An ultrasound image processing method and device based on a double-layer ultrasound catheter, and a medium.

Abstract

This invention relates to the field of ultrasound image denoising technology, and discloses an ultrasound image processing method, apparatus, and medium based on a double-layer ultrasound catheter. The method includes: acquiring an ultrasound image through a double-layer ultrasound catheter, and then inputting the ultrasound image into a trained multiplicative speckle noise denoising model to obtain a denoised ultrasound image; denoising using the trained multiplicative speckle noise denoising model can improve the denoising accuracy of the ultrasound image. The trained multiplicative speckle noise denoising model is obtained through two stages of training; the first stage involves acquiring a synthetic image dataset to train the model's initial recognition ability. During training, four loss functions are set to constrain each other, and corresponding algorithms are adjusted for each loss function to obtain the first-stage model; the second stage training uses clinical data and, through two loss function constraints, allows the model to adapt to clinical image denoising.

Description

Technical Field

[0001] This invention relates to the field of ultrasound image denoising technology, and in particular to an ultrasound image processing method, apparatus, and medium based on a double-layer ultrasound catheter. Background Technology

[0002] Ultrasound imaging, with its advantages of being non-invasive, real-time, and low-cost, has become a widely used imaging technique in clinical diagnosis. However, as the ultrasound waves emitted by the transducer propagate within human tissue, speckle noise is inevitably generated. This speckle noise manifests as randomly distributed granular textures in the image, severely reducing image contrast and resolution, blurring low-contrast tissue structures, and affecting not only the subjective visual interpretation of radiologists but also limiting the accuracy of downstream computer-aided diagnosis and image analysis tasks.

[0003] Speckle noise follows multiplicative noise, meaning that the noise intensity is proportional to the local signal intensity. Its statistical characteristics are fundamentally different from those of conventional additive Gaussian noise. Existing multiplicative noise denoising models require training with pairs of noisy and clean images. However, in clinical ultrasound imaging, it is almost impossible to obtain truly noise-free images as training labels, resulting in poor denoising accuracy of the trained models.

[0004] Therefore, a new method is needed to improve the denoising effect on multiplicative noise, thereby improving the denoising accuracy of ultrasound images. Summary of the Invention

[0005] This invention provides an ultrasound image processing method, device, and medium based on a double-layer ultrasound catheter, which is used to improve the noise reduction accuracy of ultrasound images.

[0006] The first aspect of this invention discloses an ultrasound image processing method based on a double-layer ultrasound catheter, the method comprising: Acquire an ultrasound image to be processed; the ultrasound image to be processed is acquired through a double-layer ultrasound catheter; input the ultrasound image to be processed into a trained multiplicative speckle noise denoising model to obtain a denoised ultrasound image; The trained multiplicative speckle noise denoising model is obtained in the following way: Obtain a synthetic image dataset; the synthetic image dataset is obtained by inputting the acquired noise-free ultrasound image dataset into a pre-trained noise synthesis model; The synthetic image dataset is input into a preset initial multiplicative speckle noise denoising model for training. During the training process, a preset first loss function is used for constraint to obtain an intermediate multiplicative speckle noise denoising model. Obtain a training ultrasound image dataset; input the training ultrasound image dataset into the intermediate multiplicative speckle noise denoising model for training, and use a preset second loss function for constraint during the training process to obtain a trained multiplicative speckle noise denoising model.

[0007] As an optional implementation, in the first aspect of the invention, the trained noise synthesis model is obtained in the following manner: Obtain a clinical ultrasound image dataset; extract noise from the clinical ultrasound image dataset to obtain a noise ratio map set; The noise ratio map set is input into a preset variational autoencoder for training to obtain variational autoencoder parameters; All latent vectors generated by the variational autoencoder during training are obtained to obtain a latent vector dataset; the latent vector dataset is input into a preset Gaussian mixture model to obtain a Gaussian component parameter set; the Gaussian component parameter set is associated with the tissue texture corresponding to the clinical ultrasound image dataset to obtain a texture mapping table. The trained noise synthesis model is obtained based on the variational autoencoder parameters, the Gaussian component parameter set, and the texture mapping table.

[0008] As an optional implementation, in the first aspect of the present invention, the step of inputting the synthetic image dataset into a preset initial multiplicative speckle noise denoising model for training, and using a preset first loss function for constraint during the training process to obtain an intermediate multiplicative speckle noise denoising model, includes: The synthesized image dataset is input into a preset initial multiplicative speckle noise denoising model to obtain a denoised image and a noise level map; Based on the denoised image and the noise level map, a first loss function value is obtained through a first loss function; model update parameters are obtained based on the first loss function value; the first loss function value includes pixel loss value, multi-scale structure loss value, edge loss value, and noise estimation loss value; Obtain the verification image dataset; the verification image dataset refers to the portion of the synthetic image dataset that was not used in training; Based on the model update parameters, the weights corresponding to the first loss function are monitored and updated using the verification image dataset through a preset adaptive loss balance algorithm to obtain the function update weights. Obtain the comprehensive evaluation index; the comprehensive evaluation index is obtained during the process of obtaining the function update weights; The comprehensive evaluation index is judged to obtain the model update parameters corresponding to the maximum comprehensive evaluation index; based on the model update parameters corresponding to the maximum comprehensive evaluation index, the intermediate multiplicative speckle noise denoising model is obtained.

[0009] As an optional implementation, in the first aspect of the present invention, the method further includes: The synthesized image dataset is processed using the noise loss function corresponding to the noise estimation loss value to obtain a noise level map; Based on the noise level map, a smoothed and denoised map is obtained by using the pixel loss function corresponding to the pixel loss value as a constraint. Based on the noise level map, the gradient constraint map is obtained by using the edge loss function corresponding to the edge loss value to constrain the noise level map. Based on the smoothed denoised map and the gradient constraint map, a denoised edge-preserving map is obtained; the denoised edge-preserving map is constrained using the multi-scale structural loss function corresponding to the multi-scale structural loss value to obtain a trained denoised image.

[0010] As an optional implementation, in the first aspect of the present invention, the step of using the verification image dataset to monitor and update the weights corresponding to the first loss function through a preset adaptive loss balancing algorithm to obtain the function update weights includes: Based on the verification image dataset, the peak signal-to-noise ratio change rate, structural similarity ratio change rate, edge preservation ratio change rate, and noise estimation ratio change rate are obtained. When it is determined that the edge preservation rate of change decreases in two consecutive training rounds, the weights corresponding to the edge loss function are adjusted to obtain the adjusted edge weights; When it is determined that the rate of change of the noise estimate decreases in three consecutive rounds of training, the weights corresponding to the noise loss function are adjusted to obtain the adjusted noise estimate weights. When it is determined that the peak signal-to-noise ratio change rate and the structural similarity ratio change rate are both less than a preset change threshold, the weights corresponding to the multi-scale structural loss function are adjusted to obtain the adjusted structural similarity weights. When it is determined that the peak signal-to-noise ratio change rate, the structural similarity change rate, the edge preservation change rate, and the noise estimation change rate are all less than a preset global variation threshold, the adjustment edge weight, the adjustment noise estimation weight, the adjustment structural similarity weight, and the peak signal-to-noise ratio weight corresponding to the pixel loss function are adjusted to obtain the adjustment function weight; The weights adjusted in each round are normalized to obtain the function update weights.

[0011] As an optional implementation, in the first aspect of the present invention, the step of judging the comprehensive evaluation index to obtain the model update parameters corresponding to the maximum comprehensive evaluation index includes: Based on the verification image dataset, the peak signal-to-noise ratio index and the structural similarity index are obtained; A comprehensive evaluation index is obtained based on the peak signal-to-noise ratio index and the structural similarity index; The largest comprehensive evaluation index is used as the model update index; the model update parameters are determined based on the model update index.

[0012] As an optional implementation, in the first aspect of the present invention, the step of inputting the training ultrasound image dataset into the intermediate multiplicative speckle noise denoising model for training, and using a preset second loss function for constraint during the training process to obtain the trained multiplicative speckle noise denoising model, includes: The training ultrasound image dataset is input into the intermediate multiplicative speckle noise denoising model, and a second loss function value is obtained through self-supervised training; the second loss function value includes a self-supervised loss function value and a perceptual loss function value. The second loss function value is fed back into the intermediate multiplicative speckle noise denoising model to obtain the trained multiplicative speckle noise denoising model; wherein, the self-supervised loss function corresponding to the self-supervised loss function value is used to constrain the training of the training ultrasound image dataset; the perceptual loss function corresponding to the perceptual loss function value is used to constrain the self-supervised loss function.

[0013] A second aspect of the present invention discloses an ultrasound image processing device based on a double-layer ultrasound catheter, the device comprising: An image acquisition module is used to acquire an ultrasound image to be processed; the ultrasound image to be processed is acquired through a double-layer ultrasound catheter. The image processing module is used to input the ultrasound image to be processed into a trained multiplicative speckle noise denoising model to obtain a denoised ultrasound image.

[0014] The trained multiplicative speckle noise denoising model is obtained through some or all of the steps in the ultrasound image processing method based on a double-layer ultrasound catheter as described in any of the first aspects of the present invention.

[0015] A third 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 some or all of the steps in the ultrasound image processing method based on a double-layer ultrasound catheter according to any of the first aspects of the present invention.

[0016] The fourth aspect of the present invention discloses a computer storage medium storing computer instructions, which, when invoked, are used to execute some or all of the steps in the ultrasound image processing method based on a double-layer ultrasound catheter as described in any of the first aspects of the present invention.

[0017] Compared with the prior art, the present invention has the following beneficial effects: This invention acquires ultrasound images through a double-layer ultrasound catheter, and then inputs the ultrasound images into a trained multiplicative speckle noise denoising model to obtain denoised ultrasound images. Denoising through the trained multiplicative speckle noise denoising model can improve the denoising accuracy of ultrasound images.

[0018] By acquiring a clinical ultrasound image dataset, a noise synthesis model is trained. Then, a noise-free ultrasound image dataset is input into the trained noise synthesis model to obtain a synthetic image dataset. The synthesized noise has statistically consistent characteristics with real noise, eliminating the distribution bias between synthetic and real noise, thus obtaining clean-labeled training data. Utilizing the clean-label characteristics of the synthetic image dataset, it is input into the initial multiplicative speckle noise denoising model for training. During training, the constraints of the first loss function provide the model with a clear optimization direction and a standard gradient backpropagation path, enabling the model to acquire multiplicative noise suppression capabilities, noise intensity perception capabilities, and strong denoising capabilities in uniform regions and light denoising capabilities in edge regions. Finally, a real clinical ultrasound image dataset is used as training data to train the intermediate multiplicative speckle noise denoising model obtained from the initial training. The second loss function is then used for constraints, allowing the model to adapt to real data, thereby improving the model's denoising accuracy for ultrasound images. Attached Figure Description

[0019] 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.

[0020] Figure 1 This is a schematic flowchart of an ultrasound image processing method based on a double-layer ultrasound catheter disclosed in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a double-layer ultrasonic catheter disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of another double-layer ultrasonic catheter disclosed in an embodiment of the present invention; Figure 4This is a schematic flowchart of an ultrasound image processing device based on a double-layer ultrasound catheter disclosed in an embodiment of the present invention; Figure 5 This is a schematic diagram of a device including a memory and a processor disclosed in an embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0022] 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.

[0023] 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 mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0024] This invention discloses an ultrasound image processing method, apparatus, and medium based on a double-layer ultrasound catheter, used to improve the denoising accuracy of ultrasound images. These will be described in detail below.

[0025] Example 1 Please see Figure 1 , Figure 1 This is a schematic flowchart of an ultrasound image processing method based on a double-layer ultrasound catheter, as disclosed in an embodiment of the present invention. Wherein, Figure 1 The described ultrasound image processing method based on a double-layer ultrasound catheter can be applied to an ultrasound image processing device based on a double-layer ultrasound catheter. This device can be integrated into a cloud server or a local server; the embodiments of this invention are not limited thereto. Figure 1As shown, the ultrasound image processing method based on a double-layer ultrasound catheter may include the following operations: Step 101: Obtain the ultrasound image to be processed.

[0026] In this embodiment of the invention, the ultrasound image to be processed can be a cardiac ultrasound image; the ultrasound image to be processed can be acquired through a double-layer ultrasound catheter; the double-layer ultrasound catheter is such as... Figure 2 As shown, the double-layer ultrasound catheter mainly consists of an inner sheath and an outer sheath. The micro phased array is contained within the inner sheath, while the outer sheath is in direct contact with the blood vessel. The flexible circuit passes through the cavity of the inner sheath.

[0027] In embodiments of the present invention, such as Figure 3 As shown, Figure 3 This is a schematic diagram of another double-layer ultrasonic catheter disclosed in an embodiment of the present invention. Since the phased array is connected to the outer sheath, the double-layer ultrasonic catheter can be controlled by the handle to make the catheter bend in four directions and rotate the outer sheath 360 degrees after bending, thereby driving the phased array to rotate and realize imaging at any angle.

[0028] Step 102: Input the ultrasound image to be processed into the trained multiplicative speckle noise denoising model to obtain the denoised ultrasound image.

[0029] In this embodiment of the invention, the training process of the multiplicative speckle noise denoising model can be as follows: First, a synthetic image dataset is acquired. The synthetic image dataset can be obtained by inputting the acquired noise-free ultrasound image dataset into a pre-trained noise synthesis model. The training process of the noise synthesis model can be as follows: Noise is extracted from a real clinical dataset to obtain a noise ratio map set. Then, by establishing a texture mapping table corresponding to the tissue parts, the noise synthesis model can be obtained. The acquired noise-free ultrasound image dataset is input into the noise synthesis model and multiplied with the noise generated by the noise synthesis model to obtain a synthetic image dataset. This synthetic image dataset is then used for training. Since it is a synthetic dataset, the true values ​​of the noise levels can be calculated. This allows all the labels (clean images, noise level maps) required for subsequent training to be automatically generated in one go during the synthesis process, without any manual annotation. This eliminates the distribution bias of noise in the training image dataset and improves the training accuracy of the model. The noise-free ultrasound image dataset here can be obtained by continuously acquiring more than 10 frames of images from the same sectional plane, averaging them over time after motion registration, resulting in images with low speckle noise, which can serve as a noise-free image dataset. Alternatively, it can be generated using ultrasound simulation platforms such as Field II to produce noise-free images of known tissues, which can also serve as a noise-free image dataset. For example, a synthetic image dataset could be obtained by taking a noise-free image C, dividing C into 64×64 non-overlapping patches, extracting local statistical features from each patch, obtaining the corresponding Gaussian component based on the texture type in the noise synthesis model, sampling latent vectors from the Gaussian distribution of this component, decoding the latent vectors into noise ratio patches in the decoder of the variational autoencoder, stitching them together to obtain a complete noise ratio map N, and multiplying the noise-free image C and the noise ratio map N according to the multiplicative noise characteristics to obtain a synthesized noisy image. The ground truth of the noise level in this synthesized noisy image can be obtained by applying a local sliding window to the noise ratio map N and calculating the noise level in the neighborhood of each pixel. The standard deviation gives the true value of the noise level.

[0030] In this embodiment of the invention, after acquiring the synthetic image dataset, the synthetic image dataset is input into a preset initial multiplicative speckle noise denoising model for training. During the training process, a first loss function is used for constraint to obtain an intermediate multiplicative speckle noise denoising model. Specifically, the training process of the model is constrained by the first loss function, and the corresponding function values ​​are calculated using the pixel loss function, multi-scale structure loss function, edge loss function, and noise loss function within the first loss function, thus updating the model parameters in the initial multiplicative speckle noise denoising model. Then, a preset adaptive loss balancing algorithm is used, employing the synthetic dataset that did not participate in the training, to update the weights among the pixel loss function, multi-scale structure loss function, edge loss function, and noise loss function in the first loss function, obtaining the updated weights among the four functions. A comprehensive evaluation index is obtained during the function weight update process. Based on the comprehensive evaluation index of each round, the most suitable model update parameters are selected, thereby obtaining the intermediate multiplicative speckle noise denoising model.

[0031] In this embodiment of the invention, there are corresponding relationships among the four loss functions. The noise loss function is the basis for the operation of all subsequent functions, used to see the distribution pattern of noise. Then, the pixel loss function and the edge loss function are executed in opposition to constrain the model training. Finally, the multi-scale structure loss function is used to constrain the data in the model training process to ensure that the pixel loss function and the edge loss function do not cause overall brightness shift, contrast distortion or texture mode collapse.

[0032] In this embodiment of the invention, after obtaining the intermediate multiplicative speckle noise denoising model, since the previous training was conducted using a synthetic image dataset with synthetic noise, which lacks real ultrasound images from clinical practice, the model is trained again by acquiring a training ultrasound image dataset.

[0033] In this embodiment of the invention, the training ultrasound image dataset can be a publicly available ultrasound image dataset in a database. This dataset can be ultrasound images taken multiple times at the same target point, for example, training ultrasound image data obtained by taking two or three consecutive images of the heart. The intermediate multiplicative speckle noise denoising model is trained using the training ultrasound image dataset, and a second loss function is used to constrain the model during training to obtain a well-trained multiplicative speckle noise denoising model. The constraint process of the second loss function on the model can be achieved by using a self-supervised loss function within the second loss function to constrain the training process using the training ultrasound image dataset, and by using a perceptual loss function within the second loss function to constrain the self-supervised loss function. Through the mutual constraint of the two loss functions, the accuracy of the model in denoising ultrasound images is improved.

[0034] As can be seen, the embodiments of the present invention acquire ultrasound images through a double-layer ultrasound catheter, and then input the ultrasound images into a trained multiplicative speckle noise denoising model to obtain denoised ultrasound images; denoising through the trained multiplicative speckle noise denoising model can improve the denoising accuracy of ultrasound images.

[0035] By acquiring a clinical ultrasound image dataset, a noise synthesis model is trained. Then, a noise-free ultrasound image dataset is input into the trained noise synthesis model to obtain a synthetic image dataset. The synthesized noise has statistically consistent characteristics with real noise, eliminating the distribution bias between synthetic and real noise, thus obtaining clean-labeled training data. Utilizing the clean-label characteristics of the synthetic image dataset, it is input into the initial multiplicative speckle noise denoising model for training. During training, the constraints of the first loss function provide the model with a clear optimization direction and a standard gradient backpropagation path, enabling the model to acquire multiplicative noise suppression capabilities, noise intensity perception capabilities, and strong denoising capabilities in uniform regions and light denoising capabilities in edge regions. Finally, a real clinical ultrasound image dataset is used as training data to train the intermediate multiplicative speckle noise denoising model obtained from the initial training. The second loss function is then used for constraints, allowing the model to adapt to real data, thereby improving the model's denoising accuracy for ultrasound images.

[0036] In an optional embodiment, the trained noise synthesis model can be obtained in the following way: Obtain a clinical ultrasound image dataset; extract noise from the clinical ultrasound image dataset to obtain a set of noise ratio maps; The noise ratio map set is input into a preset variational autoencoder for training to obtain the variational autoencoder parameters; All latent vectors generated during the training of the variational autoencoder are obtained to obtain the latent vector dataset; the latent vector dataset is input into a pre-defined Gaussian mixture model to obtain the Gaussian component parameter set; the Gaussian component parameter set is associated with the tissue texture corresponding to the clinical ultrasound image dataset to obtain the texture mapping table. Based on the variational autoencoder parameters, the Gaussian component parameter set, and the texture mapping table, a trained noise synthesis model is obtained.

[0037] In this optional embodiment, the noise ratio map set can be obtained by performing low-pass smoothing filtering (such as Gaussian filtering, mean filtering, or bilateral filtering) on ​​the clinical ultrasound image dataset to obtain a local mean; then dividing the clinical ultrasound image dataset by this local mean yields the noise ratio map; subsequently, the uniform tissue region of the noise ratio map is cropped to obtain individual noise patches, and the combination of all these noise patches in the noise ratio map set yields the noise ratio map set; the pixel values ​​in the noise ratio map set are distributed around 1, and the histogram reflects the distribution of the true speckle noise.

[0038] In this optional embodiment, those skilled in the art will understand that the variational autoencoder is a deep generative model consisting of two networks: an encoder and a decoder. The parameters of the variational autoencoder can be a set of noise ratio maps extracted from a real clinical ultrasound image dataset. The variational autoencoder is trained by compressing the noise ratio map set into a low-dimensional latent space and then reconstructing the low-dimensional latent space. After training, the variational autoencoder learns to encode (obtain latent vectors) and decode (recover noise patches from latent vectors) any noisy image, at which point the variational autoencoder parameters are retained.

[0039] In this optional embodiment, the latent vector dataset can be formed by compressing the entire set of noise ratio maps into latent vectors using a trained variational autoencoder. This latent vector dataset is then used for subsequent Gaussian mixture modeling. By compressing high-dimensional images into low-dimensional vectors, subsequent clustering and distribution modeling become computationally feasible.

[0040] In this optional embodiment, those skilled in the art will understand that a Gaussian mixture model is a statistical model. The Gaussian component parameter set can be obtained by inputting the latent vector dataset into the Gaussian mixture model, clustering these latent vectors into several Gaussian components, each representing a type of noise texture; for example, fine-grained noise in uniform liver parenchyma represents a Gaussian component, directional elongated texture at the edges of blood vessels represents a Gaussian component, and rough spots at the fascial interface represent a Gaussian component. The continuous distribution structure in the latent space corresponding to the latent vector dataset is transformed into several interpretable noise texture centers, each component representing a type of speckle noise. This allows sampling from any component to generate new noise that conforms to the target image. At this point, retaining the parameters of all corresponding Gaussian components constitutes the Gaussian component parameter set.

[0041] In this optional embodiment, the texture mapping table can associate the Gaussian component parameter set with its corresponding tissue texture type, enabling the synthesis stage to automatically select the matching target noise type based on the local features of the clean image. Thus, noise generation during the synthesis stage is no longer blind random sampling, but rather coordinated with the spatial features of the tissue structure, ensuring the rationality of noise matching. For example, as shown in Table 1: Table 1. Noise Correspondence to the Heart

[0042] Finally, the variational autoencoder parameters, Gaussian component parameter set, and texture mapping table are fed back into the entire model to obtain the trained noise synthesis model.

[0043] In this optional embodiment, the maximum likelihood property of the variational autoencoder itself dictates that it should theoretically cover all samples in the training data; otherwise, the reconstruction loss would penalize omissions. After training, a Gaussian mixture model is used to explicitly cluster all latent vectors in the latent space, ensuring that each Gaussian component precisely corresponds to a noise texture pattern. The EM algorithm of the Gaussian mixture model guarantees that each component is actually aggregated based on the data distribution, not assumed out of thin air. Furthermore, the variational autoencoder learns an explicit structured latent space, which the Gaussian mixture model further resolves into several independent Gaussian distributions, with each component's parameters having a clear physical meaning. We can quantitatively describe the statistical characteristics of each noise pattern and trace the origin of generated samples; for example, which component a generated noise map comes from and how far its latent vector deviates from the center of that component—this information is auditable. Finally, during the training phase, an explicit mapping table is established between each Gaussian mixture model component and its corresponding tissue texture type. During the synthesis stage, given a local patch of a clean image, the system first extracts its texture features, looks up the corresponding texture type in a table, and then samples the latent vectors from the components of the corresponding Gaussian mixture model and decodes them into a noise patch. The entire process is an explicit control of "first determining the region type, and then generating noise accordingly." This explicit mapping from texture type to noise pattern ensures the spatial consistency between the synthesized noise and the anatomical structure. Smooth solid areas will inevitably generate isotropic fine-grained noise, and directional fascial areas will inevitably generate directional elongated textures. This capability is not implicitly learned, but is guaranteed by the design of the embodiments of this invention.

[0044] As can be seen, in this optional embodiment, pure noise ratio samples are extracted from uniform tissue regions of real clinical ultrasound image datasets using a multiplicative separation method. Then, a variational autoencoder is used to compress the noise patches into latent vectors. A Gaussian mixture model is then used to cluster and model the latent space, making the noise distribution explicit into several Gaussian components. Each component represents a typical noise texture pattern, and the Gaussian component parameters corresponding to each Gaussian component are retained to obtain a set of Gaussian component parameters. The variational autoencoder ensures that the generated noise is diverse and realistic, and the Gaussian mixture model makes the noise pattern interpretable and controllable, thereby improving the accuracy of noise synthesis and thus improving the accuracy of synthesized images in the synthesized image dataset, providing a high-quality training dataset for subsequent model training.

[0045] In another optional embodiment, the synthetic image dataset is input into a preset initial multiplicative speckle noise denoising model for training. During the training process, a preset first loss function is used for constraint to obtain an intermediate multiplicative speckle noise denoising model, including: The synthetic image dataset is input into a preset initial multiplicative speckle noise denoising model to obtain a denoised image and a noise level map; Based on the denoised image and noise level map, the value of the first loss function is obtained through the first loss function; and the model update parameters are obtained based on the value of the first loss function. Obtain the validation image dataset; based on the model update parameters, use the validation image dataset to monitor and update the weights corresponding to the first loss function through a preset adaptive loss balance algorithm to obtain the function update weights; Obtain the comprehensive evaluation index; judge the comprehensive evaluation index to obtain the model update parameter corresponding to the maximum comprehensive evaluation index; and obtain the intermediate multiplicative speckle noise denoising model based on the model update parameter.

[0046] In this optional embodiment, after inputting the synthetic image dataset into a preset initial multiplicative speckle noise denoising model, the denoised image and noise level map are calculated, and then the true value of the noise level can be obtained. In this optional embodiment, the initial multiplicative speckle noise denoising model architecture can be: a CNN encoder, a CNN / Transformer hybrid bottleneck layer, a CNN decoder, and a noise estimation branch. The CNN encoder is used for downsampling, extracting multi-scale features layer by layer. In the CNN / Transformer hybrid bottleneck layer, the Transformer block captures global long-range dependencies through a self-attention mechanism, allowing the network to understand the overall structure of the image; the CNN block preserves the extraction capability of local detail features, ensuring no loss of tissue texture. The two work alternately, simultaneously modeling both global context and local details. The CNN decoder is used for upsampling, restoring resolution layer by layer and outputting a denoised image. A spatial attention module is embedded at each skip connection between the encoder and decoder, calculating attention weights based on the local gradients and texture information of the encoder features. Regions with large gradients (lesion edges, tissue boundaries) are assigned higher weights, and the decoder prioritizes preserving details in these regions; regions with small gradients (uniform tissue) are assigned lower weights, and the decoder performs stronger noise suppression in these regions. The noise estimation branch is used to create a lightweight subnetwork from the shallow features of the encoder. After passing through a few convolutional layers, it outputs a single-channel noise level map. The noise level map is the same size as the input image, and each pixel value represents an estimate of the noise intensity at that location. By designing the noise estimation branch, the model learns to infer how strong the noise is for each pixel from the input image.

[0047] In this optional embodiment, the first loss function value may include pixel loss value, multi-scale structure loss value, edge loss value, and noise estimation loss value; wherein, the pixel loss value may be calculated as follows:

[0048] In the above formula, This represents the pixel loss value. This represents the number of pixels in the vertical direction of the image. This represents the number of pixels in the horizontal direction of the image. The row index of the pixel. The column index for pixels. To denoise the image at the first Line number Column pixel values, For noise-free ultrasound images in the first... Line number The pixel values ​​of the column.

[0049] The edge loss value can be calculated as follows:

[0050] In the above formula, This represents the edge loss value. The edge map of the denoised image obtained by extracting edges from the denoised image using the Sobel operator corresponds to the edge at the th... Line number Column pixel values, The edge map of the noise-free ultrasound image obtained by extracting edges from the noise-free ultrasound image using the Sobel operator corresponds to the edge map at the 1st... Line number The pixel values ​​of the column.

[0051] The multi-scale structural loss value can be calculated as follows:

[0052] In the above formula, This represents the multi-scale structural loss value. It is a multi-scale structural similarity index used to measure the similarity between denoised images and noise-free ultrasound images.

[0053] The noise estimation loss value can be calculated as follows:

[0054] In the above formula, To estimate the loss value for noise, The noise level map at the 1st Line number The predicted value of the column, The true value of the noise level is at the th Line number The value of the column.

[0055] The first loss function value can be calculated as follows:

[0056] In the above formula, The first loss function value, For peak signal-to-noise ratio weights, For edge weights, For structural similarity weights, Weights are used to estimate noise.

[0057] In this optional embodiment, the model update parameters can be obtained by saving all the weights in the first loss function value; the validation image dataset can be a portion of the synthetic image dataset that was not used in training; the synthetic image dataset is automatically divided into two parts during training, one part for training and one part for validation.

[0058] In this optional embodiment, the function update weights can be calculated using a validation image dataset and the weights corresponding to the first loss function can be updated in real time using an adaptive loss balancing algorithm for this scheme. Specifically, the adaptive loss balancing algorithm can first calculate four indices: peak signal-to-noise ratio index, structural similarity index, edge preservation index, and noise estimation loss index, and then adjust the corresponding weights by judging the indices.

[0059] In this optional embodiment, the comprehensive evaluation index can be obtained during the weight update process of the acquisition function and used to evaluate the model training status. By selecting the weight with the highest comprehensive evaluation index score, the model with the best training state is found. Furthermore, the model update parameters corresponding to this maximum comprehensive evaluation index are fed back into the model, thus obtaining the intermediate multiplicative speckle noise denoising model.

[0060] Furthermore, because the synthetic image dataset includes noise-free ultrasound images as the standard answer, the network knows what the correct output for each pixel should be. Pixel loss drives the network output to approximate a clean image, resulting in a clear denoising direction and fast convergence. Under the supervision of the ground truth noise level, the noise estimation branch learns to infer the noise strength of each pixel from the input image. This is crucial because if the network doesn't know where the noise is strong and where it's weak, subsequent differential denoising is impossible. Edge loss works in opposition to pixel loss during training. Pixel loss aims to smooth out uniform regions, while edge loss aims to prevent lesion boundaries from being smoothed out. In this game, the network exhibits a tendency to perform strong denoising in uniform regions and weak denoising in edge regions.

[0061] As can be seen, in this optional embodiment, the initial multiplicative speckle noise denoising model is trained using a synthetic image dataset. During the training process, the model is constrained by a first loss function, which can quickly establish the model's basic denoising ability, noise perception ability, and edge protection awareness. By validating the image dataset and using an adaptive loss balancing algorithm, the weights of the first loss function are updated, which enables the model to obtain the optimal weights while improving the convergence speed of training. Finally, an intermediate multiplicative speckle noise denoising model is obtained, improving the denoising accuracy.

[0062] In yet another optional embodiment, the method may further include: The synthesized image dataset is processed using the noise loss function corresponding to the noise estimation loss value to obtain a noise level map; Based on the noise level map, the pixel loss function corresponding to the pixel loss value is used for constraint to obtain a smoothed and denoised map; Based on the noise level map, the gradient constraint map is obtained by using the edge loss function corresponding to the edge loss value to constrain the noise level map. Based on the smoothed denoised map and the gradient constraint map, a denoised edge-preserving map is obtained; the multi-scale structural loss function corresponding to the multi-scale structural loss value is used to constrain the denoised edge-preserving map to obtain the trained denoised image.

[0063] In this optional embodiment, the noise loss function can be such that, at the start of training constrained by the first loss function, the training objective is to drive the noise estimation branch within the network to learn to infer the strength of noise at each pixel location from the input image. The noise loss function minimizes the mean squared error between the predicted noise level map and the true noise level map, enabling the network to first discern the distribution pattern of noise; this step is fundamental to all subsequent differential processing. Without accurate spatial noise perception, targeted denoising cannot be achieved.

[0064] In this optional embodiment, the noise level map can be a noise intensity distribution map automatically generated during network forward propagation. This map, as a pixel-level adjustment signal, clearly marks high-noise and low-noise areas in the image. It does not directly produce denoising effects, but rather serves as a spatial attention guide map, conveying regional differentiation information to downstream pixel loss and edge loss: high-noise areas require stronger smoothing suppression, while low-noise areas require more precise edge preservation. Those skilled in the art will understand that smoothing denoising involves smoothing out messy, fluctuating interference noise and specks, preserving the true and effective trend signal.

[0065] In this optional embodiment, the smoothed denoising image can be obtained by a pixel loss function receiving an indication from a noise level map and applying a stronger smoothing force to areas of uniform tissue with high noise levels, thus obtaining a smoothed denoising image that numerically approximates a noise-free ultrasound image. The pixel loss function aims for thorough overall denoising, but if unconstrained, it can blur edges as well; therefore, an edge loss function is needed to constrain this process. Those skilled in the art will understand that obtaining a smoothed denoising image is equivalent to performing a smoothing denoising operation, which involves smoothing out jagged, fluctuating interference noise and specks while preserving the true and effective trend signal.

[0066] In this optional embodiment, the gradient constraint map can be such that the edge loss function receives the indication of the noise level map and strengthens the gradient consistency constraint in areas with weak noise (such as lesion edges and tissue boundaries) to protect the sharpness of these diagnostic key locations. It counteracts the smoothing trend of pixel loss and forms a dynamic game of "denoising-preserving edges".

[0067] Optionally, the pixel loss function and the edge loss function work together under the control of the same noise level map: in areas with strong noise, the pixel fidelity loss dominates and the denoising effect is strong; in areas with weak noise, the edge preservation loss dominates and the details are preserved. This differentiated execution based on the same perceptual signal enables the model to learn spatially adaptive denoising behavior autonomously without artificial rules.

[0068] In this optional embodiment, the denoising edge-preserving graph can be a combination of the gradient constraint graph and the smoothing denoising graph. Figure 1 By imposing constraints, we can obtain a denoised edge-preserving image; then, by applying global constraints through the multi-scale structural loss function, we can output the denoised image of the model, which is the first training denoised image.

[0069] In this optional embodiment, the multi-scale structural loss function constrains the final output from a global perspective. It compares the structural similarity between the denoised image and the noise-free ultrasound image at multiple resolution levels, ensuring that pixel loss and edge loss do not cause overall brightness shift, contrast distortion, or texture pattern collapse during local game theory. The multi-scale structural loss function acts as a constraint on the entire model training, ensuring that the visual structure and diagnostic semantics of the entire image remain stable regardless of local balance.

[0070] As can be seen, in this optional embodiment, since multiplicative speckle noise exists simultaneously in the uniform region and the edge of the lesion, but the uniform region requires heavy denoising while the edge region requires light denoising or no processing, a design is made that uses the noise loss function as the initiator for identifying noise, and the pixel loss function and the edge loss function as dual constraints, so that the model can automatically learn to process differentially. Finally, the edge loss function ensures that the pixel loss and edge loss will not cause overall brightness shift, contrast distortion or texture mode collapse when playing a local game, thereby improving the denoising accuracy of the model.

[0071] In another optional embodiment, the weights corresponding to the first loss function are monitored and updated using a preset adaptive loss balancing algorithm on the validation image dataset to obtain the function update weights, including: Based on the validation image dataset, the peak signal-to-noise ratio, structural similarity ratio, edge preservation ratio, and noise estimation ratio were obtained. When it is determined that the rate of change of edge preservation decreases in two consecutive rounds of training, the weights corresponding to the edge loss function are adjusted to obtain the adjusted edge weights. When it is determined that the rate of change of noise estimation decreases in three consecutive rounds of training, the weights corresponding to the noise loss function are adjusted to obtain the adjusted noise estimation weights. When it is determined that the peak signal-to-noise ratio change rate and the structural similarity ratio change rate are both less than the preset change threshold, the weights corresponding to the multi-scale structural loss function are adjusted to obtain the adjusted structural similarity weights. When it is determined that the peak signal-to-noise ratio change rate, structural similarity change rate, edge preservation change rate, and noise estimation change rate are all less than the preset global variation threshold, the edge weight, noise estimation weight, structural similarity weight, and peak signal-to-noise ratio weight corresponding to the pixel loss function are adjusted to obtain the adjustment function weight. The weights adjusted in each round are normalized to obtain the function update weights.

[0072] In this optional embodiment, the peak signal-to-noise ratio (PSNR) index, structural similarity index, edge preservation index, and noise estimation loss index are calculated by verifying that the image dataset is only computed without backpropagation. The PSNR index measures the optimization effect of pixel loss, the structural similarity index measures the optimization effect of multi-scale structural loss, the edge preservation index measures the optimization effect of edge loss, and the noise estimation loss index measures the optimization effect of noise loss. The PSNR index is calculated as follows:

[0073] In the above formula, Peak signal-to-noise ratio, The maximum possible value of an image pixel. To denoise the image at the first Line number Column pixel values, For noise-free ultrasound images in the first... Line number The pixel values ​​of the column.

[0074] The structural similarity index is calculated as follows:

[0075] In the above formula, It is a structural similarity index. The average pixel value of the denoised image within the current window. This represents the pixel mean of the noise-free ultrasound image within the current window. The pixel covariance between the denoised image and the noise-free ultrasound image within the current window. The standard deviation of pixels in the denoised image within the current window. The standard deviation of the pixels in the noise-free ultrasound image within the current window. The brightness stability constant, is the contrast stability constant.

[0076] The edge preservation index is calculated as follows:

[0077] In the above formula, To maintain the index at the edge, To denoise the image edge map in the first... Line number The value of the column, For noise-free ultrasound images in the first... Line number The value of the column.

[0078] The noise estimation loss index is calculated as follows:

[0079] In the above formula, To estimate the loss exponent for noise, The noise level map at the 1st Line number The predicted value of the column, The true value of the noise level is at the th Line number The value of the column.

[0080] After calculating the exponents corresponding to all loss functions, the rate of change for each loss function is obtained by calculating the rate of change of each indicator relative to the previous period. This includes the peak signal-to-noise ratio (PSNR) change rate, structural similarity change rate, edge preservation change rate, and noise estimation change rate among the four sets of functions. The method for calculating the rate of change of each indicator relative to the previous period is as follows:

[0081] In the above formula, t represents the number of training rounds. For each loss function in round t, the exponent is... is the exponent corresponding to each loss function in round t-1.

[0082] In this optional embodiment, adjusting the edge weights can be done as follows: when it is determined that the edge preservation rate of change decreases in two consecutive training rounds, it means that the edges of the denoised image are becoming blurred, and the edge preservation loss is suppressed by the pixel fidelity loss. At this time, it is necessary to adjust the edge weights corresponding to the edge loss function, increasing the weight of the edge preservation loss to counteract over-smoothing. The adjustment method can be to multiply the edge weights by a coefficient, compare them with a set value, and select the minimum value; for example, select... The minimum value, where This represents the edge weights. When the conditions are not met, the coefficient can be multiplied by 1 to keep the weights unchanged. This is the adjusted weight, also known as the adjusted edge weights.

[0083] In this optional embodiment, adjusting the noise estimation weights can be achieved by: when it is determined that the rate of change of the noise estimation decreases in three consecutive training rounds, it indicates that the noise estimation branch has converged, and continuing to force optimization is not beneficial and consumes network capacity. In this case, it is necessary to decay the noise estimation weights to free up learning capacity for other tasks. For example, the noise estimation weights can be multiplied by a coefficient less than 1 to reduce the noise estimation weights. When the condition is not met, the coefficient can be multiplied by 1 to keep the weights unchanged. This is also the adjusted weight, which is the adjusted noise estimation weight.

[0084] In this optional embodiment, adjusting the structural similarity weight can be done as follows: when it is determined that both the peak signal-to-noise ratio change rate and the structural similarity ratio change rate are less than a preset change threshold, it indicates that pixel-level denoising is approaching saturation, but the structural texture may still be damaged and not detected by the pixel loss function. In this case, it is necessary to enhance the structural similarity loss to further preserve the texture at multiple scales. The change threshold can be a specific positive integer, such as 0.5%, 0.6%, 0.7%, etc. Enhancing the structural similarity loss can be done by multiplying the structural similarity weight by a coefficient greater than 1 to increase the structural similarity weight. When the condition is not met, the coefficient can be multiplied by 1 to keep the weight unchanged. This is also the adjusted weight, which is the adjusted structural similarity weight.

[0085] In this optional embodiment, when it is determined that the peak signal-to-noise ratio change rate, structural similarity ratio change rate, edge preservation ratio change rate, and noise estimation ratio change rate are all less than a preset global variation threshold, it indicates that all tasks have stalled, and the weights of the current loss function may have deviated too far from the initial equilibrium state. A gentle weight reversion to the initial equilibrium point is needed to avoid the weights becoming fixed at a local optimum. The weights can be adjusted by multiplying all weights by a coefficient of 0.95. The global variation threshold can be a specific positive integer, such as 0.5%, 0.6%, 0.7%, etc.

[0086] In this optional embodiment, after adjusting the weights in each round, all weights are normalized to keep the sum constant; for example, the normalization calculation method can be:

[0087] Optionally, each weight can have a boundary set, for example, The peak signal-to-noise ratio (PSNR) weight boundary is [0.5, 1.5]. The edge weight boundary is [0.2, 1.0]. The structural similarity weight boundary is [0.1, 0.8]. The noise estimation weight boundary is [0.05, 0.5].

[0088] As can be seen, this optional embodiment uses an adaptive loss balancing algorithm to adjust the weights of the first loss function in real time. When the edge quality decreases, the edge weights are strengthened; when the noise estimation is saturated, the weights are weakened to release the learning capacity; when denoising stagnates, the structural similarity weights are strengthened; and when all stagnation occurs, all weights are gently shifted back to avoid local solidification. This ensures that the four objectives of denoising, preserving structure, preserving edges, and noise perception can maintain a dynamic balance throughout the entire training cycle. The adjustment process is traceable throughout, thereby improving the denoising accuracy of the model.

[0089] In another optional embodiment, the comprehensive evaluation index is determined to obtain the model update parameters corresponding to the maximum comprehensive evaluation index, including: Based on the validation image dataset, the peak signal-to-noise ratio index and structural similarity index are obtained; Based on the peak signal-to-noise ratio index and the structural similarity index, a comprehensive evaluation index is obtained; The highest comprehensive evaluation index is used as the model update index; the model update parameters are determined based on the model update index.

[0090] In this optional embodiment, the peak signal-to-noise ratio (PSNR) index and structural similarity index are obtained by calculating the verification image dataset. Since the evaluation of the entire image denoising process is based on the PSNR index and structural similarity index, the PSNR index and structural similarity index are added together to obtain a comprehensive evaluation index. The optimal model parameters during the training process are selected using the comprehensive evaluation index.

[0091] In this optional embodiment, since model training is conducted round by round, a comprehensive evaluation index is designed, and the best index among the comprehensive evaluation index data is selected as the model update index; the model update index is selected to correspond to the model update parameters, thereby obtaining the model with the best training state.

[0092] As can be seen, in this optional embodiment, in order to more intuitively understand the training status of the model, the training status of the model in different training rounds can be reflected by calculating the comprehensive evaluation index; the model parameters corresponding to the best comprehensive evaluation index are selected as the best model update parameters in this model training process, and the model update parameters at this time are fed back into the model to obtain the best training model in this round of training.

[0093] In another optional embodiment, the training ultrasound image dataset is input into the intermediate multiplicative speckle noise denoising model for training. During the training process, a preset second loss function is used for constraint to obtain the trained multiplicative speckle noise denoising model, including: The training ultrasound image dataset is input into the intermediate multiplicative speckle noise denoising model, and the second loss function value is obtained through self-supervised training; the second loss function value includes the self-supervised loss function value and the perceptual loss function value. The second loss function value is fed back into the intermediate multiplicative speckle noise denoising model to obtain the trained multiplicative speckle noise denoising model.

[0094] In this optional embodiment, the training method in this scheme can be a self-supervised training method. Those skilled in the art will understand that a self-supervised training method can be one that automatically generates supervision signals directly from the inherent structure, contextual relationships, local features, and temporal correlations of the data itself without the need for manual labeling, and relies on the data itself to construct training tasks, thereby completing the training paradigm of unlabeled pre-training and feature learning of the model.

[0095] In this optional embodiment, the training ultrasound image dataset can be achieved by taking two consecutive images of the same target object. The data of these two images are identical except for the time taken. With this image acquisition method, the noise in the first image and the second image will definitely not be completely the same.

[0096] In this optional embodiment, the self-supervised loss function corresponding to the self-supervised loss function value can be such that the input is treated as one type of noise, and the user is required to predict a completely different type of noise. The model cannot deduce the target noise from the input noise (because the two are independent and random), but can only learn to discard all noise and output the underlying clean signal. By using real data, without a standard answer, the network adapts to the subtle characteristics of real noise.

[0097] In this optional embodiment, if only the self-supervised loss function corresponding to the self-supervised loss function value is used for constraint, then under the statistical pressure of a large number of training samples, the network finds that outputting an "average value" for all uncertain noise regions is the strategy with the least long-term risk. Each pixel in the output image tends to the expected value of that location across all training samples. This results in noise in uniform regions being indeed suppressed, but lesion edges, tissue boundaries, and other structures are also smoothed out, making the image blurry and causing a loss of diagnostic information. To address this situation, this solution designs a perceptual loss function to constrain the self-supervised loss function.

[0098] In this optional embodiment, the perceptual loss function can be the LPIPS loss function, which is the learned perceptual image patch similarity loss function. Those skilled in the art will understand that the LPIPS loss function extracts feature maps from multiple intermediate layers, normalizes the channel dimension of each feature map, then uses a set of learnable weights to weight the importance of each channel, and finally calculates the weighted feature difference.

[0099] In this optional embodiment, the constraint process of the perceptual loss function on the self-supervised loss function can be as follows: by applying gradient constraints opposite to the direction of the self-supervised loss through the deep feature space, the constraints are automatically relaxed in uniform regions to allow strong denoising, and the constraints are automatically strengthened in lesion edges and texture-rich regions to prevent over-smoothing.

[0100] By constraining the denoised image to maintain consistency with the input real image in the feature space of the trained model network, unnatural smoothness or artifacts caused by self-supervised training are prevented, thus maintaining the realism of the texture.

[0101] In this optional embodiment, the second loss function may be composed of a self-supervised loss function and a perceptual loss function.

[0102] As can be seen, in this optional embodiment, the intermediate multiplicative speckle noise denoising model is trained using synthetic data, which may introduce errors in clinical practice. Therefore, this optional embodiment trains the model using a real clinical training ultrasound image dataset. During training, both a self-supervised loss function and a perceptual loss function are used to constrain the training. The self-supervised loss function promotes denoising, while the perceptual loss function prevents over-smoothing, achieving a balance and thus obtaining a well-trained multiplicative speckle noise denoising model, thereby improving the denoising accuracy of ultrasound images.

[0103] Example 2 Please see Figure 4 , Figure 4 This invention discloses an ultrasound image processing device based on a double-layer ultrasound catheter, which may include: The image acquisition module 201 is used to acquire an ultrasound image to be processed; the ultrasound image to be processed is acquired through a double-layer ultrasound catheter. The image processing module 202 is used to input the ultrasound image to be processed into a trained multiplicative speckle noise denoising model to obtain a denoised ultrasound image.

[0104] The trained multiplicative speckle noise denoising model is obtained through some or all of the steps in the ultrasound image processing method based on double-layer ultrasound catheter described in any of the embodiments.

[0105] As can be seen, the embodiments of the present invention acquire ultrasound images through a double-layer ultrasound catheter, and then input the ultrasound images into a trained multiplicative speckle noise denoising model to obtain denoised ultrasound images; denoising through the trained multiplicative speckle noise denoising model can improve the denoising accuracy of ultrasound images.

[0106] Example 3 Please see Figure 5 , Figure 5This is a schematic diagram of a structure including a memory and a processor device disclosed in an embodiment of the present invention. Figure 5 As shown, the device including memory and processor may include: Memory 301 storing executable program code; Processor 302 coupled to memory 301; The processor 302 calls the executable program code stored in the memory 301 to execute some or all of the steps in any of the ultrasound image processing methods based on double-layer ultrasound catheters in Embodiment 1 of the present invention.

[0107] Example 4 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 data processing methods for heterogeneous devices based on model adapters disclosed in Embodiment 1 of this invention.

[0108] 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.

[0109] 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.

[0110] 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. An ultrasound image processing method based on a double-layer ultrasound catheter, characterized in that, The method includes: Acquire an ultrasound image to be processed; the ultrasound image to be processed is acquired through a double-layer ultrasound catheter; input the ultrasound image to be processed into a trained multiplicative speckle noise denoising model to obtain a denoised ultrasound image; The trained multiplicative speckle noise denoising model is obtained in the following way: A synthetic image dataset is obtained; the synthetic image dataset is obtained by inputting the acquired noise-free ultrasound image dataset into a pre-trained noise synthesis model; the pre-trained noise synthesis model is obtained in the following way: Obtain a clinical ultrasound image dataset; extract noise from the clinical ultrasound image dataset to obtain a noise ratio map set; The noise ratio map set is input into a preset variational autoencoder for training to obtain variational autoencoder parameters; All latent vectors generated by the variational autoencoder during training are obtained to obtain a latent vector dataset; the latent vector dataset is input into a preset Gaussian mixture model to obtain a Gaussian component parameter set; the Gaussian component parameter set is associated with the tissue texture corresponding to the clinical ultrasound image dataset to obtain a texture mapping table. The trained noise synthesis model is obtained based on the variational autoencoder parameters, the Gaussian component parameter set, and the texture mapping table. The synthesized image dataset is input into a preset initial multiplicative speckle noise denoising model for training. During training, a preset first loss function is used for constraint to obtain an intermediate multiplicative speckle noise denoising model. The first loss function includes a noise loss function, a pixel loss function, an edge loss function, and a multi-scale structure loss function. The pixel loss function and the edge loss function are used to constrain the processing results of the noise loss function. The multi-scale structure loss function is used to constrain the pixel loss function and the edge loss function. Obtain a training ultrasound image dataset; input the training ultrasound image dataset into the intermediate multiplicative speckle noise denoising model for training, and use a preset second loss function for constraint during the training process to obtain a trained multiplicative speckle noise denoising model.

2. The ultrasound image processing method based on a double-layer ultrasound catheter according to claim 1, characterized in that, The step of inputting the synthetic image dataset into a preset initial multiplicative speckle noise denoising model for training, and using a preset first loss function for constraints during training to obtain an intermediate multiplicative speckle noise denoising model, includes: The synthesized image dataset is input into a preset initial multiplicative speckle noise denoising model to obtain a denoised image and a noise level map; Based on the denoised image and the noise level map, a first loss function value is obtained through a first loss function; model update parameters are obtained based on the first loss function value; the first loss function value includes pixel loss value, multi-scale structure loss value, edge loss value, and noise estimation loss value; Obtain the verification image dataset; the verification image dataset refers to the portion of the synthetic image dataset that was not used in training; Based on the model update parameters, the weights corresponding to the first loss function are monitored and updated using the verification image dataset through a preset adaptive loss balance algorithm to obtain the function update weights. Obtain the comprehensive evaluation index; the comprehensive evaluation index is obtained during the process of obtaining the function update weights; The comprehensive evaluation index is judged to obtain the model update parameters corresponding to the maximum comprehensive evaluation index; based on the model update parameters corresponding to the maximum comprehensive evaluation index, the intermediate multiplicative speckle noise denoising model is obtained.

3. The ultrasound image processing method based on a double-layer ultrasound catheter according to claim 2, characterized in that, The method further includes: The synthesized image dataset is processed using the noise loss function corresponding to the noise estimation loss value to obtain a noise level map; Based on the noise level map, a smoothed and denoised map is obtained by using the pixel loss function corresponding to the pixel loss value as a constraint. Based on the noise level map, the gradient constraint map is obtained by using the edge loss function corresponding to the edge loss value to constrain the noise level map. Based on the smoothed denoised map and the gradient constraint map, a denoised edge-preserving map is obtained; the denoised edge-preserving map is constrained using the multi-scale structural loss function corresponding to the multi-scale structural loss value to obtain a trained denoised image.

4. The ultrasound image processing method based on a double-layer ultrasound catheter according to claim 3, characterized in that, The step of using the verification image dataset to monitor and update the weights corresponding to the first loss function through a preset adaptive loss balance algorithm to obtain the function update weights includes: Based on the verification image dataset, the peak signal-to-noise ratio change rate, structural similarity ratio change rate, edge preservation ratio change rate, and noise estimation ratio change rate are obtained. When it is determined that the edge preservation rate of change decreases in two consecutive training rounds, the weights corresponding to the edge loss function are adjusted to obtain the adjusted edge weights; When it is determined that the rate of change of the noise estimate decreases in three consecutive rounds of training, the weights corresponding to the noise loss function are adjusted to obtain the adjusted noise estimate weights. When it is determined that the peak signal-to-noise ratio change rate and the structural similarity ratio change rate are both less than a preset change threshold, the weights corresponding to the multi-scale structural loss function are adjusted to obtain the adjusted structural similarity weights. When it is determined that the peak signal-to-noise ratio change rate, the structural similarity change rate, the edge preservation change rate, and the noise estimation change rate are all less than a preset global variation threshold, the adjustment edge weight, the adjustment noise estimation weight, the adjustment structural similarity weight, and the peak signal-to-noise ratio weight corresponding to the pixel loss function are adjusted to obtain the adjustment function weight; The weights adjusted in each round are normalized to obtain the function update weights.

5. The ultrasound image processing method based on a double-layer ultrasound catheter according to claim 4, characterized in that, The step of judging the comprehensive evaluation index to obtain the model update parameters corresponding to the maximum comprehensive evaluation index includes: Based on the verification image dataset, the peak signal-to-noise ratio index and the structural similarity index are obtained; A comprehensive evaluation index is obtained based on the peak signal-to-noise ratio index and the structural similarity index; The largest comprehensive evaluation index is used as the model update index; the model update parameters are determined based on the model update index.

6. The ultrasound image processing method based on a double-layer ultrasound catheter according to claim 5, characterized in that, The step of inputting the training ultrasound image dataset into the intermediate multiplicative speckle noise denoising model for training, and using a preset second loss function for constraint during the training process to obtain the trained multiplicative speckle noise denoising model includes: The training ultrasound image dataset is input into the intermediate multiplicative speckle noise denoising model, and a second loss function value is obtained through self-supervised training; the second loss function value includes a self-supervised loss function value and a perceptual loss function value. The second loss function value is fed back into the intermediate multiplicative speckle noise denoising model to obtain the trained multiplicative speckle noise denoising model; wherein, the self-supervised loss function corresponding to the self-supervised loss function value is used to constrain the training of the training ultrasound image dataset; the perceptual loss function corresponding to the perceptual loss function value is used to constrain the self-supervised loss function.

7. An ultrasound image processing device based on a double-layer ultrasound catheter, characterized in that, The device includes: An image acquisition module is used to acquire an ultrasound image to be processed; the ultrasound image to be processed is acquired through a double-layer ultrasound catheter. The image processing module is used to input the ultrasound image to be processed into a trained multiplicative speckle noise denoising model to obtain a denoised ultrasound image. The trained multiplicative speckle noise denoising model is obtained in the following way: A synthetic image dataset is obtained; the synthetic image dataset is obtained by inputting the acquired noise-free ultrasound image dataset into a pre-trained noise synthesis model; the pre-trained noise synthesis model is obtained in the following way: Obtain a clinical ultrasound image dataset; extract noise from the clinical ultrasound image dataset to obtain a noise ratio map set; The noise ratio map set is input into a preset variational autoencoder for training to obtain variational autoencoder parameters; All latent vectors generated by the variational autoencoder during training are obtained to obtain a latent vector dataset; the latent vector dataset is input into a preset Gaussian mixture model to obtain a Gaussian component parameter set; the Gaussian component parameter set is associated with the tissue texture corresponding to the clinical ultrasound image dataset to obtain a texture mapping table. The trained noise synthesis model is obtained based on the variational autoencoder parameters, the Gaussian component parameter set, and the texture mapping table. The synthesized image dataset is input into a preset initial multiplicative speckle noise denoising model for training. During training, a preset first loss function is used for constraint to obtain an intermediate multiplicative speckle noise denoising model. The first loss function includes a noise loss function, a pixel loss function, an edge loss function, and a multi-scale structure loss function. The pixel loss function and the edge loss function are used to constrain the processing results of the noise loss function. The multi-scale structure loss function is used to constrain the pixel loss function and the edge loss function. Obtain a training ultrasound image dataset; input the training ultrasound image dataset into the intermediate multiplicative speckle noise denoising model for training, and use a preset second loss function for constraint during the training process to obtain a trained multiplicative speckle noise denoising model.

8. An apparatus comprising a memory and a processor, characterized in that, The device includes: 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 ultrasound image processing method based on a double-layer ultrasound catheter as described in any one of claims 1-6.

9. A computer storage medium, characterized in that, The computer storage medium stores computer instructions, which, when invoked by a processor, are used to execute the ultrasound image processing method based on a double-layer ultrasound catheter as described in any one of claims 1-6.

Images