A renal fascia CT image segmentation method and system based on improved UNet network

Abstract

The present invention relates to the field of computer-aided medical diagnosis and discloses a method and system for renal fascia CT image segmentation based on an improved UNet network, comprising the following specific steps: constructing and preprocessing original renal fascia CT images to obtain a training set; in the encoder of the UNet network, replacing all convolutional layers of the Unet network encoder with an improved GA module consisting of a cascaded double-layer dynamic snake convolution and a layer of CBAM attention module to construct an improved UNet network; inputting the training set into the improved UNet network for training to obtain an optimal model; inputting the image to be predicted into the optimal model to obtain a predicted renal fascia segmentation result. The present invention solves the problem that the renal fascia is difficult to extract features due to its fuzzy boundaries and is easily affected by its surrounding tissues, and has the characteristics of being able to achieve accurate and efficient image segmentation.

Description

Technical Field

[0001] The present invention relates to the field of computer-aided medical diagnosis, and more specifically, to a renal fascia CT image segmentation method and system based on an improved UNet network. Background Art

[0002] The renal fascia is a vital structure surrounding the kidneys, and its health is closely linked to the kidneys and surrounding tissues. In the early stages of renal cancer, the renal fascia prevents tumor cell spread. However, as the disease progresses, the tumor may break through the fascia, leading to more widespread invasion. Abdominal CT scans allow doctors to visualize the relationship between the renal fascia and the renal tumor, assessing the severity of kidney disease and determining the need for surgery. However, the renal fascia appears as a fine, soft tissue density line on abdominal CT images and is easily obscured by surrounding tissue, making accurate segmentation difficult even for experienced radiologists. Using deep learning technology to effectively segment the renal fascia in abdominal CT images would aid clinical diagnosis and treatment, and provide new insights and tools for kidney disease research.

[0003] With technological advancements, deep learning technology has rapidly developed in the field of medical image segmentation. Various models, such as CNN, UNet, and FCN, have been widely used for the detection and segmentation of tumors, organs, and other diseases. However, currently available methods have limited research on renal fascia segmentation using deep learning. To address this research gap, the need for deep learning for renal fascia segmentation has become increasingly prominent.

[0004] The existing technology has a method for segmenting liver and tumor images. Through an image segmentation network based on a multi-scale feature attention mechanism, it simultaneously focuses on global spatial information, local spatial information, and inter-channel information. It can effectively utilize the limited information of the feature image, and solve the problem that the Unet network cannot effectively utilize the information of a small medical image dataset to complete accurate segmentation.

[0005] However, the existing technology cannot solve the problem that the renal fascia is easily affected by its surrounding tissues due to its fuzzy boundaries, making its features difficult to extract. Therefore, how to invent a renal fascia CT image segmentation method based on an improved UNet network is a technical problem that urgently needs to be solved in this technical field. Summary of the Invention

[0006] In order to solve the problem that the renal fascia is easily affected by its surrounding tissues due to its fuzzy boundaries, making it difficult to extract its features, the present invention provides a renal fascia CT image segmentation method and system based on an improved UNet network, which has the characteristics of being able to achieve accurate and efficient image segmentation.

[0007] In order to achieve the above-mentioned purpose of the present invention, the technical solutions adopted are as follows:

[0008] A renal fascia CT image segmentation method based on an improved UNet network includes the following specific steps:

[0009] Construct and preprocess the original renal fascia CT images to obtain the training set;

[0010] In the encoder of the UNet network, an improved GA module consisting of a cascaded two-layer dynamic snake convolution and a layer of CBAM attention module is used to replace all convolutional layers of the encoder of the UNet network to construct an improved UNet network;

[0011] The training set is input into the improved UNet network for training to obtain the best model;

[0012] The image to be predicted is input into the optimal model to obtain the predicted renal fascia segmentation result.

[0013] Preferably, the original renal fascia CT image is constructed and preprocessed, specifically in the following steps:

[0014] Obtain original renal fascia CT images;

[0015] Adjust the window width and window level of the original renal fascia CT image;

[0016] The renal fascia CT images with adjusted window width and window position were cropped;

[0017] The cropped renal fascia CT images are annotated to obtain a renal fascia dataset.

[0018] The renal fascia dataset was augmented by random rotation, random horizontal flipping, random cropping, and translation operations;

[0019] The data-enhanced renal fascia image dataset is divided into training set and test set according to the set ratio.

[0020] Furthermore, the CBAM attention module consists of a cascaded channel attention module and a spatial attention module.

[0021] Furthermore, the specific process of the GA module processing the input image is as follows:

[0022] In each GA module of the encoder, the first layer of the double-layer dynamic snake convolution is adaptively convolved according to the input image. The local features of the dynamic adjustment of its sliding direction and step size to capture the local features, output feature map :

[0023] =DSC( ,k)

[0024] Where DSC is the dynamic snake convolution operation, k is the convolution kernel size;

[0025] Will Input the second layer of dynamic snake convolution of the double-layer dynamic snake convolution to further refine the feature extraction and output the feature map :

[0026] =DSC( ,k)

[0027] Will Input to a layer of CBAM attention module, Perform channel attention mechanism processing and obtain channel descriptors through global average pooling and maximum pooling :

[0028] )

[0029] Among them, GlobalAvgPool is the global average pooling, and GlobalMaxPool is the global maximum pooling;

[0030] Descriptor Generate channel attention weights through the fully connected layer :

[0031]

[0032] in, and is the learned weight matrix, ReLU is the activation function, σ is the sigmoid activation function;

[0033] The weight Applicable to In the above example, we generate enhanced feature maps. ;

[0034] = ⊙

[0035] Where ⊙ represents element-wise multiplication;

[0036] right Perform spatial attention mechanism processing and calculate spatial features S through channel-dimensional pooling operation:

[0037] S=Concat(MaxPool( ),AvgPool( ))

[0038] Among them, Concat is a splicing operation;

[0039] Convolution is performed on the spatial feature S to generate spatial attention weights :

[0040] =σ(Conv(S, ))

[0041] in It is the convolution kernel in the spatial attention mechanism;

[0042] The weight Apply to In the example, the output feature map of the GA module is generated. :

[0043] = ⊙ .

[0044] Furthermore, when constructing the improved UNet network, an improved EE module is used to replace all decoding convolutional layers in the decoder of the UNet network; the improved EE module consists of a cascaded double-layer convolution, an SE attention module and an EMA attention module.

[0045] Furthermore, the specific process of the EE module processing the input image is as follows:

[0046] In each EE module of the decoder, the input image After the double-layer convolution operation, the output feature map :

[0047] =DoubleConv( , k3);

[0048] Among them, DoubleConv is a double-layer convolution operation, and k3 is a 3×3 convolution kernel;

[0049] Will Input into SE attention module, generate channel descriptor by global average pooling :

[0050]

[0051] Will Input fully connected layer to generate channel attention weights :

[0052]

[0053] in and is the learned weight matrix;

[0054] The weight Applicable to In the above example, we generate enhanced feature maps. :

[0055] = ⊙

[0056] Will Input EMA attention module; in the EMA attention module, the channel mean and variance of the feature map are calculated to generate a moving average channel descriptor μ and σ, perform normalization operations to generate updated channel weights :

[0057] =normalize( )

[0058] Where normalize is the normalization operation;

[0059] right Perform weighting to generate feature maps :

[0060] = ⊙

[0061] Will and Perform feature fusion to generate the output feature map of the EE module .

[0062] Furthermore, when constructing the improved UNet network, a residual module is introduced at the connection between the UNet network encoder and decoder; the residual module consists of two cascaded 3×3 convolutional layers and one 5×5 convolutional layer.

[0063] Furthermore, the training set is input into the improved UNet network for training. Specifically, the initial learning rate, training epoch, and sample batch are set. The training set is input into the improved UNet network and the model is trained using the RMSprop optimizer. During training, Dice Loss is selected as the loss function, and the learning rate of each model parameter is adaptively adjusted according to the size of the gradient.

[0064] Furthermore, after obtaining the optimal model, the model is evaluated using indicators such as the Dice coefficient. The specific steps are as follows:

[0065] The test set is input into the best model for training, the renal fascia image feature information is extracted, the renal fascia in the abdominal CT image is predicted, and the test segmentation result image is obtained;

[0066] The Dice coefficient, recall rate, precision rate, and accuracy rate are used as evaluation indicators of the model to evaluate the training results. If the evaluation result does not meet the set threshold, the training set is re-input into the improved UNet network for training.

[0067] A renal fascia CT image segmentation system based on an improved UNet network includes an image processing module, a UNet network construction module, a model training module, and an image segmentation module;

[0068] The image processing module is used to construct and preprocess the original renal fascia CT image to obtain a training set;

[0069] The UNet network construction module is used to replace all convolutional layers of the encoder of the UNet network with an improved GA module consisting of a cascaded two-layer dynamic snake convolution and a layer of CBAM attention module in the encoder of the UNet network to construct an improved UNet network;

[0070] The model training module is used to input the training set into the improved UNet network for training to obtain the best model;

[0071] The image segmentation module is used to input the image to be predicted into the optimal model to obtain the predicted renal fascia segmentation result.

[0072] The beneficial effects of the present invention are as follows:

[0073] The present invention improves the UNet network by replacing the encoding convolutional layer with an improved GA module consisting of a two-layer dynamic snake convolution and a layer of CBAM attention module. The improved UNet network can fully understand the characteristics of the renal fascia in abdominal CT images, while effectively suppressing irrelevant or redundant information, significantly improving the accuracy of renal fascia segmentation. The present invention provides a new and efficient method for automatic segmentation of the renal fascia, filling the gap in the field of renal fascia segmentation using deep learning technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 This is a schematic diagram of the specific process of the renal fascia CT image segmentation method based on the improved UNet network.

[0075] Figure 2 Schematic diagram of samples and their labels used for training and testing of the renal fascia dataset in Example 2.

[0076] Figure 3This is the overall structure diagram of the improved UNet network model in the present invention.

[0077] Figure 4 This is a structural diagram of the improved GA module in the present invention.

[0078] Figure 5 This is a structural diagram of the improved EE module in the present invention.

[0079] Figure 6 This is a structural diagram of the residual module in the present invention. DETAILED DESCRIPTION

[0080] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0081] Example 1

[0082] like Figure 1 As shown in FIG, a renal fascia CT image segmentation method based on an improved UNet network includes the following specific steps:

[0083] Construct and preprocess the original renal fascia CT images to obtain the training set;

[0084] In the encoder of the UNet network, an improved GA module consisting of a cascaded two-layer dynamic snake convolution and a layer of CBAM attention module is used to replace all convolutional layers of the encoder of the UNet network to construct an improved UNet network;

[0085] The training set is input into the improved UNet network for training to obtain the best model;

[0086] The image to be predicted is input into the optimal model to obtain the predicted renal fascia segmentation result.

[0087] Example 2

[0088] More specifically, in a specific embodiment, the renal fascia image dataset is constructed by the following steps:

[0089] Obtain original renal fascia CT images;

[0090] The window width of the original renal fascia CT image was adjusted to 300, and the window level was adjusted to -50 to observe the renal fascia more clearly;

[0091] The renal fascia CT image with adjusted window width and window position was cropped with a cropping resolution of [700,550];

[0092] The renal fascia was labeled using labelme software to obtain the renal fascia dataset.

[0093] In this embodiment, the samples and labels used for training in the renal fascia dataset are as follows: Figure 2As shown in the two left pictures in the figure, the samples and labels used for testing in the renal fascia dataset are as follows Figure 2 As shown in the two pictures on the right.

[0094] In a specific embodiment, the original renal fascia CT images are preprocessed to obtain a training set and a test set. The specific steps are as follows:

[0095] The renal fascia dataset was augmented by random rotation, random horizontal flipping, random cropping, and translation operations;

[0096] The data-enhanced renal fascia image dataset is divided into training set and test set in a ratio of 4:1.

[0097] In a specific embodiment, the CBAM attention module consists of a cascaded channel attention module and a spatial attention module.

[0098] The specific process of GA module processing input image is as follows:

[0099] In each GA module of the encoder, the first layer of the double-layer dynamic snake convolution is adaptively convolved according to the input image. The local features of the dynamic adjustment of its sliding direction and step size to capture the local features, output feature map :

[0100] =DSC( ,k)

[0101] Where DSC is the dynamic snake convolution operation, k is the convolution kernel size;

[0102] Will Input the second layer of dynamic snake convolution of the double-layer dynamic snake convolution to further refine the feature extraction and output the feature map :

[0103] =DSC( ,k)

[0104] Will Input to a layer of CBAM attention module, Perform channel attention mechanism processing and obtain channel descriptors through global average pooling and maximum pooling :

[0105] )

[0106] Among them, GlobalAvgPool is the global average pooling, and GlobalMaxPool is the global maximum pooling;

[0107] Descriptor Generate channel attention weights through the fully connected layer :

[0108]

[0109] in, and is the learned weight matrix, ReLU is the activation function, σ is the sigmoid activation function;

[0110] The weight Applicable to In the above example, we generate enhanced feature maps. ;

[0111] = ⊙

[0112] Where ⊙ represents element-wise multiplication;

[0113] right Perform spatial attention mechanism processing and calculate spatial features S through channel-dimensional pooling operation:

[0114] S=Concat(MaxPool( ),AvgPool( ))

[0115] Among them, Concat is a splicing operation;

[0116] Convolution is performed on the spatial feature S to generate spatial attention weights :

[0117] =σ(Conv(S, ))

[0118] in It is the convolution kernel in the spatial attention mechanism;

[0119] The weight Apply to In the example, the output feature map of the GA module is generated. :

[0120] = ⊙ .

[0121] In a specific embodiment, when constructing the improved UNet network, an improved EE module is used to replace all decoding convolutional layers in the decoder of the UNet network; wherein the improved EE module is composed of a cascaded double-layer convolution, an SE attention module and an EMA attention module.

[0122] In a specific embodiment, the specific process of the EE module processing the input image is as follows:

[0123] In each EE module of the decoder, the input image After the double-layer convolution operation, the output feature map :

[0124] =DoubleConv( , k3);

[0125] Among them, DoubleConv is a double-layer convolution operation, and k3 is a 3×3 convolution kernel;

[0126] Will Input into SE attention module, generate channel descriptor by global average pooling :

[0127]

[0128] Will Input fully connected layer to generate channel attention weights :

[0129]

[0130] in and is the learned weight matrix;

[0131] The weight Applicable to In the above example, we generate enhanced feature maps. :

[0132] = ⊙

[0133] Will Input EMA attention module; in the EMA attention module, the channel mean and variance of the feature map are calculated to generate a moving average channel descriptor μ and σ, perform normalization operations to generate updated channel weights :

[0134] =normalize( )

[0135] Where normalize is the normalization operation;

[0136] right Perform weighting to generate feature maps :

[0137] = ⊙

[0138] Will and Perform feature fusion to generate the output feature map of the EE module .

[0139] In a specific embodiment, when constructing the improved UNet network, a residual module is introduced at the connection between the UNet network encoder and decoder; the residual module consists of two cascaded 3×3 convolutional layers and one 5×5 convolutional layer.

[0140] In this embodiment, the structure of the improved UNet network is as follows Figure 3 and Figure 4 As shown in the figure, by replacing all the convolution layers of the encoder of the Unet network with an improved GA module consisting of a cascaded double-layer dynamic snake convolution and a layer of CBAM attention module, the characteristics of the renal fascia are emphasized while suppressing irrelevant or redundant information; the input image is first subjected to a double-layer dynamic snake convolution for feature extraction, and then the features are input into the CBAM attention module; compared with the standard convolution, the dynamic snake convolution better captures the local correlation and spatial structure of the renal fascia by introducing a more flexible convolution structure; the CBAM attention module consists of two key modules: channel attention and spatial attention. The channel attention module can evaluate the importance of each channel of the renal fascia feature map and adjust the weight of each channel accordingly; the spatial attention module focuses on the spatial position information of the renal fascia feature map, and emphasizes or suppresses features at different positions by generating an enhanced feature map;

[0141] In the decoder of the UNet network, the improved EE module is used as Figure 5 The shown method replaces the decoding convolution layer, improves the model's attention to the features of the renal fascia, enhances detail retention, and improves the quality of renal fascia segmentation; the input of the EE module is feature fused with the output after double-layer convolution, SE attention module and EMA attention module; the SE attention module can help the model better retain the edges and details of the renal fascia, and enhance the attention to the features of the renal fascia by dynamically adjusting the importance of each channel; the EMA attention module helps the model reduce its sensitivity to instantaneous noise in CT images and improve the segmentation robustness of the renal fascia in complex backgrounds.

[0142] like Figure 6 As shown in the figure, a residual module is introduced at the junction of the UNet network encoder and decoder, effectively improving the accuracy and robustness of renal fascia segmentation. The input of the residual module is combined with the output of two 3×3 convolutional layers and one 5×5 convolutional layer for feature fusion. The 3x3 convolution can capture detailed information of the renal fascia, while the 5x5 convolution can capture a wider range of contextual information, thereby enhancing the expressive power of the features.

[0143] In this embodiment, the specific process of the residual module processing the input image is as follows:

[0144] In the residual module, the input image After two layers of 3×3 convolutional layers, the output feature map :

[0145] =DoubleConv( ,k3)

[0146] Among them, DoubleConv is a double-layer convolution operation, and k3 is a 3×3 convolution kernel;

[0147] Will Input a 5×5 convolutional layer

[0148] =Conv( , k5)

[0149] Among them, Conv is the convolution operation, k5 is the 5×5 convolution kernel;

[0150] Will and Perform feature fusion to generate the output feature map of the residual module .

[0151] In a specific embodiment, the training set is input into the improved UNet network for training, specifically: the initial learning rate is set to 0.001, the training epoch is set to 40, that is, the model learns all training samples 40 times, the sample batch is 4, that is, 4 pictures are input into the model at a time; the training set is input into the improved UNet network, and the RMSprop optimizer is used for model training. During training, Dice Loss is selected as the loss function to process unbalanced data, and the learning rate of each model parameter is adaptively adjusted according to the size of the gradient.

[0152] In a specific embodiment, the image to be predicted is input into the optimal model to obtain the predicted renal fascia segmentation result. The specific steps are as follows:

[0153] The image to be predicted is input into the encoder of the best model; the feature maps generated by the GA modules of the 1st to 3rd layers are input into the GA modules of the next layer as input images, and the feature maps generated by the GA modules of the 4th layer are input into the residual module;

[0154] The fourth layer GA module inputs the feature map it generates into the residual module, which first passes through two layers of 3×3 convolutional layers;

[0155] Then pass through a 5×5 convolution layer; finally, the obtained image is fused with the input image to obtain and output the fused feature map;

[0156] Input the fused feature map into the decoder of the best model;

[0157] The fused feature map is input into the EE module of the fourth layer; the feature map generated by the EE module of the 4th to 2nd layer is input into the EE module of the previous layer; each EE module is jump-connected to its corresponding GA module; the feature map output by the EE module of the first layer is output as the predicted renal fascia segmentation result map.

[0158] In a specific embodiment, after obtaining the optimal model, the model is evaluated using indicators such as the Dice coefficient. The specific steps are as follows:

[0159] The test set is input into the best model for training, the renal fascia image feature information is extracted, the renal fascia in the abdominal CT image is predicted, and the test segmentation result image is obtained;

[0160] The Dice coefficient, recall rate, precision rate, and accuracy rate are used as evaluation indicators of the model to evaluate the training results. If the evaluation result does not meet the set threshold, the training set is re-input into the improved UNet network for training.

[0161] In this embodiment, the Dice coefficient is specifically a set similarity measurement function used to measure the similarity between the segmentation result and the true renal fascia label. The expression of the Dice coefficient is:

[0162]

[0163] The recall rate is specifically the ratio of correctly predicted positive samples to actual positive samples, that is, the ratio of the number of pixels correctly predicted to be renal fascia to the number of pixels actually being renal fascia. The expression of the recall rate is:

[0164]

[0165] The precision rate is specifically the ratio of correctly predicted positive samples to all predicted positive samples, that is, the ratio of the number of pixels correctly predicted as renal fascia to the number of pixels predicted as renal fascia. The expression of the precision rate is:

[0166]

[0167] The accuracy is specifically the ratio of correctly predicted samples to the total number of samples, that is, the ratio of the number of pixels correctly predicted as renal fascia to the number of pixels correctly predicted as non-renal fascia to the total number of pixels. The accuracy expression is as follows:

[0168]

[0169] Among them, TP represents the number of pixels correctly predicted as renal fascia, FP represents the number of pixels incorrectly predicted as renal fascia, TN represents the number of pixels correctly predicted as non-renal fascia, and FN represents the number of pixels incorrectly predicted as non-renal fascia.

[0170] Example 3

[0171] A renal fascia CT image segmentation system based on an improved UNet network includes an image processing module, a UNet network construction module, a model training module, and an image segmentation module;

[0172] The image acquisition module is used to construct a renal fascia image dataset;

[0173] The image processing module is used to construct and preprocess the original renal fascia CT image to obtain a training set;

[0174] The UNet network construction module is used to replace all convolutional layers of the encoder of the UNet network with an improved GA module consisting of a cascaded two-layer dynamic snake convolution and a layer of CBAM attention module in the encoder of the UNet network to construct an improved UNet network;

[0175] The model training module is used to input the training set into the improved UNet network for training to obtain the best model;

[0176] The image segmentation module is used to input the image to be predicted into the optimal model to obtain the predicted renal fascia segmentation result.

[0177] Obviously, the above embodiments of the present invention are merely examples for the purpose of illustrating the present invention, and are not intended to limit the embodiments of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A renal fascia CT image segmentation method based on an improved UNet network, characterized by: The specific steps include: Construct and preprocess the original renal fascia CT images to obtain the training set; In the encoder of the UNet network, an improved GA module consisting of a cascaded double-layer dynamic snake convolution and a layer of CBAM attention module is used to replace all convolutional layers of the UNet network encoder to construct an improved UNet network. When constructing the improved UNet network, an improved EE module is also used to replace all decoding convolutional layers in the decoder of the UNet network. The improved EE module consists of a cascaded double-layer convolution, an SE attention module, and an EMA attention module. The training set is input into the improved UNet network for training to obtain the best model; The image to be predicted is input into the optimal model to obtain the predicted renal fascia segmentation result.

2. The renal fascia CT image segmentation method based on the improved UNet network according to claim 1, characterized in that: Construct and preprocess the original renal fascia CT image. The specific steps are as follows: Obtain original renal fascia CT images; Adjust the window width and window level of the original renal fascia CT image; The renal fascia CT images with adjusted window width and window position were cropped; The cropped renal fascia CT images are annotated to obtain a renal fascia dataset. The renal fascia dataset was augmented by random rotation, random horizontal flipping, random cropping, and translation operations; The data-enhanced renal fascia image dataset is divided into training set and test set according to the set ratio.

3. The renal fascia CT image segmentation method based on the improved UNet network according to claim 1, characterized in that: The CBAM attention module consists of a cascaded channel attention module and a spatial attention module.

4. The renal fascia CT image segmentation method based on the improved UNet network according to claim 3, characterized in that: The specific process of GA module processing input image is as follows: In each GA module of the encoder, the first layer of the double-layer dynamic snake convolution is adaptively convolved according to the input image. The local features of the dynamic adjustment of its sliding direction and step size to capture the local features, output feature map : =DSC( ,k) Where DSC is the dynamic snake convolution operation, k is the convolution kernel size; Will Input the second layer of dynamic snake convolution of the double-layer dynamic snake convolution to further refine the feature extraction and output the feature map : =DSC( ,k) Will Input to a layer of CBAM attention module, Perform channel attention mechanism processing and obtain channel descriptors through global average pooling and maximum pooling : ) Among them, GlobalAvgPool is the global average pooling, and GlobalMaxPool is the global maximum pooling; Descriptor Generate channel attention weights through the fully connected layer : in, and is the learned weight matrix, ReLU is the activation function, σ is the sigmoid activation function; The weight Applicable to In the above example, we generate enhanced feature maps. ; = ⊙ Where ⊙ represents element-wise multiplication; right Perform spatial attention mechanism processing and calculate spatial features S through channel-dimensional pooling operation: S=Concat(MaxPool( ),AvgPool( )) Among them, Concat is a splicing operation; Convolution is performed on the spatial feature S to generate spatial attention weights : =σ(Conv(S, )) in It is the convolution kernel in the spatial attention mechanism; The weight Apply to In the example, the output feature map of the GA module is generated. : = ⊙ 。 5. The renal fascia CT image segmentation method based on the improved UNet network according to claim 4, characterized in that: The specific process of the EE module processing the input image is as follows: In each EE module of the decoder, the input image After the double-layer convolution operation, the output feature map : =DoubleConv( ,k3); Among them, DoubleConv is a double-layer convolution operation, and k3 is a 3×3 convolution kernel; Will Input into SE attention module, generate channel descriptor by global average pooling : Will Input fully connected layer to generate channel attention weights : in and is the learned weight matrix; The weight Applicable to In the above example, we generate enhanced feature maps. : = ⊙ Will Input EMA attention module; in the EMA attention module, the channel mean and variance of the feature map are calculated to generate a moving average channel descriptor μ and σ, perform normalization operations to generate updated channel weights : =normalize( ) Where normalize is the normalization operation; right Perform weighting to generate feature maps : = ⊙ Will and Perform feature fusion to generate the output feature map of the EE module .

6. The renal fascia CT image segmentation method based on the improved UNet network according to claim 1, characterized in that: When constructing the improved UNet network, a residual module is also introduced at the connection between the UNet network encoder and decoder; the residual module consists of two cascaded 3×3 convolutional layers and one 5×5 convolutional layer.

7. The renal fascia CT image segmentation method based on the improved UNet network according to claim 5, characterized in that: The training set is input into the improved UNet network for training. Specifically, the initial learning rate, training epoch, and sample batch are set. The training set is input into the improved UNet network and the model is trained using the RMSprop optimizer. During training, DiceLoss is selected as the loss function, and the learning rate of each model parameter is adaptively adjusted according to the size of the gradient.

8. The renal fascia CT image segmentation method based on the improved UNet network according to claim 4, characterized in that: After obtaining the best model, the Dice coefficient indicator is used to evaluate the model. The specific steps are as follows: The test set is input into the best model for training, the renal fascia image feature information is extracted, the renal fascia in the abdominal CT image is predicted, and the test segmentation result image is obtained; The Dice coefficient, recall rate, precision rate, and accuracy rate are used as evaluation indicators of the model to evaluate the training results. If the evaluation result does not meet the set threshold, the training set is re-input into the improved UNet network for training.

9. A renal fascia CT image segmentation system based on an improved UNet network, characterized by: Including image processing module, UNet network construction module, model training module, and image segmentation module; The image processing module is used to construct and preprocess the original renal fascia CT image to obtain a training set; The UNet network construction module is used to replace all convolutional layers of the UNet network encoder with an improved GA module consisting of a cascaded double-layer dynamic snake convolution and a layer of CBAM attention module in the UNet network encoder to construct an improved UNet network; when constructing the improved UNet network, an improved EE module is also used to replace all decoding convolutional layers in the UNet network decoder; wherein the improved EE module consists of a cascaded double-layer convolution, an SE attention module and an EMA attention module; The model training module is used to input the training set into the improved UNet network for training to obtain the best model; The image segmentation module is used to input the image to be predicted into the optimal model to obtain the predicted renal fascia segmentation result.

Images