Medical ultrasonic image noise reduction method and system based on self-supervised noise disturbance

Abstract

The invention discloses a medical ultrasonic image noise reduction method and system based on self-supervised noise disturbance, and the method comprises the steps: constructing a self-supervised ultrasonic image noise reduction network model through gradually introducing controllable synthetic noise disturbance terms; obtaining an original medical ultrasonic image containing noise, and carrying out the iterative training of the self-supervised ultrasonic image noise reduction network model, and obtaining a trained self-supervised ultrasonic image noise reduction network model; and based on the trained self-supervised ultrasonic image noise reduction network model, performing image noise reduction processing on an original noisy medical ultrasonic image acquired clinically to obtain a denoised medical ultrasonic image. According to the invention, noise and effective information of the image can be distinguished, so that the noise reduction effect of the original medical ultrasonic image is improved. The medical ultrasonic image noise reduction method and system based on self-supervised noise disturbance can be widely applied to the technical field of medical image noise reduction.

Description

Technical Field

[0001] This invention relates to the field of medical image denoising technology, and in particular to a method and system for medical ultrasound image denoising based on self-supervised noise perturbation. Background Technology

[0002] Due to physical mechanisms (such as sound wave scattering and sensor noise) and the influence of the operating environment, medical ultrasound images generally suffer from low signal-to-noise ratios and complex and irregular noise types, which seriously affect the detail recognition of tissue structure edges and small lesions in the images. Therefore, appropriately reducing noise in ultrasound images and improving the clarity of tissue structure boundaries in ultrasound images still plays a crucial role in the accuracy of doctors' visual judgment or computer image analysis. In related technologies, existing supervised learning methods rely on the combination of artificially synthesized noisy images and noise-free images for training. However, it is impossible to obtain real ideal noise-free images in clinical scenarios, resulting in the noise reduction capability of the model in practical applications not being as strong as expected in theoretical applications. In addition, most existing filtering and noise reduction models are designed for a single assumed noise type, which not only makes it difficult to work with composite noise (such as Gaussian-Poisson mixed noise, motion artifacts, etc.) in clinical applications, but also results in unsatisfactory suppression effects. Furthermore, it requires continuous experimental adjustment and selection of appropriate filtering parameters to achieve the best noise reduction effect. Summary of the Invention

[0003] To address the aforementioned technical problems, the present invention aims to provide a medical ultrasound image denoising method and system based on self-supervised noise perturbation, which can distinguish between noise and effective image information, thereby improving the denoising effect of the original medical ultrasound image.

[0004] The first technical solution adopted in this invention is: a medical ultrasound image denoising method based on self-supervised noise perturbation, comprising the following steps:

[0005] By gradually introducing controllable synthetic noise perturbation terms, a self-supervised ultrasound image denoising network model is constructed.

[0006] The self-supervised ultrasound image denoising network model is iteratively trained by acquiring the original medical ultrasound image containing noise, and the trained self-supervised ultrasound image denoising network model is obtained.

[0007] Based on a trained self-supervised ultrasound image denoising network model, image denoising is performed on the original noisy medical ultrasound images acquired clinically to obtain denoised medical ultrasound images.

[0008] Furthermore, the self-supervised ultrasound image denoising network model specifically includes an encoder and a decoder, wherein the encoder and the decoder are connected via skip connections, wherein:

[0009] Both the encoder and the decoder include four levels of convolutional blocks, and each level of convolutional block includes a convolutional layer, an activation function, and a pooling layer;

[0010] Each convolutional block of the encoder and each convolutional block of the decoder are connected by a skip connection, and at each skip connection, a controllable synthetic noise perturbation term is dynamically injected.

[0011] Furthermore, the step of iteratively training the self-supervised ultrasound image denoising network model using the original noisy medical ultrasound image to obtain the trained self-supervised ultrasound image denoising network model specifically includes:

[0012] The original medical ultrasound image containing noise is acquired and normalized to obtain the normalized original medical ultrasound image.

[0013] The normalized original medical ultrasound image is input into the self-supervised ultrasound image denoising network model;

[0014] An encoder based on a self-supervised ultrasound image denoising network model performs four levels of downsampling on the normalized original medical ultrasound image to obtain medical ultrasound image features.

[0015] Based on a self-supervised ultrasound image denoising network model, a controllable synthetic noise perturbation term is used to add simulated clinical multi-source noise interference to the medical ultrasound image features, resulting in medical ultrasound image features with added noise.

[0016] A decoder based on a self-supervised ultrasound image denoising network model gradually restores the features of medical ultrasound images after noise has been added, generating a denoised image.

[0017] Based on the minimum mean square error loss function, the difference between the denoised image and the original medical ultrasound image containing noise is calculated. If the difference does not meet the preset threshold, the self-supervised ultrasound image denoising network model is retrained until the difference meets the preset threshold, thus obtaining the trained self-supervised ultrasound image denoising network model.

[0018] Furthermore, the specific expression for adding simulated clinical multi-source noise interference to the medical ultrasound image features is as follows:

[0019]

[0020] In the above formula, F noisy This represents the features of a medical ultrasound image after adding noise perturbation, where η represents the mixing coefficient, λ represents the Poisson noise intensity, and F... encoder The encoder extracts medical ultrasound image features, where N(·) represents the standard normal distribution and Poisson(·) represents the Poisson distribution.

[0021] Furthermore, the specific expression for progressively restoring the features of the medical ultrasound image after adding noise is as follows:

[0022] I denoise =P decoder (P encoder (I noise )+F noise )

[0023] In the above formula, I denoise I represents the image after noise reduction. noise P represents a clinically acquired medical ultrasound image containing unknown noise. encoder This indicates that I is extracted using an encoder. noise Image feature operations, F noise P represents the features of a medical ultrasound image after adding known noise perturbations. decoder This indicates the operation of extracting medical ultrasound image features using a decoder.

[0024] Furthermore, the step of performing image denoising processing on the original noisy medical ultrasound images acquired clinically, based on the trained self-supervised ultrasound image denoising network model, to obtain denoised medical ultrasound images, specifically includes:

[0025] The original noisy medical ultrasound images acquired in clinical settings are obtained and normalized to obtain normalized original noisy medical ultrasound images acquired in clinical settings.

[0026] An encoder based on a self-supervised ultrasound image denoising network model extracts features from normalized original noisy medical ultrasound images acquired in clinical settings to obtain medical ultrasound image features.

[0027] A decoder based on a self-supervised ultrasound image denoising network model is used to denoise medical ultrasound image features to obtain denoised medical ultrasound images.

[0028] The second technical solution adopted in this invention is: a medical ultrasound image denoising system based on self-supervised noise perturbation, comprising:

[0029] The first module is used to construct a self-supervised ultrasound image denoising network model by gradually introducing controllable synthetic noise perturbation terms;

[0030] The second module is used to acquire the original medical ultrasound images containing noise and iteratively train the self-supervised ultrasound image denoising network model to obtain the trained self-supervised ultrasound image denoising network model.

[0031] The third module is used to perform image denoising on the original noisy medical ultrasound images acquired clinically, based on the trained self-supervised ultrasound image denoising network model, to obtain denoised medical ultrasound images.

[0032] The beneficial effects of the method and system of this invention are as follows: This invention constructs a self-supervised ultrasound image denoising network model by gradually introducing controllable synthetic noise perturbation terms. During the skip connection process, controllable synthetic noise perturbation is gradually added, and during the decoding process, noise is gradually removed and a restored image similar to the original image is generated. This results in a self-supervised denoising network with training noise feature information. The design of adding noise perturbation can, to some extent, solve the limitation of missing noise-free image data. Further, original medical ultrasound images containing noise are obtained to iteratively train the self-supervised ultrasound image denoising network model, resulting in a trained model. Finally, based on the trained model, image denoising processing is performed on original noisy medical ultrasound images acquired clinically, resulting in denoised medical ultrasound images. Using the trained self-supervised denoising network to encode and decode noisy original clinical medical ultrasound images automatically removes noise with similar features learned from the perturbation noise, achieving the purpose of denoising. This enables the distinction between noise and effective image information, thereby improving the denoising effect of the original medical ultrasound image. Attached Figure Description

[0033] Figure 1 This is a flowchart of the steps of a medical ultrasound image denoising method based on self-supervised noise perturbation according to the present invention;

[0034] Figure 2 This is a structural block diagram of a medical ultrasound image noise reduction system based on self-supervised noise perturbation according to the present invention;

[0035] Figure 3 This is a schematic diagram of the structure of the self-supervised ultrasound image denoising network model provided in a specific embodiment of the present invention;

[0036] Figure 4 This is a schematic diagram illustrating the performance evaluation results of the noise reduction model provided in a specific embodiment of the present invention;

[0037] Figure 5 This is a schematic diagram illustrating a visual comparison between the denoised image and the image before denoising, provided in a specific embodiment of the present invention. Detailed Implementation

[0038] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0039] First, it should be noted that the current mainstream noise reduction techniques can be theoretically divided into two types: traditional image noise reduction algorithms and deep learning noise reduction algorithms.

[0040] Limitations of traditional image denoising algorithms:

[0041] 1) Non-Local Means (NLM): This algorithm suppresses noise by weighting similar pixel blocks, but its search window is fixed (usually 15×15 pixels) and assumes that the noise is locally uniformly distributed. However, in actual clinical applications, noise in medical ultrasound images has non-stationary characteristics (such as a sharp drop in signal-to-noise ratio due to sound wave attenuation in deep tissues), and the fixed window cannot be dynamically adapted, resulting in noise residue or loss of high-frequency details.

[0042] 2) Wavelet thresholding for noise reduction: This method eliminates noise based on the sparsity separation of frequency domain signals. However, in practical applications, the selection of the threshold is often difficult and the computational efficiency is low. In addition, some improved basis functions (such as Daubechies wavelets) rely on experience and have limited ability to suppress dynamic motion artifacts (such as blurring caused by heartbeats), making them difficult to adapt to more complex noise scenarios.

[0043] 3) Anisotropic diffusion filtering: While it preserves edges using gradient information, under high-intensity noise (such as sensor electronic noise), gradient calculation errors accumulate, producing "staircase effect" artifacts (jagged tissue boundaries). Furthermore, diffusion coefficient adjustment relies on manual intervention, making it difficult to handle multimodal noise (such as Gaussian-Poisson mixed noise) and lacking robustness.

[0044] The bottleneck of deep learning noise reduction algorithms:

[0045] 1) Data dependency paradox: Current mainstream supervised learning model methods require precisely paired combinations of noise-free and noisy data for supervised training. However, obtaining ideal noise-free ultrasound images in clinical practice is impossible (due to patient breathing and movement causing non-repeatable scans, machine noise, etc.), resulting in poor model performance in real-world scenarios, often failing to achieve the expected results.

[0046] 2) Shortcomings in noise adaptability: Existing deep learning noise reduction models are mostly designed and developed for ultrasound images of a single part of the human body, while ultrasound images collected in actual clinical environments may include different parts such as the heart and abdomen, which leads to a decrease in the generalization performance of the model.

[0047] Therefore, traditional image denoising algorithms struggle to balance noise reduction and detail preservation, often resulting in a trade-off between the two, and they also lack adaptability to non-stationary noise. Deep learning denoising algorithms, on the other hand, are limited by the data dependency paradox and noise complexity, exhibiting weak cross-device generalization ability. The following analysis uses representative algorithms as examples to illustrate their respective limitations.

[0048] Based on this, this invention proposes a self-supervised ultrasound image denoising method, aiming to overcome the limitations of the prior art and automatically reduce noise in ultrasound images to the greatest extent possible without manually adjusting any filtering parameters. Specifically, it utilizes a dynamic noise modeling and decoupled learning framework to suppress multi-source composite noise without the need for noise-free image supervision, while preserving key anatomical details relevant to diagnosis (such as microcalcifications and vascular edges). This method significantly reduces dependence on pure, noise-free data, bypasses the analysis of specific noise types, and thus avoids the need for prior learning of noise-free ultrasound images and noise types. This improves the robustness of the denoising model in real clinical scenarios, providing a purer and more reliable image representation for accurate analysis based on medical ultrasound images.

[0049] Reference Figure 1 This invention provides a method for denoising medical ultrasound images based on self-supervised noise perturbation, the method comprising the following steps:

[0050] S100. By gradually introducing controllable synthetic noise perturbation terms, a self-supervised ultrasound image denoising network model is constructed.

[0051] Specifically, the self-supervised ultrasound image denoising network model includes an encoder and a decoder, which are connected by skip connections. Both the encoder and the decoder include four levels of convolutional blocks, each of which includes a convolutional layer, an activation function, and a pooling layer. Each level of convolutional block of the encoder and each level of convolutional block of the decoder are connected by skip connections, and at each skip connection, a controllable synthetic noise perturbation term is dynamically injected.

[0052] In this embodiment, the present invention proposes a self-supervised ultrasound image denoising network, the core architecture of which is based on an improved U-Net network, such as... Figure 3 As shown, it mainly consists of two parts: an encoder and a decoder. Skip connections are used to fuse feature maps from different stages of the encoder with feature maps from corresponding stages of the decoder. The encoder consists of multiple repeating convolutional blocks, each consisting of two 3×3 convolutional layers, followed by an activation function. Finally, a 2×2 max-pooling layer halves the spatial dimensions (width and height) of the feature map at that layer. The decoder upsamples the low-resolution feature map back to high resolution and concatenates the feature maps from the corresponding layers (with the same spatial resolution) in the encoder path onto the upsampled feature map, further refining the features through convolution.

[0053] The core innovation lies in gradually introducing controllable synthetic noise perturbations during the skip connection process, and progressively removing noise during decoding to generate a restored image similar to the original image, thereby obtaining a self-supervised denoising network with training noise feature information. Then, using this trained self-supervised denoising network to encode and decode noisy clinical ultrasound images automatically removes noise with similar features learned from the perturbation noise, achieving the purpose of noise reduction.

[0054] Adding noise perturbation to the original clinical ultrasound images with noise can, to some extent, address the limitation of missing noise-free image data. By using a multi-layer convolutional neural network to autonomously learn noise features, the threshold can be adjusted without being limited by human experience, thus enabling the distinction between noise and effective image information, thereby overcoming the two important shortcomings of the aforementioned background technologies.

[0055] S200. Obtain the original medical ultrasound image containing noise and iteratively train the self-supervised ultrasound image denoising network model to obtain the trained self-supervised ultrasound image denoising network model.

[0056] S210. Acquire the original medical ultrasound image containing noise and perform normalization processing to obtain the normalized original medical ultrasound image.

[0057] S220. Input the normalized original medical ultrasound image into the self-supervised ultrasound image denoising network model.

[0058] S230: An encoder based on a self-supervised ultrasound image denoising network model performs four-level downsampling processing on the normalized original medical ultrasound image to obtain medical ultrasound image features.

[0059] In this embodiment, the input to the encoder is a single raw ultrasound image I containing noise. noisy The image features F are then normalized and extracted hierarchically through four levels of downsampling (each level containing two convolutional layers + ReLU activation). encodr .

[0060] S240. Based on a self-supervised ultrasound image denoising network model, a controllable synthetic noise perturbation term is used to add simulated clinical multi-source noise interference to the medical ultrasound image features, resulting in medical ultrasound image features with added noise.

[0061] In this embodiment, for the noise generator, at each skip connection, controllable synthesized noise (such as Gaussian-Poisson mixed noise) that matches the statistical characteristics of real noise is dynamically injected. For each level of image feature F... encoder The expression for incorporating simulated clinical multi-source noise interference is:

[0062]

[0063] In the above formula, F noisy This represents the features of a medical ultrasound image after adding noise perturbation, where η represents the mixing coefficient, λ represents the Poisson noise intensity, and F... encoder The encoder extracts medical ultrasound image features, where N(·) represents the standard normal distribution and Poisson(·) represents the Poisson distribution.

[0064] S250, a decoder based on a self-supervised ultrasound image denoising network model, progressively restores the features of medical ultrasound images after noise has been added, generating a denoised image;

[0065] In this embodiment, the decoder gradually recovers and generates the denoised image F through four levels of upsampling (deconvolution + feature concatenation), and its expression is:

[0066] I denoise =P decoder (P encoder (I noise )+F noise )

[0067] In the above formula, I denoise I represents the image after noise reduction. noise P represents a clinically acquired medical ultrasound image containing unknown noise. encoder This indicates that I is extracted using an encoder. noise Image feature operations, F noise P represents the features of a medical ultrasound image after adding known noise perturbations. decoder This indicates the operation of extracting medical ultrasound image features using a decoder.

[0068] S260. Based on the minimum mean square error loss function, calculate the difference between the denoised image and the original medical ultrasound image containing noise. If the difference does not meet the preset threshold, retrain the self-supervised ultrasound image denoising network model until the difference meets the preset threshold, and obtain the trained self-supervised ultrasound image denoising network model.

[0069] In this embodiment, the difference between the denoised and restored image and the original image is calculated, which is the loss function. The loss function used here is the minimum mean squared error (MSE). When the loss function is minimized, the optimal denoising network parameters are obtained, and the training process ends.

[0070] S300, based on a trained self-supervised ultrasound image denoising network model, performs image denoising processing on the original noisy medical ultrasound images acquired clinically, and obtains denoised medical ultrasound images.

[0071] Specifically, the original noisy medical ultrasound images acquired clinically are acquired and normalized to obtain normalized original noisy medical ultrasound images acquired clinically; an encoder based on a self-supervised ultrasound image denoising network model extracts features from the normalized original noisy medical ultrasound images acquired clinically to obtain medical ultrasound image features; a decoder based on a self-supervised ultrasound image denoising network model denoises the medical ultrasound image features to obtain denoised medical ultrasound images.

[0072] In this embodiment, a trained image denoising network is used, with only the original noisy ultrasound image I acquired clinically as input. noisy And then normalize it.

[0073] After the encoder, features are extracted from the noise-free image, and the noise decoupling capability learned during the training phase is used to recover the denoised image, which is expressed as:

[0074] I denoise =F decoder (F encoder (I noise ))

[0075] Although the type of mixed noise contained in the original noisy ultrasound image is unknown, by using the preceding noise perturbation, the model can autonomously learn the characteristics of the noise signal and decouple and remove noise, thus still achieving the effect of self-supervised noise reduction.

[0076] Finally, the performance of the denoising model is evaluated. In the fields of image processing and signal denoising, SSIM, PSNR and SNR are key indicators for measuring algorithm performance. The denoising effect of the technical solution proposed in this embodiment is evaluated from the perspectives of numerical error and visual perception.

[0077] Signal-to-noise ratio (SNR): A measure of the power ratio of a signal to noise. The noise intensity is quantified; a higher value indicates a better noise suppression effect.

[0078] Peak signal-to-noise ratio (PSNR): Based on the mean square error (MSE), PSNR is calculated using the maximum pixel value of the image (e.g., 255) as a reference. PSNR = 10log 10 (peak) 2 / MSE)dB. A higher value indicates less distortion.

[0079] Structural Similarity (SSIM): This method evaluates the local similarity between a denoised image and the original image in terms of brightness, contrast, and structure. A comprehensive score (0-1, with values ​​closer to 1 indicating higher quality) is calculated using statistical features (mean, variance, and covariance). Its advantage lies in its alignment with the human visual system, better reflecting detail preservation and structural consistency.

[0080] Furthermore, the performance of the model constructed in this invention is evaluated, specifically as follows: Figure 4 as well as Figure 5 As shown, where, Figure 5 (a) in the image is the original noisy ultrasound image. Figure 5 Image (b) is the denoised image processed by the method of the present invention, demonstrating the noise suppression and detail preservation effects. Figure 5 (c) in the middle is Figure 5 The details of (a) in the text are shown. Figure 5 (d) in the middle is Figure 5 The details of (b) are shown in the image.

[0081] In summary, the embodiments of the present invention have the following advantages over the prior art:

[0082] 1) The technical solution proposed in this invention demonstrates good noise reduction effect in both visual and quantitative evaluation indicators, reduces the interference of noise signals, increases the proportion of diagnostically significant signals, and can provide cleaner image data for subsequent image processing and analysis.

[0083] 2) The technical solution proposed in this invention does not rely on human experience, nor does it require clean and noise-free images for training. It belongs to fully self-supervised learning, which has lower requirements for its use, is easier to implement in practical applications, and has a wider range of applications.

[0084] Reference Figure 2 A medical ultrasound image denoising system based on self-supervised noise perturbation, comprising:

[0085] The first module 201 is used to construct a self-supervised ultrasound image denoising network model by gradually introducing controllable synthetic noise perturbation terms;

[0086] The second module 202 is used to acquire the original medical ultrasound image containing noise and iteratively train the self-supervised ultrasound image denoising network model to obtain the trained self-supervised ultrasound image denoising network model.

[0087] The third module 203 is used to perform image denoising on the original noisy medical ultrasound images acquired clinically, based on the trained self-supervised ultrasound image denoising network model, to obtain denoised medical ultrasound images.

[0088] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0089] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A medical ultrasound image denoising method based on self-supervised noise perturbation, characterized in that, Includes the following steps: By gradually introducing controllable synthetic noise perturbation terms, a self-supervised ultrasound image denoising network model is constructed. The self-supervised ultrasound image denoising network model is iteratively trained by acquiring the original medical ultrasound image containing noise, and the trained self-supervised ultrasound image denoising network model is obtained. Based on a trained self-supervised ultrasound image denoising network model, image denoising is performed on the original noisy medical ultrasound images acquired clinically to obtain denoised medical ultrasound images.

2. The self-supervised noise-disturbance-based medical ultrasound image denoising method according to claim 1, wherein, The self-supervised ultrasound image denoising network model specifically includes an encoder and a decoder, with the encoder and decoder connected via skip connections, wherein: Both the encoder and the decoder include four levels of convolutional blocks, and each level of convolutional block includes a convolutional layer, an activation function, and a pooling layer; Each convolutional block of the encoder and each convolutional block of the decoder are connected by a skip connection, and at each skip connection, a controllable synthetic noise perturbation term is dynamically injected.

3. The method of claim 2, wherein, The step of iteratively training the self-supervised ultrasound image denoising network model by acquiring the original noisy medical ultrasound image to obtain the trained self-supervised ultrasound image denoising network model specifically includes: The original medical ultrasound image containing noise is acquired and normalized to obtain the normalized original medical ultrasound image. The normalized original medical ultrasound image is input into the self-supervised ultrasound image denoising network model; An encoder based on a self-supervised ultrasound image denoising network model performs four levels of downsampling on the normalized original medical ultrasound image to obtain medical ultrasound image features. Based on a self-supervised ultrasound image denoising network model, a controllable synthetic noise perturbation term is used to add simulated clinical multi-source noise interference to the medical ultrasound image features, resulting in medical ultrasound image features with added noise. A decoder based on a self-supervised ultrasound image denoising network model gradually restores the features of medical ultrasound images after noise has been added, generating a denoised image. Based on the minimum mean square error loss function, the difference between the denoised image and the original medical ultrasound image containing noise is calculated. If the difference does not meet the preset threshold, the self-supervised ultrasound image denoising network model is retrained until the difference meets the preset threshold, thus obtaining the trained self-supervised ultrasound image denoising network model.

4. The method of claim 3, wherein, The specific expression for adding simulated clinical multi-source noise interference to the medical ultrasound image features is as follows: In the above formula, F noisy denotes the medical ultrasound image feature after adding noise disturbance, η denotes the mixing coefficient, λ denotes the Poisson noise intensity, F encoder denotes the medical ultrasound image feature extracted by the encoder, N(·) denotes the standard normal distribution, and Poisson(·) denotes the Poisson distribution.

5. The self-supervised noise-disturbance-based medical ultrasound image denoising method of claim 4, wherein, The specific expression for progressively restoring the features of a medical ultrasound image after adding noise is as follows: I denoise = P decoder (P encoder (I noise ) + F noise ) In the above formula, I denoise denotes the image after noise reduction, I noise denotes the medical ultrasound image containing unknown noise collected in clinic, P encoder denotes the image feature operation of extracting I noise using the encoder, F noise denotes the medical ultrasound image feature after adding known noise disturbance, P decoder denotes the medical ultrasound image feature extraction operation using the decoder.

6. The method of claim 5, wherein, The step of using the trained self-supervised ultrasound image denoising network model to denoise the original noisy medical ultrasound images acquired clinically, and obtaining denoised medical ultrasound images, specifically includes: The original noisy medical ultrasound images acquired in clinical settings are obtained and normalized to obtain normalized original noisy medical ultrasound images acquired in clinical settings. An encoder based on a self-supervised ultrasound image denoising network model extracts features from normalized original noisy medical ultrasound images acquired in clinical settings to obtain medical ultrasound image features. The decoder of the self-supervised ultrasonic image denoising network model is used for denoising medical ultrasonic image features to obtain a denoised medical ultrasonic image.

7. A medical ultrasound image denoising system based on self-supervised noise perturbation, characterized in that, Comprise the following modules: The first module is used for constructing a self-supervised ultrasonic image denoising network model by gradually introducing controllable synthetic noise disturbance terms; The second module is used for obtaining an original medical ultrasonic image containing noise, iteratively training the self-supervised ultrasonic image denoising network model, and obtaining a trained self-supervised ultrasonic image denoising network model; The third module is used for performing image denoising processing on a clinically collected original noisy medical ultrasonic image based on the trained self-supervised ultrasonic image denoising network model to obtain a denoised medical ultrasonic image.

Images