography is provided, including:

[0052] Step S110: Obtain standard anatomical section images of the target tissue and perform pre-quality control, evaluate the uniformity of the preset measurement area, and determine the target area to be measured.

[0053] The target tissue can be the liver or other tissues. The ultrasound imaging device is set to anatomical imaging mode to acquire standard anatomical cross-sectional images of the target tissue, which are then pre-quality controlled by the system. Specifically, taking the liver as the target tissue, the system can automatically or assistedly segment the liver capsule contour, thereby pre-setting a fixed-size measurement area (e.g., a 3cm × 3cm rectangular area) below the capsule. The homogeneity of the liver parenchyma within this pre-set measurement area is automatically evaluated to ensure that it avoids major vascular structures and space-occupying lesions, thus identifying a suitable target area.

[0054] Step S120: Automatically control the ultrasound imaging equipment to switch to shear wave elastography mode via the software control interface, and acquire color-coded elastography images of the target area to be tested.

[0055] Correspondingly, the system communicates with the ultrasound imaging equipment via a software control interface, automatically sending commands to the equipment. Upon receiving the command, the ultrasound imaging equipment immediately switches from the current anatomical imaging mode to shear wave elastography mode and automatically triggers image acquisition to obtain a color-coded elastic image of the target area. Figure 2 This is a schematic diagram of the elastic pixel distribution area.

[0056] Step S130: Using a multimodal guided cascaded optimization segmentation network, the acquired color-coded elastic image is automatically evaluated for quality and segmented into regions to obtain the effective elastic signal region.

[0057] After the system controls the ultrasound imaging device to collect the color-coded elastogram of the target area to be measured, it automatically evaluates the quality of the collected color-coded elastogram. It uses a multi-modal guided cascaded optimization segmentation network to identify and segment the corresponding effective elastogram signal area, and calculates the coverage rate of the effective signals in this area. Only when the coverage rate is not lower than the preset threshold (e.g., ≥95%), it is determined that the elastogram collection is qualified this time and enters the next analysis stage; otherwise, it automatically prompts to re-collect.

[0058] Step S140: Extract the same multi-dimensional pixel value distribution features as those in the elastogram value prediction model training stage from the effective elastogram signal area.

[0059] After completing the collection and quality control of the qualified elastogram, it enters the data-driven intelligent analysis stage. This stage includes two closely related processes: model construction and application. In the model construction process, the system uses the historical data set for learning. The historical data set contains multiple groups of associated data pairs. Each group of associated data pairs includes the effective elastogram signal area with qualified quality control, and the reference elastogram value measured in the effective elastogram signal area by the gold standard method. Figure 3 Illustration of the data set for model training.

[0060] The system extracts the multi-dimensional pixel value distribution features of the effective elastogram signal area, and uses these multi-dimensional pixel value distribution features as the input and the corresponding reference elastogram value as the target, and uses the machine learning regression algorithm for training, so as to construct a prediction model that can describe the complex mapping relationship from pixel features to elastogram values. In the training process of the elastogram value prediction model, the loss function used includes the error loss term between the predicted value and the gold standard, and the constraint loss term guided by the auxiliary regularization flow, which is used to ensure that the predicted values conform to the physical laws.

[0061] In the model application process, for the newly collected and quality-controlled effective elastogram signal area, the system extracts the exactly same multi-dimensional pixel value distribution features as those in the model training stage, and can choose to smooth the multi-dimensional pixel value distribution features to enhance stability.

[0062] Step S150: Input the extracted multi-dimensional pixel value distribution features into the pre-trained elastogram value prediction model, and output the elastogram value prediction result of the corresponding target tissue.

[0063] Among them, the elastogram value prediction model adopts an attention-guided cross-scale feature fusion network structure to represent the mapping relationship from pixel features to elastogram values. The system inputs the extracted multi-dimensional pixel value distribution features into the trained elastogram value prediction model, and this elastogram value prediction model directly performs calculations and inferences, and outputs the elastogram value prediction result of the current target tissue.

[0064] The specific structure of the elasticity value prediction model is not unique. In this embodiment, the elasticity value prediction model includes: parallel branches for generating feature maps of local, intermediate and global scales, a scale attention gating module for dynamically weighting and fusing feature maps of the three scales, and parallel main prediction stream and auxiliary regularization stream. The main prediction stream outputs the elasticity value, and the auxiliary regularization stream outputs a score for constraining the physical rationality of the prediction result.

[0065] Furthermore, the specific types and quantities of multi-dimensional pixel value distribution features and elasticity value prediction results are not unique and can be set according to actual needs. In this embodiment, the multi-dimensional pixel value distribution features include at least color statistical features, texture features, and spatial distribution features, and the elasticity value prediction results include at least the average Young's modulus value, the maximum Young's modulus value, and the minimum Young's modulus value, as the final quantitative and objective measurement output.

[0066] The aforementioned automated analysis method for ultrasonic shear wave elastography transforms the traditional "manual operation - equipment calculation" model into an automated "programmed control acquisition - intelligent model analysis" model. This solves the problems of manual dependence, strong subjectivity, and poor repeatability in existing SWE measurement processes. It enables full automation and standardization from image acquisition to result analysis, and utilizes richer image information for intelligent calculation to overcome the limitations of manual operation and improve the objectivity, repeatability, and efficiency of SWE measurement.

[0067] In one embodiment, step S110 includes: using a composite evaluation strategy that integrates global grayscale statistical distribution and multi-scale frequency domain texture features to perform pre-quality control on a preset measurement region of a standard anatomical section image; when the histogram statistical index and wavelet high-frequency energy of the preset measurement region are both lower than the corresponding threshold, it is determined as the target region to be measured.

[0068] Specifically, the histogram statistical indicators include gray-level entropy and histogram skewness. In this embodiment, a composite evaluation strategy integrating global gray-level statistical distribution and multi-scale frequency domain texture features is adopted, including: acquiring a gray-level image of a preset measurement area to generate a gray-level histogram, and analyzing gray-level entropy and histogram skewness based on the gray-level histogram. If the gray-level entropy is less than or equal to a preset threshold, and the absolute value of the histogram skewness is less than or equal to a set threshold, then the histogram statistical indicators are qualified, and the preset measurement area is determined to be a region that passes the initial screening; performing discrete wavelet transform on the region that passes the initial screening to extract high-frequency sub-band coefficients containing horizontal details, vertical details, and diagonal details, calculating the average energy of each high-frequency sub-band based on the high-frequency sub-band coefficients, and if the average energy is less than or equal to the texture energy threshold, then the wavelet high-frequency energy is qualified, and the region that passes the initial screening is determined to be the target region to be tested.

[0069] In one embodiment, step S130 employs a multimodal guided cascaded optimization segmentation network to perform automatic quality assessment and region segmentation on the acquired color-coded elastic image, including:

[0070] The color elastic image and the hardness confidence map generated based on the color-hardness mapping are used as bimodal inputs. The network adopts a cascaded encoder-decoder structure. The first-level encoder learns edge and texture features from the color-coded elastic image, and the first-level decoder outputs a preliminary segmentation probability map. The second-level network uses the preliminary segmentation probability map and the hardness confidence map together to guide fine segmentation. The second-level encoder focuses on fusing spatial confidence information, and the second-level decoder is responsible for finely correcting the boundaries of the first-level result and filling internal holes. The segmentation result output by the second-level decoder is subjected to boundary adjustment by an adaptive iterative optimization module, and finally outputs a binary segmentation mask. Further, step S130 also includes: if the percentage of the binary segmentation mask output by the region segmentation to the target area is greater than or equal to a preset quality control threshold, the elastic image acquisition is qualified and determined as a valid elastic signal area.

[0071] To better understand the automated analysis method of ultrasound shear wave elastography described above, the liver will be used as the target tissue for illustration below.

[0072] like Figure 4 As shown, automated analysis of ultrasonic shear wave elastography includes a model building phase and an evaluation phase.

[0073] Model building phase:

[0074] 1. Historical Dataset Preparation: A large historical dataset is collected, containing thousands of clinically validated standardized liver shear wave elastography (SWE) examination records. Each dataset must include two parts: first, SWE images acquired in standard intercostal sections and having passed quality control; second, reference elasticity values ​​obtained by at least two senior ultrasound experts following uniform standards, manually delineating the region of interest (ROI) in the corresponding liver parenchyma region of the SWE image and measuring them, namely the mean Young's modulus, maximum Young's modulus, and minimum Young's modulus, which serve as the gold standard for model training.

[0075] 2. Feature Extraction: For each quality-controlled SWE image in the dataset, a set of pre-defined multi-dimensional pixel value distribution features are extracted from its effective elastic signal region. These multi-dimensional pixel value distribution features include color statistical features (e.g., the mean and standard deviation of the red and blue channels and their ratio), texture features (e.g., contrast and energy calculated based on the gray-level co-occurrence matrix), and spatial distribution features (e.g., the mean gradient magnitude of pixel intensity). All features are standardized after extraction to eliminate the influence of dimensions.

[0076] 3. Construction and Training of the Elasticity Value Prediction Model: This stage uses a specially designed network structure for training, which can be called an attention-guided cross-scale feature fusion network. For example... Figure 5 As shown, Conv3×3, Conv5×5, and Conv7×7 represent convolutional layers with kernel sizes of 3×3, 5×5, and 7×7, respectively; ReLU represents the ReLU activation function; H and W represent the height and width of the feature map, respectively; Concat represents the feature concatenation operation; Squeeze represents the squeeze function; FC represents a fully connected layer; softmax represents the softmax normalization function; F_local represents the f_local function; and Flatten represents the flattening operation. The construction and training of the elasticity value prediction model are implemented as follows:

[0077] Network Input and Primary Encoding: The standardized multi-dimensional pixel value distribution features are used as input. First, a fully connected encoding layer maps the features to a high-dimensional latent space (N-dimensional → D-dimensional), forming a primary feature representation.

[0078] Cross-scale feature generation and attention fusion: The network comprises three parallel processing branches, simulating attention to "local details," "medium-sized regions," and "global context" of an image, respectively. By transforming the primary feature representation in different ways (e.g., using convolutional kernels of different sizes or nonlinear projection with fully connected layers), feature maps corresponding to these three perceptual scales are generated. Subsequently, a scale-attention gating module is introduced, which automatically learns and outputs a set of ternary weights (α, β, γ) for dynamically weighted summation of the feature maps at these three scales, forming an adaptively fused comprehensive feature representation, resulting in the fused feature map. This design enables the network to autonomously determine which spatial scale information is more critical for the current input.

[0079] Dual-stream prediction and consistency constraints: The aforementioned integrated features are input into two parallel prediction streams: the main prediction stream directly outputs preliminary predictions of the average, maximum, and minimum Young's modulus values. The auxiliary regularization stream learns to predict an "elastic distribution rationality score," which does not directly correspond to specific numerical values ​​but is used to assess the physical rationality of the relationship between the preliminary prediction values ​​(e.g., the maximum value is not less than the average value). During training, a specialized loss function is designed, including an error loss term between the predicted values ​​and the gold standard, and a constraint loss term guided by the auxiliary regularization stream to ensure that the predicted values ​​conform to physical laws. This allows the output of the main prediction stream to continuously adjust towards conforming to physical laws under the guidance of the auxiliary stream.

[0080] Training process: Features extracted from historical datasets are used as input, and the corresponding gold standard elasticity values ​​are used as training targets. The network is trained end-to-end by minimizing the error between the predicted values ​​and the gold standard (e.g., smoothed L1 loss) and the aforementioned physical plausibility constraint loss. Training uses the backpropagation algorithm to optimize network parameters until the model performs stably on independent validation sets.

[0081] Evaluation phase:

[0082] During the evaluation phase, operators use a system that integrates the aforementioned models to perform fully automated testing.

[0083] 1. Standard Section Positioning: The operator first locates the standard intercostal section of the liver in the anatomical imaging mode of the ultrasound imaging equipment. The system interface guides the operator to position a preset measurement frame (e.g., 3 cm × 3 cm) at a depth of about 2 cm under the liver capsule, and automatically identifies and prompts the operator to avoid major blood vessels.

[0084] 2. One-click automatic measurement: After confirming the location is suitable, the operator clicks the "Automatic Measurement" button on the system. For the pre-quality control step, a composite evaluation strategy integrating global grayscale statistical distribution (based on histograms) and multi-scale frequency domain texture features (based on wavelet transform) is adopted to ensure that the measurement area (ROI) avoids vascular structures and strong echo interference. For example... Figure 6 As shown, the specific steps are as follows:

[0085] Step 1: Initial screening of global statistical features of the ROI image (histogram analysis). The system acquires a grayscale image of a preset measurement region (e.g., 3cm × 3cm) and calculates its grayscale histogram. Since normal liver parenchyma echoes exhibit a uniform speckle distribution, the system focuses on analyzing the grayscale entropy and skewness of the histogram.

[0086] Calculate grayscale entropy: using the formula Calculate the grayscale entropy of the region, where p(i) is the probability of grayscale level i appearing. If the grayscale entropy H is greater than a preset threshold, it indicates that the texture complexity of the region is too high (possibly containing complex blood vessel patterns), and it is judged as unqualified.

[0087] Calculate histogram skewness: This calculates the degree of skewness in the grayscale distribution. If hyperechoic vessel walls (bright) or extremely hypoechoic bile ducts (dark) are present, the histogram will show a noticeable "tailing" effect. If the absolute value of the histogram skewness, |S|, is greater than a set threshold, it is considered unqualified.

[0088] Step 2: Fine-grained local texture evaluation based on wavelet transform (frequency domain analysis). Regions that pass the initial screening proceed to the multi-scale texture analysis stage. The system performs discrete wavelet transform on the ROI image and uses wavelet bases such as Haar or Daubechies for secondary decomposition.

[0089] High-frequency subband extraction: Obtain high-frequency subband coefficients containing horizontal details (LH), vertical details (HL), and diagonal details (HH). The vessel wall and lesion edges appear as high-amplitude coefficient clusters in these high-frequency subbands, while uniform speckle noise in the liver parenchyma appears as a low-amplitude random distribution.

[0090] High-frequency energy calculation: Calculate the average energy of each high-frequency sub-band.

[0091] .

[0092] Where N is the number of high-frequency subbands, C 2 LH C 2 HL C 2 HH These are the high-frequency subband coefficients for horizontal detail (LH), vertical detail (HL), and diagonal detail (HH), respectively.

[0093] Multi-scale determination: If the energy mean E detail If the energy level is greater than the preset texture energy threshold T, it indicates that there are significant structural edges (non-uniform speckle) in the region, and the region is determined to have interfering structures.

[0094] Step 3: Fusion Decision and Automatic Feedback. The system logically fuses the evaluation results from Step 1 and Step 2. Only when the histogram statistical indicators (grayscale entropy, histogram skewness) are normal (grayscale entropy less than or equal to a preset threshold, absolute value of histogram skewness less than or equal to a set threshold), and the average energy of the high-frequency subband is less than or equal to the texture energy threshold, is the region determined to be "homogeneous liver parenchyma," and the system automatically triggers subsequent shear wave elastography mode switching. Otherwise, the system provides specific feedback to the operator based on the type of indicator that failed (such as detection of strong vascular echoes or excessively complex image texture), suggesting fine-tuning the probe position of the ultrasound imaging equipment.

[0095] 3. Automated Imaging and Region Segmentation Quality Control: The system automatically sends commands through the communication interface of the ultrasound imaging device. Upon receiving the commands, the ultrasound imaging device automatically switches to SWE mode to acquire and transmit slice images. The system immediately performs quality control segmentation on the transmitted slice images. This process is completed using a multimodal guided cascaded optimization segmentation network, such as... Figure 7 As shown, Deconv represents the transposed convolution layer. Its iterative module's internal processing flow is: a 3×3 convolution (output 16 channels) → ReLU activation → a 3×3 convolution (output 8 channels) → ReLU activation → a 1×1 convolution (output 1 channel) → Sigmoid activation. The specific steps of the quality control segmentation are as follows:

[0096] Initialization and bimodal input: The segmentation network is simultaneously input in two forms: one is the original color-coded elastic image; the other is a single-channel "hardness confidence map" generated based on the color-hardness mapping table.

[0097] Cascaded Encoder-Decoder Structure: The segmentation network itself is a cascaded structure. The first-level encoder focuses on learning edge and texture features from the color-coded elastic image, while the first-level decoder outputs a preliminary soft segmentation probability map that may contain noise and holes. The second-level network takes the probability map output from the first level and the "hardness confidence map" as input. The second-level encoder focuses on fusing spatial confidence information, while the second-level decoder is responsible for refining the boundaries of the first-level result and filling in internal holes.

[0098] Adaptive Iterative Optimization: The output of the second-level decoder does not generate the final result in one step, but instead enters a lightweight adaptive iterative optimization module. This module uses the segmentation result from the previous step and the original input as conditions to perform a few (e.g., 3) forward inference iterations. Each iteration fine-tunes the boundaries to make the segmentation contour fit the color and confidence boundaries in the image more closely, ultimately outputting a high-precision binary segmentation mask as an effective elastic signal region.

[0099] Quality control judgment: Calculate the percentage (coverage) of the area of ​​the effective elastic signal region that is finally segmented to the area of ​​the target region to be tested. If the coverage is greater than or equal to the preset quality control threshold (e.g., 95%), and the shape of the segmented region is compact, then the acquisition and segmentation are deemed qualified.

[0100] 4. Feature Extraction and Intelligent Prediction: For images that pass quality control, the system automatically extracts multi-dimensional pixel value distribution features from the segmented effective elastic signal region, which are identical to those extracted during the model building stage. These multi-dimensional pixel value distribution features are then input into a pre-deployed "attention-guided cross-scale feature fusion network" model. The model performs forward computation and directly outputs three predicted values: average Young's modulus, maximum Young's modulus, and minimum Young's modulus.

[0101] 5. Report generation: The system automatically generates a structured report that clearly lists the predicted elasticity value and the quality control status of this measurement (e.g., "Pass, coverage: 98%").

[0102] The above-mentioned automated analysis method for ultrasonic shear wave elastography has the following significant advantages:

[0103] 1. Achieve full-process automation and standardization: Automatically control equipment operation and quality control through program interfaces, completely eliminating delays, omissions and subjective judgment differences between manual operation steps, ensuring high consistency and repeatability throughout the entire process from image acquisition to analysis.

[0104] 2. Improve the objectivity and stability of results: The quality assessment and region segmentation based on full-area feature analysis are used to replace manual ROI selection, which avoids the fluctuation of results caused by the arbitrariness of ROI location and size selection, has stronger resistance to local image interference, and the measurement results are more stable.

[0105] 3. Make full use of image information and perform intelligent analysis: The model learns from the multi-dimensional features of the entire effective area, such as color, texture, and spatial distribution, and explores deep information correlations that go beyond a single physical formula, making the calculation of elasticity value more intelligent and comprehensive.

[0106] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0107] Based on the same inventive concept, this application also provides an automated analysis system for ultrasonic shear wave elastography, which implements the automated analysis method for ultrasonic shear wave elastography described above. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the automated analysis system for ultrasonic shear wave elastography provided below can be found in the limitations of the automated analysis method for ultrasonic shear wave elastography described above, and will not be repeated here.

[0108] In one embodiment, such as Figure 8 As shown, an automated analysis system for ultrasonic shear wave elastography is provided, comprising:

[0109] The first module 110 is used to acquire standard anatomical cross-sectional images of the target tissue and perform pre-quality control, evaluate the uniformity of the preset measurement area, and determine the target area to be measured.

[0110] The second module 120 is used to automatically control the ultrasound imaging equipment to switch to shear wave elastography mode through the software control interface, and to acquire color-coded elastography images of the target area to be tested.

[0111] The third module 130 is used to perform automatic quality assessment and region segmentation on the acquired color-coded elastic images using a multimodal guided cascaded optimization segmentation network to obtain effective elastic signal regions.

[0112] The fourth module 140 is used to extract the same multi-dimensional pixel value distribution features from the effective elastic signal region as those in the training phase of the elastic value prediction model.

[0113] The fifth module 150 is used to input the extracted multi-dimensional pixel value distribution features into the pre-trained elasticity value prediction model and output the elasticity value prediction results of the corresponding target tissue. The elasticity value prediction model adopts an attention-guided cross-scale feature fusion network structure to represent the mapping relationship between pixel features and elasticity values.

[0114] The modules in the aforementioned automated analysis system for ultrasonic shear wave elastography can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0115] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 9 As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements an automated analysis method for ultrasonic shear wave elastic imaging. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0116] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0117] In one embodiment, a computer device is provided. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method described above.

[0118] In one embodiment, a computer-readable storage medium is also provided. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0119] In one embodiment, a computer program product is also provided. The computer program product includes a computer program that, when executed by a processor, implements the steps of the method described above.

[0120] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0121] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0122] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. An automated analysis method for ultrasonic shear wave elastography, characterized by, include: Obtain standard anatomical cross-sectional images of the target tissue and perform pre-quality control; evaluate the uniformity of the preset measurement area to determine the target area to be measured. The ultrasonic imaging equipment is automatically switched to shear wave elastography mode via the software control interface to acquire color-coded elastography images of the target area. A multimodal guided cascaded optimization segmentation network is used to automatically assess the quality and segment the regions of the acquired color-coded elastic images to obtain effective elastic signal regions. From the effective elastic signal region, extract the same multi-dimensional pixel value distribution features as those used in the training phase of the elastic value prediction model; The extracted multi-dimensional pixel value distribution features are input into a pre-trained elasticity value prediction model, which outputs the elasticity value prediction results of the target tissue. The elasticity value prediction model adopts an attention-guided cross-scale feature fusion network structure to represent the mapping relationship between pixel features and elasticity values. The elasticity value prediction model includes: Parallel branches are used to generate feature maps at local, intermediate, and global scales. A scale attention gating module is used for dynamically weighted fusion of the three scale feature maps; The main prediction stream and the auxiliary regularization stream are parallel. The main prediction stream outputs the elasticity value, while the auxiliary regularization stream outputs a score to constrain the physical rationality of the prediction results.

2. The method of claim 1, wherein, Acquire standard anatomical section images of the target tissue and perform pre-quality control. Evaluate the uniformity of the preset measurement area to determine the target area to be tested. This includes: using a composite evaluation strategy that integrates global gray-level statistical distribution and multi-scale frequency domain texture features to perform pre-quality control on the preset measurement area of ​​the standard anatomical section image. When the histogram statistical index and wavelet high-frequency energy of the preset measurement area are both lower than the corresponding thresholds, it is determined as the target area to be tested. The histogram statistical index includes gray-level entropy and histogram skewness.

3. The method according to claim 2, characterized in that, A composite evaluation strategy integrating global gray-level statistical distribution and multi-scale frequency domain texture features includes: A grayscale image of the preset measurement area is obtained to generate a grayscale histogram. Based on the grayscale histogram, the grayscale entropy and histogram skewness are analyzed. If the grayscale entropy is less than or equal to a preset threshold and the absolute value of the histogram skewness is less than or equal to a set threshold, the histogram statistical index is qualified, and the preset measurement area is determined to be a region that passes the initial screening. Discrete wavelet transform is performed on the regions that pass the initial screening to extract high-frequency subband coefficients containing horizontal, vertical, and diagonal details. The average energy of each high-frequency subband is calculated based on the high-frequency subband coefficients. If the average energy is less than or equal to the texture energy threshold, the wavelet high-frequency energy is qualified, and the regions that pass the initial screening are determined as the target regions to be tested.

4. The method according to claim 1, characterized in that, A multimodal guided cascaded optimization segmentation network is used to automatically assess the quality and segment regions of acquired color-coded elastic images, including: The color elastic image and the hardness confidence map generated based on the color-hardness mapping are used as bimodal inputs. The network adopts a cascaded encoder-decoder structure. The first-level encoder learns edge and texture features from the color-coded elastic image, and the first-level decoder outputs a preliminary segmentation probability map. The second-level network is guided by the preliminary segmentation probability map and the hardness confidence map to perform fine segmentation. The second-level encoder is responsible for fusing spatial confidence information, and the second-level decoder is responsible for finely correcting the boundaries of the first-level result and filling internal holes. The segmentation result output by the second-level decoder is subjected to boundary adjustment by an adaptive iterative optimization module, and finally outputs a binary segmentation mask.

5. The method according to claim 4, characterized in that, After automatically assessing the quality and segmenting the acquired color-coded elastic image using a multimodal guided cascaded optimization segmentation network, the process further includes: if the percentage of the binary segmentation mask output by the region segmentation occupies the target area is greater than or equal to a preset quality control threshold, then the elastic image acquisition is qualified and it is determined to be a valid elastic signal area.

6. The method according to claim 1, characterized in that, Acquiring standard anatomical cross-sectional images of the target tissue includes: setting the ultrasound imaging device to anatomical imaging mode and acquiring standard anatomical cross-sectional images of the target tissue.

7. The method according to claim 6, characterized in that, Also includes: Learning is performed using a historical dataset, which contains multiple sets of associated data pairs. Each set of associated data pairs includes a quality-controlled effective elastic signal region and a reference elastic value obtained by measuring the effective elastic signal region using the gold standard method. Multidimensional pixel value distribution features of the effective elastic signal region are extracted. Using the multidimensional pixel value distribution features as input and the corresponding reference elastic value as the target, a machine learning regression algorithm is used for training to construct an elastic value prediction model that can characterize the mapping relationship from pixel features to elastic values.

8. The method according to claim 7, characterized in that, The training process of the elasticity value prediction model uses a loss function that includes an error loss term between the predicted value and the gold standard, as well as a constraint loss term guided by an auxiliary regularization flow to ensure that the predicted values ​​conform to physical laws.

9. The method according to any one of claims 1 to 8, characterized in that, The multi-dimensional pixel value distribution features include color statistical features, texture features, and spatial distribution features; the elasticity value prediction results include the average Young's modulus value, the maximum Young's modulus value, and the minimum Young's modulus value.

10. An automated analysis system for ultrasonic shear wave elastography, characterized in that, include: The first module is used to acquire standard anatomical cross-sectional images of the target tissue and perform pre-quality control, evaluate the uniformity of the preset measurement area, and determine the target area to be measured. The second module is used to automatically control the ultrasound imaging equipment to switch to shear wave elastography mode through the software control interface, and to acquire color-coded elastography images of the target area to be tested. The third module is used to automatically assess the quality and segment the acquired color-coded elastic images using a multimodal guided cascaded optimization segmentation network to obtain effective elastic signal regions. The fourth module is used to extract the same multi-dimensional pixel value distribution features from the effective elastic signal region as those in the training phase of the elastic value prediction model. The fifth module is used to input the extracted multi-dimensional pixel value distribution features into the pre-trained elasticity value prediction model and output the elasticity value prediction results of the corresponding target tissue. The elasticity value prediction model adopts an attention-guided cross-scale feature fusion network structure to represent the mapping relationship between pixel features and elasticity values. The elasticity value prediction model includes: Parallel branches are used to generate feature maps at local, intermediate, and global scales. A scale attention gating module is used for dynamically weighted fusion of the three scale feature maps; The main prediction stream and the auxiliary regularization stream are parallel. The main prediction stream outputs the elasticity value, while the auxiliary regularization stream outputs a score to constrain the physical rationality of the prediction results.

Images