Strain elasticity ultrasound image synthesis system and method

Abstract

The application discloses a kind of strain elasticity ultrasonic image synthesis systems and methods, the system includes: generating network module and global discrimination module, wherein, generating network module is used to obtain gray scale ultrasonic image, and gray scale ultrasonic image is handled, and synthesis strain elasticity image is obtained;Global discrimination module is connected with generating network module, and global discrimination module is used to obtain real strain elasticity image, and real strain elasticity image and synthesis strain elasticity image are discriminated to be true or false, so as to obtain prior information, and synthesis quality of synthesis strain elasticity image is improved according to prior information.The application not only can solve the problem of deep tumor location elastic imaging ultrasonic signal attenuation, but also can make the palm ultrasonic equipment lacking elasticity imaging function have strain elasticity imaging function, and has clinical gain for breast tumor benign and malignant detection.

Description

Technical Field

[0001] This invention relates to the field of computer-aided detection technology, and in particular to a strain elastic ultrasonic image synthesis system and method. Background Technology

[0002] Breast cancer, a malignant tumor, accounts for 30% of all malignant tumors in women, and its incidence is rising year by year. Breast cancer screening plays an important role in reducing the incidence and mortality of breast cancer. Clinically, traditional grayscale ultrasound combined with elastography is gradually becoming a standardized examination method. The use of elastography can effectively improve the accuracy of breast cancer screening and diagnosis.

[0003] Although elastography is a very important imaging modality, its application currently faces many challenges. First, it requires highly skilled operators, as different techniques can lead to significant imaging deviations. Second, due to limitations in its imaging principles, ultrasound signal attenuation in deeper tissues can result in inaccurate imaging. Furthermore, the hardware requirements of elastography restrict the development of portable ultrasound equipment and the scope of its applications. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art. Therefore, the first objective of the present invention is to provide a strain-elastic ultrasound image synthesis system that improves the quality of synthesized strain-elastic ultrasound images and facilitates the screening and detection of breast tumors.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0006] A strain elastic ultrasound image synthesis system includes:

[0007] A network module is used to acquire grayscale ultrasound images and process the grayscale ultrasound images to obtain a synthetic strain elastic image.

[0008] A global discrimination module is connected to the generation network module. The global discrimination module is used to acquire the real strain elastic image and to distinguish between the real strain elastic image and the synthesized strain elastic image to obtain prior information and improve the synthesis quality of the synthesized strain elastic image based on the prior information.

[0009] Optionally, the generating network module includes:

[0010] The encoder includes a convolutional input layer and a plurality of first convolutional units, each of which sequentially includes a nonlinear activation layer, a convolutional layer and a batch normalization layer. The encoder is used to encode the grayscale ultrasound image into a feature image.

[0011] A decoder, connected to the encoder, includes several second convolutional units and a convolutional output layer. Each second convolutional unit sequentially includes a nonlinear activation layer, a deconvolutional layer, and a batch normalization layer. The decoder is used to decode the feature image and synthesize the synthetic strain elastic image.

[0012] Optionally, the network structure formed by the encoder and the decoder is a U-shaped symmetrical structure.

[0013] Optionally, the system further includes a tumor discrimination module connected to the generative network module, wherein the tumor discrimination module is used to acquire region of interest images extracted from the grayscale ultrasound image and the synthetic strain elastic image, and to perform differential modeling on the region of interest images.

[0014] Optionally, the system further includes a color equalization module connected to the generating network module. The color equalization module is used to determine the color prior coefficients of the synthesized strain elastic image, and to determine the loss functions of the generating network module, the global discrimination module, and the tumor discrimination module. The loss functions are weighted and constrained by the color prior coefficients to improve the quality of the synthesized strain elastic image.

[0015] Optionally, the global discrimination module includes several third convolutional units and a convolutional output layer, each of the third convolutional units sequentially including a convolutional layer, a batch normalization layer and a nonlinear activation layer.

[0016] To achieve the above objectives, a second aspect of the present invention provides a strain-elastic ultrasound image synthesis method, applied to the aforementioned strain-elastic ultrasound image synthesis system, comprising:

[0017] Acquire grayscale ultrasound images and process the grayscale ultrasound images to obtain a synthetic strain elastic image;

[0018] Acquire real strain elastic images and distinguish between real and synthetic strain elastic images to obtain prior information, and improve the synthesis quality of synthetic strain elastic images based on the prior information.

[0019] Optionally, the step of processing the grayscale ultrasound image to obtain a synthesized strain elastic image includes:

[0020] The grayscale ultrasound image is encoded by an encoder to obtain a feature image. The encoder includes a convolutional input layer and several first convolutional units. Each first convolutional unit includes a nonlinear activation layer, a convolutional layer and a batch normalization layer in sequence.

[0021] The feature image is decoded by a decoder and the synthesized strain elastic image is obtained. The decoder includes several second convolutional units and a convolutional output layer. Each second convolutional unit includes a nonlinear activation layer, a deconvolutional layer and a batch normalization layer in sequence.

[0022] Optionally, the method further includes: acquiring a region of interest image extracted from the grayscale ultrasound image and the synthetic strain elastic image, and performing differential modeling on the region of interest image.

[0023] Optionally, the method further includes: determining the color prior coefficients of the synthesized strain elastic image, and determining the loss functions of the generation network module, the global discrimination module, and the tumor discrimination module, and applying weighted constraints to the loss functions through the color prior coefficients in order to improve the quality of the synthesized strain elastic image.

[0024] This invention has at least the following technical effects:

[0025] This invention models grayscale ultrasound images using a U-shaped network structure for efficient image feature extraction. The design of a global discrimination module and a tumor discrimination module enhances the semantic richness of the synthesized strain-elastic images. Furthermore, a color equalization module determines the color prior coefficients of the synthesized strain-elastic images, and these coefficients are used to weight and constrain the system loss function, improving the color fidelity of the synthesized strain-elastic images. This facilitates improved performance in breast tumor detection using synthesized strain-elastic images. Additionally, the model developed in this invention, as a statistical model, avoids imaging biases caused by different ultrasound physician techniques and ultrasound signal attenuation due to deeper tumor tissue. Moreover, the model, built based on images acquired by high-end ultrasound equipment, provides a gain effect in detecting benign and malignant breast tumors in handheld ultrasound devices. This allows handheld ultrasound devices to achieve strain-elastic imaging functionality without increasing hardware costs, contributing to the widespread application of handheld ultrasound.

[0026] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0027] Figure 1 This is a structural block diagram of the strain elastic ultrasonic image synthesis system provided in the first embodiment of the present invention;

[0028] Figure 2 This is a schematic diagram illustrating the working principle of a strain elastic ultrasonic image synthesis system according to an embodiment of the present invention.

[0029] Figure 3 This is a schematic diagram of the structure of a generative network module provided in an embodiment of the present invention;

[0030] Figure 4 This is a schematic diagram of the structure of a global discrimination module provided in an embodiment of the present invention;

[0031] Figure 5 This is a structural block diagram of the strain elastic ultrasonic image synthesis system provided in the second embodiment of the present invention;

[0032] Figure 6 This is a structural block diagram of the strain elastic ultrasonic image synthesis system provided in the third embodiment of the present invention;

[0033] Figure 7 This is a flowchart of a strain elastic ultrasonic image synthesis method provided in an embodiment of the present invention;

[0034] Figure 8 This is a schematic diagram comparing a synthesized elastic image with a real elastic image on an internal test set provided in an embodiment of the present invention.

[0035] Figure 9 This is a schematic diagram showing the comparison of AUC for benign and malignant breast tumor detection at various centers on a multi-center dataset provided in an embodiment of the present invention;

[0036] Figure 10 This is a schematic diagram comparing the effect of tumor depth on elastography according to an embodiment of the present invention;

[0037] Figure 11 This is a schematic diagram showing the results of testing a high-end ultrasound image training model on a handheld ultrasound device according to an embodiment of the present invention. Detailed Implementation

[0038] The following describes this embodiment in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the invention, and should not be construed as limiting the invention.

[0039] The strain elastic ultrasound image synthesis system and method of this embodiment are described below with reference to the accompanying drawings.

[0040] Figure 1 This is a structural block diagram of a strain-elastic ultrasonic image synthesis system provided in an embodiment of the present invention. Figure 1As shown, the strain-elastic ultrasound image synthesis system 10 includes a generation network module 11 and a global discrimination module 12. The generation network module 11 acquires grayscale ultrasound images and processes them to obtain a synthesized strain-elastic image. The global discrimination module 12 is connected to the generation network module 11. The global discrimination module 12 acquires a real strain-elastic image and distinguishes between the real and synthesized strain-elastic images to obtain prior information and improve the synthesis quality of the synthesized strain-elastic image based on this prior information.

[0041] like Figure 2 As shown, the network generation module 11 in this embodiment is Figure 2 In order to match the generator model structure and reduce interference from irrelevant noise signals, the generator first needs to extract the region of interest (ROI) from the ultrasound image. For retrospective data acquired from high-end ultrasound equipment, the ROI selected by the ultrasound physician can be directly extracted from the corresponding image of the retrospective data to obtain a perfectly paired grayscale ultrasound image and a real strain-elastic image. After the generator obtains the grayscale ultrasound image, the generator network module 11 processes the grayscale ultrasound image to obtain a synthetic strain-elastic image. After the global discrimination module 12 obtains the real strain-elastic image, it can distinguish between the real strain-elastic image and the synthetic strain-elastic image received from the generator network module 11 to obtain prior information for true / false discrimination and train the generator network module 11 to improve its performance. Figure 2 The quality of synthesized elastic images.

[0042] In one embodiment of the present invention, the generation network module 11 includes an encoder and a decoder. The encoder includes a convolutional input layer and several first convolutional units, each first convolutional unit sequentially including a nonlinear activation layer, a convolutional layer, and a batch normalization layer. The encoder is used to encode a grayscale ultrasound image into a feature image. The decoder is connected to the encoder and includes several second convolutional units and a convolutional output layer. Each second convolutional unit sequentially includes a nonlinear activation layer, a deconvolutional layer, and a batch normalization layer. The decoder is used to decode the feature image and synthesize a synthetic strain elastic image. The network structure formed by the encoder and decoder is a U-shaped symmetrical structure.

[0043] In this embodiment, the network generation module 11 adopts a U-shaped network structure, as shown in the figure below. Figure 2 As shown, it comprises two parts: an encoder and a decoder. The encoder performs feature encoding on the input grayscale ultrasound image to obtain image semantic information, while the decoder decodes the feature image obtained by the encoder and synthesizes it into a synthetic strain elastic image. In this embodiment, as... Figure 3As shown, the encoder includes one convolutional input layer and six convolutional units. Each convolutional unit consists of a nonlinear activation layer, a convolutional layer, and a batch normalization layer. Between every two convolutional units, a convolutional layer with a stride of 2 is used to reduce the feature resolution, achieve downsampling, and increase the receptive field of the convolutional operation. In this embodiment, the decoder also includes six convolutional units and one convolutional output layer. The convolutional unit consists of a nonlinear activation layer, a deconvolutional layer, and a batch normalization layer. The deconvolutional layer can restore the low-resolution feature image to the input size. A nonlinear activation function is used in the convolutional output layer to output a three-channel synthetic strain elastic image.

[0044] It should be noted that the encoder and decoder in this embodiment are connected across layers. This network structure design has the following advantages for strain elastic image synthesis: First, the hierarchical structure of the encoder and decoder can help the generator network module 11 learn the semantic information of the image at different scales better; second, the cross-layer connection enables the necessary detailed features in the reconstruction of the synthesized strain elastic image to be preserved, thereby improving the fidelity of the generator network module 11, i.e., the model, for high-frequency information.

[0045] In one embodiment of the present invention, the global discrimination module 12 includes a plurality of third convolutional units and a convolutional output layer, wherein each third convolutional unit includes a convolutional layer, a batch normalization layer and a nonlinear activation layer in sequence.

[0046] like Figure 4 As shown, the global discriminator module 12, or global discriminator, adopts a fully convolutional network structure, which includes four convolutional units and one convolutional output layer. Each convolutional unit contains a convolutional layer, a batch normalization layer, and a nonlinear activation layer. In this embodiment, the global discriminator adopts the paradigm of a conditional adversarial network, meaning its input simultaneously includes grayscale ultrasound images and elastography images. This design allows the global discriminator to obtain more prior information, thereby accelerating model training and improving the performance of the global discriminator. In this embodiment, the fully convolutional network structure design and the local connectivity characteristics of the convolutional operation in the convolutional output layer enable the global discriminator to effectively model the input image as a Markov random field, facilitating the reconstruction of high-frequency details in the image.

[0047] In this embodiment, by performing a true / false discrimination on the input image, an adversarial gradient can be provided to the generator during training, thereby improving the generator's capabilities. Furthermore, this embodiment employs a conditional adversarial network paradigm, where the discriminator simultaneously receives grayscale ultrasonic images and synthetic strain elastic images as input. This allows the global discriminator to acquire more prior knowledge, accelerating network training and improving the performance of the global discriminator. Additionally, the fully convolutional network structure utilizes the local connectivity and translation invariance of convolutional layers to establish a Markov random conditional field on the input image. Compared to fully connected layers, this embodiment is advantageous in preserving the local high-frequency features of the image.

[0048] like Figure 5 As shown, the system also includes a tumor discrimination module 13. The tumor discrimination module 13 is connected to the generative network module 11. The tumor discrimination module 13 is used to acquire region of interest images extracted from grayscale ultrasound images and synthetic strain elastic images, and to perform differential modeling on the region of interest images.

[0049] Specifically, for retrospective data acquired by handheld ultrasound devices, grayscale ultrasound images can be selected from the retrospective data, and regions of interest containing tumor areas can be extracted from the grayscale ultrasound images and synthetic strain elastic images. Differential modeling is then performed based on the two types of regions of interest images.

[0050] Since the entire image often includes not only the tumor region but also a large number of normal tissue regions, and the tissue elasticity of the two differs greatly, their appearance in the image also varies significantly. Therefore, in this embodiment, in addition to using a global discriminator for the entire image, a local discriminator, namely the tumor discrimination module 13, can be designed using the same network structure for the tumor region. This allows the tumor discrimination module 13 to model this difference, thereby increasing the robustness of the synthesized tumor region image. Thus, the design of global and local discriminators in this invention can improve the realism and semantic accuracy of the synthesized strain elasticity image.

[0051] like Figure 6 As shown, the system also includes a color equalization module 14. The color equalization module 14 is connected to the generator network module 11. The color equalization module 14 is used to determine the color prior coefficients of the synthesized strain elastic image, and to determine the loss functions of the generator network module 11, the global discrimination module 12, and the tumor discrimination module 13. The loss functions are weighted and constrained by the color prior coefficients in order to improve the quality of the synthesized strain elastic image.

[0052] In this embodiment, a significant characteristic of the synthesized strain elastic image is its relatively simple color distribution, with colors concentrated in red and blue. Therefore, the color distribution of the image dataset can be statistically analyzed, and the model training process can be constrained based on statistical prior information, i.e., the system loss function can be constrained using color prior coefficients. To model this color difference representation, the present invention provides a color equalization module 14, which weights the system loss function during the training process based on the color balance coefficient, i.e., the color prior coefficient, of color rarity, to improve the color realism of the synthesized image.

[0053] Specifically, the lab color space can first be quantized, and the coordinate space formed by ab can be divided into small squares at intervals of 10. Then, the frequency of all pixels in the training data falling into the small squares of the ab space is counted, and further converted into the prior probability of a color point falling into each small square. Then, the corresponding color prior coefficient for each color pixel can be calculated using the following formula:

[0054]

[0055] In the formula, γ p It is the color prior coefficient of pixel p. α is the statistical prior probability of pixel p, Q is the total number of quantized squares, and α is the color equalization parameter.

[0056] In this embodiment, by adjusting the color equalization parameter α∈[0,1], the proportion of prior information added can be controlled to avoid over-focusing on outliers and causing model training collapse, thereby adjusting the prior probability distribution relative to γ. p The impact.

[0057] Furthermore, the loss function can be weighted and constrained based on the color prior coefficients to improve the quality of the synthesized strain elastic image. The loss functions of the generator network module 11 (generator), the global discriminator module 12 (global discriminator), and the tumor discriminator module 13 (tumor discriminator) can be defined as follows:

[0058]

[0059]

[0060]

[0061]

[0062] In the formula, Let represent the loss function of the generator, λ be a coefficient used to adjust the ratio of the generator's loss to the discriminator's loss, γ be the color prior coefficient, and y be the loss function of the generator. virtual and y realThese represent the synthetic strain-elastic image data and the real strain-elastic image data, respectively. D(x,y) represents the L1 loss between the synthetic elastic image and the real elastic image. virtual ) represents the input grayscale image x and the synthesized elastic image y. virtual The output of the discriminator under the condition of x represents the cross-entropy loss between the discriminator output and the target. tumor This represents the image data of the tumor region in the input grayscale image. and These represent tumor region image data extracted from real strain-elastic images and synthetic strain-elastic images, respectively. This represents the cross-entropy loss between the discriminator output and the target. Let D(x,y) represent the loss function of the global discriminator. real () represents the input grayscale image x and the true elastic image y. real The output of the discriminator under the condition of This represents the cross-entropy loss between the discriminator output and the target. This represents the cross-entropy loss between the discriminator output and the target. This represents the loss function of the tumor discriminator. This represents the tumor region x in the input grayscale image. tumor Tumor regions and true elasticity images The output of the discriminator under the condition of This represents the cross-entropy loss between the discriminator output and the target. This indicates the tumor region x in the input grayscale image. tumor and synthetic elastic images of tumor regions The output of the discriminator under the condition of represents the cross-entropy loss between the discriminator output and the target, and Loss represents the total system loss.

[0063] like Figure 7 As shown, the present invention also provides a strain-elastic ultrasound image synthesis method, which is applied to the above-mentioned strain-elastic ultrasound image synthesis system. The method includes:

[0064] Step S1: Acquire grayscale ultrasound images and process them to obtain a synthetic strain elastic image.

[0065] In this embodiment, a grayscale ultrasound image can be obtained through a generation network module, and the grayscale ultrasound image can be processed to obtain a synthetic strain-elastic image. The step of processing the grayscale ultrasound image to obtain the synthetic strain-elastic image includes:

[0066] The grayscale ultrasound image is encoded by an encoder to obtain a feature image. The encoder includes a convolutional input layer and several first convolutional units. Each first convolutional unit includes a nonlinear activation layer, a convolutional layer and a batch normalization layer in sequence. The feature image is decoded by a decoder and synthesized to obtain a synthetic strain elastic image. The decoder includes several second convolutional units and a convolutional output layer. Each second convolutional unit includes a nonlinear activation layer, a deconvolutional layer and a batch normalization layer in sequence.

[0067] Step S2: Obtain the real strain elastic image and distinguish between the real strain elastic image and the synthesized strain elastic image to obtain prior information and improve the synthesis quality of the synthesized strain elastic image based on the prior information.

[0068] In this embodiment, the global discrimination module can distinguish between real and synthetic strain elastic images, and the tumor discrimination module can obtain the region of interest (ROI) images extracted from the grayscale ultrasound image and the synthetic strain elastic image. Then, the ROI images are modeled for differences, and the color prior coefficients of the synthetic strain elastic image are determined by the color equalization module. The loss functions of the generation network module, the global discrimination module, and the tumor discrimination module are also determined. The loss functions are weighted and constrained by the color prior coefficients to improve the quality of the synthetic strain elastic image.

[0069] It should be noted that the specific implementation method of the strain elastic ultrasound image synthesis method in this embodiment can be found in the specific implementation method of the strain elastic ultrasound image synthesis system described above. To avoid redundancy, it will not be repeated here.

[0070] As a specific example, we can first preprocess 4231 ultrasound images acquired by high-end ultrasound equipment, extract regions of interest from the original images, and adjust the paired grayscale ultrasound images and elastic ultrasound images to a size of 256*256. We can then divide the dataset into 2001 main datasets as training samples, 500 main datasets as internal test samples, and 1730 multi-center datasets as external test samples. Then, we can extract regions of interest from 349 portable ultrasound images, adjust the grayscale ultrasound images to a size of 256*256, and use them as test samples.

[0071] Furthermore, the generative network module and global discriminant module proposed in this invention are trained using training data, and then the models are tested using internal test sets, external multi-center test sets, and portable ultrasound data test sets, respectively.

[0072] Furthermore, the effectiveness of synthesized strain-elastic images was evaluated from several perspectives: image similarity index, visual assessment by ultrasound physicians with different experience levels, diagnostic efficacy for benign and malignant breast tumors, stability of synthesized elastic images at different imaging depths, and effectiveness in diagnosing benign and malignant breast tumors using portable ultrasound. Specifically, SSIM (Structural Similarity Index) and MAPE (Mean Absolute Percentage Error) were used to evaluate image similarity, while AUC (Area Under the Receiver Operating Characteristic) was used to assess the effectiveness of tumor benign and malignant detection. The evaluation results are analyzed below:

[0073] Figure 8 Visual comparison results of synthetic strain elastic images and real strain elastic images on the internal test set are presented, as well as performance comparison results in the detection of benign and malignant breast tumors. It can be seen that synthetic strain elastic images have an amplifying effect on the detection of benign and malignant breast tumors, and there is no significant difference compared with the hardware real imaging results.

[0074] Figure 9 The results show the comparison of the detection performance of benign and malignant breast tumors in various data centers on a multi-center dataset. It can be seen that the synthetic strain elastic image has good robustness in the multi-center case.

[0075] Figure 10 The results show a comparison of the effect of tumor depth on elastography. It can be seen that the synthetic strain elastography image has better detection performance than the real elastography image when the tumor depth is greater than 18 mm, and it can effectively overcome the occurrence of artifacts in image performance.

[0076] Figure 11 The results show that after training the model using data collected by high-end ultrasound equipment, the model was tested directly on a handheld ultrasound device without domain transfer. It can be seen that, on the one hand, the synthesized strain elastic image has good visual effect, and on the other hand, the introduction of the synthesized elastic image into the handheld ultrasound image has an improvement in the detection effect of benign and malignant breast tumors.

[0077] This invention employs synthetic image technology to improve the imaging quality of low-end acquisition devices, and overcomes the limitations of traditional imaging methods in various aspects such as imaging speed, imaging resolution, image modality, and slice staining techniques. Furthermore, this invention utilizes a generative adversarial model integrating a generator and a discriminator to simulate real-world data generation methods and learn the probability distribution of high-dimensional data. Through a game-like process between the generator and discriminator modules during training, this invention continuously improves the performance of both modules, thereby enhancing the quality of the synthesized data.

[0078] In summary, this invention models grayscale ultrasound images using a hierarchical U-shaped network structure for extracting image features. It retains high-frequency information in the image through cross-layer connections. Furthermore, the design of a global discrimination module and a tumor discrimination module enhances the semantic richness of the synthesized strain-elastic images. A color equalization module determines the color prior coefficients of the synthesized strain-elastic images, and these color prior coefficients, based on statistical color priors, are used to weight and constrain the system loss function, improving the color fidelity of the synthesized strain-elastic images. This facilitates improved performance in breast tumor detection using synthesized strain-elastic images. Additionally, the model developed in this invention, as a statistical model, avoids imaging biases caused by different ultrasound physician techniques and ultrasound signal attenuation due to deeper tumor tissue. Moreover, the model built based on images acquired by high-end ultrasound equipment provides a gain effect in detecting benign and malignant breast tumors in handheld ultrasound imaging, enabling handheld ultrasound to achieve strain-elastic imaging functionality without increasing hardware costs, thus promoting the application of handheld ultrasound.

[0079] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0080] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A strain-elastic ultrasonic image synthesis system, characterized in that, include: A network module is used to acquire grayscale ultrasound images and process the grayscale ultrasound images to obtain a synthetic strain elastic image. A global discrimination module is connected to the generation network module. The global discrimination module adopts a conditional adversarial network paradigm. The input includes the grayscale ultrasound image and the synthesized strain elastic image. It is used to obtain the real strain elastic image and to distinguish between the real strain elastic image and the synthesized strain elastic image to obtain prior information and improve the synthesis quality of the synthesized strain elastic image based on the prior information. A tumor discrimination module is connected to the generation network module. The tumor discrimination module is used to acquire the region of interest image extracted from the grayscale ultrasound image and the synthetic strain elastic image, and to perform differential modeling on the region of interest image based on the elastic difference between the tumor region and the normal tissue region. A color equalization module, connected to the generation network module, is used to determine the color prior coefficients of the synthesized strain elastic image, and to determine the loss functions of the generation network module, the global discrimination module, and the tumor discrimination module. The color prior coefficients are used to weight and constrain the loss functions to improve the quality of the synthesized strain elastic image. The color prior coefficients are calculated by statistically analyzing the distribution frequency of the training data after quantizing the LAB color space.

2. The strain-elastic ultrasonic image synthesis system as described in claim 1, characterized in that, The generated network module includes: The encoder includes a convolutional input layer and a plurality of first convolutional units, each of which sequentially includes a nonlinear activation layer, a convolutional layer and a batch normalization layer. The encoder is used to encode the grayscale ultrasound image into a feature image. A decoder, connected to the encoder, includes several second convolutional units and a convolutional output layer. Each second convolutional unit sequentially includes a nonlinear activation layer, a deconvolutional layer, and a batch normalization layer. The decoder is used to decode the feature image and synthesize the synthetic strain elastic image.

3. The strain-elastic ultrasonic image synthesis system as described in claim 2, characterized in that, The network structure formed by the encoder and the decoder is a U-shaped symmetrical structure.

4. The strain-elastic ultrasonic image synthesis system as described in claim 3, characterized in that, The global discrimination module includes several third convolutional units and a convolutional output layer. Each third convolutional unit includes a convolutional layer, a batch normalization layer, and a nonlinear activation layer in sequence.

5. A method for synthesizing strain-elastic ultrasound images, applied to the strain-elastic ultrasound image synthesis system as described in any one of claims 1-4, characterized in that, include: Acquire grayscale ultrasound images and process the grayscale ultrasound images to obtain a synthetic strain elastic image; Acquire real strain elastic images and distinguish between real and synthetic strain elastic images to obtain prior information, and improve the synthesis quality of synthetic strain elastic images based on the prior information.

6. The strain-elastic ultrasonic image synthesis method as described in claim 5, characterized in that, The step of processing the grayscale ultrasound image to obtain a synthesized strain elastic image includes: The grayscale ultrasound image is encoded by an encoder to obtain a feature image. The encoder includes a convolutional input layer and several first convolutional units. Each first convolutional unit includes a nonlinear activation layer, a convolutional layer and a batch normalization layer in sequence. The feature image is decoded by a decoder and the synthesized strain elastic image is obtained. The decoder includes several second convolutional units and a convolutional output layer. Each second convolutional unit includes a nonlinear activation layer, a deconvolutional layer and a batch normalization layer in sequence.

7. The strain-elastic ultrasonic image synthesis method as described in claim 6, characterized in that, The method further includes: acquiring a region of interest image extracted from the grayscale ultrasound image and the synthetic strain elastic image, and performing differential modeling on the region of interest image.

8. The strain-elastic ultrasonic image synthesis method as described in claim 7, characterized in that, The method further includes: determining the color prior coefficients of the synthesized strain elastic image, and determining the loss functions of the generation network module, the global discrimination module, and the tumor discrimination module, and weighting the loss functions using the color prior coefficients to improve the quality of the synthesized strain elastic image.

Images