Deep learning based ultrasound image quality automatic evaluation and grading method and system

Abstract

The application discloses an ultrasonic image quality automatic evaluation and grading method and system based on deep learning, and belongs to the technical field of ultrasonic image quality evaluation. The method comprises the following steps: acquiring ultrasonic imaging parameters and original echo signals, and performing beam synthesis to obtain an initial ultrasonic image; a soft thresholding module of a deep residual shrinkage network is used to perform noise suppression on multi-scale feature channels to obtain a denoised feature map; spatial attention weighted mapping is performed, structural similarity measurement is performed on a historical high-quality ultrasonic image template, and a global quality deviation index is determined; a reconstruction error is calculated for a local texture region by calling an adversarial generation network according to the global quality deviation index, and a local distortion heat map is generated; a comprehensive evaluation tensor is constructed by fusing the global deviation index and the local distortion heat map, and a quality grading label is output by adopting a reinforcement learning intelligent agent. The method can adaptively suppress ultrasonic speckle noise, finely position local texture distortion, and realize intelligent quality grading through reinforcement learning.

Description

Technical Field

[0001] This invention relates to the field of ultrasound image quality assessment technology, specifically to a method and system for automatic assessment and grading of ultrasound image quality based on deep learning. Background Technology

[0002] Ultrasound image quality assessment is a crucial step in the ultrasound testing and diagnostic process, directly impacting image interpretability and the accuracy of subsequent analysis. Current technologies often rely on manually designed feature metrics, such as image contrast, resolution, and signal-to-noise ratio, using thresholds or simple regression models for grading. Some methods introduce convolutional neural networks to directly extract global image features and output quality scores, achieving end-to-end quality prediction. However, the inherent speckle noise in ultrasound images severely interferes with feature extraction. Traditional filtering or conventional convolution operations struggle to effectively suppress noise while preserving tissue edges, leading to noise residue and feature representation degradation. This results in the quality assessment losing its sensitivity to the fidelity of the actual structure. Existing solutions lack systematic structural similarity comparisons with historically acquired high-quality ultrasound image templates, failing to quantify the degree of global structural deviation of the current image relative to high-quality standards, thus hindering an objective evaluation of the overall geometric and textural fidelity of the image. Furthermore, conventional methods generally neglect fine-grained distortion detection in local texture regions. When local artifacts or loss of subtle structures appear in the image, their information is easily overwhelmed by operations such as global average pooling, making it impossible to generate traceable local distortion distributions. Furthermore, existing grading strategies mostly employ fixed thresholds or pre-trained classifiers, failing to dynamically adjust decisions based on global and local quality information, and unable to simulate the grading logic of experts after integrating multi-dimensional information. The decision-making process lacks flexibility and self-optimization capabilities. Therefore, it is urgent to address several interrelated issues: accurately measuring global structural deviation under adaptive noise suppression, triggering fine-grained detection of local distortions based on global deviations, and intelligently integrating and grading multi-dimensional quality information. Summary of the Invention

[0003] This paper presents a method and system for automatic assessment and grading of ultrasound image quality based on deep learning. It achieves noise suppression and global structural deviation measurement through soft thresholding and spatial attention mechanism, combines adversarial generative network reconstruction error local distortion, and uses reinforcement learning agent to make decisions on the fused quality tensor. This overcomes the problems of insufficient noise suppression, lack of global template comparison, missed detection of local distortion, and rigid grading decision of existing methods, and realizes objective, accurate and intelligent grading of ultrasound image quality.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides an automatic ultrasound image quality assessment and grading method based on deep learning, comprising: acquiring imaging parameters of an ultrasound device and raw ultrasound echo signals acquired by an ultrasound probe; performing beamforming processing on the raw ultrasound echo signals to obtain an initial ultrasound image; inputting the initial ultrasound image into a deep residual shrinking network, using a soft thresholding module in the deep residual shrinking network to perform noise suppression processing on the multi-scale feature channels of the initial ultrasound image, and outputting a denoised feature map; performing spatial attention weighted mapping on the denoised feature map to obtain a weighted feature map; determining the global quality deviation index of the initial ultrasound image by measuring the structural similarity between the weighted feature map and historical high-quality ultrasound image templates; based on the global quality deviation index, calling an adversarial generative network to perform pixel-by-pixel reconstruction error calculation on the local texture regions of the initial ultrasound image to obtain a local distortion heatmap; fusing the global quality deviation index and the local distortion heatmap to construct a comprehensive evaluation tensor for ultrasound image quality; using a reinforcement learning agent to make a grade decision on the comprehensive evaluation tensor, and outputting a quality grading label for the initial ultrasound image.

[0005] The above scheme adaptively denoises multi-scale features through a soft thresholding mechanism in a deep residual shrinkage network. It can automatically determine independent thresholds based on the statistical distribution of feature channels, effectively suppressing speckle noise and artifacts in ultrasound images while preserving key structural information such as tissue boundaries. By utilizing spatial attention weighting and structural similarity measurement of historical high-quality templates, it quantifies the deviation between the current image and high-standard images from a global perspective, making the evaluation results more consistent with clinical image reading habits. When the global deviation index is triggered, an adversarial generative network is activated to perform pixel-level reconstruction error analysis on local textures. The generated local distortion heatmap can accurately locate areas with insufficient acoustic windows, acoustic shadows, or blurred textures. Finally, a reinforcement learning agent makes a grading decision by integrating global quality and local distortion information. Through interactive optimization with environmental feedback, the objectivity and consistency of the grading strategy are continuously improved, realizing automated, fine-grained evaluation and grading of ultrasound image quality.

[0006] As a preferred embodiment of the present invention, the step of inputting the initial ultrasound image into a deep residual shrinking network and using the soft thresholding module in the deep residual shrinking network to perform noise suppression processing on the multi-scale feature channels of the initial ultrasound image and outputting a denoised feature map specifically includes: performing layer-by-layer convolution operations on the initial ultrasound image using multiple convolutional layers in the deep residual shrinking network to extract the multi-scale feature channels of the initial ultrasound image under different receptive fields; and inputting the multi-scale feature channels into the global average pooling layer in the deep residual shrinking network to calculate the global statistics of each feature channel. The global statistics are input into the fully connected layer and activation function layer of the deep residual shrinkage network to generate adaptive scaling parameters for each feature channel. The adaptive scaling parameters are used to scale the corresponding feature channels to obtain scaled feature channels. The scaled feature channels are then passed through the threshold calculation subnetwork in the soft thresholding module to generate independent thresholds for each feature channel. The scaled feature channels are then soft-thresholded using the independent thresholds, setting feature values ​​with absolute values ​​less than the independent thresholds to zero and shrinking feature values ​​with absolute values ​​greater than the independent thresholds toward zero to obtain the denoised feature map.

[0007] This preferred approach learns the importance of each feature channel through global statistics and adaptive scaling parameters, and then generates an independent soft threshold for each channel through a threshold calculation sub-network. This allows the network to perform differentiated denoising based on noise intensity and feature saliency, removing noise while avoiding damage to weak but critical edge and texture features, thus improving the fidelity and discriminative power of the denoised feature map.

[0008] As another preferred embodiment of the present invention, the step of performing spatial attention weighted mapping on the denoised feature map to obtain a weighted feature map, and determining the global quality deviation index of the initial ultrasound image by measuring the structural similarity between the weighted feature map and the historical high-quality ultrasound image template, specifically includes: inputting the denoised feature map into the max pooling branch and average pooling branch in the spatial attention module respectively to obtain a first spatial attention map and a second spatial attention map; performing feature concatenation on the first spatial attention map and the second spatial attention map, and then performing fusion processing through a convolutional layer with a preset kernel size to generate a spatial attention weight map; multiplying the spatial attention weight map element-wise with the denoised feature map to obtain the weighted feature map; extracting the historical high-quality ultrasound image template that matches the current ultrasound equipment model and the current scanning site from the database, performing feature extraction on the historical high-quality ultrasound image template to obtain a template feature map; calculating the structural similarity metric between the weighted feature map and the template feature map, and determining the global quality deviation index based on the statistical distribution of the structural similarity metric across all spatial locations.

[0009] This preferred scheme utilizes a bi-branch pooling spatial attention mechanism to highlight key anatomical structures and suppress irrelevant background. Simultaneously, it compares the structure with historical high-quality templates from the same equipment and location. Statistical distribution analysis can reflect the degree of difference between the overall imaging quality and the high-standard templates, making the global deviation index device-adaptive and reducing evaluation bias caused by different parameter settings and operating habits.

[0010] As a further preferred embodiment of the present invention, the step of calling the Generative Adversarial Network (GAN) to calculate the pixel-by-pixel reconstruction error of the local texture region of the initial ultrasound image based on the global quality deviation index to obtain a local distortion heatmap specifically includes: when the global quality deviation index exceeds a preset quality threshold, inputting the initial ultrasound image into the generator in the GAN, the generator using an encoder-decoder structure to perform feature encoding and reconstruction processing on the initial ultrasound image, and outputting a reconstructed ultrasound image; simultaneously inputting the initial ultrasound image and the reconstructed ultrasound image into the discriminator in the GAN, the discriminator outputting a discrimination difference map of the initial ultrasound image and the reconstructed ultrasound image at the pixel level; calculating the reconstruction error value of each pixel position in the initial ultrasound image based on the discrimination difference map; dividing the initial ultrasound image into multiple local texture regions, aggregating and statistically analyzing the reconstruction error values ​​of all pixels in each local texture region to obtain the region reconstruction error of each local texture region; arranging the region reconstruction errors of all local texture regions according to the spatial layout of the initial ultrasound image to generate the local distortion heatmap.

[0011] This preferred scheme only triggers the adversarial generative network to perform local refinement analysis when the global deviation is large, effectively balancing computational overhead and evaluation precision. The generator and discriminator of the encoder-decoder structure work together to capture subtle texture distortion and structural aberration from pixel-level reconstruction differences. The aggregated local distortion heatmap provides accurate spatial location information for subsequent grading, which helps to distinguish different quality defect types such as global degradation and local occlusion.

[0012] As another preferred embodiment of the present invention, the step of fusing the global quality deviation index and the local distortion heatmap to construct a comprehensive evaluation tensor for ultrasound image quality, and using a reinforcement learning agent to make grade decisions on the comprehensive evaluation tensor to output the quality grade label of the initial ultrasound image, specifically includes: copying and expanding the global quality deviation index into a global deviation map with the same spatial size as the local distortion heatmap; concatenating the global deviation map and the local distortion heatmap in the channel dimension to obtain a fused feature map; performing a three-dimensional convolution operation on the fused feature map to extract the joint features of the fused feature map in the spatial and channel dimensions to generate the comprehensive evaluation tensor; inputting the comprehensive evaluation tensor into the policy network of the reinforcement learning agent, the policy network outputting a decision probability distribution corresponding to multiple quality levels; and selecting the quality level with the highest probability value as the quality grade label according to the decision probability distribution.

[0013] This preferred approach deeply integrates global deviation information and local distortion information in spatial and channel dimensions, and captures the coupling relationship between the two through three-dimensional convolution. The resulting comprehensive evaluation tensor has stronger representation capabilities. By using a reinforcement learning agent for decision-making, the grading strategy can be continuously optimized based on historical feedback, making the quality level classification more in line with clinical acceptability standards and reducing subjective differences.

[0014] As a further definition of the reinforcement learning decision-making process described above, the step of inputting the comprehensive evaluation tensor into the policy network of the reinforcement learning agent, wherein the policy network outputs a decision probability distribution corresponding to multiple quality levels, includes: extracting sequence features from the comprehensive evaluation tensor along the time dimension using multiple temporal convolutional layers in the policy network to obtain a temporal feature sequence; inputting the temporal feature sequence into a long short-term memory network unit in the policy network to update the internal hidden state of the long short-term memory network unit to obtain the decision hidden vector at the current time; inputting the decision hidden vector into a fully connected classification layer in the policy network, wherein the fully connected classification layer outputs the decision probability distribution through a flexible maximum activation function; and updating the network parameters of the policy network during the training phase of the reinforcement learning agent based on the dominance function between the decision probability distribution and the quality score reward from the environmental feedback.

[0015] This preferred approach introduces temporal convolution and long short-term memory networks to perform temporal modeling of the comprehensive evaluation tensor. It can use contextual information from multiple evaluation processes to smooth decision-making and avoid level jumps caused by accidental anomalies in a single frame image. During training, the advantage function guides the policy network to optimize towards a better classification direction, making the model more robust in classifying boundary samples.

[0016] As another preferred embodiment of the present invention, the step of acquiring the imaging parameters of the ultrasound device and the original ultrasound echo signal acquired by the ultrasound probe, and performing beamforming processing on the original ultrasound echo signal to obtain an initial ultrasound image, specifically includes: acquiring the transmission frequency, receiving gain, and dynamic range of the ultrasound device as the imaging parameters, and acquiring the original radio frequency echo signal acquired by the ultrasound probe on multiple scan lines as the original ultrasound echo signal; performing dynamic range compression processing on the original radio frequency echo signal according to the dynamic range in the imaging parameters to obtain a compressed echo signal; performing orthogonal demodulation processing on the compressed echo signal to extract the in-phase component and the quadrature component, and constructing a complex analytical signal based on the in-phase component and the quadrature component; determining the time delay compensation coefficient between the multiple scan lines according to the transmission frequency and receiving gain in the imaging parameters, and performing scan-line-by-scan delay calibration processing on the complex analytical signal using the time delay compensation coefficient to obtain a time delay calibrated signal; and performing weighted summation on the time delay calibrated signal to complete the beamforming processing and obtain the initial ultrasound image.

[0017] This preferred scheme utilizes real-time acquired imaging parameters to guide dynamic range compression and time delay calibration, enabling beamforming to adapt to echo characteristics under different scanning conditions. The resulting initial ultrasound image has a moderate dynamic range and good beam focusing, providing a basic image with a higher signal-to-noise ratio and clearer structure for subsequent quality assessment.

[0018] As a preferred embodiment of the present invention in terms of training the noise reduction network, the soft thresholding module in the deep residual shrinking network is optimized in the training phase as follows: a training set containing multiple ultrasound image samples labeled with noise levels is constructed; the ultrasound image samples in the training set are input into the deep residual shrinking network to obtain a predicted noise-reduced feature map; a first loss function value is calculated using the mean square error between the predicted noise-reduced feature map and the clean ultrasound image label corresponding to the ultrasound image sample; a second loss function value is calculated using the regularization constraint between the independent threshold output by the soft thresholding module in the deep residual shrinking network and the noise level label of the ultrasound image sample; the first loss function value and the second loss function value are weighted and summed to obtain a total loss function value; the network weights of the deep residual shrinking network are updated using the total loss function value through a backpropagation algorithm; after each backpropagation update, the scaling factor of the batch normalization layer in the deep residual shrinking network is pruned, and scaling factors smaller than a preset pruning threshold are set to zero.

[0019] This training scheme combines reconstruction error with threshold regularization to force the network to learn independent thresholds that are correlated with the real noise level, thereby enhancing the physical interpretability and generalization ability of the soft thresholding module. At the same time, sparsity pruning of the batch normalization scaling factor can implement channel-level pruning, reduce model complexity, and is beneficial for deployment at the edge of ultrasonic equipment with limited computing resources.

[0020] As a preferred embodiment of the present invention in terms of adversarial generative network structure, the encoder in the generator employs multiple downsampled convolutional blocks, the decoder employs a symmetrical number of upsampled deconvolutional blocks, and a skip connection is provided between the encoder and the decoder.

[0021] Skip connections can pass high-resolution detail features from the encoding stage to the decoding stage, enabling reconstructed ultrasound images to retain clear texture edges when restoring the global structure. This makes pixel-by-pixel reconstruction error calculation more accurate and improves the sensitivity and specificity of local distortion heatmaps.

[0022] This invention also provides a deep learning-based automatic ultrasound image quality assessment and grading system, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the aforementioned deep learning-based automatic ultrasound image quality assessment and grading method. This system can embed the above method into an executable software module, embedding it into an ultrasound equipment workstation or cloud analysis platform to perform real-time or offline quality assessment and grading of scanned images, assisting physicians in quickly screening substandard images and improving the standardization and efficiency of ultrasound examinations.

[0023] The technical effects and advantages provided by the present invention in the above technical solution are as follows: A deep residual shrinking network is used to extract multi-scale features from the initial ultrasound images. A soft thresholding module generates an independent threshold for each feature channel based on global pooling statistics, and soft thresholding is applied to zero-scale features with absolute values ​​less than the threshold and shrink features with values ​​greater than the threshold towards zero, achieving adaptive removal of speckle noise. The processed denoised feature map is fed into a spatial attention module, where two spatial description maps are generated by max pooling and average pooling branches. These maps are then concatenated and fused into a spatial attention weight map, which is element-wise multiplied with the denoised feature map to obtain a weighted feature map, enabling the network to obtain a stronger response at anatomical structure edges and tissue boundaries. Subsequently, historical high-quality ultrasound image templates from the same model of equipment and the same scanning site are retrieved from the database, and their corresponding feature maps are extracted. The structural similarity measure between the weighted feature map and the template feature map is calculated and statistically distributed across all spatial locations to form a global quality deviation index. This process automatically matches noise suppression strength with local image statistical characteristics, preventing excessive smoothing of structural details. Spatial attention weighting concentrates evaluation resources on key diagnostic regions, and historical high-quality template comparison quantifies the abstract global image quality into deviation values, which are more stable and interpretable than direct regression quality scores. An adversarial generative network is invoked based on the global quality deviation index. The generator reconstructs the initial ultrasound image through an encoder-decoder combination of skip connections. The discriminator outputs a pixel-level discrimination difference map. After calculating the reconstruction error for each pixel, the maps are aggregated according to local texture regions to form a local distortion heatmap. The global quality deviation index is replicated and expanded into a global deviation map with the same spatial size as the heatmap. After concatenation along the channel dimension, joint spatiotemporal features are extracted using 3D convolution to construct a comprehensive ultrasound image quality evaluation tensor. The policy network of the reinforcement learning agent contains temporal convolutional layers and long short-term memory network units. It extracts sequence features along the time dimension of the comprehensive evaluation tensor and updates the hidden state to obtain the decision hidden vector. Finally, a fully connected classification layer outputs the decision probability distribution for each quality level, and the level with the highest probability is selected as the quality grading label. Generative adversarial networks are highly sensitive to texture anomalies and can precisely expose local micro-distortions that are difficult to detect by conventional methods. Heatmaps can intuitively show the spatial distribution of distortions. Reinforcement learning agents continuously optimize the classification strategy through reward mechanisms in their interaction with the environment. They can integrate global and local multi-dimensional information to make intelligent classifications that are close to expert decisions, avoiding misjudgments caused by fixed thresholds and improving the overall accuracy and automation of the evaluation. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0025] Figure 1This is a flowchart of a deep learning-based automatic ultrasound image quality assessment and grading method. Figure 2 This is a schematic diagram of the ultrasound image quality assessment process based on structural similarity. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] See Figure 1 This invention provides a deep learning-based method for automatic assessment and grading of ultrasound image quality. Specifically, it includes: acquiring imaging parameters of an ultrasound device and raw ultrasound echo signals collected by an ultrasound probe; performing beamforming on the raw ultrasound echo signals to generate an initial ultrasound image; inputting the initial ultrasound image into a pre-constructed deep residual shrinking network; using a soft thresholding module in the deep residual shrinking network to suppress noise in the multi-scale feature channels of the initial ultrasound image, thereby obtaining a denoised feature map; performing spatial attention weighted mapping on the denoised feature map to obtain a weighted feature map; determining the global quality deviation index corresponding to the initial ultrasound image by calculating the structural similarity metric between the weighted feature map and historical high-quality ultrasound image templates; based on the global quality deviation index, calling an adversarial generative network to calculate the pixel-by-pixel reconstruction error of the local texture regions of the initial ultrasound image, generating a local distortion heatmap; fusing the global quality deviation index and the local distortion heatmap to construct a comprehensive ultrasound image quality evaluation tensor; and using a reinforcement learning agent to make a grade decision on the comprehensive evaluation tensor, finally outputting the quality grading label of the initial ultrasound image. Example 1:

[0028] In its implementation, the deep learning-based automatic ultrasound image quality assessment and grading method inputs the initial ultrasound image into a deep residual shrinking network. The soft thresholding module within this network suppresses noise in the multi-scale feature channels of the initial ultrasound image, outputting a denoised feature map. The process begins with feature extraction from the initial ultrasound image using the deep residual shrinking network. This network comprises multiple sequentially connected convolutional layers, each employing a 3×3 kernel followed by a batch normalization layer and a linear rectified activation function. The stack size of these convolutional layers is 16. The initial ultrasound image is processed by the first convolutional layer to obtain a first-level feature map. This first-level feature map is then input into the second convolutional layer to obtain a second-level feature map, and so on. Through layer-by-layer convolutional operations, multi-scale feature channels under different receptive fields are extracted from the initial ultrasound image. The total number of multi-scale feature channels is 64. The extracted multi-scale feature channels are input into the global average pooling layer in the deep residual shrinking network. The global average pooling layer averages the value of each feature channel across all spatial locations, and the output scalar serves as the global statistic for the corresponding feature channel, resulting in 64 global statistics. These 64 global statistics are then input into the fully connected layer and activation function layer in the deep residual shrinking network. The fully connected layer consists of two fully connected structures. The first fully connected structure compresses the dimension of the global statistics from 64 to 16, followed by a linear rectified activation function. The second fully connected structure restores the dimension from 16 to 64, followed by an sigmoid growth curve activation function, ultimately generating 64 adaptive scaling parameters, each with a value between 0 and 1. Each adaptive scaling parameter is used to perform channel-by-channel scaling on the feature channel corresponding to its name. Specifically, the entire feature map of the c-th channel in the multi-scale feature channels is multiplied by the c-th adaptive scaling parameter to obtain the scaled feature channel, which still has 64 scaled feature channels.

[0029] The scaled feature channels are fed into the soft thresholding module of the deep residual shrinking network. This module contains a threshold calculation sub-network. The structure of this sub-network is as follows: First, a global average pooling operation is performed on the input scaled feature channels to obtain a 64-dimensional vector. This 64-dimensional vector then passes through a fully connected layer that compresses the dimension from 64 to 16, a linear rectified activation function, a fully connected layer that expands the dimension from 16 to 64, and an sigmoid growth curve activation function, outputting a 64-dimensional vector. Each element in this 64-dimensional vector is multiplied by the maximum absolute value of the corresponding channel's feature map in the scaled feature channels, resulting in 64 independent thresholds. After obtaining these 64 independent thresholds, soft thresholding is performed on the scaled feature channels. Specifically, for each feature value x in the c-th channel of the scaled feature channels, if the absolute value of x is less than the c-th independent threshold... When x is greater than or equal to zero, the eigenvalue at the corresponding position is set to zero; when the absolute value of x is greater than or equal to zero... When the time comes, update the feature value at the corresponding position to ,in This indicates the sign of x. After all channels undergo soft thresholding, the denoised feature map is obtained.

[0030] During the training phase of the deep residual shrinkage network, a training set is pre-constructed containing multiple ultrasound image samples with labeled noise levels. Each ultrasound image sample in the training set corresponds to a clean ultrasound image label and a set of noise level annotations. The noise level annotation is in the form of a 64-dimensional vector, where each element corresponds to a standard noise threshold for a feature channel. c represents the channel index, which ranges from 1 to 64. The value is pre-calculated by a pre-computation network sharing a structure with the thresholding subnetwork, based on the noise level of the ultrasound image samples. During training, the ultrasound image samples are input into the deep residual shrinking network, processed through all convolutional layers, global average pooling layers, fully connected layers, and a soft thresholding module to obtain the predicted denoised feature map. The first loss function value is calculated using the predicted denoised feature map and the corresponding clean ultrasound image label of the ultrasound image sample, calculated as mean squared error. Simultaneously, 64 independent thresholds output by the soft thresholding module in the deep residual shrinking network are used. Corresponding to the noise level label of the ultrasound image sample Calculate the second loss function value using the mean squared error. Then, sum the first and second loss function values ​​using a weighted average to obtain the total loss function value. The formula for calculating the total loss function value is as follows:

[0031] in, This represents the total loss function value. This represents the weighting coefficients of the first loss function. The weights of the second loss function are represented by H and W, which represent the spatial height and width of the predicted denoised feature map, respectively. C represents the total number of feature channels, and C=64. Indicates the spatial location in the predicted denoised feature map. The pixel value of the c-th channel. This represents the pixel value at the same spatial location and channel in a clean ultrasound image label. This represents the independent threshold of the c-th feature channel. The standard noise threshold for the c-th channel in the noise level labeling of the ultrasound image sample is shown. Set to 1.0, The threshold is set to 0.1 to ensure the first loss function dominates the denoising effect, while allowing the second loss function to impose appropriate regularization constraints on the reasonableness of the threshold, preventing independent thresholds from deviating too far from the standard noise threshold. The calculated total loss function value is used to update the network weights of the deep residual shrinking network via backpropagation, including the kernel parameters of all convolutional layers, the connection weights of fully connected layers, and the learnable parameters of batch normalized layers. After each backpropagation update, the scaling factors of all batch normalized layers in the deep residual shrinking network are pruned. The preset pruning threshold is set to 0.01, and scaling factors with an absolute value less than 0.01 are directly set to zero to improve the sparsity and generalization ability of the network.

[0032] Example 2: In specific implementation, please refer to Figure 2 A deep learning-based automatic ultrasound image quality assessment and grading method performs spatial attention weighted mapping on the denoised feature map to obtain a weighted feature map. The spatial attention module contains two parallel pooling branches: max pooling and average pooling. The denoised feature map is simultaneously input into both branches. The max pooling branch performs max pooling along the channel dimension of the denoised feature map, retaining the maximum value of all channels at each spatial location, and outputting a first spatial attention map with the same spatial size and one channel. The average pooling branch performs average pooling along the channel dimension of the denoised feature map, calculating the average value of all channels at each spatial location, and outputting a second spatial attention map with the same spatial size and one channel. The first and second spatial attention maps are then concatenated along the channel dimension to obtain a concatenated attention map with the same spatial size and two channels. The concatenated attention maps are fused using a 7×7 convolutional layer. The 7×7 kernel size is chosen to balance the aggregation range of local neighborhood information with computational overhead, enabling the fusion of two pooling information types within a larger receptive field. This generates a spatial attention weight map with one channel and an unchanged spatial size. The spatial attention weight map is then multiplied element-wise with the denoised feature map. The multiplication operation multiplies the feature values ​​of all channels in the denoised feature map by the corresponding weight values ​​in the spatial attention weight map at each spatial location, resulting in a weighted feature map.

[0033] The system extracts historical high-quality ultrasound image templates that match the current ultrasound equipment model and the current scanning site from the database. The database pre-stores multiple sets of key-value pairs. The key consists of the ultrasound equipment model code and the scanning site code, and the value corresponds to historical ultrasound image templates acquired by the ultrasound equipment model at that scanning site and rated as high-quality by experts. During extraction, the system reads the current ultrasound equipment model code and the scanning site code set by the operator, and searches the database for historical high-quality ultrasound image templates whose keys completely match the current ultrasound equipment model code and the current scanning site code. The extracted historical high-quality ultrasound image templates are input into a pre-trained feature extraction network. The structure of this feature extraction network is completely identical to the convolutional layer group structure retained after removing the soft thresholding module in the deep residual shrinking network used to generate the denoised feature map. The network parameters are learned synchronously during the training of the deep residual shrinking network. After the historical high-quality ultrasound image templates are processed layer by layer by multiple convolutional layers in the feature extraction network, a template feature map is output. The number of channels in the template feature map is the same as the number of channels in the weighted feature map, both being 64, and the spatial dimensions are also the same as the weighted feature map.

[0034] Calculate the structural similarity metric between the weighted feature map and the template feature map. The structural similarity metric is calculated independently at each spatial location in both the weighted and template feature maps. For spatial location (r, s), the 64-dimensional feature vector of the weighted feature map at that location is denoted as... The 64-dimensional feature vector of the template feature map at this location is denoted as... Structural similarity measure The calculation method is as follows: Calculate mean and standard deviation ,calculate mean and standard deviation ,calculate and Covariance between And introduce a stability constant. and To avoid a denominator of zero, structural similarity metrics The expression is:

[0035] in, This represents the structural similarity metric at spatial coordinates (r, s), where r is the row index, ranging from 1 to the height R of the weighted feature map, and s is the column index, ranging from 1 to the width S of the weighted feature map. This represents the mean of the 64-dimensional feature vectors of the weighted feature map at position (r,s); This represents the standard deviation of the 64-dimensional eigenvectors of the weighted feature map at position (r,s); This represents the standard deviation of the 64-dimensional feature vector of the template feature map at position (r,s); This represents the standard deviation of the 64-dimensional feature vector of the template feature map at position (r,s); This represents the covariance of the 64-dimensional feature vectors of the weighted feature map and the template feature map at position (r,s); and It is the stability constant. , Stability constant and The above values ​​are set based on the fact that within the dynamic range of 8-bit image brightness values, the two constants can effectively prevent instability in the structural similarity measurement values ​​of weak texture areas.

[0036] Get all Structural similarity measure at each spatial location Then, the global quality deviation index is determined based on the statistical distribution of structural similarity metrics across all spatial locations. The mean of all structural similarity metrics is calculated. and standard deviation Set adjustment factor Global quality deviation index Determined by the following formula:

[0037] in, This represents the global quality deviation index, which typically ranges from 0 to 1. A larger value indicates a greater degree of deviation of the initial ultrasound image from the historical high-quality ultrasound image template. This represents the mean of all spatial location structural similarity measures. The standard deviation of all spatial location structural similarity measures; The adjustment factor is set to 0.3, based on the selection of the value that maximizes the accuracy of quality grading within the range of 0.1 to 0.5 through experimental grid search. Example 3:

[0038] In practice, a preset quality threshold of 0.5 is set. This threshold is determined by analyzing the global quality deviation index distribution of historical ultrasound image samples, setting 0.5 as the critical value that can clearly distinguish normal images from images with local texture anomalies. When the global quality deviation index exceeds 0.5, an adversarial generative network is invoked to calculate the pixel-by-pixel reconstruction error of the local texture regions of the initial ultrasound image, generating a local distortion heatmap.

[0039] The Generative Adversarial Network (GAN) consists of a generator and a discriminator. The generator employs an encoder-decoder architecture. The encoder comprises four downsampled convolutional blocks connected sequentially. Each downsampled convolutional block contains a 3×3 kernel layer with a stride of 2, followed by a batch normalization layer and a leaky rectified linear activation function (RCEM) with a negative slope of 0.2. The output channels of the four downsampled convolutional blocks are 64, 128, 256, and 512, respectively. The decoder consists of four upsampled deconvolutional blocks connected sequentially, symmetrical in number to the downsampled convolutional blocks. Each upsampled deconvolutional block contains a 3×3 kernel layer with a stride of 2, followed by a batch normalization layer and an RCEM. The output channels of the four upsampled deconvolutional blocks are 256, 128, 64, and 3, respectively, with the last upsampled deconvolutional block having 3 output channels, corresponding to the three color channels of the reconstructed ultrasound image. Between the encoder and decoder, the output of the first downsampled convolutional block is concatenated with the input of the fourth upsampled deconvolutional block in the channel dimension via a skip connection. The output of the second downsampled convolutional block is concatenated with the input of the third upsampled deconvolutional block, and the output of the third downsampled convolutional block is concatenated with the input of the second upsampled deconvolutional block. The skip connection enables the decoder to fuse shallow texture detail information from the encoder.

[0040] The initial ultrasound image is input into the generator; the spatial dimension of the initial ultrasound image is height. , The number of channels is 3. The encoder performs feature encoding on the initial ultrasound image step by step, and the decoder performs upsampling reconstruction on the deepest feature map output by the encoder, and finally outputs the reconstructed ultrasound image. The spatial size of the reconstructed ultrasound image is the same as that of the initial ultrasound image.

[0041] The discriminator employs a fully convolutional structure. The input is a stitched result of two images in the channel dimension: an initial ultrasound image and a reconstructed ultrasound image. The stitched result is an input tensor with 6 channels. The discriminator consists of five sequentially connected convolutional layers. The first four layers use 4×4 kernels with a stride of 2 and padding of 1, followed by batch normalization layers and a leaky linear rectified activation function with a negative slope of 0.2. The output channel numbers are 64, 128, 256, and 512, respectively. The fifth convolutional layer uses a 4×4 kernel with a stride of 1 and padding of 1, outputting 1 channel, without batch normalization or activation. The discriminator processes the stitched input tensor and outputs a tensor with a spatial dimension of... Discriminant difference plot, for Round down to one-sixteenth. for Rounded down to one-sixteenth. The value at each spatial location in the discriminant difference map represents the discriminator's judgment of the degree of consistency between the initial ultrasound image and the reconstructed ultrasound image in the corresponding local region of that pixel. The larger the absolute value, the more obvious the difference.

[0042] The reconstruction error value at each pixel location in the initial ultrasound image is calculated based on the discriminant difference map, and bilinear interpolation is used to reconstruct the image. The size discrimination difference map is upsampled to the same size as the initial ultrasound image. Size, position of each pixel after upsampling The value at that location is denoted as The absolute value of this value is defined as the reconstruction error value of the corresponding pixel position in the initial ultrasound image. ,Right now ,in For pixel row index, the value ranges from 1 to , For pixel column indexes, the value range is from 1 to .

[0043] The initial ultrasound image was divided into multiple local texture regions. A fixed-size sliding window (32×32 pixels, 16 pixels step) was used for division. Starting from the top left corner of the initial ultrasound image, the window was cropped sequentially from left to right and top to bottom to generate multiple local texture regions. Each local texture region corresponded to a rectangular sub-region, which was identified by its starting row number. End line number Starting column number End column number Delineation. The reconstruction error values ​​of all pixels within each local texture region are aggregated and statistically analyzed by averaging. This represents the region reconstruction error. The calculation formula is:

[0044] in, This represents the region reconstruction error of a local texture area. Indicates the starting row number of the local texture region in the initial ultrasound image; Indicates the end row number of the local texture region in the initial ultrasound image; Indicates the starting column number of the local texture region in the initial ultrasound image; This represents the end column number of the local texture region in the initial ultrasound image; p is the row number variable, with values ​​ranging from... arrive ; q is the column index variable, and its value range is from arrive ; Indicates the pixel position The reconstruction error value at that location.

[0045] The reconstruction errors of all local texture regions are arranged according to the spatial layout of the initial ultrasound image to generate a local distortion heatmap. The arrangement is as follows: a center pixel position is determined for each local texture region, with the center position's row coordinates being... Column coordinates are The region reconstruction error of this local texture area The local distortion heatmap is obtained by filling the corresponding center position and the surrounding rectangular area of ​​a two-dimensional matrix with the same spatial size as the initial ultrasound image. For pixel positions where multiple local texture regions overlap, the maximum value of the region reconstruction error is taken as the final value. The value of each pixel in the local distortion heatmap represents the severity of local texture distortion at that location.

[0046] Example 4: In practice, the global quality deviation index is copied and expanded into a global deviation map with the same spatial size as the local distortion heatmap. The copying and expansion operation is achieved by filling the scalar value of the global quality deviation index into each pixel position of a two-dimensional matrix. The spatial height and spatial width of the two-dimensional matrix are consistent with those of the local distortion heatmap. The two-dimensional matrix obtained after filling is the global deviation map, and the value of all pixel positions in the global deviation map is equal to the global quality deviation index.

[0047] The global deviation map and the local distortion heatmap are concatenated along the channel dimension. The global deviation map is regarded as a single-channel matrix, and the local distortion heatmap is regarded as a single-channel matrix. They are stacked along the channel dimension to obtain a fused feature map with 2 channels and unchanged spatial size. The first channel of the fused feature map corresponds to the global deviation map, and the second channel of the fused feature map corresponds to the local distortion heatmap.

[0048] A 3D convolution operation is performed on the fused feature map to extract joint features in the spatial and channel dimensions, generating a comprehensive ultrasound image quality assessment tensor. The 3D convolutional layer containing the 3D convolution operation comprises 16 3D convolutional kernels, each with a size of 3×3×2. The first two 3×3 dimensions operate on the spatial dimension, and the third 2 dimensions operate on the channel dimension of the fused feature map. The stride is set to 1, the padding is set to 1, and the activation function is a linear rectified function. The 3D convolution operation transforms the fused feature map into a 3D tensor containing 16 feature planes, with the spatial dimensions of the tensor being the same as the fused feature map. The 3D tensor is then divided into multiple spatial slices along the height dimension. Each spatial slice is a 2D matrix with a size equal to the number of channels × the width. All spatial slices are arranged from top to bottom along the height direction, forming a comprehensive ultrasound image quality assessment tensor with height, channel, and width dimensions. The temporal dimension of the comprehensive ultrasound image quality assessment tensor is equal to its spatial height, and the input feature dimension at each time step is the product of the number of channels and the width.

[0049] The ultrasound image quality assessment tensor is input into the policy network of the reinforcement learning agent. The policy network outputs decision probability distributions corresponding to multiple quality levels. The policy network of the reinforcement learning agent consists of multiple sequentially connected temporal convolutional layers, a long short-term memory network unit, and a fully connected classification layer. Specifically, the multiple temporal convolutional layers include three one-dimensional temporal convolutional layers. Each one-dimensional temporal convolutional layer extracts sequence features along the temporal dimension of the ultrasound image quality assessment tensor. The first one-dimensional temporal convolutional layer has the feature dimension as its input channel, 32 output channels, a kernel size of 3, a stride of 1, and padding of 1. The second one-dimensional temporal convolutional layer has 32 input channels, 64 output channels, a kernel size of 3, a stride of 1, and padding of 1. The third one-dimensional temporal convolutional layer has 64 input channels, 128 output channels, a kernel size of 3, a stride of 1, and padding of 1. Each one-dimensional temporal convolutional layer is followed by a batch normalization layer and a linear rectified activation function. After processing through three one-dimensional temporal convolutional layers, a temporal feature sequence is obtained. The length of the temporal feature sequence is the same as the temporal dimension length of the ultrasound image quality comprehensive evaluation tensor, and the feature vector dimension of each time step is 128.

[0050] The temporal feature sequence is input into a Long Short-Term Memory (LSTM) network unit within the policy network. The hidden layer dimension of each LSM unit is set to 256. The LSM unit updates its internal hidden state sequentially along the time steps of the temporal feature sequence. At each time step, it receives the 128-dimensional feature vector of the current time step and outputs a 256-dimensional hidden vector, passing the updated cell state and hidden state to the next time step. After processing at all time steps, the 256-dimensional hidden vector of the last time step is used as the decision hidden vector for the current time step.

[0051] The latent decision vector is input into the fully connected classification layer in the policy network. The input dimension of the fully connected classification layer is 256, and the output dimension is the number of quality levels. The number of quality levels is set to 3, corresponding to the three levels of "excellent", "good", and "poor". The fully connected classification layer outputs the decision probability distribution through the flexible maximum activation function. The decision probability distribution is a 3-dimensional vector, and each element value in the vector represents the probability of selecting the corresponding quality level.

[0052] Based on the decision probability distribution, the quality level with the highest probability value is selected as the initial ultrasound image quality grading label. During the inference phase, the quality level with the highest probability value is directly used as the output result.

[0053] During the training phase of a reinforcement learning agent, the agent learns through interaction with its environment. The environment receives the decision probability distribution output by the agent and samples a quality-level action 'a' according to the probability distribution. The environment returns a quality score reward 'r' based on a predefined scoring rule: when the sampled action 'a' matches the pre-labeled true quality level, the quality score reward 'r' is +1; when the sampled action 'a' does not match the true quality level, the quality score reward 'r' is -1. During the training phase, a value network is introduced in parallel with the policy network. The structure of the value network is identical to the temporal feature extraction part of the policy network, and the output of the value network is a scalar state value. , This represents the expected cumulative discount reward of the current input ultrasound image quality comprehensive evaluation tensor. Advantage function. Calculated using the following formula:

[0054] in, The value of the advantage function for performing action a in state s is represented by r; r represents the quality score reward returned by the environment. This represents the discount factor, set to 0.99. The basis for this setting is to give the rewards of future steps a higher weight in the current decision, but not entirely equivalent to the immediate reward. Indicates the next state Corresponding state value, next state This represents the state of the agent after processing the next ultrasound image. This represents the state value corresponding to the current state s. The loss function of the policy network is constructed based on the advantage function, and the policy gradient method is used when updating the policy network parameters to maximize the expected log probability weighted by the advantage function; the loss function of the value network is the state value. With target value The mean squared error between the two networks is calculated, and the parameters of the two networks are updated synchronously in each iteration.

[0055] Example 5: In practical implementation, a deep learning-based automatic ultrasound image quality assessment and grading method acquires the ultrasound device's transmission frequency, receiving gain, and dynamic range as imaging parameters, and obtains the raw radio frequency echo signals collected by the ultrasound probe on multiple scan lines as the raw ultrasound echo signals. The ultrasound device outputs the three imaging parameters—transmission frequency, receiving gain, and dynamic range—in its current operating state through its control interface. The transmission frequency is obtained by reading the nominal center frequency of the ultrasound probe; the receiving gain is obtained by reading the current gain setting of the ultrasound device's receiving amplifier; and the dynamic range is obtained by reading the difference between the upper and lower limits of the current dynamic range of the ultrasound device's logarithmic compression module. The ultrasound probe collects raw radio frequency echo signals on multiple scan lines. The total number of scan lines is determined by the number of array elements and the scanning mode of the ultrasound probe. Each scan line corresponds to a sequence of radio frequency voltage values ​​arranged chronologically. The raw radio frequency echo signals from all scan lines are synchronously acquired and stored in a buffer memory, constituting the raw ultrasound echo signal.

[0056] Based on the dynamic range in the imaging parameters, the original radio frequency echo signal is subjected to dynamic range compression processing to obtain the compressed echo signal. Dynamic range compression is performed on each time sampling point of each scan line in the original radio frequency echo signal. The process is as follows: first, the absolute value of the original radio frequency echo signal amplitude at the current sampling point is calculated; then, the absolute value is divided by the upper limit of the dynamic range and the logarithm is taken to base 10; finally, it is multiplied by the dynamic range value; amplitudes exceeding the lower limit of the dynamic range are truncated to the lower limit, and the compressed amplitude is output. The compressed echo signal retains the relative relationship between the amplitudes of the original radio frequency echo signal while limiting the amplitude range within the dynamic range.

[0057] The compressed echo signal is subjected to quadrature demodulation to extract the in-phase and quadrature components. A complex analytic signal is then constructed based on these components. The quadrature demodulation is implemented using digital quadrature demodulation. The compressed echo signal is multiplied by a pair of quadrature reference signals with frequencies equal to the ultrasonic device's transmission frequency. These reference signals are a cosine reference signal and a sine reference signal. The cosine reference signal multiplied by the compressed echo signal is then low-pass filtered to obtain the in-phase component, and the sine reference signal multiplied by the compressed echo signal is then low-pass filtered to obtain the quadrature component. The in-phase component is used as the real part of the complex number, and the quadrature component as the imaginary part, to construct the complex analytic signal. Each sampling point in the complex analytic signal is a complex value. The modulus of the complex value represents the envelope amplitude of the echo signal at that sampling point, and the argument of the complex value represents the instantaneous phase of the echo signal at that sampling point.

[0058] Based on the transmission frequency and receiving gain in the imaging parameters, the time delay compensation coefficient between multiple scan lines is determined. This time delay compensation coefficient is then used to perform time delay calibration on the complex analytical signal line by line, yielding the time-delay calibrated signal. The method for determining the time delay compensation coefficient is as follows: The wavelength of the corresponding ultrasound wave in human tissue is calculated based on the transmission frequency. The wavelength is equal to the average propagation speed of ultrasound in human soft tissue, 1540 meters per second, divided by the transmission frequency. Based on the geometry of the ultrasound probe and the deflection angle of the scan lines, the acoustic path difference between each array element on each scan line and the reference array element is calculated. The acoustic path difference is divided by the wavelength and then by the ultrasound propagation speed to obtain the time delay compensation coefficient, which is a time offset in seconds. After determining the time delay compensation coefficient, the complex analytic signal is subjected to time delay calibration processing line by line. The time delay calibration processing is implemented by a digital time delay filter. The digital time delay filter adopts the fractional time delay filtering method. The finite impulse response filter designed by the window function method applies a fractional delay corresponding to the corresponding time delay compensation coefficient to each sampling point of the complex analytic signal. After the time delay calibration processing, the echo signals of each channel are aligned on the time axis to obtain the time delay calibrated signal.

[0059] The time-delay calibrated signals are weighted and summed to complete beamforming processing, yielding the initial ultrasound image. During the weighted summation process, the time-delay calibrated signal on each scan line is assigned a weighting coefficient. The weighting window function formed by these coefficients uses a Hamming window, and the formula for calculating the Hamming window coefficients is as follows:

[0060] in, This represents the weighting coefficient corresponding to the nth scan line, where n is the scan line index, and the value of n ranges from 0 to... An integer, N, representing the total number of scan lines. Pi is a constant. The coefficients of the Hamming window are set to achieve a balance between main lobe width and side lobe suppression, effectively reducing artifact interference introduced by side lobes in beamforming. The signal after time delay calibration for each scan line is compared with the corresponding weighting coefficients. Multiply the signals and then sum them point by point according to their corresponding spatial positions to obtain the gray value of each pixel in the initial ultrasound image. The gray values ​​of all pixels constitute the initial ultrasound image.

[0061] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for automatic assessment and grading of ultrasound image quality based on deep learning, characterized in that, include: The imaging parameters of the ultrasound equipment and the raw ultrasound echo signal collected by the ultrasound probe are obtained, and the raw ultrasound echo signal is processed by beamforming to obtain an initial ultrasound image. The initial ultrasound image is input into a deep residual shrinkage network, and the soft thresholding module in the deep residual shrinkage network is used to perform noise suppression processing on the multi-scale feature channels of the initial ultrasound image, and output a noise-reduced feature map. Spatial attention weighted mapping is performed on the denoised feature map to obtain a weighted feature map. The global quality deviation index of the initial ultrasound image is determined by measuring the structural similarity between the weighted feature map and the historical high-quality ultrasound image template. Based on the global quality deviation index, the adversarial generative network is invoked to perform pixel-by-pixel reconstruction error calculation on the local texture region of the initial ultrasound image, resulting in a local distortion heatmap; By integrating the global quality deviation index and the local distortion heatmap, a comprehensive evaluation tensor for ultrasound image quality is constructed. A reinforcement learning agent is then used to make a grade decision on the comprehensive evaluation tensor, and the quality grade label of the initial ultrasound image is output.

2. The method for automatic assessment and grading of ultrasound image quality based on deep learning according to claim 1, characterized in that, The process involves inputting the initial ultrasound image into a deep residual shrinking network, using a soft thresholding module within the deep residual shrinking network to perform noise suppression processing on the multi-scale feature channels of the initial ultrasound image, and outputting a denoised feature map, including: The initial ultrasound image is subjected to layer-by-layer convolution operation using multiple convolutional layers in the deep residual shrinkage network to extract multi-scale feature channels of the initial ultrasound image under different receptive fields. The multi-scale feature channels are input into the global average pooling layer in the deep residual shrinkage network to calculate the global statistics for each feature channel. The global statistics are input into the fully connected layer and activation function layer in the deep residual shrinkage network to generate adaptive scaling parameters for each feature channel; The corresponding feature channels are scaled using the adaptive scaling parameters to obtain scaled feature channels; The scaled feature channels are passed through the threshold calculation sub-network in the soft thresholding module to generate an independent threshold for each feature channel; The scaled feature channels are soft-thresholded using the independent threshold, setting feature values ​​with absolute values ​​less than the independent threshold to zero and shrinking feature values ​​with absolute values ​​greater than the independent threshold toward zero to obtain the denoised feature map.

3. The method for automatic assessment and grading of ultrasound image quality based on deep learning according to claim 2, characterized in that, The process involves performing spatial attention-weighted mapping on the denoised feature map to obtain a weighted feature map. Then, by measuring the structural similarity between the weighted feature map and historical high-quality ultrasound image templates, the global quality deviation index of the initial ultrasound image is determined, including: The denoised feature map is input into the max pooling branch and average pooling branch in the spatial attention module to obtain the first spatial attention map and the second spatial attention map. After feature concatenation of the first spatial attention map and the second spatial attention map, they are fused through a convolutional layer with a preset kernel size to generate a spatial attention weight map. The spatial attention weight map is multiplied element-wise with the denoised feature map to obtain the weighted feature map; Extract the historical high-quality ultrasound image templates that match the current ultrasound equipment model and the current scanning area from the database, and perform feature extraction on the historical high-quality ultrasound image templates to obtain template feature maps; Calculate the structural similarity metric between the weighted feature map and the template feature map, and determine the global quality deviation index based on the statistical distribution of the structural similarity metric across all spatial locations.

4. The method for automatic assessment and grading of ultrasound image quality based on deep learning according to claim 3, characterized in that, The step of calling an adversarial generative network to perform pixel-by-pixel reconstruction error calculation on the local texture region of the initial ultrasound image based on the global quality deviation index, to obtain a local distortion heatmap, includes: When the global quality deviation index exceeds a preset quality threshold, the initial ultrasound image is input into the generator in the adversarial generative network. The generator uses an encoder-decoder structure to perform feature encoding and reconstruction processing on the initial ultrasound image and outputs a reconstructed ultrasound image. The initial ultrasound image and the reconstructed ultrasound image are simultaneously input into the discriminator in the generative adversarial network, and the discriminator outputs a pixel-level discriminant difference map between the initial ultrasound image and the reconstructed ultrasound image. Based on the discriminant difference map, the reconstruction error value of each pixel position in the initial ultrasound image is calculated; The initial ultrasound image is divided into multiple local texture regions. The reconstruction error values ​​of all pixels in each local texture region are aggregated and statistically analyzed to obtain the region reconstruction error of each local texture region. The region reconstruction errors of all local texture regions are arranged according to the spatial layout of the initial ultrasound image to generate the local distortion heatmap.

5. The method for automatic assessment and grading of ultrasound image quality based on deep learning according to claim 4, characterized in that, The process involves fusing the global quality deviation index and the local distortion heatmap to construct a comprehensive ultrasound image quality assessment tensor. A reinforcement learning agent is then used to make a grade decision on this comprehensive assessment tensor, outputting the initial ultrasound image quality grade label, including: The global quality deviation index is replicated and expanded into a global deviation map with the same spatial size as the local distortion heatmap; The global deviation map and the local distortion heatmap are concatenated along the channel dimension to obtain a fused feature map. Perform a 3D convolution operation on the fused feature map to extract the joint features of the fused feature map in the spatial dimension and the channel dimension, and generate the comprehensive evaluation tensor; The comprehensive evaluation tensor is input into the policy network of the reinforcement learning agent, and the policy network outputs a decision probability distribution corresponding to multiple quality levels; Based on the decision probability distribution, the quality level with the highest probability value is selected as the quality grading label.

6. The method for automatic assessment and grading of ultrasound image quality based on deep learning according to claim 5, characterized in that, The process of inputting the comprehensive evaluation tensor into the policy network of the reinforcement learning agent, wherein the policy network outputs a decision probability distribution corresponding to multiple quality levels, includes: The comprehensive evaluation tensor is subjected to sequence feature extraction along the time dimension using multiple temporal convolutional layers in the policy network to obtain a temporal feature sequence. The temporal feature sequence is input into the long short-term memory network unit in the policy network to update the internal hidden state of the long short-term memory network unit and obtain the decision hidden vector at the current time. The decision latent vector is input into the fully connected classification layer in the policy network, and the fully connected classification layer outputs the decision probability distribution through a flexible maximum activation function. During the training phase of the reinforcement learning agent, the network parameters of the policy network are updated according to the advantage function between the decision probability distribution and the quality score reward of the environmental feedback.

7. The method for automatic assessment and grading of ultrasound image quality based on deep learning according to claim 1, characterized in that, The process of acquiring the imaging parameters of the ultrasound equipment and the raw ultrasound echo signal collected by the ultrasound probe, and performing beamforming processing on the raw ultrasound echo signal to obtain an initial ultrasound image includes: The transmission frequency, receiving gain, and dynamic range of the ultrasound device are obtained as the imaging parameters, and the original radio frequency echo signals collected by the ultrasound probe on multiple scan lines are obtained as the original ultrasound echo signals. Based on the dynamic range in the imaging parameters, the original radio frequency echo signal is subjected to dynamic range compression processing to obtain the compressed echo signal. The compressed echo signal is subjected to quadrature demodulation to extract the in-phase component and the quadrature component, and a complex analytical signal is constructed based on the in-phase component and the quadrature component. Based on the transmission frequency and receiving gain in the imaging parameters, the time delay compensation coefficient between the plurality of scan lines is determined, and the time delay compensation coefficient is used to perform time delay calibration processing on the complex analytical signal line by line to obtain the time delay calibrated signal; The time-delay calibrated signal is weighted and summed to complete the beamforming process and obtain the initial ultrasound image.

8. The method for automatic assessment and grading of ultrasound image quality based on deep learning according to claim 1, characterized in that, The soft thresholding module in the deep residual shrinkage network is optimized during the training phase in the following ways: A training set containing multiple ultrasound image samples with labeled noise levels is constructed. The ultrasound image samples in the training set are input into the deep residual shrinkage network to obtain the predicted denoised feature map. The first loss function value is calculated using the mean square error between the predicted denoised feature map and the clean ultrasound image label corresponding to the ultrasound image sample. The second loss function value is calculated by utilizing the regularization constraint between the independent threshold output by the soft thresholding module in the deep residual shrinkage network and the noise level label of the ultrasound image sample; The first loss function value and the second loss function value are weighted and summed to obtain the total loss function value. The total loss function value is then used to update the network weights of the deep residual shrinkage network through the backpropagation algorithm. After each backpropagation update, the scaling factor of the batch normalization layer in the deep residual shrinkage network is pruned, and scaling factors smaller than a preset pruning threshold are set to zero.

9. The method for automatic assessment and grading of ultrasound image quality based on deep learning according to claim 4, characterized in that, The encoder in the generator employs multiple downsampled convolutional blocks, the decoder employs a symmetrical number of upsampled deconvolutional blocks, and a skip connection is provided between the encoder and the decoder.

10. A deep learning-based automatic ultrasound image quality assessment and grading system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the deep learning-based automatic assessment and grading method for ultrasound image quality as described in any one of claims 1 to 9.

Images