Ultrasonic image processing method and ultrasonic image processing device

Abstract

The invention discloses an ultrasonic image processing method and an ultrasonic image processing device, and belongs to the field of medical image diagnos.The processing method comprises the following specific steps that P1, high-frequency and low-frequency ultrasonic waves are emitted through a dynamic frequency modulation probe, a composite ultrasonic image is constructed according to feedback ultrasonic signals, and then noise in the image is recognized and separated; according to the method, the final image can be ensured to have high resolution, low noise and high focus visibility, the diagnosis accuracy is improved, accurate focus identification and classification are realized, the reliability of clinical decision is improved, the image fusion quality is improved, and richer diagnosis information can be provided; the imaging quality of an operator is improved, the reduction of the image quality caused by manual operation errors is reduced, the stability and the repeatability of ultrasonic examination are improved, the method can adapt to different patients and different examination scenes, and the stability and the applicability of ultrasonic imaging are improved.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging diagnosis, and particularly to an ultrasonic image processing method and an ultrasonic image processing device. Background Art

[0002] Ultrasonic imaging technology has been widely used in the fields of medical imaging diagnosis, interventional therapy, industrial non-destructive testing, etc. due to its advantages such as non-invasiveness, real-time performance, and low cost. It plays an important role especially in clinical examinations of cardiovascular, abdominal, musculoskeletal system, breast, thyroid, etc. However, traditional ultrasonic imaging still faces many challenges, such as limited imaging resolution, insufficient penetration ability, interference of noise and artifacts, dependence on manual experience for lesion recognition, and limited single-modal image information. These problems seriously affect the diagnostic accuracy and automation level of ultrasonic images. In response to these challenges, an ultrasonic image processing method and an ultrasonic image processing device are proposed in this paper.

[0003] After retrieval, Chinese Patent No. CN119559080A discloses an ultrasonic image processing method and an ultrasonic image processing device. Although this invention can effectively enhance the image clarity of ultrasonic images, it cannot ensure that the final image has high resolution, low noise, and high lesion visibility, reducing the diagnostic accuracy, and cannot provide richer diagnostic information. In addition, the existing ultrasonic image processing methods and ultrasonic image processing devices are prone to image quality degradation caused by human operation errors, reducing the imaging quality of operators, with poor stability and repeatability of ultrasonic examinations, and unable to adapt to different patients and different examination scenarios. Therefore, an ultrasonic image processing method and an ultrasonic image processing device are proposed. Summary of the Invention

[0004] The purpose of the present invention is to solve the defects existing in the prior art, and to propose an ultrasonic image processing method and an ultrasonic image processing device.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions:

[0006] An ultrasonic image processing method, and the specific steps of the processing method are as follows:

[0007] P1. By dynamically modulating the probe frequency, high-frequency and low-frequency ultrasonic waves are emitted, and based on the feedback ultrasonic signals, a composite ultrasonic image is constructed, and then the noise in the image is identified and separated;

[0008] P2. Based on the processed composite ultrasonic image, a blood flow-hardness correlation model is established, and abnormal regions are identified, and at the same time, risk grading is performed on each abnormal region;

[0009] P3. The lesion information of the ultrasonic data is segmented, and the scanning trajectory and pressure of the operator are monitored in real time, and at the same time, a best scanning path prompt is dynamically generated;

[0010] P4. Evaluate the real-time constructed ultrasound images, provide immediate optimization suggestions, and then fuse the ultrasound images with the patient's CT or MRI images to generate a multi-modal fusion heat map;

[0011] P5. Based on clinical guidelines and historical case libraries, establish a dynamic diagnostic decision tree, and generate a corresponding list of differential diagnoses for the patient according to the multi-modal fusion heat map.

[0012] As a further solution of the present invention, the specific steps of constructing a composite ultrasound image as described in P1 and then identifying and separating the noise in the image are as follows:

[0013] S1.1: Synchronously transmit high-frequency and low-frequency ultrasonic waves through a dynamically frequency-modulated probe, and then the dynamically frequency-modulated probe receives the high-frequency and low-frequency ultrasonic signals reflected from the tissue, and converts the high-frequency and low-frequency ultrasonic echo signals into corresponding high-frequency ultrasonic images and low-frequency ultrasonic images respectively;

[0014] S1.2: According to the information correlation between the high-frequency ultrasonic image and the low-frequency ultrasonic image, spatially align the two sets of ultrasonic images through image registration technology, and then use the Fourier transform to convert the high-frequency ultrasonic image and the low-frequency ultrasonic image from the spatial domain to the frequency domain;

[0015] S1.3: Calculate the adaptive fusion weights of each ultrasonic image according to the information contribution ratios of different frequency bands of the high-frequency ultrasonic image and the low-frequency ultrasonic image, perform frequency-domain fusion based on the weights of the high-frequency ultrasonic image and the low-frequency ultrasonic image, and then convert the fused composite ultrasonic image back to the spatial domain through the inverse Fourier transform, and perform edge sharpening processing on the finally generated composite ultrasonic image.

[0016] As a further solution of the present invention, the specific steps of identifying and separating the noise in the image as described in P1 are as follows:

[0017] S2.1: Normalize the composite ultrasonic image, and then input the normalized composite ultrasonic image into the convolutional neural network CNN. The CNN receives the input composite ultrasonic image and performs forward propagation layer by layer on the composite ultrasonic image;

[0018] S2.2: After the convolutional layer in the CNN receives the composite ultrasonic image, by using multiple small-sized convolutional kernels, extract the local edge, texture, and local feature maps of the basic tissue structure of the input composite ultrasonic image, enhance the non-linear expression ability of each local feature map through the ReLU activation function, and then perform normalization processing on the extracted local feature maps;

[0019] S2.3: Input each local feature map generated through multi-layer convolution processing into the depth feature layer. The depth feature layer uses convolution kernels of different sizes in parallel on each local feature map to obtain feature information of different scales, and then fuses the feature information of different scales to generate a fused feature map;

[0020] S2.4: Use the fused feature map output by the depth feature layer to generate an enhanced deep feature map through residual connection. Calculate the noise probability of each pixel point using the deep feature map generated by CNN, and based on the noise probability map, separate the noise component from the input composite ultrasound image. Then reconstruct the denoised composite ultrasound image and use adaptive enhancement to boost the lesion signal.

[0021] As a further solution of the present invention, the specific steps for P2 to identify abnormal regions and perform risk grading on each abnormal region are as follows:

[0022] S3.1: Analyze tissue hardness and tissue hardness variability through shear wave elastography, and then use color Doppler ultrasound and pulsed Doppler to evaluate the blood flow velocity, flow direction, and blood flow pattern in the lesion area, and calculate the blood flow resistance index of the corresponding patient;

[0023] S3.2: Perform normalization processing on tissue hardness and blood flow velocity respectively, and calculate the blood flow-hardness correlation index through the normalized tissue hardness and blood flow velocity. Construct a blood flow-hardness spatial mapping diagram based on the composite ultrasound image, and then set the abnormal threshold according to statistical analysis and historical clinical data;

[0024] S3.3: Perform pixel-level scanning on the composite ultrasound image, identify the coordinates of the regions higher than the abnormal threshold, and at the same time mark the identified abnormal regions. Then calculate the risk score of the abnormal regions according to the blood flow-hardness correlation index, blood flow resistance index, and tissue hardness variability, and mark the abnormal regions as low risk, medium risk, or high risk according to the preset range standard.

[0025] An ultrasound image processing device includes a dynamic frequency modulation probe, a signal fusion module, a denoising and enhancement module, a tissue imaging analysis module, a hemodynamic analysis module, a lesion segmentation module, a three-dimensional reconstruction module, an image real-time optimization module, an image fusion module, and a diagnostic decision-making module;

[0026] The dynamic frequency modulation probe is used to synchronously transmit high-frequency and low-frequency ultrasonic waves during scanning and receive the echo signals of high-frequency and low-frequency ultrasonic waves;

[0027] The signal fusion module is used to fuse the received echo signals of high-frequency and low-frequency ultrasonic waves and establish the corresponding composite ultrasonic image;

[0028] The denoising and enhancement module is used to remove speckle noise and artifacts in the compound ultrasonic image and enhance the echo signal of tiny lesions;

[0029] The tissue imaging analysis module is based on the enhanced compound ultrasonic image and analyzes the elastic distribution of tissues in real time;

[0030] The hemodynamic analysis module measures the blood flow velocity and direction in real time based on the analysis results of the tissue imaging analysis module;

[0031] The 3D reconstruction module generates a 3D reconstruction image using the compound ultrasonic image and performs lesion segmentation on the 3D ultrasonic reconstruction data;

[0032] The image real-time optimization module evaluates the imaging quality of the current compound ultrasonic image in real time and provides optimization guidance;

[0033] The image fusion module is used to fuse the collected compound ultrasonic image with CT or MRI images to generate multi-modal images;

[0034] The diagnostic decision-making module analyzes ultrasonic features according to clinical guidelines and the historical case database, generates a differential diagnosis list, and recommends subsequent examination items.

[0035] As a further solution of the present invention, the specific steps for the image real-time optimization module to evaluate the imaging quality of the current compound ultrasonic image in real time and provide optimization guidance are as follows:

[0036] S4.1: Extract the edge sharpness, contrast, and noise level of the current compound ultrasonic image, use the collected data as quality evaluation parameters, and then collect the scanning trajectory of the dynamic frequency modulation probe, including the position and rotation angle of the dynamic frequency modulation probe in the 3D space at different times, and measure the pressure applied by the probe in real time, and calculate its mean and standard deviation;

[0037] S4.2: Establish a corresponding state space according to the quality evaluation parameters, scanning trajectory, and pressure data at different times, establish a corresponding action space based on adjusting the scanning angle, changing the probe pressure, and optimizing the probe moving speed, and establish a corresponding reward function according to the preset pressure range and scanning path;

[0038] S4.3: Set up a Q network and a target network, and initialize the experience pool. At the beginning of the iteration, at each time step, obtain the state corresponding to the current compound ultrasonic image at the current time, and select an action that can be taken in the current state from the action space through the ε-greedy strategy;

[0039] S4.4: Execute the selected action, generate a new state, calculate the reward value of the new state through the reward function, integrate the state, selected action, new state, and the reward value of the new state at the current time step into an experience dataset, store it in the experience pool, and then randomly select multiple groups of experience datasets from the experience pool;

[0040] S4.5: Input multiple groups of experience datasets into the Q-network, calculate and update the target Q-value and the loss value through the Bellman equation, then based on the backpropagation algorithm, import the loss value into the Q-network and update the Q-network parameters, record the training rounds. When the preset training rounds are reached, update the target network parameters based on the Q-network parameters of the current round;

[0041] S4.6: After each round of training, reduce the ε value. When the ε value converges within the preset threshold, stop training, and then continuously monitor the operator's scanning state in real time, including quality assessment parameters, scanning trajectory, and pressure data, to form the current scanning state;

[0042] S4.7: Input the current scanning state into the trained Q-network, calculate the Q-values of all actions in the action space, select the action with the largest Q-value as the optimal adjustment strategy, and provide real-time feedback to the operator according to the selection result, indicating the best scanning method, including arrow tips, pressure feedback, and image quality scores. After each new scanning adjustment is executed, update the scanning state and recalculate the optimal adjustment strategy.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] 1. This ultrasonic image processing method synchronously transmits and receives high-frequency and low-frequency ultrasonic signals through a dynamically frequency-modulated probe, generates high- and low-frequency ultrasonic images respectively, aligns them through image registration technology, converts the images to the frequency domain using Fourier transform, calculates the adaptive fusion weight based on the information contribution ratio, and performs frequency-domain fusion. Subsequently, the fused ultrasonic image is restored to the spatial domain through inverse Fourier transform, and edge sharpening and normalization processing are carried out. The fused ultrasonic image is input into a CNN. The local features are extracted through the CNN convolutional layer, the multi-scale information is fused in the deep feature layer, and enhanced features are generated through residual connections. The noise probability is calculated, the noise is separated and the denoised image is reconstructed. At the same time, the lesion signal is adaptively enhanced. The tissue hardness is analyzed through shear wave elastography, the hemodynamic characteristics are evaluated by combining color and pulsed Doppler ultrasound, and the blood flow-hardness correlation index is calculated. A blood flow-hardness spatial mapping diagram is constructed. By setting the abnormal threshold through statistical analysis, the high-risk area is identified, and the risk score is calculated based on the blood flow-hardness correlation index, blood flow resistance index and tissue hardness variability. Finally, the abnormal area is marked with risk grading, which can ensure that the final image has high resolution, low noise and high lesion visibility, improve the diagnostic accuracy, achieve accurate lesion identification and grading, improve the reliability of clinical decision-making, improve the image fusion quality, and can provide richer diagnostic information.

[0045] 2. This ultrasonic image processing device extracts the edge sharpness, contrast and noise level of the composite ultrasonic image, collects the probe scanning trajectory, spatial position, rotation angle and applied pressure, establishes a state space. At the same time, an action space based on the scanning angle, probe pressure and movement mode is defined, and a reward function is constructed. A deep Q network is adopted, the experience pool is initialized, and the current state is obtained at each time step. The action is selected using the ε-greedy strategy. After execution, a new state is generated and the reward value is calculated. The experience is stored in the experience pool, the target Q value is updated through the Bellman equation, and the Q network parameters are optimized using backpropagation. After the training converges, the trained Q network is used to monitor the scanning state in real time, calculate the optimal adjustment strategy, and provide arrow guidance, pressure feedback and image quality score to the operator, improve the imaging quality of the operator, reduce the image quality degradation caused by human operation errors, improve the stability and repeatability of ultrasonic examination, and can adapt to different patients and different examination scenarios, enhancing the stability and applicability of ultrasonic imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, and do not constitute a limitation to the present invention.

[0047] Figure 1 It is a flowchart of an ultrasonic image processing method proposed by the present invention;

[0048] Figure 2 System block diagram of an ultrasonic image processing device proposed by the present invention. Specific implementation mode

[0049] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0050] Embodiment 1

[0051] Refer to Figure 1 , this embodiment discloses an ultrasonic image processing method, and the specific steps of this processing method are as follows:

[0052] By dynamically frequency modulating the probe, high-frequency and low-frequency ultrasonic waves are emitted, and based on the feedback ultrasonic signals, a composite ultrasonic image is constructed, and then the noise in the separated image is identified.

[0053] Specifically, high-frequency and low-frequency ultrasonic waves are synchronously emitted by the dynamically frequency modulated probe. Then, the dynamically frequency modulated probe receives the high-frequency and low-frequency ultrasonic signals reflected from the tissue, and converts the high-frequency and low-frequency ultrasonic echo signals into corresponding high-frequency ultrasonic images and low-frequency ultrasonic images respectively. According to the information correlation between the high-frequency ultrasonic image and the low-frequency ultrasonic image, the two sets of ultrasonic images are spatially aligned through image registration technology. Then, the high-frequency ultrasonic image and the low-frequency ultrasonic image are converted from the spatial domain to the frequency domain by Fourier transform. Based on the information contribution ratio of different frequency bands of the high-frequency ultrasonic image and the low-frequency ultrasonic image, the adaptive fusion weight of each ultrasonic image is calculated, and frequency domain fusion is performed based on the weights of the high-frequency ultrasonic image and the low-frequency ultrasonic image. Then, the fused composite ultrasonic image is converted back to the spatial domain by inverse Fourier transform, and edge sharpening processing is performed on the finally generated composite ultrasonic image.

[0054] Specifically, normalize the composite ultrasound image, and then input the normalized composite ultrasound image into a convolutional neural network (CNN). The CNN receives the input composite ultrasound image and performs forward propagation layer by layer on the composite ultrasound image. After the convolutional layer in the CNN receives the composite ultrasound image, it extracts local edge, texture, and local feature maps of the basic tissue structure of the input composite ultrasound image by using multiple small-sized convolutional kernels, and enhances the non-linear expression ability of each local feature map through the ReLU activation function. Then, perform normalization processing on the extracted local feature maps. Input the local feature maps generated through multi-layer convolutional processing into the deep feature layer. The deep feature layer uses convolutional kernels of different sizes in parallel on each local feature map to obtain feature information of different scales. Then, fuse the feature information of different scales to generate a fused feature map. Connect the fused feature map output by the deep feature layer through a residual connection to generate an enhanced deep feature map. Calculate the noise probability of each pixel point using the deep feature map generated by the CNN, and based on the noise probability map, separate the noise component from the input composite ultrasound image. Then, reconstruct the denoised composite ultrasound image and use adaptive enhancement to boost the lesion signal.

[0055] Based on the processed composite ultrasound image, establish a blood flow-hardness correlation model, identify abnormal regions, and simultaneously perform risk grading on each abnormal region.

[0056] Specifically, analyze tissue hardness and tissue hardness variability through shear wave elastography, and then use color Doppler ultrasound and pulsed Doppler to evaluate the blood flow velocity, flow direction, and blood flow pattern in the lesion area, and calculate the blood flow resistance index of the corresponding patient. Normalize the tissue hardness and blood flow velocity respectively, and calculate the blood flow-hardness correlation index based on the normalized tissue hardness and blood flow velocity. Construct a blood flow-hardness spatial mapping diagram based on the composite ultrasound image, and then set an abnormal threshold according to statistical analysis and historical clinical data. Perform pixel-level scanning on the composite ultrasound image, identify the coordinates of the areas higher than the abnormal threshold, and at the same time mark the identified abnormal areas. Then, calculate the risk score of the abnormal area based on the blood flow-hardness correlation index, blood flow resistance index, and tissue hardness variability, and mark the abnormal area as low risk, medium risk, or high risk according to the preset range standard.

[0057] Segment the lesion information of the ultrasound data, and real-time monitor the operator's scanning trajectory and pressure, and simultaneously dynamically generate a best scanning path prompt.

[0058] Evaluate the real-time constructed ultrasound image, provide immediate optimization suggestions, and then fuse the ultrasound image with the patient's CT or MRI images to generate a multi-modal fusion heat map.

[0059] Based on clinical guidelines and historical case databases, a dynamic diagnostic decision tree is established, and a corresponding list of differential diagnoses for patients is generated according to the multi-modal fusion heat map.

[0060] Example 2

[0061] Refer to Figure 2 , this example discloses an ultrasonic image processing device, including a dynamic frequency modulation probe, a signal fusion module, a denoising and enhancement module, a tissue imaging analysis module, a hemodynamic analysis module, a lesion segmentation module, a three-dimensional reconstruction module, an image real-time optimization module, an image fusion module, and a diagnostic decision module;

[0062] The dynamic frequency modulation probe is used to synchronously transmit high-frequency and low-frequency ultrasonic waves during scanning and receive the echo signals of high-frequency and low-frequency ultrasonic waves; the signal fusion module is used to fuse the received echo signals of high-frequency and low-frequency ultrasonic waves to establish a corresponding composite ultrasonic image; the denoising and enhancement module is used to remove speckle noise and artifacts in the composite ultrasonic image and enhance the echo signals of small lesions.

[0063] The tissue imaging analysis module is based on the enhanced composite ultrasonic image to analyze the elastic distribution of tissues in real time; the hemodynamic analysis module is based on the analysis results of the tissue imaging analysis module to measure the blood flow velocity and direction in real time; the three-dimensional reconstruction module uses the composite ultrasonic image to generate a three-dimensional reconstruction image and performs lesion segmentation on the three-dimensional ultrasonic reconstruction data.

[0064] The image real-time optimization module evaluates the imaging quality of the current composite ultrasonic image in real time and provides optimization guidance.

[0065] Specifically, extract the edge sharpness, contrast, and noise level of the current compound ultrasonic image, and use the collected data as quality assessment parameters. Then collect the scanning trajectory of the dynamic frequency modulation probe, including the position and rotation angle of the dynamic frequency modulation probe in 3D space at different times, and measure the pressure applied by the probe in real time. At the same time, calculate its mean and standard deviation. Based on the quality assessment parameters, scanning trajectory, and pressure data at different times, establish the corresponding state space. Based on adjusting the scanning angle, changing the probe pressure, and optimizing the probe movement speed, establish the corresponding action space. And based on the preset pressure range and scanning path, establish the corresponding reward function. Set up the Q-network and the target network, and initialize the experience pool. At the beginning of each iteration, at each time step, obtain the state corresponding to the compound ultrasonic image at the current time, and select an action that can be taken in the current state from the action space through the ε-greedy strategy. Execute the selected action and generate a new state. At the same time, calculate the reward value of the new state through the reward function, and integrate the state, selected action, new state, and new state reward value at the current time step into an experience data set and store it in the experience pool. Then randomly select multiple groups of experience data sets from the experience pool, input the multiple groups of experience data sets into the Q-network, calculate and update the target Q value and the loss value through the Bellman equation. Then, based on the backpropagation algorithm, import the loss value into the Q-network and update the Q-network parameters. Record the number of training rounds. When the preset number of training rounds is reached, update the target network parameters based on the Q-network parameters of the current round. After each round of training, update the ε value. When the ε value converges within the preset threshold, stop training. Then, in real time, monitor the operator's scanning state, including quality assessment parameters, scanning trajectory, and pressure data, to form the current scanning state. Input the current scanning state into the trained Q-network, calculate the Q values of all actions in the action space, and select the action with the largest Q value as the best adjustment strategy. And according to the selection result, provide real-time feedback to the operator, indicating the best scanning method, including arrow tips, pressure feedback, and image quality scores. Then, after each new scan adjustment is executed, update the scanning state and recalculate the optimal adjustment strategy.

[0066] The image fusion module is used to fuse the collected compound ultrasonic image with CT or MRI images to generate multi-modal images; the diagnostic decision module analyzes the ultrasonic features based on clinical guidelines and the historical case database, generates a differential diagnosis list, and recommends subsequent examination items.

Claims

1. An ultrasonic image processing method, characterized in that, The specific steps of this processing method are as follows: P1. Transmit high-frequency and low-frequency ultrasonic waves through a dynamic frequency modulation probe, construct a composite ultrasonic image based on the feedback ultrasonic signal, and then identify and separate the noise in the image; P2. Establish a blood flow-hardness correlation model based on the processed composite ultrasonic image, identify abnormal regions, and simultaneously perform risk grading on each abnormal region; P3. Segment the lesion information of the ultrasonic data, monitor the operator's scanning trajectory and pressure in real time, and dynamically generate a prompt for the optimal scanning path; P4. Evaluate the ultrasonographic image constructed in real time, provide immediate optimization suggestions, and then fuse the ultrasonographic image with the patient's CT or MRI images to generate a multimodal fusion thermogram; P5. Establish a dynamic diagnostic decision tree based on clinical guidelines and the historical case database, and generate a corresponding list of differential diagnoses for the patient according to the multimodal fusion thermogram.

2. The ultrasonic image processing method according to claim 1, characterized in that The specific steps of constructing the composite ultrasonic image and then identifying and separating the noise in the image described in P1 are as follows: S1.1: Synchronously transmit high-frequency and low-frequency ultrasonic waves through a dynamic frequency modulation probe. Then, the dynamic frequency modulation probe receives the high-frequency and low-frequency ultrasonic signals reflected from the tissue, and converts the high-frequency and low-frequency ultrasonic echo signals into corresponding high-frequency ultrasonic images and low-frequency ultrasonic images respectively; S1.2: According to the information correlation between the high-frequency ultrasonic image and the low-frequency ultrasonic image, spatially align the two sets of ultrasonic images through image registration technology, and then use the Fourier transform to convert the high-frequency ultrasonic image and the low-frequency ultrasonic image from the spatial domain to the frequency domain; S1.3: Calculate the adaptive fusion weights of each ultrasonic image based on the information contribution ratios of different frequency bands of the high-frequency ultrasonic image and the low-frequency ultrasonic image, perform frequency domain fusion based on the weights of the high-frequency ultrasonic image and the low-frequency ultrasonic image, then convert the fused composite ultrasonic image back to the spatial domain through the inverse Fourier transform, and perform edge sharpening processing on the finally generated composite ultrasonic image.

3. The ultrasonic image processing method according to claim 2, wherein The specific steps of identifying and separating the noise in the image described in P1 are as follows: S2.1: Normalize the composite ultrasonic image, and then input the normalized composite ultrasonic image into the convolutional neural network CNN. The CNN receives the input composite ultrasonic image and performs forward propagation layer by layer on the composite ultrasonic image; S2.2: After the convolutional layer in the CNN receives the composite ultrasonic image, by using multiple small-sized convolutional kernels, extract the local edge, texture, and local feature maps of the basic tissue structure of the input composite ultrasonic image, enhance the non-linear expression ability of each local feature map through the ReLU activation function, and then perform normalization processing on the extracted local feature maps; S2.3: Input the local feature maps generated by multi-layer convolutional processing into the deep feature layer. The deep feature layer uses convolutional kernels of different sizes in parallel on each local feature map to obtain feature information of different scales, and then fuse the feature information of different scales to generate a fused feature map; S2.4: Generate an enhanced deep feature map from the fused feature map output by the deep feature layer through residual connection. Calculate the noise probability of each pixel point using the deep feature map generated by CNN, and based on the noise probability map, separate the noise component from the input composite ultrasound image. Then reconstruct the denoised composite ultrasound image and use adaptive enhancement to boost the lesion signal.

4. An ultrasonic image processing method according to claim 3, characterized in that, The specific steps for identifying the abnormal regions and grading the risks of each abnormal region as described in P2 are as follows: S3.1: Analyze tissue hardness and tissue hardness variability through shear wave elastography, and then use color Doppler ultrasound and pulsed Doppler to evaluate the blood flow velocity, flow direction, and blood flow pattern in the lesion area, and calculate the blood flow resistance index of the corresponding patient; S3.2: Normalize tissue hardness and blood flow velocity respectively, and calculate the blood flow-hardness correlation index using the normalized tissue hardness and blood flow velocity. Construct a blood flow-hardness spatial mapping diagram based on the composite ultrasound image, and then set the abnormal threshold according to statistical analysis and historical clinical data; S3.3: Perform pixel-level scanning on the composite ultrasound image, identify the coordinates of the regions higher than the abnormal threshold, and mark the identified abnormal regions. Then calculate the risk score of the abnormal regions according to the blood flow-hardness correlation index, blood flow resistance index, and tissue hardness variability, and mark the abnormal regions as low risk, medium risk, or high risk according to the preset range criteria.

5. An ultrasonic image processing device for implementing the ultrasonic image processing method according to any one of claims 1-4, characterized in that, It includes a dynamic frequency modulation probe, a signal fusion module, a denoising and enhancement module, a tissue imaging analysis module, a blood flow dynamics analysis module, a three-dimensional reconstruction module, an image real-time optimization module, an image fusion module, and a diagnostic decision module; The dynamic frequency modulation probe is used to synchronously transmit high-frequency and low-frequency ultrasonic waves during scanning and receive the echo signals of high-frequency and low-frequency ultrasonic waves; The signal fusion module is used to fuse the received echo signals of high-frequency and low-frequency ultrasonic waves and establish the corresponding composite ultrasonic image; The denoising and enhancement module is used to remove speckle noise and artifacts in the composite ultrasonic image and enhance the echo signals of small lesions; The tissue imaging analysis module is used to analyze the elastic distribution of tissues in real time based on the enhanced composite ultrasonic image; The blood flow dynamics analysis module is used to measure the blood flow velocity and direction in real time based on the analysis results of the tissue imaging analysis module; The three-dimensional reconstruction module is used to generate a three-dimensional reconstruction image using the composite ultrasonic image and segment the lesions from the three-dimensional ultrasound reconstruction data; The image real-time optimization module is used to evaluate the imaging quality of the current composite ultrasonic image in real time and provide optimization guidance; The image fusion module is used to fuse the acquired composite ultrasonic image with CT or MRI images to generate multi-modal images; The diagnostic decision module analyzes the ultrasonic features based on clinical guidelines and historical case databases, generates a differential diagnosis list, and recommends subsequent examination items.

6. An ultrasonic image processing apparatus according to claim 5, characterized in that, The specific steps for the image real-time optimization module to evaluate the imaging quality of the current composite ultrasonic image in real time and provide optimization guidance are as follows: S4.1: Extract the edge sharpness, contrast, and noise level of the current compound ultrasonic image. Take the collected data as quality assessment parameters. Then collect the scanning trajectory of the dynamic frequency modulation probe, including the position and rotation angle of the dynamic frequency modulation probe in 3D space at different times, and measure the pressure applied by the probe in real time. At the same time, calculate its mean and standard deviation; S4.2: Establish a corresponding state space based on the quality assessment parameters, scanning trajectory, and pressure data at different times. Establish a corresponding action space based on adjusting the scanning angle, changing the probe pressure, and optimizing the probe movement speed. And establish a corresponding reward function according to the preset pressure range and scanning path; S4.3: Set up the Q-network and the target network, and initialize the experience pool. At the beginning of each iteration, at each time step, obtain the state corresponding to the compound ultrasonic image at the current time, and select an action that can be taken in the current state from the action space through the ε-greedy strategy; S4.4: Execute the selected action and generate a new state. At the same time, calculate the reward value of the new state through the reward function, and integrate the state, selected action, new state, and new state reward value at the current time step into an experience data set, and store it in the experience pool. Then randomly select multiple groups of experience data sets from the experience pool; S4.5: Input multiple groups of experience data sets into the Q-network, calculate and update the target Q value and the loss value through the Bellman equation. Then, based on the backpropagation algorithm, import the loss value into the Q-network and update the Q-network parameters. Record the training rounds. When the preset training rounds are reached, update the target network parameters based on the Q-network parameters of the current round; S4.6: After each round of training, update the ε value. When the ε value converges within the preset threshold, stop training. Then, monitor the operator's scanning state in real time, including quality assessment parameters, scanning trajectory, and pressure data, to form the current scanning state; S4.7: Input the current scanning state into the trained Q-network, calculate the Q values of all actions in the action space, and select the action with the largest Q value as the best adjustment strategy. And provide real-time feedback to the operator according to the selection result, indicating the best scanning method, including arrow tips, pressure feedback, and image quality scores. Then, after each new scan adjustment, update the scanning state and recalculate the optimal adjustment strategy.

Images