Medical ultrasonic image segmentation method and system based on wave fusion UNet

Abstract

The invention provides a medical ultrasonic image segmentation method and system based on wave fusion UNet. The method comprises the following steps: acquiring an ultrasonic image sample, and preprocessing the ultrasonic image sample; a segmentation model is constructed based on a U-Net model, the segmentation model comprises an encoder and a decoder, and the preprocessed samples are adopted to train the segmentation model: the encoder extracts multi-scale features, high-resolution spatial information is directly transmitted to the decoder by adopting jump connection between the encoder and the decoder, and the high-resolution spatial information is extracted by the decoder. Frequency domain features and spatial features of high-resolution features are fused in the shallow layer of the decoder, and high-resolution spatial information is screened and enhanced through an attention gating mechanism; then the decoder reconstructs a high-resolution segmentation mask; and segmenting the to-be-segmented image by using the trained segmentation model. According to the method, the problems of noise, low contrast, difficulty in reserving edge details and the like in ultrasonic image segmentation are effectively solved, and the segmentation accuracy is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and particularly relates to a medical ultrasound image segmentation method and system based on wave fusion UNet. Background Art

[0002] Ultrasonic imaging (US) is an important medical diagnostic tool, which has advantages such as portability, real-time imaging, and cost-effectiveness. However, due to its inherent noise, low contrast, and irregular anatomical texture, the accurate segmentation of ultrasonic images remains a major challenge. These problems often lead to blurred edges and difficult-to-distinguish details, restricting the effectiveness of traditional segmentation methods (including deep learning-based models such as U-Net).

[0003] Currently, although deep learning techniques have made significant progress in the field of medical image segmentation, the particularity of ultrasonic images makes it difficult for existing methods to achieve ideal segmentation results. For example, the traditional U-Net architecture performs poorly in processing noisy and low-contrast images, especially in terms of the segmentation accuracy of edge regions. In addition, although existing methods based on attention mechanisms and frequency-domain feature analysis have improved the segmentation performance to a certain extent, there are still problems such as insufficient feature fusion, high computational cost, and inaccurate segmentation of small targets when dealing with ultrasonic images. Summary of the Invention

[0004] In order to overcome the defects existing in the above-mentioned prior art, the purpose of the present invention is to provide a medical ultrasound image segmentation method and system based on wave fusion UNet for improving the accuracy of ultrasonic image segmentation.

[0005] To achieve the above object of the present invention, the present invention provides a medical ultrasound image segmentation method based on wave fusion UNet, including the following steps:

[0006] Obtain an ultrasonic image sample and perform preprocessing on it;

[0007] Build a segmentation model based on the U-Net model. The segmentation model includes an encoder and a decoder. Use the preprocessed sample to train the segmentation model: the encoder extracts multi-scale features, and a skip connection is used between the encoder and the decoder to directly transfer high-resolution features to the decoder. In the shallow layer of the decoder, fuse the frequency-domain features and spatial features of the high-resolution features, and also screen and enhance the high-resolution features through an attention gating mechanism; then the decoder reconstructs a high-resolution segmentation mask;

[0008] Use the trained segmentation model to segment the image to be segmented.

[0009] The present invention effectively solves the problems existing in ultrasonic image segmentation, such as noise, low contrast, and difficulty in retaining edge details, and significantly improves the accuracy of segmentation; the segmentation model of the present invention has excellent segmentation performance on multiple ultrasonic image datasets and has good generalization ability.

[0010] In an alternative embodiment of the medical ultrasonic image segmentation method based on wave fusion UNet, the WaveFusion module fuses the frequency domain features and spatial features of high-resolution features in the shallow layer of the decoder:

[0011] The WaveFusion module first performs dimensionality reduction on the input high-resolution features, and decomposes the features after dimensionality reduction into high-frequency components and low-frequency components by wavelet transform; enhances the high-frequency components and low-frequency components at different resolutions and frequencies, and finally combines the results of the enhancement process.

[0012] This alternative embodiment effectively solves the problem of maintaining edge information in noisy images, improves the sensitivity of the segmentation model to edge details, and thus better retains the contour information of the target region.

[0013] In an alternative embodiment of the medical ultrasonic image segmentation method based on wave fusion UNet, the WaveFusion module includes a first convolutional module for processing the high-resolution features output by the encoder, and a first wavelet transform module for decomposing the high-resolution features into high-frequency components and low-frequency components; a second convolutional module for processing the low-resolution features output by the decoder, and a second wavelet transform module for decomposing the low-resolution features into high-frequency components and low-frequency components;

[0014] The WaveFusion module further includes a first convolutional kernel generator, a second convolutional kernel generator, a first carafe operator, and a second carafe operator; the first convolutional kernel generator generates a convolutional kernel for the first carafe operator based on the low-frequency component output by the first wavelet transform module and the low-frequency component output by the second wavelet transform module, and the first carafe operator enhances the low-frequency component at different resolutions based on this convolutional kernel; the second convolutional kernel generator generates a convolutional kernel for the second carafe operator based on the high-frequency component output by the first wavelet transform module and the high-frequency component output by the second wavelet transform module, and the second carafe operator enhances the high-frequency component at different resolutions based on this convolutional kernel;

[0015] Combine the enhanced high-frequency components and low-frequency components.

[0016] In this alternative solution, the utilization of features is maximized at different resolutions, while spectral bias is alleviated. Compared with traditional methods, CARAFE can aggregate context information from a wider receptive field, achieve more comprehensive feature fusion, maximize the utilization of features at different resolutions, and alleviate spectral bias at the same time.

[0017] Optionally, both the first wavelet transform module and the second wavelet transform module use multi-scale two-dimensional discrete wavelet transform with a scale of 3. The multi-scale property of DWT in this alternative solution facilitates the examination of features at multiple scales, capable of capturing both rough details and fine details.

[0018] In an alternative solution of this wave fusion-based UNet medical ultrasound image segmentation method, the SGAG module is used to screen and enhance the high-resolution features in the shallow layer of the decoder:

[0019] The SGAG module generates a gating signal using the high-level feature g from the encoder input, and controls the information flow X from the input of the previous layer of the decoder through this gating signal to dynamically screen and enhance the high-resolution features.

[0020] In an alternative solution of this wave fusion-based UNet medical ultrasound image segmentation method, the SGAG module includes two Shuffle convolutions. The two Shuffle convolutions perform convolution operations on the high-level feature g and the information flow x respectively, then perform batch normalization processing respectively, add the results of the normalization processing, and use ReLU activation to obtain the joint attention coefficient of the high-level feature g and the information flow x;

[0021] The SGAG module further includes a one-dimensional convolutional layer. The one-dimensional convolutional layer performs one-dimensional convolution on the feature after ReLU activation, then uses Sigmoid activation, and applies the attention coefficient to the information flow x for feature enhancement.

[0022] Optionally, the Shuffle convolution is a 3×3 Shuffle Group Convolution. This alternative solution uses 3×3 Shuffle Group Convolution instead of the standard 2D convolution operation, which not only reduces the computational cost but also enables the model to capture a larger spatial context.

[0023] In an alternative solution of this wave fusion-based UNet medical ultrasound image segmentation method, a hybrid loss function combining Dice loss and cross-entropy loss is used to supervise the training of the segmentation model.

[0024] This alternative uses a hybrid loss function that combines Dice loss and cross-entropy loss, which can better guide edge segmentation. By using the deep supervision technique to generate outputs at each decoder layer, a supervision signal is provided at each level, enabling the network to understand the data more accurately and achieve more precise segmentation.

[0025] In an alternative of this medical ultrasound image segmentation method based on wave fusion UNet, the hybrid loss function is:

[0026] L = αL CE + βL DSC

[0027] where L CE is the cross-entropy loss, specifically:

[0028]

[0029] The Dice loss expression is:

[0030]

[0031] y is the predicted segmentation mask, m is the ground truth segmentation mask, α and β are weight coefficients, and n refers to the total number of pixels in the image.

[0032] The present invention also proposes a medical ultrasound image segmentation system, including a data acquisition module, a processing module, and a storage module for acquiring ultrasound image samples and images to be segmented;

[0033] The data acquisition module is communicatively connected to the processing module and sends the ultrasound image samples and the images to be segmented to it. The processing module is communicatively connected to the storage module. The storage module is used to store at least one executable instruction, and the executable instruction causes the processing module to perform the operations corresponding to the above-mentioned medical ultrasound image segmentation method based on wave fusion UNet on the images to be segmented.

[0034] The beneficial effects of the present invention are:

[0035] The present invention first introduces wavelet transform into the U-Net architecture. Through wavelet decomposition and enhancement operations, it effectively integrates high-frequency and multi-scale edge details and spatial features, solves the problem of difficult retention of edge information in ultrasound image segmentation, and significantly improves the sensitivity of the segmentation model to edge details and the segmentation accuracy.

[0036] In the present invention, the SGAG module is proposed. Based on the gating mechanism of Shuffle convolution, combining the attention mechanism and Shuffle convolution, it not only enhances the accuracy of feature selection, but also reduces the computational cost and improves the efficiency of the model. In addition, through the calculation of channel attention and spatial attention and the channel shuffle operation, the SGAG module further enhances the model's ability to express channel features, effectively suppresses irrelevant features, and improves the segmentation accuracy.

[0037] The segmentation model in the present invention integrates the WaveFusion module and the SGAG module in the U-Net architecture, making the present invention significantly superior to the prior art in the segmentation performance of multiple ultrasound image datasets. Especially when dealing with ultrasound images with noise, low contrast, and complex anatomical structures, it performs outstandingly. At the same time, the model also has good generalization ability and can adapt to different medical image segmentation tasks.

[0038] Additional aspects and advantages of the present invention will be given in part in the following description, will become apparent in part from the following description, or will be learned through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The above and / or additional aspects and advantages of the present invention will become apparent and easy to understand from the description of the embodiments in conjunction with the following drawings, where:

[0040] Figure 1 is the structural schematic diagram of the present invention;

[0041] Figure 2 is the structural schematic diagram of the WaveFusion module;

[0042] Figure 3 is the structural schematic diagram of the SGAG module;

[0043] Figure 4 is the structural schematic diagram of Shuffle Group Convolution.

[0044] Wherein, Figure 2 X in l represents the low-resolution feature of the l-th layer of the decoder, while Y l+1 represents the high-resolution feature of the (l + 1)-th layer of the encoder. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] Embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the drawings, where the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the drawings are exemplary and are only used to explain the present invention and should not be construed as a limitation of the present invention.

[0046] In the description of the present invention, unless otherwise specified and defined, it should be noted that the terms "installation", "connection", and "linkage" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the communication inside two components. It can be directly connected or indirectly connected through an intermediate medium. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0047] Embodiment 1

[0048] The present invention provides an embodiment of a medical ultrasound image segmentation method based on wave fusion UNet. The steps of this embodiment are as follows:

[0049] Obtain ultrasound image samples and preprocess them.

[0050] In this embodiment, the ultrasound image samples used are from multiple ultrasound image datasets (such as BUSI, TN3K) and skin lesion datasets (such as ISIC17).

[0051] The preprocessing of the ultrasound image samples includes operations such as normalization, cropping, and data augmentation to improve the robustness and generalization ability of the model.

[0052] Based on the U-Net model, a segmentation model is constructed. In this embodiment, this segmentation model is called WF-UNet, and its backbone network adopts an encoder-decoder structure similar to U-Net. The encoder gradually reduces the spatial resolution of the feature map through convolution and downsampling (such as max pooling), while increasing the number of channels, and extracts multi-scale features of the image through residual blocks. A skip connection is used between the encoder and the decoder to directly transfer high-resolution features to the decoder, that is, the part represented by the dotted line in the figure. The decoder uses upsampling (such as transposed convolution or interpolation) and convolution operations to reconstruct the high-resolution segmentation mask, gradually restoring the spatial resolution of the feature map, and combining high-resolution features to retain detailed information. Here, the high-resolution features refer to the features extracted by the encoder in the shallow layer (that is, the layer with less downsampling). These features have a higher spatial resolution and retain more detailed information (such as edges, textures, etc.). It should be noted that, as Figure 1 shown, in this embodiment, a WaveFusion module is introduced in the shallow layer of the decoder to fuse the frequency domain features and spatial features of the high-resolution features transmitted by the encoder, enhancing edge retention; an SGAG module is also integrated in the shallow layer of the decoder. Through the attention gating mechanism of the SGAG module, the high-resolution features are screened and enhanced, optimizing feature selection, reducing false alarms, and improving the segmentation accuracy.

[0053] Meanwhile, in this embodiment, the bottleneck layer located between the encoder and the decoder is set as the zeroth layer. The bottleneck layer is the deepest feature extraction layer, responsible for capturing global semantic information; the input and output layers are the highest layers, that is, the number of network layers extends from the bottleneck layer to both sides of the encoder and the decoder.

[0054] In this embodiment, the WaveFusion module uses wavelet transform to capture and integrate high-frequency and multi-scale edge details and spatial features. The high-frequency and multi-scale edge details refer to the high-resolution features output by the encoder, and the spatial features refer to the features extracted by the convolutional neural network, which are opposite to the frequency domain features obtained by wavelet transform. The WaveFusion module first performs dimensionality reduction on the high-resolution features. Subsequently, wavelet transform is used to decompose the high-resolution features into high-frequency components and low-frequency components. Then, the high-frequency components and low-frequency components are enhanced at different resolutions and frequencies, and finally the results of the enhancement processing are merged to improve the segmentation performance. This is specifically implemented through the following structure of the WaveFusion module.

[0055] The WaveFusion module includes a first convolutional module, a first wavelet transform module, a second convolutional module, a second wavelet transform module, a first convolutional kernel generator, a second convolutional kernel generator, a first CARAFE operator, a second CARAFE operator, and a concat module; the first convolutional module performs dimensionality reduction on the high-resolution features output by the encoder, and the first wavelet transform module decomposes the high-resolution features into the frequency domain to obtain the high-frequency and low-frequency components of the high-resolution features; the second convolutional module performs dimensionality reduction on the low-resolution features output by the decoder, and the second wavelet transform module decomposes the low-resolution features into the frequency domain to obtain the high-frequency and low-frequency components of the low-resolution features. The first convolutional kernel generator generates a convolutional kernel for the first CARAFE operator based on the low-frequency component output by the first wavelet transform module and the low-frequency component output by the second wavelet transform module. This convolutional kernel uses the first CARAFE operator to refine the dimensionality-reduced low-resolution features, enhance the low-frequency information. During this operation, if the size of the low-frequency component of the low-resolution features is inconsistent with the size of the low-frequency component of the high-resolution features, interpolation is performed. "The enhanced low-frequency features are combined with the linearly interpolated low-frequency features to form fused low-pass features, and these features are further processed by the first CARAFE operator to enhance the high-pass features of the low resolution." The first CARAFE operator performs upsampling dynamically according to different inputs to achieve more accurate feature reconstruction. The second convolutional kernel generator generates a convolutional kernel for the second CARAFE operator based on the high-frequency component output by the first wavelet transform module and the high-frequency component output by the second wavelet transform module. This convolutional kernel uses the second CARAFE operator to refine the dimensionality-reduced high-resolution features, enhance the high-frequency information. During this operation, the high-frequency component of the low-resolution features is interpolated to make its size consistent with the high-frequency component of the high-resolution features. "The enhanced high-frequency features are combined with the linearly interpolated high-frequency features to form fused high-pass features, and these features are further processed by the CARAFE operator to enhance the low-pass features of the low high resolution." The low-frequency and high-frequency features obtained by the first CARAFE operator and the second CARAFE operator are merged in the concat module.

[0056] In this embodiment, both the first wavelet transform module and the second wavelet transform module use multi-scale two-dimensional discrete wavelet transform (2D-DWT) with a scale of 3. The mean values of the high-frequency components in different directions are used to represent the high-frequency components, that is, the average of the high-frequency components in three directions is calculated, and the frequency information at different scales is enhanced separately, and finally the sum is obtained to get the finally enhanced features. The multi-scale property of DWT facilitates examining features at multiple scales, being able to capture both rough details and fine details.

[0057] In this embodiment, the SGAG module is used to screen and enhance high-resolution features in the shallow layer of the decoder. Specifically, the SGAG module uses the high-level feature g from the encoder input to generate a gating signal, and controls the information flow X from the input of the previous layer of the decoder ( Figure 1 the part of the ResBlock pointing to SGAG, the lower-resolution layer), to dynamically screen and enhance high-resolution features. This is specifically implemented through the following structure of the SGAG module.

[0058] As Figure 3 shown, the SGAG module includes two Shuffle convolutions. First, feature extraction: the two Shuffle convolutions perform convolution operations on the high-level feature g and the information flow x respectively; Channel attention: by calculating the joint attention coefficient of the feature g and the information flow x to determine important features. In this embodiment, after the convolution operations on the high-level feature g and the information flow x, batch normalization is performed respectively, the results of the normalization are added, and ReLU activation is used to achieve this; Finally, the attention coefficient is applied to the information flow x to enhance important features and suppress unimportant features for feature enhancement. In this embodiment, this is achieved through the following structure and operations: The SGAG module also includes a one-dimensional convolutional layer, which performs one-dimensional convolution on the ReLU-activated feature, and then uses Sigmoid activation.

[0059] The output Y of the SGAG module can be expressed as:

[0060] Y = x · σ(BN(C(qatt(g, x))))

[0061] where qatt(g, x) = ReLU(BN(SCx(x)) + BN(SCg(g))), σ is the Sigmoid activation function, BN is batch normalization, C is a 1×1 convolution operation, and SCx and SCg are Shuffle Group convolution operations on x and g respectively.

[0062] As Figure 4 shown, in this embodiment, the Shuffle convolution selected is a 3×3 Shuffle Group Convolution (SGC), which not only reduces the computational cost, but also enables the model to capture a larger spatial context.

[0063] In addition, through the calculation of channel attention and spatial attention, and subsequent channel shuffling operations, the SGAG module enhances the model's ability to express channel features and further improves the accuracy of feature selection.

[0064] The pre - processed samples are used to train the segmentation model, and a hybrid loss function combining Dice loss and cross - entropy loss is used to supervise the training of the segmentation model.

[0065] The hybrid loss function is:

[0066] L = αL CE + βL DSC

[0067] Among them, L CE is the cross - entropy loss, specifically:

[0068]

[0069] The Dice loss expression is:

[0070]

[0071] y is the predicted segmentation mask, m is the ground - truth segmentation mask, α and β are weight coefficients, and n refers to the total number of pixels in the image.

[0072] After training, the model is evaluated. In this embodiment, metrics such as the Dice similarity coefficient (Dice), intersection - over - union (IoU), and 95th percentile Hausdorff distance (HD95) are used to evaluate the segmentation performance of the model to verify the effectiveness and superiority of the model.

[0073] The experimental results show that WF - UNet is significantly superior to existing state - of - the - art methods in terms of metrics such as Dice, IoU, and HD95. For example, as shown in Table 1, on the BUSI dataset, the Dice score of WF - UNet reaches 72.43%, the IoU is 64.59%, and the HD95 is 91.48; on the TN3K dataset, the Dice score is 79.66%, the IoU is 69.90%, and the HD95 is 90.80; on the ISIC17 dataset, the Dice score is 94.06%, the IoU is 89.24%, and the HD95 is 11.37. These results fully demonstrate that the WF - UNet model has significant advantages in processing ultrasound image segmentation tasks with noise, low contrast, and complex anatomical structures, and also shows its good generalization ability.

[0074] Table 1

[0075]

[0076] In application, the trained segmentation model WF - UNet is applied to actual ultrasound image segmentation tasks to provide accurate segmentation results for medical diagnosis and assist doctors in disease diagnosis and treatment planning.

[0077] Example Two

[0078] This embodiment provides a medical ultrasound image segmentation system, including a data acquisition module, a processing module, and a storage module for acquiring ultrasound image samples and images to be segmented; the data acquisition module is communicatively connected to the processing module and sends the ultrasound image samples and the images to be segmented to it; the processing module is communicatively connected to the storage module, and the storage module is used to store at least one executable instruction, and the executable instruction enables the processing module to perform operations corresponding to the medical ultrasound image segmentation method based on wave fusion UNet as described in Embodiment 1 on the images to be segmented.

[0079] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0080] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A medical ultrasound image segmentation method based on wave fusion UNet, characterized in that It includes the following steps: Obtain an ultrasonic image sample and preprocess it; Build a segmentation model based on the U-Net model. The segmentation model includes an encoder and a decoder. Use the preprocessed sample to train the segmentation model: the encoder extracts multi-scale features, and a skip connection is used between the encoder and the decoder to directly transfer high-resolution features to the decoder. In the shallow layer of the decoder, the frequency-domain features and spatial features of the high-resolution features are fused, and the high-resolution features are also screened and enhanced through an attention gating mechanism; then the decoder reconstructs a high-resolution segmentation mask; Use the trained segmentation model to segment the image to be segmented.

2. The medical ultrasound image segmentation method based on wave fusion UNet according to claim 1, wherein, Use the WaveFusion module to fuse the frequency-domain features and spatial features of the high-resolution features in the shallow layer of the decoder: The WaveFusion module first performs dimensionality reduction on the input high-resolution features, and uses wavelet transform to decompose the dimensionality-reduced features into high-frequency components and low-frequency components; the high-frequency components and low-frequency components are enhanced at different resolutions and frequencies, and finally the results of the enhancement processing are merged.

3. The medical ultrasound image segmentation method based on wave fusion UNet according to claim 2, wherein The WaveFusion module includes a first convolutional module for processing the high-resolution features output by the encoder, and a first wavelet transform module for decomposing the high-resolution features into high-frequency components and low-frequency components; a second convolutional module for processing the low-resolution features output by the decoder, and a second wavelet transform module for decomposing the low-resolution features into high-frequency components and low-frequency components; The WaveFusion module also includes a first convolution kernel generator, a second convolution kernel generator, a first carafe operator, and a second carafe operator; the first convolution kernel generator generates a convolution kernel for the first carafe operator based on the low-frequency component output by the first wavelet transform module and the low-frequency component output by the second wavelet transform module, and the first carafe operator enhances the low-frequency component at different resolutions based on this convolution kernel; The second convolution kernel generator generates a convolution kernel for the second carafe operator based on the high-frequency component output by the first wavelet transform module and the high-frequency component output by the second wavelet transform module, and the second carafe operator enhances the high-frequency component at different resolutions based on this convolution kernel; Merge the enhanced high-frequency components and low-frequency components.

4. The method for segmenting medical ultrasound images based on wave fusion UNet according to claim 3, wherein, Both the first wavelet transform module and the second wavelet transform module use multi-scale two-dimensional discrete wavelet transform with a scale of 3.

5. The medical ultrasound image segmentation method based on wave fusion UNet according to claim 1, characterized in that Use the SGAG module to screen and enhance the high-resolution features in the shallow layer of the decoder: The SGAG module uses the high-level features g from the encoder input to generate a gating signal, and controls the information flow X from the input of the previous layer of the decoder through this gating signal to dynamically screen and enhance the high-resolution features.

6. The medical ultrasonic image segmentation method based on wave fusion UNet according to claim 5, characterized in that, The SGAG module includes two Shuffle convolutions. The two Shuffle convolutions respectively perform convolution operations on the high-level feature g and the information flow x, then perform batch normalization processing respectively, add the results of the normalization processing, and use ReLU activation to obtain the joint attention coefficient of the high-level feature g and the information flow x; The SGAG module further includes a one-dimensional convolutional layer. The one-dimensional convolutional layer performs one-dimensional convolution on the feature after ReLU activation, then uses Sigmoid activation, applies the attention coefficient to the information flow x, and performs feature enhancement.

7. The medical ultrasound image segmentation method based on wave fusion UNet according to claim 6, wherein The Shuffle convolution is a 3×3 Shuffle Group Convolution.

8. The medical ultrasound image segmentation method based on wave fusion UNet according to claim 1, wherein, A hybrid loss function combining Dice loss and cross-entropy loss is used to supervise the training of the segmentation model.

9. The method for segmenting medical ultrasound images based on wave fusion UNet according to claim 8, wherein, The hybrid loss function is: L = αL CE + βL DSC Among them, L CE is the cross-entropy loss, specifically: The expression of Dice loss is: y is the predicted segmentation mask, m is the true segmentation mask, α and β are weight coefficients, and n refers to the total number of pixels in the image.

10. A medical ultrasound image segmentation system, characterized in that, It includes a data acquisition module, a processing module, and a storage module for acquiring ultrasonic image samples and images to be segmented; The data acquisition module is communicatively connected to the processing module and sends the ultrasonic image samples and the images to be segmented to it; the processing module is communicatively connected to the storage module, and the storage module is used to store at least one executable instruction, and the executable instruction enables the processing module to perform the operations corresponding to the medical ultrasonic image segmentation method based on wave fusion UNet as described in any one of claims 1-9 based on the ultrasonic image samples, and segment the images to be segmented.

Images