Method for segmenting cardiac ultrasound image through multi-scale convolutional network based on boundary perception

Abstract

A method for segmenting a cardiac ultrasound image by a multi-scale convolutional network based on boundary perception relates to the technical field of image segmentation, and integrates a multi-level edge feature enhancement module, a feature fusion module and a multi-scale feature channel enhancement module in the multi-scale convolutional network based on boundary perception. And the accuracy and efficiency of cardiac ultrasound image segmentation are obviously improved. And the feature fusion module deeply mines and integrates information from different network layers, so that the recognition capability of the model on a complex heart structure is improved. The multi-scale feature channel enhancement module provides a wider receptive field and richer scale information for the network, the scale variation problem existing in the heart image is effectively solved, and then heart MRI image segmentation with higher robustness and high adaptability is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of image segmentation, and in particular to a method for segmenting cardiac ultrasound images based on a boundary-aware multi-scale convolutional network. Background Art

[0002] Cardiovascular disease has long been the leading cause of death among all kinds of diseases. Therefore, it is of great significance to use computer-aided diagnosis systems to determine whether the heart has lesions. In the field of medical image processing, accurate segmentation of cardiac images is crucial for doctors to conduct quantitative analysis of cardiac structure and function, and plays an indispensable role in the diagnosis and treatment of cardiovascular diseases. However, the complex nature of cardiac images and the accompanying noise often make it difficult for existing image segmentation technologies to accurately divide the contours of the heart.

[0003] Common methods for cardiac ultrasound image segmentation can be divided into two categories: traditional segmentation techniques and deep learning-based methods. Traditional techniques include threshold processing, edge and region-based processing, and model-based segmentation. These methods require extensive expertise and strict evaluation criteria to ensure accuracy, so they have certain dependencies and limitations in application. Although deep learning-based segmentation methods have significant advantages, they also face challenges such as high data demand, blurred edge features, noise and artifacts in the segmentation process, and limited generalization capabilities. Therefore, it is necessary to consider how to improve segmentation accuracy on limited cardiac ultrasound image datasets. Summary of the invention

[0004] In order to overcome the deficiencies of the above technologies, the present invention provides a method for cardiac ultrasound image segmentation based on a boundary-aware multi-scale convolutional network, which improves the accuracy and efficiency of cardiac ultrasound image segmentation.

[0005] The technical solution adopted by the present invention to overcome the technical problems is:

[0006] A method for segmenting cardiac ultrasound images based on a boundary-aware multi-scale convolutional network, comprising:

[0007] a) Obtain X cardiac ultrasound images to obtain a cardiac ultrasound image dataset N, where N = {N1, N2, ..., N i ,...,N X},N i is the i-th cardiac ultrasound image, i∈{1,...,X};

[0008] b) Take the i-th cardiac ultrasound image N i Preprocessing is performed to obtain the preprocessed cardiac ultrasound image S i , all X preprocessed cardiac ultrasound images constitute a preprocessed cardiac ultrasound image dataset S, S = {S1, S2, ..., Si ,...,S X};

[0009] c) dividing the preprocessed cardiac ultrasound image dataset S into a training set, a validation set, and a test set;

[0010] d) establishing a segmentation network model consisting of an encoder, a skip connection unit, and a decoder;

[0011] e) The i-th preprocessed cardiac ultrasound image S in the training set i Input into the encoder of the segmentation network model and output the feature maps respectively Feature Map Feature Map Feature Map Feature Map

[0012] f) The feature map Feature Map Feature Map Feature Map Feature Map Input into the jump connection unit of the segmentation network model, and output the feature maps respectively Feature Map Feature Map Feature Map

[0013] g) The feature map Feature Map Feature Map Feature Map Input into the decoder of the segmentation network model and output the segmentation result image

[0014] h) Calculate the total loss L loss , through the total loss L loss Train the segmentation network model to obtain an optimized segmentation network model;

[0015] Preferably, in step a), X cardiac ultrasound images are obtained from the CAMUS cardiac segmentation ultrasound image dataset, where X is 100; in step c), the preprocessed cardiac ultrasound image dataset S is divided into a training set, a validation set, and a test set in a ratio of 7:1:2.

[0016] Further, step b) comprises the following steps:

[0017] b-1) Use the center_crop() function in OpenCV to crop the i-th cardiac ultrasound image N i Perform a center cropping operation with a size of 256×256 to obtain the i-th cardiac ultrasound image N after croppingi ′;

[0018] b-2) Using Gaussian filtering to crop the i-th cardiac ultrasound image N i ′ is used for denoising, and the denoised i-th cardiac ultrasound image N is obtained. i ″;

[0019] b-3) The denoised i-th cardiac ultrasound image N i ″Perform z-score normalization operation to obtain the standardized i-th cardiac ultrasound image N i ″′;

[0020] b-4) Use L2 norm to normalize the i-th cardiac ultrasound image N i ″′ performs normalization operation to obtain the preprocessed cardiac ultrasound image S i .

[0021] Further, step e) comprises the following steps:

[0022] e-1) The encoder of the segmentation network model is composed of a first depth-separable convolution block, a first downsampling block, a second depth-separable convolution block, a second downsampling block, a third depth-separable convolution block, a third downsampling block, a fourth depth-separable convolution block, a fourth downsampling block, and a fifth depth-separable convolution block;

[0023] e-2) The first depth-wise separable convolutional block of the encoder is composed of a point-by-point convolutional layer, a ReLU activation function, and a BN layer. The i-th preprocessed cardiac ultrasound image S in the training set is i Input into the first depth separable convolution block and output the feature map

[0024] e-3) The first downsampling block of the encoder is composed of ReLU activation function, BN layer, and average pooling layer in sequence. Input into the first downsampling block and output the feature map

[0025] e-4) The second depth-wise separable convolutional block of the encoder is composed of a point-by-point convolutional layer, a ReLU activation function, and a BN layer. Input into the second depth separable convolution block, and output the feature map

[0026] e-5) The second downsampling block of the encoder is composed of ReLU activation function, BN layer, and average pooling layer in sequence. Input into the second downsampling block and output the feature map

[0027] e-6) The third depth-wise separable convolutional block of the encoder is composed of a point-by-point convolutional layer, a ReLU activation function, and a BN layer. Input into the third depth separable convolution block and output the feature map

[0028] e-7) The third downsampling block of the encoder is composed of ReLU activation function, BN layer, and average pooling layer in sequence. Input into the third downsampling block and output the feature map

[0029] e-8) The fourth depth-wise separable convolutional block of the encoder is composed of a point-by-point convolutional layer, a ReLU activation function, and a BN layer in sequence. Input into the fourth depth separable convolution block, and output the feature map

[0030] e-9) The fourth downsampling block of the encoder is composed of ReLU activation function, BN layer, and average pooling layer in sequence. Input to the fourth downsampling block and output the feature map

[0031] e-10) The fifth depth-wise separable convolutional block of the encoder is composed of a point-by-point convolutional layer, a ReLU activation function, and a BN layer. Input into the fifth depth separable convolution block and output the feature map Preferably, in step e-2), the convolution kernel size of the point-by-point convolution layer of the first depth-separable convolution block is 1×1, the padding is the same, and the step size is 1; in step e-3), the pooling kernel size of the average pooling layer of the first downsampling block is 2×2; in step e-4), the convolution kernel size of the point-by-point convolution layer of the second depth-separable convolution block is 1×1, the padding is the same, and the step size is 1; in step e-5), the pooling kernel size of the average pooling layer of the second downsampling block is 2×2; in step e-6), the point-by-point convolution of the third depth-separable convolution block The convolution kernel size of the layer is 1×1, the padding is the same, and the step size is 1; the pooling kernel size of the average pooling layer of the third downsampling block in step e-7) is 2×2; the convolution kernel size of the point-by-point convolution layer of the fourth depth-wise separable convolution block in step e-8) is 1×1, the padding is the same, and the step size is 1; the pooling kernel size of the average pooling layer of the fourth downsampling block in step e-9) is 2×2; the convolution kernel size of the point-by-point convolution layer of the fifth depth-wise separable convolution block in step e-10) is 1×1, and the padding is the same.

[0032] Further, step f) comprises the following steps:

[0033] f-1) The jump connection unit of the segmentation network model is composed of a first multi-level edge feature enhancement module MLEFE, a second multi-level edge feature enhancement module MLEFE, a third multi-level edge feature enhancement module MLEFE, and a multi-scale adaptive channel enhancement module MS-ACE;

[0034] f-2) The first multi-level edge feature enhancement module MLEFE of the jump connection unit is composed of a first branch and a second branch. The first branch is composed of a first convolutional layer, a BN layer, a ReLU activation function, and a second convolutional layer in sequence. Input into the first branch and output the feature map The second branch consists of a channel attention mechanism, a multi-head attention mechanism, a spatial attention mechanism, and an upsampling layer. Input into the channel attention mechanism of the second branch, and output the feature map The feature map Input into the multi-head attention mechanism of the second branch, and output the feature map The feature map Input into the spatial attention mechanism of the second branch, and output the feature map The feature map Feature Map Feature Map Perform element-by-element multiplication to obtain the feature map The feature map Input to the upsampling layer of the second branch, and output the feature map The feature map feature map Feature Map Feature Map Perform addition operation to obtain feature map

[0035] f-3) The feature map Replace the feature map in step f-2) Then repeat step f-2) to get the feature map

[0036] f-4) The feature map With feature map Perform addition operation to obtain feature map

[0037] f-5) The second multi-level edge feature enhancement module MLEFE of the jump connection unit is composed of a first branch and a second branch. The first branch is composed of a first convolutional layer, a BN layer, a ReLU activation function, and a second convolutional layer in sequence. Input into the first branch and output the feature map The second branch consists of a channel attention mechanism, a multi-head attention mechanism, a spatial attention mechanism, and an upsampling layer. Input into the channel attention mechanism of the second branch, and output the feature map The feature map Input into the multi-head attention mechanism of the second branch, and output the feature map The feature map Input into the spatial attention mechanism of the second branch, and output the feature map The feature map Feature Map Feature Map Perform element-by-element multiplication to obtain the feature map The feature map Input to the upsampling layer of the second branch, and output the feature map The feature map feature map Feature Map Feature Map Perform addition operation to obtain feature map

[0038] f-6) The feature map With feature map Perform addition operation to obtain feature map

[0039] f-7) The third multi-level edge feature enhancement module MLEFE of the jump connection unit is composed of a first branch and a second branch. The first branch is composed of a first convolutional layer, a BN layer, a ReLU activation function, and a second convolutional layer in sequence. Input into the first branch and output the feature map The second branch consists of a channel attention mechanism, a multi-head attention mechanism, a spatial attention mechanism, and an upsampling layer. Input into the channel attention mechanism of the second branch, and output the feature map The feature map Input into the multi-head attention mechanism of the second branch, and output the feature map The feature map Input into the spatial attention mechanism of the second branch, and output the feature map The feature map Feature Map Feature Map Perform element-by-element multiplication to obtain the feature map The feature map Input to the upsampling layer of the second branch, and output the feature map The feature map feature map Feature Map Feature Map Perform addition operation to obtain feature map

[0040] f-8) The feature map With feature map Perform addition operation to obtain feature map

[0041] f-9) The multi-scale adaptive channel enhancement module MS-ACE of the jump connection unit is composed of the first branch, the second branch, the third branch, the convolution layer, and the ReLU activation function. The first branch is composed of the first convolution unit and the second convolution unit. The first convolution unit of the first branch is composed of the first convolution layer, the second convolution layer, the ReLU activation function, the BN layer, and the GAP layer in sequence. The feature map Input into the first convolution unit of the first branch, and output the feature map The second convolution unit of the first branch is composed of the first convolution layer, BN layer, second convolution layer, and Softmax function in sequence. Input into the second convolution unit of the first branch, and output the feature map The second branch is composed of the first convolution unit and the second convolution unit. The first convolution unit of the second branch is composed of the first convolution layer, the second convolution layer, the ReLU activation function, the BN layer, and the GAP layer in sequence. Input into the first convolution unit of the second branch, and output the feature map The second convolution unit of the second branch is composed of the first convolution layer, the BN layer, the second convolution layer, and the Softmax function in sequence. Input into the second convolution unit of the second branch, and output the feature map The third branch is composed of the first convolution unit and the second convolution unit. The first convolution unit of the third branch is composed of the first convolution layer, the second convolution layer, the ReLU activation function, the BN layer, and the GAP layer in sequence. Input into the first convolution unit of the third branch, and output the feature map The second convolution unit of the third branch is composed of the first convolution layer, the BN layer, the second convolution layer, and the Softmax function in sequence. Input into the second convolution unit of the first branch, and output the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The feature map Feature Map Feature Map After the addition operation, it is sequentially input into the convolution layer and ReLU activation function of the multi-scale adaptive channel enhancement module MS-ACE, and the feature map is output.

[0042] Preferably, in step f-2), the convolution kernel size of the first convolution layer of the first branch of the first multi-level edge feature enhancement module MLEFE is 3×3, padding=1, and stride=1, and the convolution kernel size of the second convolution layer of the first branch of the first multi-level edge feature enhancement module MLEFE is 1×1, padding=0, and stride=0; in step f-5), the convolution kernel size of the first convolution layer of the first branch of the second multi-level edge feature enhancement module MLEFE is 3×3, padding=1, and stride=1, and the convolution kernel size of the second convolution layer of the first branch of the second multi-level edge feature enhancement module MLEFE is 1×1, padding=0, and stride=0. tride=0;In step f-7), the convolution kernel size of the first convolution layer of the first branch of the third multi-level edge feature enhancement module MLEFE is 3×3, padding=1, and stride=1, and the convolution kernel size of the second convolution layer of the first branch of the third multi-level edge feature enhancement module MLEFE is 1×1, padding=0, and stride=0; In step f-9), the convolution kernel size of the first convolution layer of the first convolution unit of the first branch is 3×3, padding=0, and stride=0, the convolution kernel size of the second convolution layer of the first convolution unit of the first branch is 3×3, padding=0, and stride=0, and the convolution kernel size of the first convolution layer of the second convolution unit of the first branch is 3×3, padding=0, and stride=0, and the convolution kernel size of the first convolution layer of the second convolution unit of the first branch is 3×3, padding=0, and stride=0, and the convolution kernel size of the first convolution layer of the second convolution unit of the first branch is 1×1, padding=0, and stride=0, and the convolution kernel size of the first convolution layer of the second convolution unit of the first branch is 1×1, padding=0, and stride=0, and the convolution kernel size of the first convolution layer of the first ... The convolution kernel size of the convolution layer is 3×3, padding=1, and stride=1. The convolution kernel size of the second convolution layer of the second convolution unit of the first branch is 1×1, padding=0, and the stride is 0. The convolution kernel size of the first convolution layer of the first convolution unit of the second branch is 3×3, padding=0, and the stride is 0. The convolution kernel size of the second convolution layer of the first convolution unit of the second branch is 1×1, padding=0, and the stride is 0. The convolution kernel size of the first convolution layer of the second convolution unit of the second branch is 3×3, padding=1, and stride=1. The convolution kernel size of the second convolution layer of the second convolution unit of the second branch is 1×1, padding=0, and the stride is 0. ding=0, stride is 0, the convolution kernel size of the first convolution layer of the first convolution unit of the third branch is 3×3, padding=0, and stride is 0, the convolution kernel size of the second convolution layer of the first convolution unit of the third branch is 1×1, padding=0, and stride=0, the convolution kernel size of the first convolution layer of the second convolution unit of the third branch is 3×3, padding=1, and stride=1, the convolution kernel size of the second convolution layer of the second convolution unit of the third branch is 1×1, padding=0, and stride is 0, and the convolution kernel size of the convolution layer of the multi-scale adaptive channel enhancement module MS-ACE is 1×1, padding=0, and stride=0.

[0043] Further, step g) comprises the following steps:

[0044] g-1) The decoder of the segmentation network model is composed of a first decoding layer, a first upsampling layer, a first feature fusion module FFM, a second decoding layer, a second upsampling layer, a third decoding layer, a third upsampling layer, a second feature fusion module FFM, a fourth decoding layer, and a fourth upsampling layer;

[0045] g-2) The feature map With feature map Input to the first decoding layer of the decoder for concatenation to obtain the feature map

[0046] g-3) The feature map Input to the first upsampling layer of the decoder, and output the feature map

[0047] g-4) The first feature fusion module FFM of the decoder consists of the first branch, the second branch, the convolution layer, and the ReLU activation function. The first branch consists of the first small branch, the second small branch, and the ReLU activation function. The second branch consists of the first small branch, the second small branch, the convolution layer, and the ReLU activation function. The first small branch of the first branch consists of the transposed convolution layer, the convolution layer, and the BN layer in sequence. The feature map Input into the first small branch of the first branch, and output the feature map The second small branch of the first branch is composed of convolutional layers and LN layers in sequence. Input into the first small branch of the first branch, and output the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The first small branch of the second branch is composed of a transposed convolution layer, a convolution layer, a LN layer, and a ReLU activation function. Input into the first small branch of the second branch, and output the feature map The second branch of the second branch is composed of convolutional layer, BN layer, and ReLU activation function in sequence. Input into the second small branch of the second branch, and output the feature map The feature map With feature map Perform element-by-element addition to obtain the feature map The feature map Input into the ReLU activation function of the first branch, and output the feature map The feature map Input into the convolution layer and ReLU activation function of the second branch in turn, and output the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The feature map The convolutional layer and ReLU activation function of the first feature fusion module FFM are input in sequence, and the feature map is output.

[0048] g-5) The feature map With feature map Input to the second decoding layer of the decoder for concatenation to obtain the feature map

[0049] g-6) The feature map Input to the second upsampling layer of the decoder, and output the feature map

[0050] g-7) The feature map With feature map Input to the third decoding layer of the decoder for concatenation to obtain the feature map

[0051] g-8) The feature map Input to the third upsampling layer of the decoder, and output the feature map

[0052] g-9) The second feature fusion module FFM of the decoder is composed of the first branch, the second branch, the convolution layer, and the ReLU activation function. The first branch is composed of the first small branch, the second small branch, and the ReLU activation function. The second branch is composed of the first small branch, the second small branch, the convolution layer, and the ReLU activation function. The first small branch of the first branch is composed of the transposed convolution layer, the convolution layer, and the BN layer in sequence. The feature map Input into the first small branch of the first branch, and output the feature map The second small branch of the first branch is composed of convolutional layers and LN layers in sequence. Input into the first small branch of the first branch, and output the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The first small branch of the second branch is composed of a transposed convolution layer, a convolution layer, a LN layer, and a ReLU activation function. Input into the first small branch of the second branch, and output the feature map The second branch of the second branch is composed of convolutional layer, BN layer, and ReLU activation function in sequence. Input into the second small branch of the second branch, and output the feature map The feature map With feature map Perform element-by-element addition to obtain the feature map The feature map Input into the ReLU activation function of the first branch, and output the feature map The feature map Input into the convolution layer and ReLU activation function of the second branch in turn, and output the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The feature map The convolutional layer and ReLU activation function of the second feature fusion module FFM are input in sequence, and the feature map is output.

[0053] g-10) The feature map With feature map Input to the fourth decoding layer of the decoder for splicing operation to obtain the feature map

[0054] g-11) The feature map Input to the fourth upsampling layer of the decoder, and output the feature map The feature map With feature map Perform stitching operation to obtain the segmentation result image

[0055] Preferably, in step g-4), the convolution kernel size of the transposed convolution layer of the first small branch of the first branch is 2×2, the padding is the same, and the step size is 1, the convolution kernel size of the convolution layer of the first small branch of the first branch is 3×3, the padding is 0, and the step size is 0, the convolution kernel size of the convolution layer of the second small branch of the first branch is 3×3, the padding is 0, and the step size is 0, the convolution kernel size of the transposed convolution layer of the first small branch of the second branch is 2×2, the padding is the same, and the step size is 1, the convolution kernel size of the convolution layer of the first small branch of the second branch is 3×3, the padding is 0, and the step size is 0, the convolution kernel size of the convolution layer of the second small branch of the second branch is 3×3, the padding is 0, and the step size is 0; in step g-9), the convolution kernel size of the transposed convolution layer of the first small branch of the first branch is 2×2, the padding is the same, and the step length is 1, the convolution kernel size of the convolution layer of the first small branch of the first branch is 3×3, the padding is 0, and the step length is 0, the convolution kernel size of the convolution layer of the second small branch of the first branch is 3×3, the padding is 0, and the step length is 0, the convolution kernel size of the transposed convolution layer of the first small branch of the second branch is 2×2, the padding is the same, and the step length is 1, the convolution kernel size of the convolution layer of the first small branch of the second branch is 3×3, the padding is 0, and the step length is 0, and the convolution kernel size of the convolution layer of the second small branch of the second branch is 3×3, the padding is 0, and the step length is 0. Further, step h) includes the following steps:

[0056] h-1) through the formula L loss =W1L Dice +W2L BCE Calculate the total loss L loss , where L Dice is the Dice loss, L BCE is the binary cross entropy loss, W1 and W2 are weight values;

[0057] h-2) Use the Adam optimizer to use the total loss L loss Train the segmentation network model to obtain an optimized segmentation network model.

[0058] The beneficial effects of the present invention are as follows: by integrating a multi-level edge feature enhancement module, a feature fusion module and a multi-scale feature channel enhancement module in a multi-scale convolutional network based on boundary perception, the accuracy and efficiency of cardiac ultrasound image segmentation are significantly improved. In particular, the multi-level edge feature enhancement module can accurately capture the subtle differences between the heart edge and the surrounding tissues, and optimize the presentation of shape and structural information. At the same time, the feature fusion module deeply mines and integrates information from different network layers, improving the model's ability to recognize complex cardiac structures. In addition, the multi-scale feature channel enhancement module provides the network with a wider receptive field and richer scale information, effectively coping with the scale variation problem existing in cardiac images, thereby achieving more robust and adaptable cardiac MRI image segmentation. The application of this method is expected to greatly improve the efficiency and accuracy of cardiovascular disease diagnosis and treatment, which has important practical and long-term significance for the medical field, and the accuracy of segmentation is improved through the design of this network structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a structural diagram of the segmentation network model of the present invention;

[0060] Figure 2 A structural diagram of the multi-level edge feature enhancement MLEFE of the present invention;

[0061] Figure 3 It is a structural diagram of the multi-scale adaptive channel enhancement module MS-ACE of the present invention;

[0062] Figure 4 This is a structural diagram of the feature fusion module FFM of the present invention. DETAILED DESCRIPTION

[0063] The following is combined with Figure 1 To Attachment Figure 4 The present invention is further described.

[0064] A method for segmenting cardiac ultrasound images based on a boundary-aware multi-scale convolutional network, comprising:

[0065] a) Obtain X cardiac ultrasound images to obtain a cardiac ultrasound image dataset N, where N = {N1, N2, ..., N i ,...,N X},N i is the i-th cardiac ultrasound image, i∈{1,...,X}.

[0066] b) Take the i-th cardiac ultrasound image N i Preprocessing is performed to obtain the preprocessed cardiac ultrasound image S i , all X preprocessed cardiac ultrasound images constitute a preprocessed cardiac ultrasound image dataset S, S = {S1, S2, ..., Si ,...,S X}.

[0067] c) Divide the preprocessed cardiac ultrasound image dataset S into a training set, a validation set, and a test set.

[0068] d) Establish a segmentation network model consisting of an encoder, a skip connection unit, and a decoder.

[0069] e) The i-th preprocessed cardiac ultrasound image S in the training set i Input into the encoder of the segmentation network model and output the feature maps respectively Feature Map Feature Map Feature Map Feature Map

[0070] f) The feature map Feature Map Feature Map Feature Map Feature Map Input into the jump connection unit of the segmentation network model, and output the feature maps respectively Feature Map Feature Map Feature Map

[0071] g) The feature map Feature Map Feature Map Feature Map Input into the decoder of the segmentation network model and output the segmentation result image

[0072] h) Calculate the total loss L loss , through the total loss L loss Train the segmentation network model to obtain an optimized segmentation network model.

[0073] Table 1 Comparison of different experimental methods:

[0074]

[0075] As can be seen from Table 1, a variety of classic segmentation methods are compared, namely U-Net, TransUNet, ResUNet, Att-UNet, and FCT. Each method segmented the CAMUS dataset, and used four evaluation indicators, namely Dice coefficient, Hausdorff distance, precision (Precision), and pixel accuracy (PA), to calculate the accuracy of cardiac image segmentation at the end of systole (ES) and end of diastole (ED). The lower the Hausdorff distance score, the better the performance, and the higher the Dice coefficient, precision (Precision), and pixel accuracy (PA) scores, the better the performance. It can be seen from the data in the table that the average Dice evaluation index of the end-systolic period (ES) of the cardiac image segmentation method of the present invention is improved by 1.25% compared with the previous optimal method, the average precision (Precision) evaluation index is improved by 0.49% compared with the previous optimal method, and the pixel accuracy (PA) evaluation index is improved by 0.09% compared with the previous optimal method. The average Dice evaluation index of the end-diastolic period (ED) of the cardiac image segmentation is improved by 0.13% compared with the previous optimal method, the average Hausdorff distance evaluation index is improved by 0.03% compared with the previous optimal method, and the pixel accuracy (PA) evaluation index is improved by 0.09% compared with the previous optimal method, which is better than the previous method, indicating that the method of the present invention has certain advantages and the segmentation effect is relatively ideal.

[0076] In one embodiment of the present invention, in step a), X cardiac ultrasound images are obtained from the CAMUS cardiac segmentation ultrasound image dataset, where X is 100; in step c), the preprocessed cardiac ultrasound image dataset S is divided into a training set, a validation set, and a test set in a ratio of 7:1:2.

[0077] In one embodiment of the present invention, step b) comprises the following steps:

[0078] b-1) Use the center_crop() function in OpenCV to crop the i-th cardiac ultrasound image N i Perform a center cropping operation with a size of 256×256 to obtain the i-th cardiac ultrasound image N after cropping i ′.

[0079] b-2) Using Gaussian filtering to crop the i-th cardiac ultrasound image N i ′ is used for denoising, and the denoised i-th cardiac ultrasound image N is obtained. i ″.

[0080] b-3) The denoised i-th cardiac ultrasound image N i ″Perform z-score normalization operation to obtain the standardized i-th cardiac ultrasound image Ni ″′.

[0081] b-4) Use L2 norm to normalize the i-th cardiac ultrasound image N i ″′ performs normalization operation to obtain the preprocessed cardiac ultrasound image S i .

[0082] In one embodiment of the present invention, step e) comprises the following steps:

[0083] e-1) The encoder of the segmentation network model is composed of a first depth-separable convolution block, a first downsampling block, a second depth-separable convolution block, a second downsampling block, a third depth-separable convolution block, a third downsampling block, a fourth depth-separable convolution block, a fourth downsampling block, and a fifth depth-separable convolution block.

[0084] e-2) The first depth-wise separable convolutional block of the encoder is composed of a point-by-point convolutional layer, a ReLU activation function, and a BN layer. The i-th preprocessed cardiac ultrasound image S in the training set is i Input into the first depth separable convolution block and output the feature map

[0085] e-3) The first downsampling block of the encoder is composed of ReLU activation function, BN layer, and average pooling layer in sequence. Input into the first downsampling block and output the feature map

[0086] e-4) The second depth-wise separable convolutional block of the encoder is composed of a point-by-point convolutional layer, a ReLU activation function, and a BN layer. Input into the second depth separable convolution block, and output the feature map

[0087] e-5) The second downsampling block of the encoder is composed of ReLU activation function, BN layer, and average pooling layer in sequence. Input into the second downsampling block and output the feature map

[0088] e-6) The third depth-wise separable convolutional block of the encoder is composed of a point-by-point convolutional layer, a ReLU activation function, and a BN layer. Input into the third depth separable convolution block and output the feature map

[0089] e-7) The third downsampling block of the encoder is composed of ReLU activation function, BN layer, and average pooling layer in sequence. Input into the third downsampling block and output the feature map

[0090] e-8) The fourth depth-wise separable convolutional block of the encoder is composed of a point-by-point convolutional layer, a ReLU activation function, and a BN layer in sequence. Input into the fourth depth separable convolution block, and output the feature map

[0091] e-9) The fourth downsampling block of the encoder is composed of ReLU activation function, BN layer, and average pooling layer in sequence. Input to the fourth downsampling block and output the feature map

[0092] e-10) The fifth depth-wise separable convolutional block of the encoder is composed of a point-by-point convolutional layer, a ReLU activation function, and a BN layer. Input into the fifth depth separable convolution block and output the feature map In this embodiment, in step e-2), the convolution kernel size of the point-by-point convolution layer of the first depth-separable convolution block is 1×1, the padding is the same, and the step size is 1; in step e-3), the pooling kernel size of the average pooling layer of the first downsampling block is 2×2; in step e-4), the convolution kernel size of the point-by-point convolution layer of the second depth-separable convolution block is 1×1, the padding is the same, and the step size is 1; in step e-5), the pooling kernel size of the average pooling layer of the second downsampling block is 2×2; in step e-6), the point-by-point convolution kernel size of the third depth-separable convolution block is 2×2. The convolution kernel size of the convolution layer is 1×1, the padding is the same, and the step size is 1; the pooling kernel size of the average pooling layer of the third downsampling block in step e-7) is 2×2; the convolution kernel size of the point-by-point convolution layer of the fourth depth-separable convolution block in step e-8) is 1×1, the padding is the same, and the step size is 1; the pooling kernel size of the average pooling layer of the fourth downsampling block in step e-9) is 2×2; the convolution kernel size of the point-by-point convolution layer of the fifth depth-separable convolution block in step e-10) is 1×1, and the padding is the same.

[0093] In one embodiment of the present invention, step f) comprises the following steps:

[0094] f-1) The jump connection unit of the segmentation network model is composed of a first multi-level edge feature enhancement module MLEFE, a second multi-level edge feature enhancement module MLEFE, a third multi-level edge feature enhancement module MLEFE, and a multi-scale adaptive channel enhancement module MS-ACE.

[0095] f-2) The first multi-level edge feature enhancement module MLEFE of the jump connection unit is composed of a first branch and a second branch. The first branch is composed of a first convolutional layer, a BN layer, a ReLU activation function, and a second convolutional layer in sequence. Input into the first branch and output the feature map The second branch consists of a channel attention mechanism, a multi-head attention mechanism, a spatial attention mechanism, and an upsampling layer. Input into the channel attention mechanism of the second branch, and output the feature map The feature map Input into the multi-head attention mechanism of the second branch, and output the feature map The feature map Input into the spatial attention mechanism of the second branch, and output the feature map The feature map Feature Map Feature Map Perform element-by-element multiplication to obtain the feature map The feature map Input to the upsampling layer of the second branch, and output the feature map The feature map feature map Feature Map Feature Map Perform addition operation to obtain feature map

[0096] f-3) The feature map Replace the feature map in step f-2) Then repeat step f-2) to get the feature map

[0097] f-4) The feature map With feature map Perform addition operation to obtain feature map

[0098] f-5) The second multi-level edge feature enhancement module MLEFE of the jump connection unit is composed of a first branch and a second branch. The first branch is composed of a first convolutional layer, a BN layer, a ReLU activation function, and a second convolutional layer in sequence. Input into the first branch and output the feature map The second branch consists of a channel attention mechanism, a multi-head attention mechanism, a spatial attention mechanism, and an upsampling layer. Input into the channel attention mechanism of the second branch, and output the feature map The feature map Input into the multi-head attention mechanism of the second branch, and output the feature map The feature map Input into the spatial attention mechanism of the second branch, and output the feature map The feature map Feature Map Feature Map Perform element-by-element multiplication to obtain the feature map The feature map Input to the upsampling layer of the second branch, and output the feature map The feature map feature map Feature Map Feature Map Perform addition operation to obtain feature map

[0099] f-6) The feature map With feature map Perform addition operation to obtain feature map

[0100] f-7) The third multi-level edge feature enhancement module MLEFE of the jump connection unit is composed of a first branch and a second branch. The first branch is composed of a first convolutional layer, a BN layer, a ReLU activation function, and a second convolutional layer in sequence. Input into the first branch and output the feature map The second branch consists of a channel attention mechanism, a multi-head attention mechanism, a spatial attention mechanism, and an upsampling layer. Input into the channel attention mechanism of the second branch, and output the feature map The feature map Input into the multi-head attention mechanism of the second branch, and output the feature map The feature map Input into the spatial attention mechanism of the second branch, and output the feature map The feature map Feature Map Feature Map Perform element-by-element multiplication to obtain the feature map The feature map Input to the upsampling layer of the second branch, and output the feature map The feature map feature map Feature Map Feature Map Perform addition operation to obtain feature map

[0101] f-8) The feature map With feature map Perform addition operation to obtain feature map

[0102] f-9) The multi-scale adaptive channel enhancement module MS-ACE of the jump connection unit is composed of the first branch, the second branch, the third branch, the convolution layer, and the ReLU activation function. The first branch is composed of the first convolution unit and the second convolution unit. The first convolution unit of the first branch is composed of the first convolution layer, the second convolution layer, the ReLU activation function, the BN layer, and the GAP layer in sequence. The feature map Input into the first convolution unit of the first branch, and output the feature map The second convolution unit of the first branch is composed of the first convolution layer, BN layer, second convolution layer, and Softmax function in sequence. Input into the second convolution unit of the first branch, and output the feature map The second branch is composed of the first convolution unit and the second convolution unit. The first convolution unit of the second branch is composed of the first convolution layer, the second convolution layer, the ReLU activation function, the BN layer, and the GAP layer in sequence. Input into the first convolution unit of the second branch, and output the feature map The second convolution unit of the second branch is composed of the first convolution layer, the BN layer, the second convolution layer, and the Softmax function in sequence. Input into the second convolution unit of the second branch, and output the feature map The third branch is composed of the first convolution unit and the second convolution unit. The first convolution unit of the third branch is composed of the first convolution layer, the second convolution layer, the ReLU activation function, the BN layer, and the GAP layer in sequence. Input into the first convolution unit of the third branch, and output the feature map The second convolution unit of the third branch is composed of the first convolution layer, the BN layer, the second convolution layer, and the Softmax function in sequence. Input into the second convolution unit of the first branch, and output the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The feature map Feature Map Feature Map After the addition operation, it is sequentially input into the convolution layer and ReLU activation function of the multi-scale adaptive channel enhancement module MS-ACE, and the feature map is output. In this embodiment, in step f-2), the convolution kernel size of the first convolution layer of the first branch of the first multi-level edge feature enhancement module MLEFE is 3×3, padding=1, and stride=1, and the convolution kernel size of the second convolution layer of the first branch of the first multi-level edge feature enhancement module MLEFE is 1×1, padding=0, and stride=0; in step f-5), the convolution kernel size of the first convolution layer of the first branch of the second multi-level edge feature enhancement module MLEFE is 3×3, padding=1, and stride=1, and the convolution kernel size of the second convolution layer of the first branch of the second multi-level edge feature enhancement module MLEFE is 1×1, padding=0. , stride=0; in step f-7), the convolution kernel size of the first convolution layer of the first branch of the third multi-level edge feature enhancement module MLEFE is 3×3, padding=1, and stride=1, and the convolution kernel size of the second convolution layer of the first branch of the third multi-level edge feature enhancement module MLEFE is 1×1, padding=0, and stride=0; in step f-9), the convolution kernel size of the first convolution layer of the first convolution unit of the first branch is 3×3, padding=0, and the stride is 0, the convolution kernel size of the second convolution layer of the first convolution unit of the first branch is 3×3, padding=0, and the stride is 0, and the convolution kernel size of the second convolution layer of the first convolution unit of the first branch is 3×3, padding=0, and the stride is 0, and the convolution kernel size of the first convolution unit of the first branch is 3×3, padding=0, and the stride is 0, and the convolution kernel size of the first convolution unit of the first branch is The convolution kernel size of the convolution layer is 3×3, padding=1, and stride=1. The convolution kernel size of the second convolution layer of the second convolution unit of the first branch is 1×1, padding=0, and the stride is 0. The convolution kernel size of the first convolution layer of the first convolution unit of the second branch is 3×3, padding=0, and the stride is 0. The convolution kernel size of the second convolution layer of the first convolution unit of the second branch is 1×1, padding=0, and the stride is 0. The convolution kernel size of the first convolution layer of the second convolution unit of the second branch is 3×3, padding=1, and stride=1. The convolution kernel size of the second convolution layer of the second convolution unit of the second branch is 1×1, padding=0, and the stride is 0. dding=0, stride is 0, the convolution kernel size of the first convolution layer of the first convolution unit of the third branch is 3×3, padding=0, and stride is 0, the convolution kernel size of the second convolution layer of the first convolution unit of the third branch is 1×1, padding=0, and stride=0, the convolution kernel size of the first convolution layer of the second convolution unit of the third branch is 3×3, padding=1, and stride=1, the convolution kernel size of the second convolution layer of the second convolution unit of the third branch is 1×1, padding=0, and stride is 0, and the convolution kernel size of the convolution layer of the multi-scale adaptive channel enhancement module MS-ACE is 1×1, padding=0, and stride=0.

[0103] In one embodiment of the present invention, step g) comprises the following steps:

[0104] g-1) The decoder of the segmentation network model is composed of a first decoding layer, a first upsampling layer, a first feature fusion module FFM, a second decoding layer, a second upsampling layer, a third decoding layer, a third upsampling layer, a second feature fusion module FFM, a fourth decoding layer, and a fourth upsampling layer.

[0105] g-2) The feature map With feature map Input to the first decoding layer of the decoder for concatenation to obtain the feature map

[0106] g-3) The feature map Input to the first upsampling layer of the decoder, and output the feature map

[0107] g-4) The first feature fusion module FFM of the decoder consists of the first branch, the second branch, the convolution layer, and the ReLU activation function. The first branch consists of the first small branch, the second small branch, and the ReLU activation function. The second branch consists of the first small branch, the second small branch, the convolution layer, and the ReLU activation function. The first small branch of the first branch consists of the transposed convolution layer, the convolution layer, and the BN layer in sequence. The feature map Input into the first small branch of the first branch, and output the feature map The second small branch of the first branch is composed of convolutional layers and LN layers in sequence. Input into the first small branch of the first branch, and output the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The first small branch of the second branch is composed of a transposed convolution layer, a convolution layer, a LN layer, and a ReLU activation function. Input into the first small branch of the second branch, and output the feature map The second branch of the second branch is composed of convolutional layer, BN layer, and ReLU activation function in sequence. Input into the second small branch of the second branch, and output the feature map The feature map With feature map Perform element-by-element addition to obtain the feature map The feature map Input into the ReLU activation function of the first branch, and output the feature map The feature map Input into the convolution layer and ReLU activation function of the second branch in turn, and output the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The feature map The convolutional layer and ReLU activation function of the first feature fusion module FFM are input in sequence, and the feature map is output.

[0108] g-5) The feature map With feature map Input to the second decoding layer of the decoder for concatenation to obtain the feature map

[0109] g-6) The feature map Input to the second upsampling layer of the decoder, and output the feature map

[0110] g-7) The feature map With feature map Input to the third decoding layer of the decoder for concatenation to obtain the feature map

[0111] g-8) The feature map Input to the third upsampling layer of the decoder, and output the feature map

[0112] g-9) The second feature fusion module FFM of the decoder is composed of the first branch, the second branch, the convolution layer, and the ReLU activation function. The first branch is composed of the first small branch, the second small branch, and the ReLU activation function. The second branch is composed of the first small branch, the second small branch, the convolution layer, and the ReLU activation function. The first small branch of the first branch is composed of the transposed convolution layer, the convolution layer, and the BN layer in sequence. The feature map Input into the first small branch of the first branch, and output the feature map The second small branch of the first branch is composed of convolutional layers and LN layers in sequence. Input into the first small branch of the first branch, and output the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The first small branch of the second branch is composed of a transposed convolution layer, a convolution layer, a LN layer, and a ReLU activation function. Input into the first small branch of the second branch, and output the feature map The second branch of the second branch is composed of convolutional layer, BN layer, and ReLU activation function in sequence. Input into the second small branch of the second branch, and output the feature map The feature map With feature map Perform element-by-element addition to obtain the feature map The feature map Input into the ReLU activation function of the first branch, and output the feature map The feature map Input into the convolution layer and ReLU activation function of the second branch in turn, and output the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The feature map The convolutional layer and ReLU activation function of the second feature fusion module FFM are input in sequence, and the feature map is output.

[0113] g-10) The feature map With feature map Input to the fourth decoding layer of the decoder for splicing operation to obtain the feature map

[0114] g-11) The feature map Input to the fourth upsampling layer of the decoder, and output the feature map The feature map With feature map Perform stitching operation to obtain the segmentation result image In this embodiment, in step g-4), the convolution kernel size of the transposed convolution layer of the first small branch of the first branch is 2×2, the padding is the same, and the step size is 1, the convolution kernel size of the convolution layer of the first small branch of the first branch is 3×3, the padding is 0, and the step size is 0, the convolution kernel size of the convolution layer of the second small branch of the first branch is 3×3, the padding is 0, and the step size is 0, the convolution kernel size of the transposed convolution layer of the first small branch of the second branch is 2×2, the padding is the same, and the step size is 1, the convolution kernel size of the convolution layer of the first small branch of the second branch is 3×3, the padding is 0, and the step size is 0, the convolution kernel size of the convolution layer of the second small branch of the second branch is 3×3, the padding is 0, and the step size is 3×3. length is 0; in step g-9), the convolution kernel size of the transposed convolution layer of the first small branch of the first branch is 2×2, the padding is the same, and the step size is 1, the convolution kernel size of the convolution layer of the first small branch of the first branch is 3×3, the padding is 0, and the step size is 0, the convolution kernel size of the convolution layer of the second small branch of the first branch is 3×3, the padding is 0, and the step size is 0, the convolution kernel size of the transposed convolution layer of the first small branch of the second branch is 2×2, the padding is the same, and the step size is 1, the convolution kernel size of the convolution layer of the first small branch of the second branch is 3×3, the padding is 0, and the step size is 0, and the convolution kernel size of the convolution layer of the second small branch of the second branch is 3×3, the padding is 0, and the step size is 0.

[0115] In one embodiment of the present invention, step h) comprises the following steps:

[0116] h-1) through the formula L loss =W1L Dice +W2L BCE Calculate the total loss L loss , where L Dice is the Dice loss, L BCE is the binary cross entropy loss, W1 and W2 are weight values.

[0117] h-2) Use the Adam optimizer to use the total loss L loss Train the segmentation network model to obtain the optimized segmentation network model. The learning rate during training is le -4 , the batch size is set to 10.

[0118] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for segmenting cardiac ultrasound images based on a boundary-aware multi-scale convolutional network, characterized in that: include: a) Obtain X cardiac ultrasound images to obtain a cardiac ultrasound image dataset N, where N = {N1, N2, ..., N i ,...,N X },N i is the i-th cardiac ultrasound image, i∈{1,...,X}; b) Take the i-th cardiac ultrasound image N i Preprocessing is performed to obtain the preprocessed cardiac ultrasound image S i , all X preprocessed cardiac ultrasound images constitute a preprocessed cardiac ultrasound image dataset S, S = {S1, S2, ..., S i ,...,S X }; c) dividing the preprocessed cardiac ultrasound image dataset S into a training set, a validation set, and a test set; d) establishing a segmentation network model consisting of an encoder, a skip connection unit, and a decoder; e) The i-th preprocessed cardiac ultrasound image S in the training set i Input into the encoder of the segmentation network model and output the feature maps respectively Feature Map Feature Map Feature Map Feature Map f) The feature map Feature Map Feature Map Feature Map Feature Map Input into the jump connection unit of the segmentation network model, and output the feature maps respectively Feature Map Feature Map Feature Map g) The feature map Feature Map Feature Map Feature Map Input into the decoder of the segmentation network model and output the segmentation result image h) Calculate the total loss L loss , through the total loss L loss Train the segmentation network model to obtain an optimized segmentation network model; 2. The method for segmenting cardiac ultrasound images based on a boundary-aware multi-scale convolutional network according to claim 1, characterized in that: In step a), X cardiac ultrasound images are obtained from the CAMUS cardiac segmentation ultrasound image dataset, where X is 100; in step c), the preprocessed cardiac ultrasound image dataset S is divided into a training set, a validation set, and a test set in a ratio of 7:1:

2.

3. The method for segmenting cardiac ultrasound images based on a boundary-aware multi-scale convolutional network according to claim 1, characterized in that: Step b) comprises the following steps: b-1) Use the center_crop() function in OpenCV to crop the i-th cardiac ultrasound image N i Perform a center cropping operation with a size of 256×256 to obtain the i-th cardiac ultrasound image N after cropping i ′; b-2) Using Gaussian filtering to crop the i-th cardiac ultrasound image N i ′ is used for denoising, and the denoised i-th cardiac ultrasound image N is obtained. i ″; b-3) The denoised i-th cardiac ultrasound image N i ″Perform z-score normalization operation to obtain the standardized i-th cardiac ultrasound image N i ″′; b-4) Use L2 norm to normalize the i-th cardiac ultrasound image N i ″′ performs normalization operation to obtain the preprocessed cardiac ultrasound image S i .

4. The method for segmenting cardiac ultrasound images based on a boundary-aware multi-scale convolutional network according to claim 1, characterized in that: Step e) comprises the following steps: e-1) The encoder of the segmentation network model is composed of a first depth-separable convolution block, a first downsampling block, a second depth-separable convolution block, a second downsampling block, a third depth-separable convolution block, a third downsampling block, a fourth depth-separable convolution block, a fourth downsampling block, and a fifth depth-separable convolution block; e-2) The first depth-wise separable convolutional block of the encoder is composed of a point-by-point convolutional layer, a ReLU activation function, and a BN layer. The i-th preprocessed cardiac ultrasound image S in the training set is i Input into the first depth separable convolution block, and output the feature map e-3) The first downsampling block of the encoder is composed of ReLU activation function, BN layer, and average pooling layer in sequence. Input into the first downsampling block and output the feature map e-4) The second depth-wise separable convolutional block of the encoder is composed of a point-by-point convolutional layer, a ReLU activation function, and a BN layer. Input into the second depth separable convolution block, and output the feature map e-5) The second downsampling block of the encoder is composed of ReLU activation function, BN layer, and average pooling layer in sequence. Input into the second downsampling block and output the feature map e-6) The third depth-wise separable convolutional block of the encoder is composed of a point-by-point convolutional layer, a ReLU activation function, and a BN layer. Input into the third depth separable convolution block and output the feature map e-7) The third downsampling block of the encoder is composed of ReLU activation function, BN layer, and average pooling layer in sequence. Input into the third downsampling block and output the feature map e-8) The fourth depth-wise separable convolutional block of the encoder is composed of a point-by-point convolutional layer, a ReLU activation function, and a BN layer in sequence. Input into the fourth depth separable convolution block, and output the feature map e-9) The fourth downsampling block of the encoder is composed of ReLU activation function, BN layer, and average pooling layer in sequence. Input to the fourth downsampling block and output the feature map e-10) The fifth depth-wise separable convolutional block of the encoder is composed of a point-by-point convolutional layer, a ReLU activation function, and a BN layer. Input into the fifth depth separable convolution block and output the feature map 5. The method for segmenting cardiac ultrasound images based on a boundary-aware multi-scale convolutional network according to claim 4, characterized in that: In step e-2), the convolution kernel size of the point-by-point convolution layer of the first depth-separable convolution block is 1×1, the padding is the same, and the step size is 1; in step e-3), the pooling kernel size of the average pooling layer of the first downsampling block is 2×2; in step e-4), the convolution kernel size of the point-by-point convolution layer of the second depth-separable convolution block is 1×1, the padding is the same, and the step size is 1; in step e-5), the pooling kernel size of the average pooling layer of the second downsampling block is 2×2; in step e-6), the point-by-point convolution layer of the third depth-separable convolution block is 2×2. The convolution kernel size is 1×1, the padding is the same, and the step size is 1; in step e-7), the pooling kernel size of the average pooling layer of the third downsampling block is 2×2; in step e-8), the convolution kernel size of the point-by-point convolution layer of the fourth depth-wise separable convolution block is 1×1, the padding is the same, and the step size is 1; in step e-9), the pooling kernel size of the average pooling layer of the fourth downsampling block is 2×2; in step e-10), the convolution kernel size of the point-by-point convolution layer of the fifth depth-wise separable convolution block is 1×1, and the padding is the same.

6. The method for segmenting cardiac ultrasound images based on a boundary-aware multi-scale convolutional network according to claim 1, characterized in that: Step f) comprises the following steps: f-1) The jump connection unit of the segmentation network model is composed of a first multi-level edge feature enhancement module MLEFE, a second multi-level edge feature enhancement module MLEFE, a third multi-level edge feature enhancement module MLEFE, and a multi-scale adaptive channel enhancement module MS-ACE; f-2) The first multi-level edge feature enhancement module MLEFE of the jump connection unit is composed of a first branch and a second branch. The first branch is composed of a first convolutional layer, a BN layer, a ReLU activation function, and a second convolutional layer in sequence. Input into the first branch and output the feature map The second branch consists of a channel attention mechanism, a multi-head attention mechanism, a spatial attention mechanism, and an upsampling layer. Input into the channel attention mechanism of the second branch, and output the feature map The feature map Input into the multi-head attention mechanism of the second branch, and output the feature map The feature map Input into the spatial attention mechanism of the second branch, and output the feature map The feature map Feature Map Feature Map Perform element-by-element multiplication to obtain the feature map The feature map Input to the upsampling layer of the second branch, and output the feature map The feature map feature map Feature Map Feature Map Perform addition operation to obtain feature map f-3) The feature map Replace the feature map in step f-2) Then repeat step f-2) to get the feature map f-4) The feature map With feature map Perform addition operation to obtain feature map f-5) The second multi-level edge feature enhancement module MLEFE of the jump connection unit is composed of a first branch and a second branch. The first branch is composed of a first convolutional layer, a BN layer, a ReLU activation function, and a second convolutional layer in sequence. Input into the first branch and output the feature map The second branch consists of a channel attention mechanism, a multi-head attention mechanism, a spatial attention mechanism, and an upsampling layer. Input into the channel attention mechanism of the second branch, and output the feature map The feature map Input into the multi-head attention mechanism of the second branch, and output the feature map The feature map Input into the spatial attention mechanism of the second branch, and output the feature map The feature map Feature Map Feature Map Perform element-by-element multiplication to obtain the feature map The feature map Input to the upsampling layer of the second branch, and output the feature map The feature map feature map Feature Map Feature Map Perform addition operation to obtain feature map f-6) The feature map With feature map Perform addition operation to obtain feature map f-7) The third multi-level edge feature enhancement module MLEFE of the jump connection unit is composed of a first branch and a second branch. The first branch is composed of a first convolutional layer, a BN layer, a ReLU activation function, and a second convolutional layer in sequence. Input into the first branch and output the feature map The second branch consists of a channel attention mechanism, a multi-head attention mechanism, a spatial attention mechanism, and an upsampling layer. Input into the channel attention mechanism of the second branch, and output the feature map The feature map Input into the multi-head attention mechanism of the second branch, and output the feature map The feature map Input into the spatial attention mechanism of the second branch, and output the feature map The feature map Feature Map Feature Map Perform element-by-element multiplication to obtain the feature map The feature map Input to the upsampling layer of the second branch, and output the feature map The feature map feature map Feature Map Feature Map Perform addition operation to obtain feature map f-8) The feature map With feature map Perform addition operation to obtain feature map f-9) The multi-scale adaptive channel enhancement module MS-ACE of the jump connection unit is composed of the first branch, the second branch, the third branch, the convolution layer, and the ReLU activation function. The first branch is composed of the first convolution unit and the second convolution unit. The first convolution unit of the first branch is composed of the first convolution layer, the second convolution layer, the ReLU activation function, the BN layer, and the GAP layer in sequence. The feature map Input into the first convolution unit of the first branch, and output the feature map The second convolution unit of the first branch is composed of the first convolution layer, BN layer, second convolution layer, and Softmax function in sequence. Input into the second convolution unit of the first branch, and output the feature map The second branch is composed of the first convolution unit and the second convolution unit. The first convolution unit of the second branch is composed of the first convolution layer, the second convolution layer, the ReLU activation function, the BN layer, and the GAP layer in sequence. Input into the first convolution unit of the second branch, and output the feature map The second convolution unit of the second branch is composed of the first convolution layer, the BN layer, the second convolution layer, and the Softmax function in sequence. Input into the second convolution unit of the second branch, and output the feature map The third branch is composed of the first convolution unit and the second convolution unit. The first convolution unit of the third branch is composed of the first convolution layer, the second convolution layer, the ReLU activation function, the BN layer, and the GAP layer in sequence. Input into the first convolution unit of the third branch, and output the feature map The second convolution unit of the third branch is composed of the first convolution layer, the BN layer, the second convolution layer, and the Softmax function in sequence. Input into the second convolution unit of the first branch, and output the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The feature map Feature Map Feature Map After the addition operation, it is sequentially input into the convolution layer and ReLU activation function of the multi-scale adaptive channel enhancement module MS-ACE, and the feature map is output.

7. The method for segmenting cardiac ultrasound images based on a boundary-aware multi-scale convolutional network according to claim 6, characterized in that: In step f-2), the convolution kernel size of the first convolution layer of the first branch of the first multi-level edge feature enhancement module MLEFE is 3×3, padding=1, and stride=1, and the convolution kernel size of the second convolution layer of the first branch of the first multi-level edge feature enhancement module MLEFE is 1×1, padding=0, and stride=0; in step f-5), the convolution kernel size of the first convolution layer of the first branch of the second multi-level edge feature enhancement module MLEFE is 3×3, padding=1, and stride=1, and the convolution kernel size of the second convolution layer of the first branch of the second multi-level edge feature enhancement module MLEFE is 1×1, padding=0, and stride=0; in step f-7), the first convolution kernel size of the first branch of the third multi-level edge feature enhancement module MLEFE The convolution kernel size of the convolution layer is 3×3, padding=1, and stride=1, and the convolution kernel size of the second convolution layer of the first branch of the third multi-level edge feature enhancement module MLEFE is 1×1, padding=0, and stride=0; in step f-9), the convolution kernel size of the first convolution layer of the first convolution unit of the first branch is 3×3, padding=0, and the stride is 0, the convolution kernel size of the second convolution layer of the first convolution unit of the first branch is 3×3, padding=0, and stride=0, the convolution kernel size of the first convolution layer of the second convolution unit of the first branch is 3×3, padding=1, and stride=1, and the convolution kernel size of the second convolution layer of the second convolution unit of the first branch is 1×1, padding=0, and the stride is 0. The convolution kernel size of the first convolution layer of the first convolution unit of the second branch is 3×3, padding=0, and stride=0. The convolution kernel size of the second convolution layer of the first convolution unit of the second branch is 1×1, padding=0, and stride=0. The convolution kernel size of the first convolution layer of the second convolution unit of the second branch is 3×3, padding=1, and stride=1. The convolution kernel size of the second convolution layer of the second convolution unit of the second branch is 1×1, padding=0, and stride=0. The convolution kernel size of the first convolution layer of the first convolution unit of the third branch is 3 ×3, padding=0, and stride=0, the convolution kernel size of the second convolution layer of the first convolution unit of the third branch is 1×1, padding=0, and stride=0, the convolution kernel size of the first convolution layer of the second convolution unit of the third branch is 3×3, padding=1, and stride=1, the convolution kernel size of the second convolution layer of the second convolution unit of the third branch is 1×1, padding=0, and stride=0, and the convolution kernel size of the convolution layer of the multi-scale adaptive channel enhancement module MS-ACE is 1×1, padding=0, and stride=0.

8. According to the method for segmenting cardiac ultrasound images using a boundary-aware multi-scale convolutional network as described in claim 1, step g) comprises the following steps: g-1) The decoder of the segmentation network model is composed of a first decoding layer, a first upsampling layer, a first feature fusion module FFM, a second decoding layer, a second upsampling layer, a third decoding layer, a third upsampling layer, a second feature fusion module FFM, a fourth decoding layer, and a fourth upsampling layer; g-2) The feature map With feature map Input to the first decoding layer of the decoder for concatenation to obtain the feature map g-3) The feature map Input to the first upsampling layer of the decoder, and output the feature map g-4) The first feature fusion module FFM of the decoder consists of the first branch, the second branch, the convolution layer, and the ReLU activation function. The first branch consists of the first small branch, the second small branch, and the ReLU activation function. The second branch consists of the first small branch, the second small branch, the convolution layer, and the ReLU activation function. The first small branch of the first branch consists of the transposed convolution layer, the convolution layer, and the BN layer in sequence. The feature map Input into the first small branch of the first branch, and output the feature map The second small branch of the first branch is composed of convolutional layers and LN layers in sequence. Input into the first small branch of the first branch, and output the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The first small branch of the second branch is composed of a transposed convolution layer, a convolution layer, a LN layer, and a ReLU activation function. Input into the first small branch of the second branch, and output the feature map The second branch of the second branch is composed of convolutional layer, BN layer, and ReLU activation function in sequence. Input into the second small branch of the second branch, and output the feature map The feature map With feature map Perform element-by-element addition to obtain the feature map The feature map Input into the ReLU activation function of the first branch, and output the feature map The feature map Input into the convolution layer and ReLU activation function of the second branch in turn, and output the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The feature map The convolutional layer and ReLU activation function of the first feature fusion module FFM are input in sequence, and the feature map is output. g-5) The feature map With feature map Input to the second decoding layer of the decoder for concatenation to obtain the feature map g-6) The feature map Input to the second upsampling layer of the decoder, and output the feature map g-7) The feature map With feature map Input to the third decoding layer of the decoder for concatenation to obtain the feature map g-8) The feature map Input to the third upsampling layer of the decoder, and output the feature map g-9) The second feature fusion module FFM of the decoder is composed of the first branch, the second branch, the convolution layer, and the ReLU activation function. The first branch is composed of the first small branch, the second small branch, and the ReLU activation function. The second branch is composed of the first small branch, the second small branch, the convolution layer, and the ReLU activation function. The first small branch of the first branch is composed of the transposed convolution layer, the convolution layer, and the BN layer in sequence. The feature map Input into the first small branch of the first branch, and output the feature map The second small branch of the first branch is composed of convolutional layers and LN layers in sequence. Input into the first small branch of the first branch, and output the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The first small branch of the second branch is composed of a transposed convolution layer, a convolution layer, a LN layer, and a ReLU activation function. Input into the first small branch of the second branch, and output the feature map The second branch of the second branch is composed of convolutional layer, BN layer, and ReLU activation function in sequence. Input into the second small branch of the second branch, and output the feature map The feature map With feature map Perform element-by-element addition to obtain the feature map The feature map Input into the ReLU activation function of the first branch, and output the feature map The feature map Input into the convolution layer and ReLU activation function of the second branch in turn, and output the feature map The feature map With feature map Perform element-by-element multiplication to obtain the feature map The feature map The convolutional layer and ReLU activation function of the second feature fusion module FFM are input in sequence, and the feature map is output. g-10) The feature map With feature map Input to the fourth decoding layer of the decoder for splicing operation to obtain the feature map g-11) The feature map Input to the fourth upsampling layer of the decoder, and output the feature map The feature map With feature map Perform the splicing operation to obtain the segmentation result image 9. The method for segmenting cardiac ultrasound images based on a boundary-aware multi-scale convolutional network according to claim 8, characterized in that: In step g-4), the convolution kernel size of the transposed convolution layer of the first small branch of the first branch is 2×2, the padding is the same, and the step size is 1, the convolution kernel size of the convolution layer of the first small branch of the first branch is 3×3, the padding is 0, and the step size is 0, the convolution kernel size of the convolution layer of the second small branch of the first branch is 3×3, the padding is 0, and the step size is 0, the convolution kernel size of the transposed convolution layer of the first small branch of the second branch is 2×2, the padding is the same, and the step size is 1, the convolution kernel size of the convolution layer of the first small branch of the second branch is 3×3, the padding is 0, and the step size is 0, and the convolution kernel size of the convolution layer of the second small branch of the second branch is 3×3, the padding is 0, and the step size is 0; In step g-9), the convolution kernel size of the transposed convolution layer of the first small branch of the first branch is 2×2, the padding is the same, and the step size is 1, the convolution kernel size of the convolution layer of the first small branch of the first branch is 3×3, the padding is 0, and the step size is 0, the convolution kernel size of the convolution layer of the second small branch of the first branch is 3×3, the padding is 0, and the step size is 0, the convolution kernel size of the transposed convolution layer of the first small branch of the second branch is 2×2, the padding is the same, and the step size is 1, the convolution kernel size of the convolution layer of the first small branch of the second branch is 3×3, the padding is 0, and the step size is 0, and the convolution kernel size of the convolution layer of the second small branch of the second branch is 3×3, the padding is 0, and the step size is 0.

10. The method for segmenting cardiac ultrasound images based on a boundary-aware multi-scale convolutional network according to claim 1, characterized in that: Step h) comprises the following steps: h-1) through the formula L loss =W1L Dice +W2L BCE Calculate the total loss L loss , where L Dice is the Dice loss, L BCE is the binary cross entropy loss, W1 and W2 are weight values; h-2) Use the Adam optimizer to use the total loss L loss Train the segmentation network model to obtain an optimized segmentation network model.

Images