Ultrasonic image denoising method and system fusing speckle distribution constraint and structural consistency constraint

Abstract

The invention discloses an ultrasonic image denoising method and system fusing speckle distribution constraint and structural consistency constraint. The method comprises the following steps: firstly, acquiring an ultrasonic video sequence, preprocessing an image, and inputting the preprocessed image into a DVDNet video denoising network to generate a low-noise ultrasonic image; secondly, pairing the low-noise ultrasonic image with an original single-frame ultrasonic image to form a training data set of a U-Net single-frame denoising network; then, a U-Net single-frame denoising network is constructed, and training is carried out; obtaining a to-be-denoised single-frame ultrasonic image again, preprocessing the to-be-denoised single-frame ultrasonic image, inputting the preprocessed to-be-denoised single-frame ultrasonic image into the trained U-Net single-frame denoising network, and outputting the denoised single-frame ultrasonic image; and finally, post-processing the single-frame ultrasonic image, and outputting the final de-noised single-frame ultrasonic image. According to the method, a pseudo-true value is generated by adopting DVDNet, and U-Net double-constraint training is combined, so that noise suppression and anatomical structure fidelity are realized, the consistency of speckle statistical rules is ensured, and an excessive smooth edge structure is avoided.

Description

An ultrasound image denoising method and system that integrates speckle distribution constraints and structural consistency constraints Technical Field

[0001] This invention belongs to the field of medical image processing and computer vision technology, and particularly relates to an ultrasound image denoising method and system that integrates speckle distribution constraints and structural consistency constraints. Background Technology

[0002] Ultrasound imaging technology, with its outstanding advantages such as no radiation, good real-time performance, low cost and portability, has become an indispensable diagnostic tool in modern clinical medicine, playing a vital role, especially in obstetric examination, cardiac function assessment and abdominal organ diagnosis.

[0003] However, the quality of ultrasound images is inherently limited by the physical imaging principles of the device. During imaging, as emitted ultrasound waves propagate through biological tissue, they interact with microscopic structures much smaller than the wavelength, generating numerous backscattered echoes. These echoes coherently superimpose at the receiving probe, forming a multiplicative noise known as "speckle." This speckle noise not only reduces the signal-to-noise ratio and contrast of the image but also often obscures or blurs fine structures and tissue features. This not only introduces significant subjectivity and uncertainty into the doctor's visual interpretation but also severely interferes with the accuracy and reliability of subsequent automated identification, segmentation, and measurement by computer-aided diagnostic systems.

[0004] Currently, ultrasound image denoising techniques can be mainly divided into traditional image processing methods and deep learning-based methods.

[0005] Traditional image processing methods, such as anisotropic diffusion filtering, nonlocal means, block matching, and 3D filtering, suppress noise by modeling the statistical properties of images in the spatial or transform domains. However, these methods are essentially based on general, content-independent image statistical priors, lacking a specific understanding of the complex anatomical structures and pathological features in medical images that are crucial for diagnosis. When processing fetal ultrasound images with extremely low signal-to-noise ratios, a fundamental trade-off often arises: over-filtering leads to blurred tissue edges and loss of detail, while insufficient filtering fails to effectively suppress noise, making it difficult to meet the demands of high-precision diagnosis.

[0006] Deep learning-based methods, particularly convolutional neural networks, have demonstrated the potential to surpass traditional approaches in image denoising through end-to-end learning. Early deep learning methods typically followed a supervised learning paradigm, requiring a large number of precisely paired "noise-clean" image pairs as training data. The model was trained by minimizing the error between the network's predicted image and the "clean" ground truth image. However, this paradigm faces an almost insurmountable data bottleneck in the field of medical ultrasound imaging. Since ultrasound images are essentially the result of deterministic speckle patterns generated by tissue scattering and system noise, obtaining absolutely "clean" and noise-free ultrasound images is physically impossible. While approximate data pairs can be constructed through simulation synthesis or by using data from other imaging modalities, this approach introduces a severe domain shift problem—the "clean" images seen by the model during the learning process are fundamentally different from the speckle patterns that always exist in the real clinical environment. This leads to poor generalization ability and performance on real clinical data, and may even introduce unrealistic artifacts, misleading diagnoses. Summary of the Invention

[0007] To address the problems existing in the above-mentioned background technology, the present invention proposes an ultrasound image denoising method and system that integrates speckle distribution constraints and structural consistency constraints.

[0008] This invention combines the video denoising network DVDNet with a dedicated U-Net deep network, enabling the generation of visually clean ultrasound images while ensuring they conform to the physical and statistical laws of ultrasound speckle, thus protecting the integrity of key anatomical structures. Through supervision of pseudo-ground values ​​generated by the video denoising network, a composite optimization strategy is introduced, incorporating pixel reconstruction loss, speckle distribution constraint loss, and structural consistency constraint loss. This achieves the synergistic preservation of ultrasound speckle characteristics and anatomical structures without requiring real, noise-free annotations, providing highly reliable image input for clinical diagnosis.

[0009] In a first aspect, the present invention provides an ultrasound image denoising method that integrates speckle distribution constraints and structural consistency constraints, comprising the following steps:

[0010] The original ultrasound video sequence was acquired, and each frame of the ultrasound video sequence was preprocessed.

[0011] The preprocessed ultrasound video sequence is input into the DVDNet video denoising network, and adjacent multi-frame images are jointly processed to generate low-noise ultrasound images.

[0012] The denoised low-noise ultrasound image is used as a coarse pseudo-true value sample and paired with the original single-frame ultrasound image to form the training dataset of the U-Net single-frame denoising network.

[0013] A U-Net single-frame denoising network is constructed. The network weights are pre-initialized based on the U-Net structure. The U-Net single-frame denoising network with speckle distribution constraints and structural consistency constraints is trained using the training dataset. The original single-frame ultrasound image is used as input. The encoder-decoder structure is used for feature extraction and reconstruction. The denoised ultrasound image is output. The network weights are updated by combining the composite loss function calculation and backpropagation. The optimal model weights on the validation set are saved.

[0014] Acquire a single-frame ultrasound image to be denoised, and perform the preprocessing described above;

[0015] The preprocessed single-frame ultrasound image is input into the trained U-Net single-frame denoising network. Feature extraction and feature reconstruction are performed through forward propagation of the encoding and decoding paths, and the denoised single-frame ultrasound image is output.

[0016] The single-frame ultrasound image is post-processed to output the final denoised single-frame ultrasound image.

[0017] Secondly, the present invention provides an ultrasound image denoising system that integrates speckle distribution constraints and structural consistency constraints, comprising:

[0018] The preprocessing module is used to preprocess the acquired raw ultrasound video sequence and preprocess each frame of the ultrasound video sequence.

[0019] The DVDNet video denoising module is used to input the preprocessed ultrasound video sequence into the DVDNet video denoising network, jointly process adjacent multiple frames of images, and generate low-noise ultrasound images.

[0020] The dataset construction module is used to pair the denoised low-noise ultrasound images as coarse pseudo-true value samples with the original single-frame ultrasound images to form the training dataset of the U-Net single-frame denoising network.

[0021] The U-Net single-frame denoising network training module is used to construct the U-Net single-frame denoising network. Based on the U-Net structure, the network weights are pre-initialized. The U-Net single-frame denoising network with speckle distribution constraints and structural consistency constraints is trained using the training dataset. The original single-frame ultrasound image is used as input. The encoder-decoder structure is used for feature extraction and reconstruction. The denoised ultrasound image is output. The network weights are updated by combining the composite loss function calculation and backpropagation. The optimal model weights on the validation set are saved.

[0022] The preprocessing module is used to acquire a single-frame ultrasound image to be denoised and perform the preprocessing.

[0023] The U-Net single-frame denoising module is used to input the preprocessed single-frame ultrasound image into the trained U-Net single-frame denoising network, and perform feature extraction and feature reconstruction through forward propagation of the encoding and decoding paths, and output the denoised single-frame ultrasound image.

[0024] The post-processing module is used to perform post-processing on the single-frame ultrasound image and output the final denoised single-frame ultrasound image.

[0025] Thirdly, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described ultrasonic image denoising method that fuses speckle distribution constraints and structural consistency constraints.

[0026] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described ultrasonic image denoising method that fuses speckle distribution constraints and structural consistency constraints.

[0027] The beneficial effects of this invention are as follows: This invention achieves a balance between noise suppression and anatomical fidelity by using DVDNet to generate pseudo-ground values ​​and U-Net for dual-constraint training. This invention maintains medical physical consistency; speckle distribution constraints ensure that the output image conforms to the statistical laws of real speckle, avoiding the generation of non-physical textures; and structural consistency constraints significantly reduce excessive smoothing of edge structures. Furthermore, this invention can be implemented using a purely software solution, requiring no additional hardware, reducing costs while simultaneously improving medical quality and efficiency. Attached Figure Description

[0028] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. These drawings constitute a part of the specification and are used together with the embodiments of the present invention to explain the present invention, but do not constitute a limitation thereof. Those skilled in the art should understand that these drawings are not necessarily drawn to scale. In the drawings:

[0029] Figure 1 is a flowchart of the overall process of the method of the present invention;

[0030] Figure 2 shows the structure of U-Net, which integrates speckle distribution constraint loss and structural consistency constraint loss;

[0031] Figure 3 is a comparison diagram of the original ultrasound image and the ultrasound image obtained by the method of the present invention; Detailed Implementation

[0032] 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.

[0033] Example 1

[0034] As shown in Figure 1, this embodiment of the invention provides an ultrasound image denoising method that integrates speckle distribution constraints and structural consistency constraints, specifically including the following steps:

[0035] S1: Construct a coarse-level pseudo-real value training set

[0036] Furthermore, in this embodiment of the invention, the open-source video denoising network DVDNet is used to process the ultrasound dynamic sequence to generate a low-noise image with small temporal fluctuations, which serves as the supervision signal for the second-stage network. S1 includes:

[0037] S11: Ultrasound sequence preprocessing

[0038] Furthermore, in this embodiment of the invention, the original ultrasound B-mode video sequence is acquired from the ultrasound scanner of the hospital, and each frame is preprocessed. First, black borders and displayed device information, patient privacy information, etc. are removed. Then, bilinear interpolation is used to uniformly scale the image to 512×512 pixels, and then the pixel values ​​are normalized to the [0,1] interval.

[0039] S12: Generate pseudo-real values ​​using a video denoising network

[0040] Furthermore, in this embodiment of the invention, the preprocessed ultrasound video sequence is input into a video denoising network. Adjacent frames in the video sequence are jointly processed to generate a low-noise reference image for the corresponding center frame. The video denoising network utilizes the temporal consistency characteristics between adjacent frames to suppress random noise components that vary significantly over time, thereby obtaining an ultrasound image with enhanced temporal stability. This low-noise image is used as a coarse pseudo-true value for supervised training of the subsequent single-frame ultrasound image denoising network, rather than as the final denoising output.

[0041] S13: Generate the data pairs needed for training

[0042] Furthermore, in this embodiment of the invention, the low-noise ultrasound image output after denoising is used as a coarse pseudo-true value sample, which, together with the corresponding original single-frame ultrasound image, constitutes a training sample pair, forming the training dataset for the second-stage deep denoising network.

[0043] S2: Training a U-Net denoising network with dual constraints

[0044] Furthermore, in this embodiment of the invention, a U-Net backbone network is constructed, and a composite loss function containing speckle distribution constraints and structural consistency constraints is designed, enabling the network to simultaneously learn the statistical regularity of speckle and the ability to protect anatomical structures during training. Specifically, S2 includes:

[0045] S21: Design of a Deep Denoising Network Based on U-Net

[0046] Furthermore, in this embodiment of the invention, a deep denoising network with U-Net as the backbone is constructed, and a typical encoder-decoder structure is adopted, including encoding path, decoding path and skip connections between corresponding layers.

[0047] The encoding path consists of multiple downsampling modules, each including convolution operations, nonlinear activation functions, and downsampling operations, used to extract multi-scale feature information from ultrasound images layer by layer. The decoding path consists of multiple upsampling modules, each including upsampling operations, convolution operations, and nonlinear activation functions, used to gradually restore the spatial resolution of the image. Skip connections between corresponding layers directly transmit shallow spatial detail features to the decoding path to enhance the preservation of tissue boundaries and structural information.

[0048] S22: Design of a composite loss function for speckle noise modeling and structural protection

[0049] Furthermore, embodiments of the present invention employ a composite loss function. Including pixel reconstruction loss Speckle distribution constraint loss and structural consistency constraint loss Its overall form is expressed as:

[0050]

[0051] (1) Pixel-level reconstruction loss

[0052] Furthermore, the pixel reconstruction loss is used to constrain the similarity between the network output and the coarse pseudo-true value at the pixel level. The mean squared error loss is preferably used, and its expression is:

[0053]

[0054] in, This represents the corresponding coarse-level pseudo-truth image. This represents the image output by the network.

[0055] (2) Speckle distribution constraint loss

[0056] Since the speckle noise in the B-mode grayscale image output by the ultrasound imaging system exhibits a multiplicative distribution characteristic, this embodiment of the invention performs logarithmic transformation processing on the ultrasound image output by the denoising network to convert the multiplicative noise into an additive form, and constrains the local intensity distribution of the image in the logarithmic domain to guide the network output result to conform to the statistical characteristics of real ultrasound speckle.

[0057] Specifically, this involves performing a logarithmic transformation on the network output image:

[0058]

[0059] in, To prevent the use of small, unstable constants, intensity statistics are performed on the denoised image after logarithmic transformation within a local window of a preset size. The statistical results are then compared with the statistical characteristics of the corresponding coarse-level pseudo-true image at the same window position. By constructing a distribution consistency constraint term, the speckle distribution characteristics of the network output image in the local region are restricted to maintain consistency with the real ultrasound image.

[0060] (3) Structural consistency constraint loss

[0061] Furthermore, in order to protect tissue boundaries and anatomical structures, this embodiment of the invention introduces structural consistency constraints into the loss function.

[0062] Specifically, by performing gradient operator processing on the denoised image output by the network and the corresponding coarse-level pseudo-ground image respectively, edge and structural information in the image is extracted, and consistency constraints are applied to the gradient response results of the two, thereby limiting the offset of tissue boundaries and anatomical structures during the denoising process. The expression is as follows:

[0063]

[0064] in, This represents the gradient operator, used to extract edge and structural information from an image. This constraint effectively prevents blurring of tissue boundaries and structural collapse during the denoising process.

[0065] S23: Network Training Process

[0066] First, pre-initialized network weights based on the U-Net structure are loaded as the initial state of the model. These initial weights are used to accelerate network convergence and improve training stability. Then, the U-Net denoising network, which integrates speckle distribution constraints and structural consistency constraints, is trained using the training samples constructed in step S13.

[0067] During training, the model takes a single frame of raw ultrasound image as input and performs feature extraction and reconstruction step by step through the encoding and decoding paths. The encoding path maps the input image to a multi-scale feature space, while the decoding path gradually restores the spatial resolution through upsampling operations and introduces shallow features of the corresponding scale by combining skip connections, thereby obtaining a denoised prediction image with the same size as the input.

[0068] Furthermore, in this embodiment of the invention, the denoised image output by the model and the corresponding coarse pseudo-true value image are used together to calculate the loss function.

[0069] In each training epoch, the model performs forward prediction on samples in the training set, calculates the loss value according to the aforementioned loss function, and updates the network weights using the backpropagation algorithm. Simultaneously, the learning rate and training parameters are dynamically adjusted based on changes in the performance of the validation set. After training, the model weights with the best performance on the validation set are saved for subsequent denoising inference on single-frame ultrasound images.

[0070] S3: Single-frame ultrasound image denoising inference training process

[0071] Furthermore, in this embodiment of the invention, after completing the network architecture and composite loss function design, the U-Net denoising network is trained using the training samples. A single frame of original ultrasound image is input into the network, the network output is calculated, and the network parameters are updated via backpropagation based on the composite loss function until the model converges or reaches the preset number of training rounds. Specifically, S3 includes:

[0072] S31: Input Image Preprocessing

[0073] Acquire a single-frame ultrasound B-mode image to be denoised, and perform the same preprocessing as in S11 on the single-frame ultrasound image to be processed: remove irrelevant regions, scale to 512×512, and normalize pixel values.

[0074] S32: Network forward propagation processing

[0075] Furthermore, in this embodiment of the invention, the preprocessed single-frame ultrasound image is input into the trained U-Net denoising network, and feature extraction and reconstruction are performed through the forward propagation process of the encoding and decoding paths. The encoding path is used to extract multi-scale contextual features, and the decoding path is used to restore spatial resolution step by step, and shallow structural features are introduced by combining skip connections, thereby maintaining the continuity of tissue boundaries and anatomical structures while suppressing speckle noise, resulting in a denoised prediction image.

[0076] S33: Post-processing and Output

[0077] Furthermore, in this embodiment of the invention, post-processing operations are performed on the denoised image output by the network. The pixel values ​​of the denoised image are reverse mapped from the normalized space to the original image intensity range. Then, the image format is converted or saved as needed, and the final denoised single-frame ultrasound image is output.

[0078] Example 2

[0079] To clearly illustrate the present invention, the following detailed description is provided in conjunction with specific embodiments.

[0080] The steps of the method of the present invention, in conjunction with specific embodiments, are as follows:

[0081] S1: Construct a coarse-level pseudo-real value training set

[0082] Furthermore, S1 specifically includes:

[0083] S11: Ultrasound sequence preprocessing

[0084] Furthermore, this embodiment uses fetal ultrasound B-mode video data collected from the hospital's radiology department as the raw input. The dataset contains 30 clinical videos, each approximately 60 seconds long, with a frame rate of 25fps and an original resolution of 720×576 pixels.

[0085] The process involves preprocessing the acquired raw video: First, the raw ultrasound video is cropped frame by frame to remove black borders, device logos, parameter displays, and patient privacy information, retaining only the effective imaging area. Then, bilinear interpolation is used to uniformly scale the cropped images to 512×512 pixels. Finally, the image pixel grayscale values ​​are linearly normalized to the [0,1] interval to eliminate intensity differences caused by different devices and scanning parameters, resulting in a standardized ultrasound video sequence.

[0086] S12: Pseudo-real value generation based on DVDNet

[0087] Furthermore, after preprocessing, the standardized ultrasound video sequence is input into the DVDNet video denoising network. DVDNet utilizes the temporal consistency between adjacent frames to jointly model the target frame, thereby effectively suppressing random noise components that vary significantly over time.

[0088] In this embodiment, five consecutive frames of ultrasound images are used as the input window of DVDNet, and the denoising result of the corresponding center frame is output. Since ultrasound speckle exhibits strong statistical stability over a short timescale, the temporal fluctuations of the image processed by DVDNet are significantly reduced, making it a near-low-noise reference image. This output result is defined as a coarse-level pseudo-ground value image and used for supervised training of the subsequent single-frame denoising network.

[0089] S13: Training data pairs are generated

[0090] Through the above processing, each ultrasound video sequence yields a pair of data: an original single-frame ultrasound image and a corresponding coarse-level pseudo-true value image. The original image... Its corresponding DVDNet denoising output Pairing them up to form training sample pairs ( , Ultimately, approximately 45,000 training pairs can be obtained, with the training set and validation set divided in a 4:1 ratio. These sample pairs will serve as supervision signals for the second-stage dedicated denoising network.

[0091] S2: Training a U-Net denoising network with dual constraints

[0092] Furthermore, S2 specifically includes:

[0093] S21: U-Net Network Architecture Design

[0094] Further, as shown in Figure 2, this embodiment uses the classic U-Net architecture as the base network. This network consists of an encoder and a decoder, with skip connections fusing shallow and deep features. The encoder contains four downsampling blocks, each consisting of two 3×3 convolutional layers and one 2×2 max-pooling layer; the decoder contains four upsampling blocks, each consisting of one 2×2 deconvolutional layer and two 3×3 convolutional layers. The network input is a single-channel grayscale image (512×512), and the output is also a single-channel grayscale image (512×512). The number of channels is set to a multiple of 32 to ensure computational efficiency. The network parameters are initialized using the Kaiming initialization method, the activation function is uniformly ReLU, and batch normalization layers are applied after each convolutional layer.

[0095] S22: Design of Composite Loss Function

[0096] Furthermore, to guide the network in learning the physical characteristics of ultrasonic speckle noise and protect tissue structure, this embodiment designs a composite loss function, which includes three terms: pixel-level reconstruction loss. Speckle distribution constraint loss and structural consistency constraint loss The overall loss function is defined as:

[0097]

[0098] in, and For weighting coefficients. In the specific implementation, the speckle distribution constraint loss is achieved by comparing the statistical features of the network output image and the pseudo-ground image within a local window, and the structural consistency constraint loss is achieved by introducing gradient consistency or edge preservation terms, thereby suppressing noise while protecting key anatomical structures.

[0099] S23: Network Training Process

[0100] Furthermore, in this embodiment, pre-trained weights based on the U-Net structure are first loaded as the initial state of the denoising model, and the model is trained using the training set constructed in S11. The training set consists of the original single-frame ultrasound image and its corresponding coarse-level pseudo-ground value image, wherein the original ultrasound image serves as the network input and the coarse-level pseudo-ground value image serves as the supervision signal.

[0101] The model employs an encoder-decoder architecture to process the input image. The encoder extracts features step-by-step through a four-level downsampling module, mapping the input single-channel ultrasound image into four sets of feature maps at different scales. The decoder restores the spatial resolution step-by-step through a four-level upsampling module and concatenates these maps with the corresponding feature maps from the encoder, thus preserving local structural details while maintaining global contextual information. Finally, the model outputs a single-channel denoised ultrasound image with the same size as the input image.

[0102] During model training, the output denoised image and the corresponding coarse pseudo-true image are used for loss function calculation in S22.

[0103] The network was trained using the PyTorch framework with the Adam optimizer, an initial learning rate of 10⁻⁴, a batch size of 8, and 200 training epochs. At the end of each epoch, the model performance was evaluated using a validation set, with metrics including PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index). Training was stopped early if the PSNR on the validation set did not improve for five consecutive epochs. The model weights with the highest PSNR on the validation set were then retained as the final training result.

[0104] S3: Single-frame ultrasound image denoising inference process

[0105] Furthermore, S3 specifically includes:

[0106] S31: Input Image Preprocessing

[0107] Furthermore, this embodiment acquires a single-frame ultrasound B-mode image to be processed from a clinical ultrasound scanning device. First, the original image is cropped to remove black borders, device parameter display areas, and non-imaging areas, retaining only the effective imaging area. Then, bilinear interpolation is used to uniformly scale the cropped image to 512×512 pixels. After size normalization, the image grayscale values ​​are linearly normalized, mapping pixel intensity to the [0,1] interval to ensure that the input data distribution remains consistent with the network training phase. After the above processing, a standardized single-frame ultrasound image that meets the model input requirements is obtained.

[0108] S32: Network forward propagation processing

[0109] Furthermore, in this embodiment, the preprocessed single-frame ultrasound image is fed as input into the trained U-Net denoising network. The model input is a multi-channel image with a size of 512×512, and the output is a single-channel denoised prediction image of uniform size.

[0110] During forward inference, the input image sequentially passes through a 4-level encoding module and a 4-level decoding module. The encoding module extracts local texture features and global structural features step by step through convolution and downsampling operations. The decoding module restores spatial resolution step by step through upsampling operations and concatenates the feature maps with those from the corresponding encoding layers to enhance the recovery of structural information. Finally, a denoised ultrasound image is output through a 1×1 convolutional layer.

[0111] S33: Post-processing and Output

[0112] Furthermore, as shown in Figure 3, this embodiment performs necessary numerical truncation and grayscale inverse normalization on the network output results to meet the requirements for ultrasound image display, ultimately outputting a single-frame denoised ultrasound image. This denoising result effectively suppresses speckle noise while maintaining the integrity of fetal tissue boundaries and anatomical structures, and can be directly used for subsequent identification, measurement, or clinical diagnostic analysis.

[0113] Example 3

[0114] In one exemplary embodiment, an ultrasound image denoising system that integrates speckle distribution constraints and structural consistency constraints is provided, comprising:

[0115] The preprocessing module is used to preprocess the acquired raw ultrasound video sequence and preprocess each frame of the ultrasound video sequence.

[0116] The DVDNet video denoising module is used to input the preprocessed ultrasound video sequence into the DVDNet video denoising network, jointly process adjacent multiple frames of images, and generate low-noise ultrasound images.

[0117] The dataset construction module is used to pair the denoised low-noise ultrasound images as coarse pseudo-true value samples with the original single-frame ultrasound images to form the training dataset of the U-Net single-frame denoising network.

[0118] The U-Net single-frame denoising network training module is used to construct the U-Net single-frame denoising network. Based on the U-Net structure, the network weights are pre-initialized. The U-Net single-frame denoising network with speckle distribution constraints and structural consistency constraints is trained using the training dataset. The original single-frame ultrasound image is used as input. The encoder-decoder structure is used for feature extraction and reconstruction. The denoised ultrasound image is output. The network weights are updated by combining the composite loss function calculation and backpropagation. The optimal model weights on the validation set are saved.

[0119] The preprocessing module is used to acquire a single-frame ultrasound image to be denoised and perform the preprocessing.

[0120] The U-Net single-frame denoising module is used to input the preprocessed single-frame ultrasound image into the trained U-Net single-frame denoising network, and perform feature extraction and feature reconstruction through forward propagation of the encoding and decoding paths, and output the denoised single-frame ultrasound image.

[0121] The post-processing module is used to perform post-processing on the single-frame ultrasound image and output the final denoised single-frame ultrasound image.

[0122] Example 4

[0123] In one exemplary embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the ultrasound image denoising method described above that fuses speckle distribution constraints and structural consistency constraints.

[0124] Example 5

[0125] In one exemplary embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the ultrasound image denoising method that fuses speckle distribution constraints and structural consistency constraints as described above.

[0126] Finally, it should be noted that the above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for denoising ultrasound images that integrates speckle distribution constraints and structural consistency constraints, characterized in that, Includes the following steps: The original ultrasound video sequence is acquired, and each frame of the ultrasound video sequence is preprocessed. The preprocessed ultrasound video sequence is input into the DVDNet video denoising network to jointly process adjacent multi-frame images to generate low-noise ultrasound images. The denoised low-noise ultrasound images are used as coarse pseudo-ground value samples and paired with the original single-frame ultrasound images to form the training dataset for the U-Net single-frame denoising network. The U-Net single-frame denoising network is constructed, and the network weights are pre-initialized based on the U-Net structure. The training dataset is used to refine the U-Net single-frame denoising network that incorporates speckle distribution constraints and structural consistency constraints. Training is performed using the original single-frame ultrasound image as input. An encoder-decoder structure is used for feature extraction and reconstruction, and a denoised ultrasound image is output. The network weights are updated by combining a composite loss function and backpropagation, and the optimal model weights on the validation set are saved. A single-frame ultrasound image to be denoised is obtained and the preprocessing is performed. The preprocessed single-frame ultrasound image is input into the trained U-Net single-frame denoising network, and feature extraction and reconstruction are performed through forward propagation of the encoding and decoding paths, outputting a denoised single-frame ultrasound image. The single-frame ultrasound image is post-processed to output the final denoised single-frame ultrasound image.

2. The method according to claim 1, characterized in that, The preprocessing includes: first, removing black borders, displayed device information, and patient privacy information; second, using bilinear interpolation to uniformly scale to 512×512 pixels; and finally, normalizing the pixel values ​​to the [0,1] range.

3. The method according to claim 1 or 2, characterized in that, The DVDNet video denoising network utilizes the temporal consistency feature between adjacent frames to suppress random noise and enhance the temporal stability of ultrasound images.

4. The method according to claim 1, characterized in that, The encoder-decoder structure includes an encoding path, a decoding path, and skip connections. The encoding path consists of multi-level downsampling modules, each level including convolution, activation function, and downsampling operation, used to extract multi-scale feature information of ultrasound images. The decoding path consists of multi-level upsampling modules, each level including upsampling, convolution, and activation function, used to restore the spatial resolution of the image. The skip connections directly transmit shallow spatial detail features to the decoding path, enhancing the ability to preserve tissue boundaries and structural information.

5. The method according to claim 1 or 4, characterized in that, The composite loss function includes pixel reconstruction loss, speckle distribution constraint loss, and structural consistency constraint loss. The pixel reconstruction loss is used to constrain the similarity between the network output and the coarse-level pseudo-true value at the pixel level, and the mean square error loss is preferred. The speckle distribution constraint loss ensures that the local speckle distribution characteristics of the denoising network output image are consistent with the real ultrasound image. The structural consistency constraint loss extracts the edge structure information of the denoised image and the coarse-level pseudo-true value image through the gradient operator and constrains the consistency of their gradient responses, limiting the offset of tissue boundaries and anatomical structures during the denoising process.

6. The method according to claim 1, characterized in that, Logarithmic transformation is performed on the noisy ultrasound image output by the U-Net single-frame denoising network to convert multiplicative noise into additive form, and the local intensity distribution is constrained in the logarithmic domain so that the output conforms to the statistical characteristics of real ultrasound speckle.

7. The method according to claim 1 or 6, characterized in that, The post-processing reverses the pixel values ​​from the normalized space to the original image intensity range, and then converts or saves the image format.

8. An ultrasound image denoising system that integrates speckle distribution constraints and structural consistency constraints, characterized in that, include: The preprocessing module is used to preprocess the acquired raw ultrasound video sequence and preprocess each frame of the ultrasound video sequence. The DVDNet video denoising module is used to input the preprocessed ultrasound video sequence into the DVDNet video denoising network, jointly process adjacent multiple frames of images, and generate low-noise ultrasound images. The dataset construction module is used to pair the denoised low-noise ultrasound images as coarse pseudo-true value samples with the original single-frame ultrasound images to form the training dataset of the U-Net single-frame denoising network. The U-Net single-frame denoising network training module is used to construct the U-Net single-frame denoising network, pre-initialize the network weights based on the U-Net structure, and train the U-Net single-frame denoising network with speckle distribution constraints and structural consistency constraints using the training dataset. The original single-frame ultrasound image is used as input, and an encoder-decoder structure is used for feature extraction and reconstruction. The denoised ultrasound image is output, and the network weights are updated by combining the composite loss function calculation and backpropagation. The optimal model weights on the validation set are saved. The preprocessing module is used to acquire a single-frame ultrasound image to be denoised and perform the preprocessing; the U-Net single-frame denoising module is used to input the preprocessed single-frame ultrasound image into the trained U-Net single-frame denoising network, extract and reconstruct features through forward propagation of the encoding and decoding paths, and output the denoised single-frame ultrasound image. The post-processing module is used to perform post-processing on the single-frame ultrasound image and output the final denoised single-frame ultrasound image.

9. An electronic device, characterized in that: The method includes a memory and a processor, wherein the memory stores a computer program executable by the processor, and the processor executes the computer program to implement the ultrasound image denoising method according to any one of claims 1-7, which integrates speckle distribution constraints and structural consistency constraints.

10. A storage medium, characterized in that: The storage medium stores a computer program, which, when read and executed, implements the ultrasonic image denoising method that integrates speckle distribution constraints and structural consistency constraints as described in any one of claims 1-7.

Images