Noise prediction model training and ultrasonic image denoising method, related device and medium

Abstract

The invention discloses a noise prediction model training and ultrasonic image denoising method, a related device and a medium, and relates to the field of noise prediction.The method comprises the steps that a first noise adding image and a second noise adding image are obtained through a first noise adding algorithm and a second noise adding algorithm, and then a first loss function and a second loss function are generated; wherein the first loss function and the second loss function are mutually independent and jointly constrain the training process of the initial noise prediction model; through constraint of two loss functions on a model training process, when sample signals have tiny differences, the model preferentially captures two groups of common anatomical structure features, accurately distinguishes noise and signals, reduces structural damage, and significantly improves noise adaptability and signal retention performance of a traditional method; therefore, a better noise prediction model is obtained, and the noise prediction accuracy of the noise prediction model is improved.

Description

Technical Field

[0001] This invention relates to the field of ultrasound imaging, and more particularly to a noise prediction model training method, related apparatus, and medium for ultrasound image denoising. Background Technology

[0002] With the development of social technology, ultrasound imaging is increasingly being applied in the medical field. In medicine, ultrasound imaging, due to its low cost, non-invasiveness, and real-time capability, has become one of the most important advancements in cardiac imaging. In recent years, deep learning has become a research hotspot in medical imaging, and ultrasound image analysis is no exception. Thanks to its powerful feature learning capabilities, deep neural networks can automatically extract complex distribution features from large-scale data in ultrasound images, effectively modeling and suppressing noise, and significantly improving image contrast and structural fidelity. However, current mainstream methods mostly rely on supervised training with pairs of noise-clean images. In ultrasound imaging scenarios, this data requirement faces insurmountable bottlenecks. Training the noise prediction model in diffusion models typically requires numerous iterative sampling steps, resulting in slow inference speed, high computational resource consumption during training, and low accuracy in predicting noise.

[0003] Practice has shown that existing noise prediction models have low accuracy in predicting noise. Therefore, it is necessary to propose a new method for training noise prediction models. By using a new training method to train the model, a well-trained noise prediction model can be obtained, thereby improving the accuracy of noise prediction models in predicting noise. Summary of the Invention

[0004] The purpose of this invention is to provide a noise prediction model training method to improve the accuracy of noise prediction.

[0005] To address the aforementioned technical problems, the first aspect of this invention discloses a method for training a noise prediction model, the method comprising: Obtain a set of original ultrasound images. For each original ultrasound image in the set: add noise to the original ultrasound image based on a first noise-adding algorithm to obtain a first noisy image; add noise to the original ultrasound image based on a second noise-adding algorithm to obtain a second noisy image. The first noise-adding algorithm and the second noise-adding algorithm are different. All the first noisy images are used as training datasets to train a preset initial noise prediction model. During the training of the initial noise prediction model, a first loss function is determined based on the initial predicted noise obtained by the initial noise prediction model for the first noisy image and the first noisy algorithm corresponding to the first noisy image; a second loss function is determined based on the initial predicted noise obtained by the initial noise prediction model for the first noisy image and the second noisy image corresponding to the first noisy image; the training of the initial noise prediction model is constrained by the first loss function and the second loss function to obtain a trained noise prediction model, which is used to identify noise in ultrasound images.

[0006] As an optional implementation, in the first aspect of the present invention, the first noise-adding algorithm This is achieved through the following formula:

[0007] in,

[0008]

[0009] In the above formula, This is the initial, un-noiseed image. This is the first noisy sample image at time t. This is the preset first noise control hyperparameter. This is the preset second noise control hyperparameter. The first type of added noise, The identity matrix in the covariance matrix. The Gaussian distribution function is... This is the first noisy image; And, the second noise-adding algorithm This is achieved through the following formula:

[0010] in,

[0011]

[0012] In the above formula, This is the second noisy sample image at time t. For the second noisy image, This is the second type of noise.

[0013] As an optional implementation, in the first aspect of the present invention, the first loss function The calculation method is as follows:

[0014] in,

[0015] In the above formula, This represents the initial prediction noise.

[0016] As an optional implementation, in the first aspect of the invention, the second loss function The calculation method is as follows:

[0017] in,

[0018] In the above formula, This is the second type of noise.

[0019] As an optional implementation, in the first aspect of the present invention, the step of using a first loss function and a second loss function to jointly constrain the training of the initial noise prediction model to obtain a trained noise prediction model includes: The first convergence condition is obtained based on the convergence criterion of the first loss function, and the second convergence condition is obtained based on the convergence criterion of the second loss function. The initial noise prediction model training process is fed back based on the first and second convergence conditions during training; If either the first convergence condition or the second convergence condition fails to meet the preset convergence criterion, the noise control parameters in the initial noise prediction model are adjusted, and the initial noise prediction model is trained again. When both the first and second convergence conditions meet the preset convergence criteria, a well-trained noise prediction model is obtained.

[0020] A second aspect of this invention discloses an ultrasonic noise reduction method, the method comprising: Input the ultrasound image to be denoised into any of the trained noise prediction models described in the first aspect of the present invention, obtain the output result of the noise prediction model, and obtain the target predicted noise on the ultrasound image to be denoised based on the output result. The ultrasound image to be denoised is subjected to inverse denoising operation based on the target predicted noise to obtain the denoised target ultrasound image.

[0021] As an optional implementation, in a second aspect of the invention, the reverse denoising operation... The calculation method is as follows:

[0022] in,

[0023]

[0024] In the above formula, This is the preset first noise control hyperparameter. This is the first noisy image at time t. Let be the mean of the first noisy image at time t. Let be the covariance matrix of the first noisy image at time t. Let be the predicted noise of the first noisy image at time t. The Gaussian distribution function is... This is the initial denoised sample image at time T.

[0025] A third aspect of the present invention discloses a noise prediction model training device, the device comprising: The data acquisition module is used to acquire a set of original ultrasound images. For each original ultrasound image in the set of original ultrasound images: the original ultrasound image is denoised based on a first denoising algorithm to obtain a first denoised image, and the original ultrasound image is denoised based on a second denoising algorithm to obtain a second denoised image, wherein the first denoising algorithm and the second denoising algorithm are different. The model training module is used to train a preset initial noise prediction model using all the first noisy images as training datasets. During the training process of the initial noise prediction model, a first loss function is determined based on the initial predicted noise obtained by the initial noise prediction model for the first noisy image and the first noisy algorithm corresponding to the first noisy image; a second loss function is determined based on the initial predicted noise obtained by the initial noise prediction model for the first noisy image and the second noisy image corresponding to the first noisy image; the training of the initial noise prediction model is constrained by the first loss function and the second loss function to obtain a trained noise prediction model, which is used to identify noise on ultrasound images.

[0026] As an optional implementation, in a third aspect of the present invention, the first noise-adding algorithm This is achieved through the following formula:

[0027] in,

[0028]

[0029] In the above formula, This is the initial, un-noiseed image. This is the first noisy sample image at time t. This is the preset first noise control hyperparameter. This is the preset second noise control hyperparameter. The first type of added noise, The identity matrix in the covariance matrix. The Gaussian distribution function is... This is the first noisy image; And, the second noise-adding algorithm This is achieved through the following formula:

[0030] in,

[0031]

[0032] In the above formula, This is the second noisy sample image at time t. For the second noisy image, This is the second type of noise.

[0033] As an optional implementation, in a third aspect of the invention, the second loss function The calculation method is as follows:

[0034] in,

[0035] In the above formula, This is the second type of noise.

[0036] As an optional implementation, in a third aspect of the present invention, the model training module uses a first loss function and a second loss function to jointly constrain the training of the initial noise prediction model to obtain a trained noise prediction model. The specific operation includes: The first convergence condition is obtained based on the convergence criterion of the first loss function, and the second convergence condition is obtained based on the convergence criterion of the second loss function. The initial noise prediction model training process is fed back based on the first and second convergence conditions during training; If either the first convergence condition or the second convergence condition fails to meet the preset convergence criterion, the noise control parameters in the initial noise prediction model are adjusted, and the initial noise prediction model is trained again. When both the first and second convergence conditions meet the preset convergence criteria, a well-trained noise prediction model is obtained.

[0037] A fourth aspect of the present invention discloses an ultrasonic image denoising apparatus, the apparatus comprising: The noise prediction acquisition module is used to input the ultrasound image to be denoised into any of the trained noise prediction models described in the first aspect of the present invention, obtain the output result of the noise prediction model, and obtain the target predicted noise on the ultrasound image to be denoised based on the output result. The target image acquisition module is used to perform an inverse denoising operation on the ultrasound image to be denoised based on the target predicted noise, so as to obtain the denoised target ultrasound image.

[0038] As an optional implementation, in a fourth aspect of the invention, the reverse denoising operation... The calculation method is as follows:

[0039] in,

[0040]

[0041] In the above formula, This is the preset first noise control hyperparameter. This is the first noisy image at time t. Let be the mean of the first noisy image at time t. Let be the covariance matrix of the first noisy image at time t. Let be the predicted noise of the first noisy image at time t. The Gaussian distribution function is... This is the initial denoised sample image at time T.

[0042] A fifth aspect of the present invention discloses an apparatus comprising a memory and a processor, the apparatus comprising: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the noise prediction model training method disclosed in the first aspect of the present invention or the ultrasound image denoising method disclosed in the second aspect of the present invention.

[0043] The sixth aspect of the present invention discloses a computer storage medium storing computer instructions, which, when invoked by a processor, are used to execute the noise prediction model training method disclosed in the first aspect of the present invention or the ultrasound image denoising method disclosed in the second aspect of the present invention.

[0044] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: This invention obtains a first noisy image and a second noisy image through a first noise-adding algorithm and a second noise-adding algorithm, and then generates a first loss function and a second loss function. The first loss function and the second loss function are independent of each other, but together constrain the training process of the initial noise prediction model. By constraining the model training process through the two loss functions, when there are subtle differences in the sample signals, the model prioritizes capturing two sets of common anatomical structural features, accurately distinguishing noise from signals, reducing structural damage, and significantly improving the noise adaptability and signal preservation performance of traditional methods. This results in a better noise prediction model and improves the accuracy of the noise prediction model in predicting noise. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart illustrating a noise prediction model training method disclosed in an embodiment of the present invention; Figure 2 This is a schematic flowchart of another ultrasound image denoising method disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a noise prediction model training device disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram of another ultrasonic image denoising device disclosed in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of another ultrasonic image denoising device disclosed in an embodiment of the present invention; Detailed Implementation To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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.

[0047] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0048] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0049] This invention discloses a noise prediction model training method that can comprehensively predict noise in ultrasound images, thereby improving the accuracy of noise prediction models. The following sections provide detailed explanations.

[0050] Example 1 Please see Figure 1 , Figure 1 This is a flowchart illustrating a noise prediction model training method disclosed in an embodiment of the present invention. Figure 1 The described noise prediction model training method can be applied to a noise prediction model training device, which can be integrated into a cloud server or a local server; this embodiment of the invention does not impose any limitations. Figure 1 As shown, the noise prediction model training method may include the following operations: Step 101: Obtain the set of original ultrasound images. For each original ultrasound image in the set: add noise to the original ultrasound image based on the first noise-adding algorithm to obtain the first noise-adding image. Add noise to the original ultrasound image based on the second noise-adding algorithm to obtain the second noise-adding image.

[0051] In the embodiments of this invention, those skilled in the art will understand that the noise-adding algorithm is an algorithm that adds noise to an image after processing it. Any noise-adding algorithm can be used, including but not limited to Gaussian distribution noise algorithm, salt-and-pepper distribution noise algorithm, Poisson distribution noise algorithm, impulse distribution noise algorithm, mixed noise algorithm, Laplacian noise-adding algorithm, and exponential mechanism noise-adding algorithm.

[0052] In this embodiment of the invention, the first noise-adding algorithm is different from the second noise-adding algorithm. Therefore, the noise on the first and second noisy images is also different. Thus, either the first or second noisy image can be used as training data, and the other noisy image can be used as training reference data. This allows the model to rely on the independence of the two sets of noise to require the predicted noise to match the other set of real noise; and to rely on the lack of correlation between the noise to prompt the model to eliminate noise and focus on the two sets of common stable anatomical structures.

[0053] In this embodiment of the invention, the original ultrasound image set can be a collection of multiple acquired ultrasound image data that have been processed and then grouped together. The ultrasound image data can be ultrasound video data containing multiple frames of ultrasound images, a single ultrasound image, or a collection of multiple ultrasound images used to identify the same object. For example, the ultrasound image data can be ultrasound image data of the heart.

[0054] Optionally, before acquiring the set of ultrasound images, the method includes processing the acquired ultrasound image data, which may involve uniformizing the size of all ultrasound images.

[0055] In this embodiment of the invention, the first noise-adding algorithm and the second noise-adding algorithm can be obtained by performing two independent forward noise-adding operations on the same initial image to obtain two different noise-adding images.

[0056] Step 102: Use all the first noisy images as training datasets to train the preset initial noise prediction model.

[0057] In this embodiment of the invention, the training dataset can be all the noisy images obtained through the first independent forward noisy process, and the initial noise prediction model that has just been set can be used to train the dataset.

[0058] Step 103: During the training process of the initial noise prediction model, determine the first loss function based on the initial predicted noise obtained by the initial noise prediction model for the first noisy image and the first noisy algorithm corresponding to the first noisy image.

[0059] In this embodiment of the invention, the first loss function is used to constrain the training process of the initial noise prediction model. The first noisy image is input into the initial noise prediction model, and the model predicts the noise to obtain the predicted noise. The first loss function is calculated using the first noisy image and the predicted noise to obtain the first loss function value. The first loss function value is used to evaluate whether the model meets the requirements under the constraint of the first loss function. The first loss function is the core of the inverse denoising of the denoising diffusion probability model. Through the first loss function, the model can accurately predict noise to ensure the denoising ability of subsequent models. In this way, the first loss function allows the model to master the ability to predict a single set of noise.

[0060] Step 104: Determine the second loss function based on the initial predicted noise obtained from the first noisy image and the second noisy image corresponding to the first noisy image, according to the initial noise prediction model.

[0061] In this embodiment of the invention, the second loss function is used to constrain the training process of the initial noise prediction model. The first noisy image is input into the initial noise prediction model, which predicts the noise to obtain a predicted noisy image. The second loss function is then calculated using the first noisy image, the predicted noise, and the second noisy image to obtain a second loss function value. This second loss function value is used to evaluate whether the model meets the requirements under the constraint of the second loss function. The second loss function generates paired noisy samples using the first and second noisy algorithms. These paired noisy samples are the first and second noisy images. These two noisy images jointly constrain the model to ensure that the noise prediction of any one set of noisy samples remains consistent with the real noise of the other set, forcing the model to focus on the common cardiac anatomical features of the two sets of samples, thus distinguishing noise from key structures. In this way, the second loss function can rely on the independence of the two sets of noise to require the predicted noise to match the other set of real noise; and rely on the lack of correlation between the noise to prompt the model to eliminate noise and focus on the common stable anatomical structures of the two sets.

[0062] Step 105: Use the first loss function and the second loss function together to constrain the training of the initial noise prediction model, and obtain the trained noise prediction model.

[0063] In this embodiment of the invention, the first loss function and the second loss function are independent of each other and jointly constrain the initial noise prediction model. Through the constraint between the two loss functions, when there are subtle differences in the sample signals, the model can preferentially capture two sets of common anatomical structural features, accurately distinguish noise from signals, reduce structural damage, and significantly improve the noise adaptability and signal preservation performance of traditional methods. The noise prediction model is used to identify noise in ultrasound images.

[0064] As can be seen, the embodiments of the present invention obtain a first noisy image and a second noisy image through a first noise-adding algorithm and a second noise-adding algorithm, and then generate a first loss function and a second loss function. The first loss function and the second loss function are independent of each other and jointly constrain the training process of the initial noise prediction model. By constraining the model training process through the two loss functions, when there are subtle differences in the sample signals, the model can prioritize capturing two sets of common anatomical structural features, accurately distinguishing noise from signals, reducing structural damage, and significantly improving the noise adaptability and signal preservation performance of traditional methods. This results in a better noise prediction model and improves the accuracy of the noise prediction model in predicting noise.

[0065] In an optional embodiment, the above-described method of adding noise to the original ultrasound image based on the first noise-adding algorithm to obtain a first noisy image may include: For obtaining the first noisy image, the process can be as follows: first, obtain the initial unnoisy image; then, calculate the first noisy sample image based on the initial unnoisy image using the first noisy intermediate algorithm; then, calculate the first noisy sample image set based on the first noisy sample image using the first noisy interval algorithm; and finally, calculate the first noisy image based on the first noisy sample image set using the first noisy algorithm. This calculation process can be achieved using the following formula:

[0066] in,

[0067]

[0068] In the above formula, This is the initial, un-noiseed image. This is the first noisy sample image at time t. This is the preset first noise control hyperparameter. This is the preset second noise control hyperparameter. The first type of added noise, The identity matrix in the covariance matrix. The Gaussian distribution function is... This is the first noisy image; The multiplication operation from t=1 to t=T represents multiplying the conditional probabilities of each step to obtain the joint probability of the entire reverse sequence. Typically, it's an increasing sequence that directly controls the noise level at step t, such as linear scheduling.

[0069] The preset first noise control hyperparameter ; Initial unnoised image It could be the starting point of the diffusion process, the raw data without any added noise; or the first noisy sample image at time t. This could be the first intermediate sample obtained after t steps of noise addition, representing the gradual fusion of the initial sample and noise. The preset first noise level. It can follow a normal distribution with a mean of 0 and a variance of 1, which is the core source of the randomness in the diffusion process and is independent of the initial sample.

[0070] Furthermore, the above-mentioned method of adding noise to the original ultrasound image based on the second noise-adding algorithm to obtain a second noise-added image may include: To obtain the second noisy image, one can first obtain the initial unnoisy image, then obtain the second noisy sample image based on the initial unnoisy image using the second noisy intermediate algorithm, then obtain the second noisy sample image set based on the second noisy sample image using the second noisy interval algorithm, and finally obtain the second noisy image based on the second noisy sample image set using the second noisy algorithm.

[0071] The above calculation method can be achieved through the following formula:

[0072] in,

[0073]

[0074] In the above formula, This is the second noisy sample image at time t. For the second noisy image, This is the second type of noise.

[0075] As can be seen, in this optional embodiment, a first noisy image and a second noisy image are obtained through a first noise-adding algorithm and a second noise-adding algorithm, and then a first loss function and a second loss function are generated. The first loss function and the second loss function are independent of each other and jointly constrain the training process of the initial noise prediction model. By constraining the model training process with the two loss functions, when there are subtle differences in the sample signals, the model can prioritize capturing two sets of common anatomical structural features, accurately distinguish noise from signals, reduce structural damage, and significantly improve the noise adaptability and signal preservation performance of traditional methods, thus further improving the accuracy of noise prediction.

[0076] In another alternative embodiment, the first loss function can be obtained in the following manner: First, the first denoised sample image and the initial undenoised image are used to obtain the first denoised noise using a first noise algorithm. Then, the first loss function value is obtained based on the first denoised noise using a first loss algorithm. That is, it can be achieved using the following formula:

[0077] in,

[0078] In the above formula, For initial prediction noise, Based on the first noisy image The noise predicted at time t.

[0079] As can be seen, in this optional embodiment, the loss function value obtained for the first loss function can be compared with the model's convergence criterion to better constrain the model's training process, enabling the model to more accurately predict noise to ensure its denoising ability, thereby further improving the accuracy of noise prediction.

[0080] In yet another alternative embodiment, the second loss function can be obtained in the following manner: First, the second noise is obtained from the first noise using the second noise algorithm. Then, the second loss function value is obtained from the second noise using the second loss algorithm.

[0081] That is, it can be achieved through the following formula:

[0082] in,

[0083] In the above formula, The second noise is calculated from the difference between the first and second noisy sample images and the relationship between the second noise and the first noise. The second loss algorithm is based on the initial prediction noise. Second noise The relationship between them is calculated.

[0084] As can be seen, in this optional embodiment, the loss function value obtained by the second loss function can better constrain the training process of the model through the model's convergence criterion. The second loss function generates pairs of noisy samples by adding noise twice independently, constraining the model to keep the noise prediction of one set of noisy samples consistent with the real noise of the other set, forcing the model to focus on the cardiac anatomical features common to the two sets of samples, and achieving the distinction between noise and key structures.

[0085] In another optional embodiment, using a first loss function and a second loss function to jointly constrain the training of the initial noise prediction model to obtain a trained noise prediction model may include: The first convergence condition is obtained based on the convergence criterion of the first loss function, and the second convergence condition is obtained based on the convergence criterion of the second loss function. Feedback is provided on the initial noise prediction model training process based on the first and second convergence conditions during training; If either the first convergence condition or the second convergence condition fails to meet the preset convergence criteria, the noise control parameters in the initial noise prediction model are adjusted, and the initial noise prediction model is trained again. When both the first and second convergence conditions meet the preset convergence criteria, a well-trained noise prediction model is obtained.

[0086] In this optional embodiment, those skilled in the art will understand that convergence conditions include, but are not limited to, the state where model parameters gradually stabilize during training and performance no longer significantly improves during training. In such cases, the model can be considered to have reached convergence. For example, during training, if the model's loss function value changes by less than a preset threshold for multiple consecutive rounds, or the L2 norm of the model parameter gradient is less than a certain threshold, or task-related evaluation metrics show no significant improvement, such as classification accuracy showing no significant improvement on the validation set for multiple consecutive rounds.

[0087] In this optional embodiment, feedback on the initial noise prediction model training process can be provided by determining whether the model has reached the convergence threshold based on the output loss function value. In this optional embodiment, if either the first convergence condition or the second convergence condition fails to meet the preset convergence criterion, it can be that the first convergence condition meets the second convergence condition but does not, the first convergence condition fails but the second convergence condition meets, or the first convergence condition fails and the second convergence condition does not meet either.

[0088] As can be seen, this optional embodiment can effectively constrain the model's training process by setting convergence conditions. The entire model is considered converged only when both convergence conditions are met. Feeding the two convergence conditions back to the model training process can improve the accuracy of the trained model in predicting noise.

[0089] Example 2 Please see Figure 2 , Figure 2 This is a schematic flowchart of an ultrasound image denoising method disclosed in an embodiment of the present invention. Figure 2 The described ultrasound image denoising method can be applied to an ultrasound image denoising device, which can be integrated into a cloud server or a local server; this embodiment of the invention is not limited thereto. Figure 2 As shown, the ultrasound image denoising method may include the following operations: Step 201: Input the ultrasound image to be denoised into any of the trained noise prediction models in Example 1, obtain the output result of the noise prediction model, and obtain the target predicted noise on the ultrasound image to be denoised based on the output result.

[0090] In this embodiment of the invention, the ultrasound image to be denoised can be an abdominal ultrasound image, a cardiovascular ultrasound image, or a cranial ultrasound image, including but not limited to ultrasound images of the liver, gallbladder, pancreas, spleen, kidneys, abdominal cavity, heart, and blood vessels.

[0091] In this embodiment of the invention, the output result can be a set of sample images containing noise after the noise prediction model processes the image to be denoised. The set of sample images includes images containing noise at different times.

[0092] Step 202: Perform inverse denoising operation on the ultrasound image to be denoised based on the target predicted noise to obtain the denoised target ultrasound image.

[0093] As can be seen, the embodiments of the present invention obtain a first noisy image and a second noisy image through a first noise-adding algorithm and a second noise-adding algorithm, and then generate a first loss function and a second loss function. The first loss function and the second loss function are independent of each other, jointly constraining the training process of the initial noise prediction model. By constraining the model training process through the two loss functions, when there are subtle differences in the sample signals, the model prioritizes capturing two sets of common anatomical structural features, accurately distinguishing noise from signal, reducing structural damage, and significantly improving the noise adaptability and signal preservation performance of traditional methods. This results in a better noise prediction model and improves the accuracy of the noise prediction model in predicting noise. Subsequently, the noise prediction model is used to obtain the image to be denoised, and the image to be denoised is then denoised, converting the predicted noise into a denoised target ultrasound image, thus allowing the target predicted noise to be applied to the actual image denoising operation.

[0094] In an optional embodiment, performing an inverse denoising operation on the ultrasound image to be denoised based on the target predicted noise to obtain a denoised target ultrasound image may include: A first noisy image and its predicted noise are obtained. A first denoised image is obtained by using a first denoising algorithm based on the first noisy image and its predicted noise. A first denoised sample image set is obtained by using a second denoising algorithm based on the first denoised image. A denoised target ultrasound image is obtained by using a multiplicative algorithm based on the first denoised sample image set.

[0095] Specifically, the above solution can be implemented using the following formula:

[0096] in,

[0097]

[0098] In the above formula, This is the preset first noise control hyperparameter. This is the first noisy image at time t. Let be the mean of the first noisy image at time t. Let be the covariance matrix of the first noisy image at time t. Let be the predicted noise of the first noisy image at time t. The Gaussian distribution function is... This is the initial denoised sample image at time T.

[0099] Multiplicative algorithms include, but are not limited to, iterative methods, recursive methods, fast exponentiation algorithms, divide-and-conquer methods, modular arithmetic-optimized cumulative multiplication, and high-precision cumulative multiplication methods. The preset first noise control hyperparameter... , Typically, it's an increasing sequence that directly controls the noise level at step t; the noisy sample image at time t. It can be an intermediate sample obtained after t steps of noise addition, or the result of the gradual fusion of the initial sample and noise.

[0100] As can be seen, in this optional embodiment, the speckle noise in ultrasound images, caused by backscattered signals, exhibits multiplicative noise characteristics with a signal-correlated distribution, appearing as a granular, black-and-white dot texture structure on the image. Speckle noise is unavoidable and degrades visual quality, limiting the interpretability and diagnostic accuracy of medical images. Therefore, effectively suppressing speckle noise while preserving structural and textural details remains a key challenge in ultrasound image processing. This optional embodiment therefore designs a denoising method based on a multiplicative algorithm. The multiplicative algorithm integrates and calculates noisy images from different time points, enabling the ultrasound denoising model to obtain denoised images with clear structural and textural details, thus improving the model's denoising capability.

[0101] Example 3 Please see Figure 3 , Figure 3 This invention discloses a noise prediction model training device, which may include: The data acquisition module 301 is used to acquire a set of original ultrasound images. For each original ultrasound image in the set of original ultrasound images: the original ultrasound image is denoised based on a first denoising algorithm to obtain a first denoised image, and the original ultrasound image is denoised based on a second denoising algorithm to obtain a second denoised image. The first denoising algorithm and the second denoising algorithm are different. The model training module 302 is used to train a preset initial noise prediction model using all the first noisy images as training datasets. During the training process of the initial noise prediction model, a first loss function is determined based on the initial predicted noise obtained by the initial noise prediction model for the first noisy image and the first noisy algorithm corresponding to the first noisy image; a second loss function is determined based on the initial predicted noise obtained by the initial noise prediction model for the first noisy image and the second noisy image corresponding to the first noisy image; the training of the initial noise prediction model is constrained by the first loss function and the second loss function together to obtain a trained noise prediction model, which is used to identify noise on ultrasound images.

[0102] As can be seen, the noise prediction model training device in this embodiment of the invention can obtain a first noisy image and a second noisy image through a first noise-adding algorithm and a second noise-adding algorithm, and then generate a first loss function and a second loss function. The first loss function and the second loss function are independent of each other and jointly constrain the initial noise prediction model training process. By constraining the model training process through the two loss functions, when there are subtle differences in the sample signals, the model can prioritize capturing two sets of common anatomical structural features, accurately distinguishing noise from signals, reducing structural damage, and significantly improving the noise adaptability and signal preservation performance of traditional methods. This results in a better noise prediction model and improves the accuracy of the noise prediction model in predicting noise.

[0103] In an optional embodiment, the data acquisition module 301 applies a first noise-adding algorithm to the noise prediction model. This can be achieved through the following formula:

[0104] in,

[0105]

[0106] In the above formula, This is the initial, un-noiseed image. This is the first noisy sample image at time t. This is the preset first noise control hyperparameter. This is the preset second noise control hyperparameter. The first type of added noise, The identity matrix in the covariance matrix. The Gaussian distribution function is... This is the first noisy image; And the aforementioned second noise-adding algorithm This is achieved through the following formula:

[0107] in,

[0108]

[0109] In the above formula, This is the second noisy sample image at time t. This is the second noisy image.

[0110] As can be seen, this optional embodiment obtains different noises through a noise addition algorithm, thereby requiring the predicted noise to match another set of real noises based on the independence of the two sets of noises; and relying on the uncorrelatedness of the noises, it prompts the model to exclude noise and focus on the two sets of common stable anatomical structures, which can improve the accuracy of noise prediction.

[0111] In another alternative embodiment, the model training module 302 can obtain the first loss function on the noise prediction model in the following way:

[0112] in,

[0113] In the above formula, This represents the initial prediction noise.

[0114] As can be seen, this optional embodiment, based on the loss function value obtained from the first loss function and the model's convergence criterion, can better constrain the model's training process, enabling the model to accurately predict noise and ensure its denoising capability, thereby improving the accuracy of noise prediction.

[0115] In yet another optional embodiment, the model training module 302 can obtain the second loss function on the noise prediction model in the following manner:

[0116] in,

[0117] In the above formula, This is the second type of noise.

[0118] As can be seen, this optional embodiment, based on the loss function value obtained by the second loss function and the model's convergence criterion, can better constrain the model's training process. The second loss function generates pairs of noisy samples by adding noise twice independently, constraining the model to keep the noise prediction of one set of noisy samples consistent with the real noise of the other set. This forces the model to focus on the common cardiac anatomical features of the two sets of samples, thereby distinguishing noise from key structures and improving the accuracy of noise prediction.

[0119] In another optional embodiment, the model training module 302 uses a first loss function and a second loss function to jointly constrain the training of the initial noise prediction model to obtain a trained noise prediction model, which may include: The first convergence condition is obtained based on the convergence criterion of the first loss function, and the second convergence condition is obtained based on the convergence criterion of the second loss function. Feedback is provided on the initial noise prediction model training process based on the first and second convergence conditions during training; If either the first convergence condition or the second convergence condition fails to meet the preset convergence criteria, the noise control parameters in the initial noise prediction model are adjusted, and the initial noise prediction model is trained again. When both the first and second convergence conditions meet the preset convergence criteria, a well-trained noise prediction model is obtained.

[0120] As can be seen, this optional embodiment can effectively constrain the model training process by setting corresponding convergence conditions for the model. The two convergence conditions are used to provide feedback on the model training process, which can improve the accuracy of the trained model in predicting noise.

[0121] Example 4 Please see Figure 4 , Figure 4 This invention discloses an ultrasonic image denoising device, which may include: The noise prediction acquisition module 401 is used to input the ultrasound image to be denoised into the noise prediction model device trained in Example 3, obtain the output result of the noise prediction device, and obtain the target predicted noise on the ultrasound image to be denoised based on the output result. The target image acquisition module 402 is used to perform an inverse denoising operation on the ultrasound image to be denoised based on the target predicted noise, so as to obtain the denoised target ultrasound image.

[0122] As can be seen, the embodiments of the present invention verify the accuracy of the predicted noise by denoising the image to be denoised and converting the predicted noise into a denoised target ultrasound image.

[0123] In an optional embodiment, the target image acquisition module 402 performs the reverse denoising operation calculation on the ultrasound image denoising device as follows:

[0124] in,

[0125]

[0126] In the above formula, This is the preset first noise control hyperparameter. This is the first noisy image at time t. Let be the mean of the first noisy image at time t. Let be the covariance matrix of the first noisy image at time t. Let be the predicted noise of the first noisy image at time t. The Gaussian distribution function is... This is the initial denoised sample image at time T.

[0127] As can be seen, in this optional embodiment, the multiplicative algorithm can improve the denoising capability of the model, enabling the ultrasonic denoising model to obtain a denoised image with clear structural and texture details.

[0128] Example 5 Please see Figure 5 , Figure 5 This is a schematic diagram of the structure of another ultrasonic image denoising device disclosed in an embodiment of the present invention. For example... Figure 5 As shown, the ultrasound image denoising device may include: Memory 501 storing executable program code; Processor 502 coupled to memory 501; The processor 502 calls the executable program code stored in the memory 501 to execute some or all of the steps in the noise prediction model training and ultrasound image denoising method described in Embodiment 1 or Embodiment 2 of the present invention.

[0129] Example 6 This invention discloses a computer storage medium storing computer instructions. When these computer instructions are invoked, they are used to execute some or all of the steps in any of the noise prediction model training and ultrasound image denoising methods disclosed in Embodiment 1 or Embodiment 2 of this invention.

[0130] The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0131] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0132] Finally, it should be noted that the above embodiments are merely preferred embodiments of the present invention and are only used to illustrate the technical solutions of the present invention, not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for training a noise prediction model, characterized in that, The method includes: Obtain a set of original ultrasound images. For each original ultrasound image in the set: add noise to the original ultrasound image based on a first noise-adding algorithm to obtain a first noisy image; add noise to the original ultrasound image based on a second noise-adding algorithm to obtain a second noisy image. The first noise-adding algorithm and the second noise-adding algorithm are different. All the first noisy images are used as training datasets to train a preset initial noise prediction model. During the training of the initial noise prediction model, a first loss function is determined based on the initial predicted noise obtained by the initial noise prediction model for the first noisy image and the first noisy algorithm corresponding to the first noisy image; a second loss function is determined based on the initial predicted noise obtained by the initial noise prediction model for the first noisy image and the second noisy image corresponding to the first noisy image; the training of the initial noise prediction model is constrained by the first loss function and the second loss function to obtain a trained noise prediction model, which is used to identify noise in ultrasound images.

2. The noise prediction model training method according to claim 1, characterized in that, First noise-adding algorithm This is achieved through the following formula: in, In the above formula, This is the initial, un-noiseed image. The first noisy sample image at time t. The preset first noise control hyperparameter, This is the preset second noise control hyperparameter. The first type of added noise, The identity matrix is ​​the covariance matrix. The function is a Gaussian distribution. This is the first noisy image; And, the second noise-adding algorithm This is achieved through the following formula: in, In the above formula, This is the second noisy sample image at time t. For the second noisy image, This is the second type of noise.

3. The noise prediction model training method according to claim 2, characterized in that, The first loss function The calculation method is as follows: in, In the above formula, This represents the initial prediction noise.

4. The noise prediction model training method according to claim 3, characterized in that, The second loss function The calculation method is as follows: in, In the above formula, This is the second type of noise.

5. The noise prediction model training method according to claim 4, characterized in that, The process of using a first loss function and a second loss function to jointly constrain the training of the initial noise prediction model to obtain a trained noise prediction model includes: The first convergence condition is obtained based on the convergence criterion of the first loss function, and the second convergence condition is obtained based on the convergence criterion of the second loss function. The initial noise prediction model training process is fed back based on the first and second convergence conditions during training; If either the first convergence condition or the second convergence condition fails to meet the preset convergence criterion, the noise control parameters in the initial noise prediction model are adjusted, and the initial noise prediction model is trained again. When both the first and second convergence conditions meet the preset convergence criteria, a well-trained noise prediction model is obtained.

6. A method for denoising ultrasound images, characterized in that, The method includes: Input the ultrasound image to be denoised into the trained noise prediction model according to any one of claims 1-5, obtain the output result of the noise prediction model, and obtain the target predicted noise on the ultrasound image to be denoised based on the output result; The ultrasound image to be denoised is subjected to inverse denoising operation based on the target predicted noise to obtain the denoised target ultrasound image.

7. The ultrasound image denoising method according to claim 6, characterized in that, The reverse denoising operation The calculation method is as follows: in, In the above formula, The preset first noise control hyperparameter, This is the first noisy image at time t. Let be the mean of the first noisy image at time t. Let be the covariance matrix of the first noisy image at time t. Let be the predicted noise of the first noisy image at time t. The function is a Gaussian distribution. This is the initial denoised sample image at time T.

8. A noise prediction model training device, characterized in that, The device includes: The data acquisition module is used to acquire a set of original ultrasound images. For each original ultrasound image in the set of original ultrasound images: the original ultrasound image is denoised based on a first denoising algorithm to obtain a first denoised image, and the original ultrasound image is denoised based on a second denoising algorithm to obtain a second denoised image, wherein the first denoising algorithm and the second denoising algorithm are different. The model training module is used to train a preset initial noise prediction model using all the first noisy images as training datasets. During the training process of the initial noise prediction model, a first loss function is determined based on the initial predicted noise obtained by the initial noise prediction model for the first noisy image and the first noisy algorithm corresponding to the first noisy image; a second loss function is determined based on the initial predicted noise obtained by the initial noise prediction model for the first noisy image and the second noisy image corresponding to the first noisy image; the training of the initial noise prediction model is constrained by the first loss function and the second loss function to obtain a trained noise prediction model, which is used to identify noise on ultrasound images.

9. An ultrasonic image denoising device, characterized in that, The device includes: The noise prediction acquisition module is used to input the ultrasound image to be denoised into the trained noise prediction model according to any one of claims 1-5, obtain the output result of the noise prediction model, and obtain the target predicted noise on the ultrasound image to be denoised based on the output result. The target image acquisition module is used to perform an inverse denoising operation on the ultrasound image to be denoised based on the target predicted noise, so as to obtain the denoised target ultrasound image.

10. An apparatus comprising a memory and a processor, characterized in that, The apparatus includes: a memory storing executable program code, and a processor coupled to the memory; the processor calls the executable program code stored in the memory to execute the noise prediction model training method as described in any one of claims 1-5 or the ultrasound image denoising method as described in any one of claims 6-7.

11. A computer storage medium, characterized in that, The computer storage medium stores computer instructions, which, when invoked by the processor, are used to execute the noise prediction model training method as described in any one of claims 1-5 or the ultrasound image denoising method as described in any one of claims 6-7.

Images