An unsupervised ultrasound image conversion method based on continuous learning

Abstract

The application discloses a kind of based on continuous learning's unsupervised ultrasonic image conversion method, it is related to computer vision ultrasonic image technical field, the method includes: with CycleGAN framework as foundation, introduce double-branch residual module, through the explicit decoupling of content branch and style branch, realize the structural information of input medical image and the respective modeling of ultrasonic speckle texture, combined with LoRA fine-tuning mechanism, only by adding a small amount of low-rank trainable parameters in content branch can realize new task learning, the dynamic LoRA adapter designed can automatically select optimal parameter configuration according to task characteristics, realize the adaptive conversion of different modal to ultrasonic image.The present application can realize the continuous learning of multi-task under the condition of no paired data and without retraining the whole model, effectively alleviate the problem of catastrophic forgetting.The generated image compared with prior art, texture reality, structure consistency and parameter efficiency etc.are all significantly improved, with high flexibility and practical value.

Description

Technical Field

[0001] This invention relates to the field of computer vision ultrasound image technology, and specifically to an unsupervised ultrasound image conversion method based on continuous learning. Background Technology

[0002] Ultrasound imaging, as an important medical imaging diagnostic technique, is widely used in clinical examinations, interventional treatments, and surgical procedures due to its non-invasiveness, real-time nature, and convenience. It plays an irreplaceable role, especially in the diagnosis of obstetrics and gynecology, cardiology, and musculoskeletal disorders. However, ultrasound operation is highly demanding, requiring strict skill levels from the operator, particularly in precise scanning and diagnostic scenarios, often necessitating extensive clinical experience and excellent hand-eye coordination. For beginners, mastering the skills typically takes a considerable amount of time, and the lack of sufficient training opportunities may increase diagnostic and treatment risks when dealing with rare cases.

[0003] The emergence of ultrasound simulation technology offers a new approach to solving the aforementioned problems. By constructing virtual environments or using simulators, learners can simulate various clinical scenarios, including different anatomical structures, pathological features, and ultrasound operation procedures, within a controlled, safe, and repeatable training platform, thereby improving the systematic nature and effectiveness of training. However, due to the scarcity of real ultrasound case data, existing simulation training still falls short in terms of case diversity.

[0004] In recent years, deep learning-based ultrasound image synthesis methods have emerged, effectively reducing the reliance on real ultrasound data and enabling researchers to synthesize ultrasound images (US) using other types of medical or tissue images, such as magnetic resonance imaging (MRI) and computed tomography (CT). While these methods have made significant progress in synthesis quality, their training processes are typically limited to a single type of medical image data and often require additional data for guidance. When migrating to other modalities or new anatomical structures, model retraining is often necessary, leading to high training costs, significant storage pressure, and the added burden of data preparation. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide an unsupervised ultrasound image conversion method based on continuous learning.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] This invention discloses an unsupervised ultrasound image conversion method based on continuous learning, comprising the following steps:

[0008] S1. Construct the original image domain and the ultrasound image domain, and define a first mapping and a second mapping. The first mapping is from the original image domain to the ultrasound image domain, and the second mapping is from the ultrasound image domain to the original image domain.

[0009] S2. Based on the CycleGAN framework, a generator, a restorer, and a discriminator are created. The generator is used to convert the original image into an ultrasound image, and the restorer is used to reconstruct the original image from the ultrasound image.

[0010] S3. Set up a dual-branch residual module in the intermediate feature extraction layer between the generator and the reducer. Based on the explicit decoupling of the content branch and the style branch, model the structural features and texture features of the input feature map respectively.

[0011] S4. For the initial task, train the generator to obtain the initial basic parameter model; for subsequent tasks, load the pre-trained parameters of the initial basic parameter model and freeze its backbone weights. At the content branch of the dual-branch residual module, add a trainable convolutional low-rank adaptation layer to achieve parameter fine-tuning.

[0012] S5. Optimize the generator and discriminator using the adversarial loss, identity loss and cycle consistency loss of the CycleGAN framework;

[0013] S6. Input the training dataset into the dynamic LoRA adapter, calculate and save the normalized covariance matrix and feature prototypes for each task.

[0014] S7. During the testing phase, the test image is input into the dynamic LoRA adapter, and its correlation with the normalized covariance matrix and feature prototype corresponding to each task is calculated. The best matching task number is determined, and the corresponding Conv-LoRA layer parameters are loaded into the optimized generator.

[0015] S8. Input the test image into the optimized generator to obtain the converted ultrasound image.

[0016] Further, step S1 specifically includes: constructing the original image domain. and ultrasound image domain ,in Represents the original image domain The first original image in the image, Represents the ultrasound image domain The first ultrasound image in the image defines the first mapping. Second mapping Then the original image domain and ultrasound image domain The images in the dataset are divided into a training set and a test set in an 8:2 ratio, with the original image domain being the training set. and ultrasound image domain All images in the dataset are constructed as unpaired datasets.

[0017] Preferably, step S2 specifically includes: creating a generator G, a reducer F, and two discriminators using a bidirectional generative adversarial network architecture based on the CycleGAN framework. and .

[0018] Preferably, step S3 specifically includes: setting a dual-branch residual module in the intermediate feature extraction layer between the generator G and the reducer F, wherein the dual-branch residual module includes a content branch and a style branch; the content branch uses a 3×3 convolutional layer, and the input feature map is... The output after content branching is ,in The content transformation function is represented; the style branch uses a depthwise separable convolutional structure with a 7×7 convolutional kernel, and the input feature map is... The output after style branching is ,in The style transformation function is represented; a channel attention mechanism is introduced to weight the outputs of the two branches, followed by feature fusion through 1×1 convolution and residual connection to finally obtain the output of the dual-branch residual module. , ,in and This indicates a channel attention-weighted operation. This represents feature fusion and residual mapping. This indicates feature splicing.

[0019] Preferably, step S4 specifically includes: for the initial task 0, without setting a Conv-LoRA layer, training the generator to obtain the initial basic parameter model. For subsequent tasks ,load The pre-trained parameters are then frozen and their backbone weights are used in the generator. G At the content branch of the dual-branch residual module, a trainable convolutional low-rank adaptation layer is added to enable parameter fine-tuning.

[0020] Preferably, step S5 specifically includes the following steps:

[0021] S51. Perform data preprocessing on the images in the training set, using the first original image... Input to generator G In the process, a second ultrasound image was obtained. The first ultrasound image Input to restorer F In the process, the second original image is obtained. ;Will Input generator G Obtain the second intermediate result ,Will Input restorer F In the middle, the first intermediate result is obtained. ; the first original image Input to restorer F In the middle, the third intermediate result was obtained. The first ultrasound image Input to generator G In the middle, the fourth intermediate result was obtained. ;

[0022] S52. Calculate Identity Loss and cycle consistency loss :

[0023] ;

[0024] ;

[0025] in, This indicates that for all original image domains The first original image in Take the average of the calculation results. This indicates that for all ultrasound image domains The first ultrasound image in Take the average of the calculation results. Represents the L1 norm;

[0026] S53, transfer the second ultrasound image and the first ultrasound image Input to discriminator Computational ultrasound image domain The losses of the confrontation , ;

[0027] The second original image and the first original image Input to discriminator Calculate the original image domain The losses of the confrontation , The generator and discriminator are optimized using the adversarial loss, identity loss, and cycle consistency loss of the CycleGAN framework.

[0028] Preferably, step S6 specifically includes: selecting from the training set the data belonging to the original image domain. N images are input into a pre-trained ResNet-18 convolutional neural network to extract its 512-dimensional feature vectors. Then, a graph basis power step transformation is performed on it to obtain a 512-dimensional feature vector after the graph basis power step transformation. , And calculate subsequent tasks. The average value of the 512-dimensional eigenvectors after power-law step transformation of N graph bases. The first feature prototype is obtained. ,in Represents the original image domain in the training set The One sample.

[0029] Preferably, step S7 specifically includes the following steps:

[0030] S71. The 512-dimensional feature vector and the first feature prototype after power-law step transformation using N graph bases. Correlation calculations were performed to obtain the first covariance matrix. The first covariance matrix Shape If the value of N is less than 512, then the first covariance matrix needs to be adjusted. Perform shrinkage to obtain the shrunken covariance matrix. , ,in, Describe the first covariance matrix Mean diagonal variance Describe the first covariance matrix The mean off-diagonal covariance, It is the identity matrix. and This is the scaling factor;

[0031] S72, Regarding the first covariance matrix Perform correlation matrix normalization. ,in Describe the first covariance matrix In the The standard deviation of the dimensional components is calculated using the following formula: , Describe the first covariance matrix In the The standard deviation of the dimensional components is calculated using the following formula: , Describe the first covariance matrix The elements are used to obtain the normalized covariance matrix. ; Represents the normalized covariance matrix Element;

[0032] S73. For initial task 0, save the generator. G The parameters, and the second feature prototype Second covariance matrix For subsequent tasks Save Conv-LoRA layer parameters and the first feature prototype and the third covariance matrix ;

[0033] S74, Loader G pre-trained model G 0 ;

[0034] S75, through formula Calculate Mahalanobis distance T represents transpose; then, using the formula... Calculate the dataset with the minimum Mahalanobis distance value For the initial task 0, there is no need to load the Conv-LoRA layer parameters; for subsequent tasks... The Conv-LoRA layer parameters corresponding to the dataset with the smallest Mahalanobis distance value. Load into generator G middle.

[0035] Preferably, step S8 specifically includes: dividing the original image domain The first original image belonging to the test set Input to the optimized generator G In this process, the original image domain is obtained. The first original image The converted ultrasound image.

[0036] The beneficial effects of this invention are:

[0037] 1) This application enables high-quality conversion of different types of medical images (such as CT and MRI) to ultrasound images without relying on additional paired data and without retraining the entire model; compared with the prior art, the present invention has significantly improved the texture realism, structural consistency and parameter efficiency of the generated images. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the training process of an unsupervised ultrasound image conversion method based on continuous learning according to an embodiment of the present invention.

[0039] Figure 2 This is a schematic diagram of the test process for an unsupervised ultrasound image conversion method based on continuous learning, according to an embodiment of the present invention.

[0040] Figure 3 This is a schematic diagram of the framework of an unsupervised ultrasound image conversion method based on continuous learning, according to an embodiment of the present invention.

[0041] Figure 4 This is a schematic diagram of the DBRB module of an unsupervised ultrasound image conversion method based on continuous learning, according to an embodiment of the present invention.

[0042] Figure 5 This is a schematic diagram illustrating the data, training process, and conversion results of an unsupervised ultrasound image conversion method based on continuous learning, according to an embodiment of the present invention. Detailed Implementation

[0043] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0044] See Figures 1-4 This invention discloses an unsupervised ultrasound image conversion method based on continuous learning. To simplify the training process for multimodal image-to-ultrasound image conversion and reduce the storage pressure of learning new modal data, while avoiding reliance on additional auxiliary data, this application introduces a dual-branch residual block (DBRB) based on the CycleGAN framework. Through explicit decoupling of content and style branches, it achieves separate modeling of anatomical structures and ultrasound speckle textures, thereby significantly improving the structural fidelity and texture realism of the generated images. Furthermore, by combining a LoRA (Low-Rank Adaptation) fine-tuning mechanism, by fine-tuning the content branch of the DBRB, only a small number of low-rank parameters need to be updated to complete the learning of new tasks, greatly reducing the storage and computational overhead of the model. Simultaneously, a dynamic LoRA adapter module is designed to automatically select the optimal parameter configuration based on task characteristics, effectively enhancing the flexibility and scalability of the method. The method includes the following steps:

[0045] S1. Construct the original image domain and the ultrasound image domain, and define a first mapping and a second mapping. The first mapping is from the original image domain to the ultrasound image domain, and the second mapping is from the ultrasound image domain to the original image domain.

[0046] S2. Based on the CycleGAN framework, a generator, a restorer, and a discriminator are created. The generator is used to convert the original image into an ultrasound image, and the restorer is used to reconstruct the original image from the ultrasound image.

[0047] S3. Multiple dual-branch residual blocks (DBRBs) are set in the intermediate feature extraction layer between the generator and the reducer. Based on the explicit decoupling of the content branch and the style branch, the structural features and texture features of the input feature map are modeled respectively. The dual-branch residual block (DBRB) includes two independent feature extraction paths, namely the content branch and the style branch. The content branch uses a 3×3 convolutional layer to extract local structural features. The style branch uses a 7×7 depthwise separable convolution to extract ultrasonic speckle and global texture features. The outputs of the two branches are weighted and fused by the channel attention mechanism, and then feature fusion and residual connection are performed through a 1×1 convolution.

[0048] S4. For the initial task, the generator is trained to obtain the initial basic parameter model. For subsequent tasks, the pre-trained parameters of the initial basic parameter model are loaded and its backbone weights are frozen. At the content branch of the dual-branch residual module, a trainable convolutional low-rank adaptation layer (Conv-LoRA layer) is added to achieve parameter fine-tuning. The Conv-LoRA layer adopts a low-rank structure of "channel dimensionality reduction + small-channel convolution" to fine-tune the convolutional layer parameters of the content branch of the initial basic parameter model to adapt to the subsequent ultrasound conversion task.

[0049] S5. Optimize the generator and discriminator using the adversarial loss, identity loss and cycle consistency loss of the CycleGAN framework;

[0050] S6. Input the training dataset into the dynamic LoRA adapter, calculate and save the normalized covariance matrix and feature prototype corresponding to each task; the dynamic LoRA adapter uses ResNet-18 to extract features and performs Tukey's Ladder of Power Transformation (Tukey) to calculate the feature prototype and covariance matrix of each task, realize correlation matrix normalization and covariance shrinkage, determine the task category of the input image through Mahalanobis distance, dynamically load the LoRA parameters of the corresponding task, and realize cross-modal adaptation;

[0051] S7. During the testing phase, the test image is input into the dynamic LoRA adapter, and its correlation with the normalized covariance matrix and feature prototype corresponding to each task is calculated. The best matching task number is determined, and the corresponding Conv-LoRA layer parameters are loaded into the optimized generator.

[0052] S8. Input the test image into the optimized generator to obtain the converted ultrasound image.

[0053] For example, step S1 specifically includes: constructing the original image domain and ultrasound image domain ,in Represents the original image domain The first original image in the image, Represents the ultrasound image (US image) domain The first ultrasound image in the image defines the first mapping. Second mapping Then the original image domain and ultrasound image domain The images in the dataset are divided into a training set and a test set in an 8:2 ratio, with the original image domain being the training set. and ultrasound image domain All images in the dataset are constructed as unpaired datasets.

[0054] For example, step S2 specifically includes: creating a generator G, a reducer F, and two discriminators using a bidirectional generative adversarial network architecture based on the CycleGAN framework. and The generator G is only fully trained in the initial task to obtain the basic parameter model, which is then used as a pre-trained model for other tasks.

[0055] For example, step S3 specifically includes: setting up multiple dual-branch residual modules in the intermediate feature extraction layer of generator G and reducer F to replace traditional ResNet residual blocks, wherein the dual-branch residual modules include content branches and style branches; the content branches adopt 3×3 convolutional layers, and the input feature maps are... The output after content branching is ,in The content transformation function is represented; the style branch uses a depthwise separable convolutional structure with a 7×7 convolutional kernel, and the input feature map is... The output after style branching is ,in The style transformation function is represented; a channel attention mechanism is introduced to weight the outputs of the two branches, followed by feature fusion through 1×1 convolution and residual connection to improve training stability and expressive power, ultimately obtaining the output of the dual-branch residual module. , ,in and This indicates a channel attention-weighted operation. This represents feature fusion and residual mapping. This indicates feature splicing.

[0056] For example, step S4 specifically includes: for the initial task 0, without setting the Conv-LoRA layer, training the generator to obtain the initial basic parameter model. For subsequent tasks ,load The pre-trained parameters are then frozen and their backbone weights are used in the generator. G At the content branch of the dual-branch residual module, a trainable convolutional low-rank adaptation layer is added to achieve parameter fine-tuning. The Conv-LoRA layer employs "channel dimensionality reduction plus small channels". The structure of "convolution" allows for small random initialization. It is a 1×1 convolution with weights At the same time, initialize with zero. for Convolution, its weights Then for the input feature map The formula for the low-rank increment is: , For subsequent tasks LoRA-adjusted weights for: ,in, For the initial task model G 0 Content branch convolutional layer parameters, These are the parameters for the Conv-LoRA layer.

[0057] For example, step S5 specifically includes the following steps:

[0058] S51. Perform data preprocessing on the images in the training set, using the first original image... Input to generator G In the process, a second ultrasound image was obtained. The first ultrasound image Input to restorer F In the process, the second original image is obtained. ;Will Input generator G Obtain the second intermediate result ,Will Input restorer F In the middle, the first intermediate result is obtained. ; the first original image Input to restorer F In the middle, the third intermediate result was obtained. The first ultrasound image Input to generator G In the middle, the fourth intermediate result was obtained. ;

[0059] S52. Calculate Identity Loss and cycle consistency loss :

[0060] ;

[0061] ;

[0062] in, This indicates that for all original image domains The first original image in Take the average of the calculation results. This indicates that for all ultrasound image domains The first ultrasound image in Take the average of the calculation results. Represents the L1 norm;

[0063] S53, transfer the second ultrasound image and the first ultrasound image Input to discriminator Computational ultrasound image domain The losses of the confrontation , ; the second original image and the first original image Input to discriminator Calculate the original image domain The losses of the confrontation , Using the adversarial loss, identity loss, and cycle consistency loss of the CycleGAN framework, the overall loss function expression is as follows: ,in, As a weighting parameter, in this embodiment Optimize the generator and discriminator.

[0064] For example, step S6 specifically includes: selecting from the training set the data belonging to the original image domain. N images are input into a pre-trained ResNet-18 convolutional neural network to extract its 512-dimensional feature vectors. Then, it is subjected to Tukey's Ladder of Power Transformation (Tukey) to obtain the 512-dimensional feature vector after the Tukey power ladder transformation. , In this embodiment And calculate subsequent tasks. The average value of the 512-dimensional eigenvectors after power-law step transformation of N graph bases. The first feature prototype is obtained. ,in Represents the original image domain in the training set The One sample.

[0065] For example, step S7 specifically includes the following steps:

[0066] S71. The 512-dimensional feature vector and the first feature prototype after power-law step transformation using N graph bases. Correlation calculations were performed to obtain the first covariance matrix. The first covariance matrix Shape If the value of N is less than 512, then the first covariance matrix needs to be adjusted. Perform shrinkage to obtain the shrunken covariance matrix. , ,in, Describe the first covariance matrix Mean diagonal variance Describe the first covariance matrix The mean off-diagonal covariance, It is the identity matrix. and This is the scaling factor;

[0067] S72, Regarding the first covariance matrix Correlation matrix normalization is performed to adjust all diagonal elements to 1. ,in Describe the first covariance matrix In the The standard deviation of the dimensional components is calculated using the following formula: , Describe the first covariance matrix In the The standard deviation of the dimensional components is calculated using the following formula: ; Obtain the normalized covariance matrix ; Represents the normalized covariance matrix Element; Describe the first covariance matrix The element is calculated using the following formula: , It is the i-th dimension component of the feature vector after the nth sample is transformed. It is the i-th dimension component of the mean vector of this task. When hour, Let be the variance of the i-th dimension component, when hour, Let be the covariance of the i-th and j-th dimension components;

[0068] S73. For initial task 0, save the generator. G The parameters, and the second feature prototype Second covariance matrix For subsequent tasks Save Conv-LoRA layer parameters and the first feature prototype and the third covariance matrix ;

[0069] S74, Loader G pre-trained model G 0 ;

[0070] S75, through formula Calculate Mahalanobis distance T represents transpose; then, using the formula... Calculate the dataset with the minimum Mahalanobis distance value For the initial task 0, there is no need to load the Conv-LoRA layer parameters; for subsequent tasks... The Conv-LoRA layer parameters corresponding to the dataset with the smallest Mahalanobis distance value. Load into generator G The framework diagram of this application is shown in the figure below. Figure 3 As shown, where, This represents the normalized covariance matrix at t=1 for subsequent tasks. This represents the first feature prototype at t=1 for the subsequent task. This represents the normalized covariance matrix at t=2 for the subsequent task. This represents the first feature prototype at t=2 for the subsequent task. This represents the normalized covariance matrix for subsequent tasks at t=n. This represents the first feature prototype for subsequent tasks at t=n. and This represents the parameters of the two convolutional layers in the Conv-LoRA layer; This represents the image generated by generator G.

[0071] For example, step S8 specifically includes: dividing the original image domain The first original image belonging to the test set Input to the optimized generator G In this process, the original image domain is obtained. The first original image The converted ultrasound image.

[0072] For example, the model was trained using the Adam optimizer with default hyperparameter settings, where The initial learning rate is The weight decay is 0. The batch size is 2, the epoch is 100, and the deep learning framework used is PyTorch 2.0.

[0073] For example, the corresponding data, training process diagram, and transformation result diagram are as follows: Figure 5 As shown: the two rows of images represent the relevant images for Task 0 and Task 1, respectively. Compared with previous methods, as shown... Figure 5 As shown, the ultrasound image conversion method based on continuous learning of this invention exhibits excellent performance in both conversion quality and forgetting rate. Regarding conversion quality, the generated US image not only realistically simulates ultrasound speckle features but also effectively preserves the semantic information of the original image. In terms of forgetting rate, the method of this invention maintains the conversion quality of Task 0 even after completing Task 1 learning. Furthermore, this invention has high parameter utilization in its model structure design, requiring only approximately 0.3M trainable parameters to adapt to new tasks without retraining the entire model, thus significantly reducing storage overhead and training complexity.

[0074] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. An unsupervised ultrasound image conversion method based on continuous learning, characterized in that, Includes the following steps: S1. Construct the original image domain and the ultrasound image domain, and define a first mapping and a second mapping. The first mapping is from the original image domain to the ultrasound image domain, and the second mapping is from the ultrasound image domain to the original image domain. S2. Based on the CycleGAN framework, a generator, a restorer, and a discriminator are created. The generator is used to convert the original image into an ultrasound image, and the restorer is used to reconstruct the original image from the ultrasound image. S3. Set up a dual-branch residual module in the intermediate feature extraction layer between the generator and the reducer. Based on the explicit decoupling of the content branch and the style branch, model the structural features and texture features of the input feature map respectively. S4. For the initial task, train the generator to obtain the initial basic parameter model; for subsequent tasks, load the pre-trained parameters of the initial basic parameter model and freeze its backbone weights. At the content branch of the dual-branch residual module, add a trainable convolutional low-rank adaptation layer to achieve parameter fine-tuning. S5. Optimize the generator and discriminator using the adversarial loss, identity loss and cycle consistency loss of the CycleGAN framework; S6. Input the training dataset into the dynamic LoRA adapter, calculate and save the normalized covariance matrix and feature prototypes for each task. S7. During the testing phase, the test image is input into the dynamic LoRA adapter, and its correlation with the normalized covariance matrix and feature prototype corresponding to each task is calculated. The best matching task number is determined, and the corresponding Conv-LoRA layer parameters are loaded into the optimized generator. S8. Input the test image into the optimized generator to obtain the converted ultrasound image.

2. The unsupervised ultrasound image conversion method based on continuous learning according to claim 1, characterized in that, Step S1 specifically includes: constructing the original image domain and ultrasound image domain ,in Represents the original image domain The first original image in the image, Represents the ultrasound image domain The first ultrasound image in the image defines the first mapping. Second mapping Then the original image domain and ultrasound image domain The images in the dataset are divided into a training set and a test set in an 8:2 ratio, with the original image domain being the training set. and ultrasound image domain All images in the dataset are constructed as unpaired datasets.

3. The unsupervised ultrasound image conversion method based on continuous learning according to claim 2, characterized in that, Step S2 specifically includes: creating a generator G, a reducer F, and two discriminators using a bidirectional generative adversarial network architecture based on the CycleGAN framework. and .

4. The unsupervised ultrasound image conversion method based on continuous learning according to claim 3, characterized in that, Step S3 specifically includes: setting a dual-branch residual module in the intermediate feature extraction layer between the generator G and the reducer F, wherein the dual-branch residual module includes a content branch and a style branch; the content branch uses a 3×3 convolutional layer, and the input feature map is... The output after content branching is ,in The content transformation function is represented; the style branch uses a depthwise separable convolutional structure with a 7×7 convolutional kernel, and the input feature map is... The output after style branching is ,in The style transformation function is represented; a channel attention mechanism is introduced to weight the outputs of the two branches, followed by feature fusion through 1×1 convolution and residual connection to finally obtain the output of the dual-branch residual module. , ,in and This indicates a channel attention-weighted operation. This represents feature fusion and residual mapping. This indicates feature splicing.

5. The unsupervised ultrasound image conversion method based on continuous learning according to claim 4, characterized in that, Step S4 specifically includes: For the initial task 0, without setting a Conv-LoRA layer, training the generator to obtain the initial basic parameter model. For subsequent tasks ,load The pre-trained parameters are then frozen and their backbone weights are used in the generator. G At the content branch of the dual-branch residual module, a trainable convolutional low-rank adaptation layer is added to enable parameter fine-tuning.

6. The unsupervised ultrasound image conversion method based on continuous learning according to claim 5, characterized in that, Step S5 specifically includes the following steps: S51. Perform data preprocessing on the images in the training set, using the first original image... Input to generator G In the process, a second ultrasound image was obtained. The first ultrasound image Input to restorer F In the process, the second original image is obtained. ;Will Input generator G Obtain the second intermediate result ,Will Input restorer F In the middle, the first intermediate result is obtained. ; the first original image Input to restorer F In the middle, the third intermediate result was obtained. The first ultrasound image Input to generator G In the middle, the fourth intermediate result was obtained. ; S52. Calculate Identity Loss and cycle consistency loss : ； ； in, This indicates that for all original image domains The first original image in Take the average of the calculation results. This indicates that for all ultrasound image domains The first ultrasound image in Take the average of the calculation results. Represents the L1 norm; S53, transfer the second ultrasound image and the first ultrasound image Input to discriminator Computational ultrasound image domain The losses of the confrontation , ; The second original image and the first original image Input to discriminator Calculate the original image domain The losses of the confrontation , The generator and discriminator are optimized using the adversarial loss, identity loss, and cycle consistency loss of the CycleGAN framework.

7. The unsupervised ultrasound image conversion method based on continuous learning according to claim 6, characterized in that, Step S6 specifically includes: selecting from the training set elements belonging to the original image domain... N images are input into a pre-trained ResNet-18 convolutional neural network to extract its 512-dimensional feature vectors. Then, a graph basis power step transformation is performed on it to obtain a 512-dimensional feature vector after the graph basis power step transformation. , And calculate subsequent tasks. The average value of the 512-dimensional eigenvectors after power-law step transformation of N graph bases. The first feature prototype is obtained. ,in Represents the original image domain in the training set The One sample.

8. The unsupervised ultrasound image conversion method based on continuous learning according to claim 7, characterized in that, Step S7 specifically includes the following steps: S71. The 512-dimensional feature vector and the first feature prototype after power-law step transformation using N graph bases. Correlation calculations were performed to obtain the first covariance matrix. The first covariance matrix Shape If the value of N is less than 512, then the first covariance matrix needs to be adjusted. Perform shrinkage to obtain the shrunken covariance matrix. , ,in, Describe the first covariance matrix Mean diagonal variance Describe the first covariance matrix The mean off-diagonal covariance, It is the identity matrix. and This is the scaling factor; S72, Regarding the first covariance matrix Perform correlation matrix normalization. ,in Describe the first covariance matrix In the The standard deviation of the dimensional components is calculated using the following formula: , Describe the first covariance matrix In the The standard deviation of the dimensional components is calculated using the following formula: , Describe the first covariance matrix The elements are used to obtain the normalized covariance matrix. ; Represents the normalized covariance matrix Element; S73. For initial task 0, save the generator. G The parameters, and the second feature prototype Second covariance matrix For subsequent tasks Save Conv-LoRA layer parameters and the first feature prototype and the third covariance matrix ; S74, Loader G pre-trained model G 0 ; S75, through formula Calculate Mahalanobis distance T represents transpose; then, using the formula... Calculate the dataset with the minimum Mahalanobis distance value For the initial task 0, there is no need to load the Conv-LoRA layer parameters; for subsequent tasks... The Conv-LoRA layer parameters corresponding to the dataset with the smallest Mahalanobis distance value. Load into generator G middle.

9. The unsupervised ultrasound image conversion method based on continuous learning according to claim 8, characterized in that, Step S8 specifically includes: dividing the original image domain The first original image belonging to the test set Input to the optimized generator G In this process, the original image domain is obtained. The first original image The converted ultrasound image.

Images