Ultrasonic image synthesis method and system
Through the ultrasound image synthesis method, the potential diffusion model is used to realize the diffusion process in the latent space, and high-quality synthetic ultrasound images are generated, which solves the privacy protection problem in medical data sharing and realizes secure sharing and big data utilization.
Patent Information
- Application Number
- CN202411674091.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-21
- Publication Date
- 2025-04-15
AI Technical Summary
The existing technology has problems with patient privacy protection in medical data sharing, and it is difficult to generate high-quality synthetic medical image data without revealing privacy, which limits the secure sharing of medical data and the effective utilization of big data.
Ultrasonic image synthesis methods are adopted, including image preprocessing, potential diffusion model construction and training, synthetic image generation and evaluation, and the forward and reverse diffusion processes are realized in the latent space through the potential diffusion model, synthetic ultrasonic images are generated, and the quality, privacy protection performance and practicality of the synthetic image are evaluated through the evaluation module.
It effectively reduces computing costs, supports condition generation, ensures flexibility and controllability of image synthesis, improves image synthesis quality, ensures privacy protection performance, and supports the secure sharing of medical data and the effective utilization of big data.
Smart Images
Figure CN119832101A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of ultrasonic technology, and more particularly to an ultrasonic image synthesis method and system. Background Art
[0002] With the rapid development of medical big data, medical imaging data plays an increasingly important role in clinical diagnosis and medical research. However, the issue of patient privacy protection seriously restricts the possibility of cross-center data sharing and the construction of diverse and large-scale data sets. Existing data sharing methods mainly rely on data de-identification and differential privacy technologies, but these methods have the problems of re-identification risk and decreased data utility. Although federated learning can train models without sharing original data, it still faces the risk of privacy leakage during model updates. Therefore, there is an urgent need for a method that can generate high-quality synthetic medical imaging data while protecting patient privacy, so as to promote the secure sharing of medical data and the effective use of big data.
[0003] Based on this, the present invention provides an ultrasonic image synthesis method and system. Summary of the invention
[0004] The purpose of the present invention is to provide an ultrasound image synthesis method and system in order to overcome the defects of the above-mentioned prior art, and to solve the above-mentioned data sharing limitations through a system including an image preprocessing module, an image generation module, and a synthetic image evaluation module.
[0005] The purpose of the present invention can be achieved by the following technical solutions:
[0006] The first object of the present invention is to provide an ultrasonic image synthesis method, comprising the following steps:
[0007] S1. Preprocessing the ultrasonic image, wherein the preprocessing includes size unification and image normalization to obtain a preprocessed ultrasonic image;
[0008] S2. constructing a potential diffusion model based on the preprocessed ultrasound image, and training a deep generative model based on the potential diffusion model to generate a synthetic ultrasound image;
[0009] S3. Evaluate the synthesis quality, privacy protection performance and practicality of the synthesized ultrasound image.
[0010] Further, step S2 includes the following process:
[0011] S21: Constructing a potential diffusion model:
[0012] S211: The preprocessed ultrasound image is input to the encoder. After being processed by the encoder, the input image x0 is mapped to the latent space to obtain the latent vector z0. The encoding process can be expressed as formula (1): z0 = f encode (x0);
[0013] S212: Forward-noise the latent vector z0 obtained in step S211 in the latent space to obtain the noisy latent vector z t , the noise adding process can be expressed as formula (2): in, is a hyperparameter, ∈ t is Gaussian noise, t=1,…,T;
[0014] S213: In the latent space, the noisy latent vector z obtained in step S212 is calculated. t The back diffusion process is performed to gradually predict and remove the noise added at each time step using a neural network, and the Breast Imaging Reporting and Data System (BI-RADS) classification labels of breast ultrasound images are introduced as the input of the denoising module to construct a latent diffusion model training set.
[0015] S22: Train the latent diffusion model until the model error converges, and obtain the deep generative model through training;
[0016] S23: Using the deep generative model in step S22, the latent vector of the synthetic ultrasound image is obtained, and the synthetic ultrasound image is obtained through the decoder. The decoding process of the decoder can be expressed as formula (3): c =f decode (z c ), where z c is a latent vector that meets the conditions, x c The synthetic ultrasound images meet the requirements.
[0017] Furthermore, step S1 includes the following process:
[0018] Size normalization: scaling the ultrasound image in equal proportions, and then cropping the center of the scaled ultrasound image into a square to obtain a size normalized ultrasound image;
[0019] Image normalization: The size-normalized ultrasound image data is normalized to obtain a preprocessed ultrasound image.
[0020] Furthermore, step S1 includes the following process:
[0021] Resize: Use the Resize function in the torchvision.transforms library to scale the ultrasound image so that its shorter side is 512 pixels. Then use the CenterCrop function to crop the center of the scaled ultrasound image into a 512*512 pixel square to obtain a resized ultrasound image.
[0022] Image normalization: Use the Normalize function in the torchvision.transforms library to normalize the resized ultrasound image data to obtain the preprocessed ultrasound image.
[0023] Furthermore, in step S213, the denoising module uses DiT (Diffusion Transformer) to predict the noise in the reverse diffusion process, and the noisy latent vector z t Pachify the image and get a 2*2 image block embedding, which is used as the input of DiT together with the embedding vector of the time step and BI-RADS classification label.
[0024] Furthermore, in step S22, the loss function of the model training is the mean square error loss of the predicted noise and the actual added noise at all time steps. The mean square error loss of each time step is expressed as formula (4): Among them, ∈ θ (z t ) is the prediction noise of the current time step, ∈ t is the noise actually added in the current time step, t=1,…,T.
[0025] Furthermore, in step S23, the process of obtaining the latent vector of the synthetic ultrasound image by using the diffusion model trained in step S22 is as follows:
[0026] A denoising diffusion implicit model (DDIM) sampler is used, and classifier-free guidance control conditional generation is used to generate synthetic ultrasound images of different BI-RADS classifications. No explicit classifier is required for classifier-free guidance control conditional generation, and only the unconditional noise prediction module and the conditional noise prediction module need to be trained simultaneously.
[0027] Furthermore, the ultrasound image is a breast ultrasound image, and the synthetic ultrasound image is a synthetic breast ultrasound image.
[0028] Furthermore, in step S23, (after training is completed) during inference, the adjusted noise prediction without classifier guidance is used, which is expressed as formula (5): Among them, ω is the intensity of conditional control, ∈θ (z t ) is the unconditional noise prediction, ∈ θ (z t ,c) is the conditional noise prediction, is the adjusted noise prediction without classifier guidance.
[0029] Further, step S3 includes the following process:
[0030] Synthesis quality assessment: quantitatively assess the quality of synthesized ultrasound images using an algorithm that measures image authenticity; assess the authenticity and BI-RADS classification of synthesized ultrasound images;
[0031] Privacy protection performance evaluation: extract synthetic ultrasound images, use the encoder to extract their 512-dimensional feature vectors, calculate the cosine similarity of each synthetic ultrasound image with all the synthetic ultrasound images in the model training set in step S21 to determine its nearest neighbor in the model training set, and visually evaluate each synthetic ultrasound image and its nearest neighbor to determine whether each synthetic ultrasound image and its nearest neighbor are visually consistent;
[0032] Practical evaluation: A breast ultrasound image classification model was trained using synthetic ultrasound images and real ultrasound images, and the classification performance was compared with that of a breast ultrasound image classification model that only used real ultrasound images as a training set.
[0033] Furthermore, algorithms for measuring the authenticity of images include Inception Score (IS) and Fréchet Inception Distance (FID).
[0034] Furthermore, several ultrasound residents and several senior ultrasound physicians with at least 10 years of clinical experience evaluated the authenticity and BI-RADS classification of the synthetic ultrasound images.
[0035] Furthermore, one or more senior ultrasound physicians with at least 10 years of clinical experience visually evaluated each synthetic ultrasound image and its nearest neighbors.
[0036] A second object of the present invention is to provide an ultrasonic image synthesis system for implementing the ultrasonic image synthesis method, comprising:
[0037] An image preprocessing module is used for preprocessing ultrasound images, and the preprocessing steps include size unification and image normalization;
[0038] An image generation module, connected to the data preprocessing module, for constructing a potential diffusion model, training a deep generative model based on the potential diffusion model, and generating a synthetic ultrasound image;
[0039] The synthetic image evaluation module is connected to the image generation module and is used to evaluate the synthesis quality, privacy protection performance and practicality of the synthetic image.
[0040] Furthermore, the image generation module includes a denoising module.
[0041] Compared with the prior art, the present invention has the following beneficial effects:
[0042] 1) The ultrasonic image synthesis method and system provided by the present technical solution adopt a latent diffusion model framework to realize forward and reverse diffusion processes in the latent space, which effectively reduces the computational cost and supports conditional generation to ensure the flexibility and controllability of image synthesis.
[0043] 2) The present technical solution provides an ultrasound image synthesis method and system. Compared with the traditional potential diffusion model, the present invention uses DiT instead of UNet as the backbone structure of the noise prediction module. Transformer is known for its excellent long-range dependency modeling ability, which enables the model to generate more consistent and detailed images, effectively improving the image synthesis quality.
[0044] 3) The present invention provides an ultrasound image synthesis method and system. The present invention adopts a classifier-free guidance strategy in the model inference stage to achieve conditional generation of breast ultrasound images of different BI-RADS classifications. In the medical field, this function is particularly important. Generating images suitable for specific clinical scenarios helps support accurate diagnosis and treatment planning under the medical data sharing framework and provides a reliable data source.
[0045] 4) The ultrasound image synthesis method and system provided by the present technical solution is a systematic synthetic medical image evaluation method, including synthesis quality, privacy protection performance and practicality evaluation, which is conducive to ensuring the effectiveness and security of the synthetic image. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 is a structural block diagram of the ultrasonic image synthesis system of the present invention;
[0047] Figure 2 This is a working principle diagram of the image generation module of the present invention;
[0048] Figure 3 The DiT module working principle diagram of the denoising module in the image generation module of the present invention;
[0049] Figure 4 A breast ultrasound image generated in an embodiment of the present invention;
[0050] Figure 5 This is a flow chart of a doctor's evaluation of the quality of a synthesized image in an embodiment of the present invention. DETAILED DESCRIPTION
[0051] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention.
[0052] Features such as component models, material names, connection structures, control methods, etc. that are not clearly stated in this technical solution shall be deemed as common technical features disclosed in the prior art.
[0053] Example 1
[0054] The present invention provides an ultrasonic image synthesis system and method.
[0055] The structural block diagram of the ultrasound image synthesis system is as follows: Figure 1 As shown, including:
[0056] An image preprocessing module is used for preprocessing ultrasound images, and the preprocessing steps include size unification and image normalization;
[0057] an image generation module, connected to the data preprocessing module, for constructing a deep generative model, training the model based on the preprocessed ultrasound image, and generating a synthetic ultrasound image, Figure 2 The working principle of the image generation module of the present invention is shown. The input image is processed by the encoder to obtain the latent vector z0. The forward diffusion process includes T steps of noise addition, and z0 is obtained from z0. T During the back diffusion process, the denoising module predicts and removes the noise added at each time step, and the noise z sampled during inference s ′ After S reasoning steps, we get z0 ′ . Finally, z0 ′ Decode through decoder D and output the synthesized image;
[0058] The synthetic image evaluation module is connected to the image generation module and is used to evaluate the synthesis quality, privacy protection performance and practicality of the synthetic image.
[0059] The ultrasonic image synthesis method comprises the following steps:
[0060] Preprocessing ultrasound images, including size unification and image normalization;
[0061] Build a deep generative model, train the model based on preprocessed ultrasound images, and generate synthetic ultrasound images;
[0062] The synthesis quality, privacy protection performance and practicality of the synthesized image are evaluated.
[0063] The ultrasonic image synthesis method specifically comprises the following steps:
[0064] S1. Preprocessing the ultrasound image, wherein the preprocessing includes size unification and image normalization to obtain a preprocessed ultrasound image, specifically including the following process:
[0065] Resize: Use the Resize function in the torchvision.transforms library to scale the ultrasound image so that its shorter side is 512 pixels. Then use the CenterCrop function to crop the center of the image to 512*512 pixels to obtain a resized ultrasound image.
[0066] Image normalization: Use the Normalize function in the torchvision.transforms library to normalize the resized ultrasound image data.
[0067] S2. Based on the preprocessed ultrasound image, a potential diffusion model is constructed, and based on the potential diffusion model, a deep generative model is trained to generate a synthetic ultrasound image, which specifically includes the following process:
[0068] S21: Constructing a potential diffusion model:
[0069] S211: The preprocessed ultrasound image is input to the encoder. After being processed by the encoder, the input image x0 is mapped to the latent space to obtain the latent vector z0. The encoding process can be expressed as formula (1): z0 = f encode (x0);
[0070] S212: Forward-noise the latent vector z0 obtained in step S211 in the latent space to obtain the noisy latent vector z t , the noise adding process can be expressed as formula (2): in, is a hyperparameter, ∈ t is Gaussian noise, t=1,…,T;
[0071] S213: In the latent space, the noisy latent vector z obtained in step S212 is calculated. t The back diffusion process is performed, and the neural network is used to gradually predict and remove the noise added at each time step. The Breast Imaging Reporting and Data System (BI-RADS) classification labels of breast ultrasound images are introduced as the input of the denoising module to construct a latent diffusion model training set. In the back diffusion process, the denoising module uses DiT (Diffusion Transformer) to predict noise and the noisy latent vector z is tPachify the image blocks to get 2*2 image block embeddings, which are used as the input of DiT together with the embedding vectors of the time step and BI-RADS classification label.
[0072] Figure 3 The working principle diagram of the DiT module of the denoising module in the image generation module of the present invention is shown in Figure 1. Each DiT module uses adaptive layer normalization and initializes the layer normalization parameters to zero (Adaptive Layer Norm-Zero, AdaLN-Zero). The input of this module includes image block embedding and conditional embedding, where the conditional embedding contains the embedding information of the BI-RADS label and time step.
[0073] S22: Train the latent diffusion model until the model error converges, obtain the deep generative model through training, and save the trained deep generative model;
[0074] In this embodiment, when performing step S22, the loss function of the model training is the mean square error loss of the predicted noise and the actual added noise at all time steps. The mean square error loss of each time step is expressed as formula (3): Among them, ∈ θ (z t ) is the prediction noise of the current time step, ∈ t is the noise actually added in the current time step, t=1,…,T.
[0075] S23: Using the deep generative model in step S22, the latent vector of the synthetic ultrasound image is obtained, and the synthetic ultrasound image is obtained through the decoder. The decoding process of the decoder can be expressed as formula (3): c =f decode (z c ), where z c is a latent vector that meets the conditions, x c The synthetic ultrasound images are eligible;
[0076] In this embodiment, when performing step S23, (after training is completed) during inference, a denoised diffusion implicit model (DDIM) sampler is used, and classifier-free guidance control condition generation is used to generate synthetic breast ultrasound images of different BI-RADS classifications. No explicit classifier is required when generating classifier-free guidance control conditions, and only the unconditional noise prediction module and the conditional noise prediction module need to be trained simultaneously. During inference, the adjusted noise prediction under classifier-free guidance is used, which is expressed as formula (5): Among them, ω is the intensity of conditional control, ∈ θ (z t ) is the unconditional noise prediction, ∈ θ (zt ,c) is the conditional noise prediction, is the adjusted noise prediction without classifier guidance.
[0077] In this embodiment, Figure 4 The breast ultrasound images (synthesized ultrasound images) of different BI-RADS classifications synthesized through steps S21, S22 and S23 are shown.
[0078] S3, evaluating the synthesis quality, privacy protection performance and practicality of the synthesized ultrasound image, specifically including the following process:
[0079] Synthetic quality assessment: IS and FID were used to quantitatively evaluate the quality of the synthesized ultrasound images. In addition, three ultrasound residents and three senior ultrasound doctors with at least 10 years of clinical work experience evaluated the authenticity and BI-RADS classification of the synthesized ultrasound images.
[0080] In this embodiment, the calculation method of IS is as shown in formula (6): Among them, x~p g Represents the generated image distribution p g The sampled image, D KL (p||q) represents the KL divergence between distributions p and q, p(y|x) is the conditional distribution, and p(y) = ∫ x p(y|x)p g (x) is the marginal distribution. To facilitate numerical comparison, this embodiment omits the exponential operation in the original formula and directly uses ln(IS(G)), which has no effect on the comparison of the results.
[0081] In this embodiment, the calculation method of FID is as follows: based on the Inception v3 network, the feature representation of the real image and the synthetic image is extracted at the last pooling layer, and then the FID is calculated, as shown in formula (7): 2 (F,G)=|μ X -μ Y | 2 +tr[∑ X +∑ Y -2(∑ X ∑ Y ) 1 / 2 ], where μ X , μ Y are the means of the real image set F and the synthetic image set G, ∑ X and∑ Y are their respective covariance matrices, and take the positive square root of the covariance matrix. n The above formula holds true when is a normal distribution on .
[0082] In this embodiment, Figure 5 The workflow of evaluating the authenticity and BI-RADS classification of synthetic ultrasound images by 3 ultrasound residents and 3 senior ultrasound doctors with at least 10 years of clinical work experience is presented. In the doctor evaluation experiment 1, 3 ultrasound residents and 3 senior ultrasound doctors with at least 10 years of clinical work experience conducted full-image evaluation on a dataset of 500 real ultrasound images and 500 synthetic ultrasound images, and judged each image as a real image or a synthetic image; in the doctor evaluation experiment 2, the 500 real ultrasound images and 500 synthetic ultrasound images used in the doctor evaluation experiment 1 were cropped to obtain the cropped images of the lesion area, and the 3 ultrasound residents and 3 senior ultrasound doctors with at least 10 years of clinical work experience judged each image as a real image or a synthetic image; in the doctor evaluation experiment 3, 3 senior ultrasound doctors with at least 10 years of clinical work experience conducted full-image evaluation on another 500 real ultrasound images and 500 synthetic ultrasound images, and judged each image as BI-RADS level 3 or BI-RADS level 4-5. It should be noted that the data sets used in each evaluation experiment were obtained by random sampling, and the ratio of BI-RADS 3 images to BI-RADS 4-5 images was 1:1.
[0083] Privacy protection performance evaluation: 500 synthetic breast ultrasound images are randomly selected, and their 512-dimensional feature vectors are extracted using an encoder. The cosine similarity of each synthetic ultrasound image with all images in the model training set in step S213 is calculated to determine its nearest neighbor in the training set. A senior ultrasound doctor with at least 10 years of clinical work experience visually evaluates each synthetic ultrasound image and its nearest neighbor to determine whether they are visually consistent.
[0084] Practical evaluation: A breast ultrasound image classification model was trained using synthetic ultrasound images (synthetic breast ultrasound images) and real ultrasound images (real breast ultrasound images), and the classification performance was compared with that of a classification model that only used real ultrasound images as a training set.
[0085] In this embodiment, the specific process of evaluating the practicality of synthetic ultrasound images is as follows: two different data sets are used to train the breast ultrasound image classification model, training set 1 contains 800 real breast ultrasound images, and training set 2 consists of 400 synthetic breast ultrasound images and 400 real breast ultrasound images randomly selected from training set 1. The ratio of BI-RADS 3 images to BI-RADS 4-5 images in each training set is 1:1. The breast ultrasound image classifier adopts the ResNet-50 architecture to predict the BI-RADS category (BI-RADS 3 or BI-RADS 4-5) of breast ultrasound images. During the training process, the parameters of ResNet-50 are iteratively updated through back propagation, and the cross entropy loss function is adopted. The training is set to 50 cycles, the batch size is 32, and the learning rate is fixed to 1×10 -4 , no weight decay was applied. The AdamW algorithm was used for optimization. To evaluate the performance stability of the model on different data subsets, a five-fold cross validation was used, with four folds used for training and one fold used for internal testing. Subsequently, 400 real breast ultrasound images were used as an external test set, and the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and F1 score were calculated to evaluate the model performance.
[0086] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and this document does not limit this.
[0087] The above description of the embodiments is to facilitate the understanding and use of the invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to these embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be within the scope of protection of the present invention.
Claims
1. A method for synthesizing an ultrasonic image, characterized in that: The method comprises the following steps: S1. Preprocessing the ultrasonic image, wherein the preprocessing includes size unification and image normalization to obtain a preprocessed ultrasonic image; S2. constructing a potential diffusion model based on the preprocessed ultrasound image, and training a deep generative model based on the potential diffusion model to generate a synthetic ultrasound image; S3. evaluating the synthesis quality, privacy protection performance and practicality of the synthesized ultrasound image; Step S2 includes the following process: S21: Constructing a potential diffusion model: S211: The preprocessed ultrasound image is input to the encoder. After being processed by the encoder, the input image x0 is mapped to the latent space to obtain the latent vector z0. The encoding process is expressed as formula (1): z0 = f encode (x0); S212: Forward-noise the latent vector z0 obtained in step S211 in the latent space to obtain the noisy latent vector z t , the noise adding process is expressed as formula (2): in, is a hyperparameter, ∈ t is Gaussian noise, t=1,…,T; S213: In the latent space, the noisy latent vector z obtained in step S212 is calculated. t Perform the reverse diffusion process, use the neural network to gradually predict and remove the noise added at each time step, and introduce the BI-RADS classification label as the input of the denoising module to build the potential diffusion model training set; S22: Train the latent diffusion model until the error of the model converges, and obtain the deep generative model through training; S23: Using the deep generative model in step S22, the latent vector of the synthetic ultrasound image is obtained, and the synthetic ultrasound image is obtained through the decoder. The decoding process of the decoder can be expressed as formula (3): c =f decode (z c ), where z c is a latent vector that meets the conditions, x c The synthetic ultrasound images meet the requirements.
2. The ultrasonic image synthesis method according to claim 1, characterized in that: Step S1 includes the following process: Size normalization: scaling the ultrasound image in equal proportions, and then cropping the center of the scaled ultrasound image into a square to obtain a size normalized ultrasound image; Image normalization: The size-normalized ultrasound image data is normalized to obtain a preprocessed ultrasound image.
3. The ultrasonic image synthesis method according to claim 1, characterized in that: In step S213, the denoising module uses DiT to predict the noise in the back diffusion process, and the noisy latent vector z t The image blocks are processed in blocks to obtain image block embedding, which is used as the input of DiT together with the embedding vector of the time step and BI-RADS classification label.
4. The ultrasonic image synthesis method according to claim 1, characterized in that: In step S22, the loss function of model training is the mean square error loss of the predicted noise and the actual added noise at all time steps. The mean square error loss of each time step is expressed as formula (4): Among them, ∈ θ (z t ) is the prediction noise of the current time step, ∈ t is the noise actually added in the current time step, t=1,…,T.
5. The ultrasonic image synthesis method according to claim 1, characterized in that: In step S23, a denoising diffusion implicit model sampler is used and a classifier-free guided control condition is used to generate synthetic ultrasound images of different BI-RADS classifications.
6. The ultrasonic image synthesis method according to claim 5, characterized in that: In step S23, during inference, the adjusted noise prediction without classifier guidance is used, which is expressed as formula (5): Among them, ω is the intensity of conditional control, ∈ θ (z t ) is the unconditional noise prediction, ∈ θ (z t ,c) is the conditional noise prediction, is the adjusted noise prediction without classifier guidance.
7. The ultrasonic image synthesis method according to claim 1, characterized in that: Step S3 includes the following process: Synthesis quality assessment: quantitatively assess the quality of synthesized ultrasound images using an algorithm that measures image authenticity; assess the authenticity and BI-RADS classification of synthesized ultrasound images; Privacy protection performance evaluation: extract synthetic ultrasound images, use the encoder to extract their 512-dimensional feature vectors, calculate the cosine similarity of each synthetic ultrasound image with all the synthetic ultrasound images in the model training set in step S21 to determine its nearest neighbor in the model training set, and visually evaluate each synthetic ultrasound image and its nearest neighbor to determine whether each synthetic ultrasound image and its nearest neighbor are visually consistent; Practical evaluation: A breast ultrasound image classification model was trained using synthetic ultrasound images and real ultrasound images, and the classification performance was compared with that of a breast ultrasound image classification model that only used real ultrasound images as a training set.
8. An ultrasonic image synthesis system for implementing the method according to any one of claims 1 to 7, characterized in that: The system comprises: An image preprocessing module is used for preprocessing ultrasound images, and the preprocessing steps include size unification and image normalization; An image generation module, connected to the data preprocessing module, for constructing a potential diffusion model, training a deep generative model based on the potential diffusion model, and generating a synthetic ultrasound image; The synthetic image evaluation module is connected to the image generation module and is used to evaluate the synthesis quality, privacy protection performance and practicality of the synthetic image.