CN117078705B - A CT image segmentation method based on Bhattacharyya coefficient active contour attention - Google Patents

Abstract

The invention relates to the technical field of medical image segmentation, and in particular to a CT image segmentation method based on Babbitt coefficient active contour attention. The steps are as follows: the data set is CT images of 3D kidneys and renal tumors; the nibabel library is called to process the volumes in the data set. The data is sliced in 2D png format, and 10 slices are selected from each volume data to obtain the data set D'. The training set in D' is data enhanced to obtain the data set D; the network structure includes an encoding part and a decoding part; Use the Dice loss function to calculate the loss function; use the SGD optimizer to adjust the weights and biases in the network through backpropagation; save the optimal weights and biases in a newly created file; read the images in the test set Complete the segmentation and save the result as a jpg format file. This invention can focus on the target structure and can also obtain better segmentation results for pictures with uneven gray value distribution.

Description

A CT image segmentation method based on Bhattacharyya coefficient active contour attention

Technical Field

The present invention relates to the technical field of medical image segmentation, and in particular to a CT image segmentation method based on Bhattacharyya coefficient active contour attention.

Background Art

CT image segmentation is an important task in the field of medical imaging. It aims to segment the target structure in the image so that doctors can diagnose the disease and formulate treatment plans. CT image segmentation methods can be divided into traditional segmentation methods and deep learning segmentation methods. Traditional segmentation methods include classic methods based on image processing technology such as graph-based segmentation, threshold segmentation, and region growing. Their effects may be affected by image quality, background noise, etc. Deep learning segmentation methods include U-Net and its variants, DeepLab series and other methods. These deep learning methods have achieved good performance in medical image segmentation with the help of large-scale data and powerful computing power. However, for CT images, due to the influence of different tissue density and different image acquisition parameters, the grayscale is uneven, which greatly increases the difficulty of segmentation.

Therefore, in order to solve the above problems, a CT image segmentation method based on Bhattacharyya coefficient active contour attention is proposed to solve the above problems.

Summary of the invention

In view of the shortcomings of the prior art, the present invention develops a CT image segmentation method based on Bhattacharyya coefficient active contour attention, which can focus on the target structure and obtain better segmentation results for images with uneven grayscale value distribution.

The technical solution of the present invention to solve the technical problem is:

A CT image segmentation method based on Bhattacharyya coefficient active contour attention, the steps are as follows:

S1. The dataset is 3D CT images of kidneys and kidney tumors from MICCAI KiTS19. The training set of the dataset includes 210 3D-CT images, and the test set includes 90 3D-CT images.

S2. Call the nibabel library to process the volume data in the dataset to obtain 2D .png format slices, select 10 slices from each volume data to obtain dataset D’, perform data enhancement on the training set in dataset D’ to obtain dataset D;

S3. The network structure includes an encoding part and a decoding part. The encoding part uses the VGG16 pre-trained network to extract features from the input image, and the decoding part uses the Bhattacharyya coefficient active contour attention structure and the attention depth separable convolution block;

S4. Calculate the loss function Loss: The loss function Loss uses the Dice loss function;

S5. Use the SGD optimizer to adjust the weights and biases in the network through backpropagation;

S6. During the training process of the data set D, set the variable for saving the optimal evaluation index to store the optimal weight and bias, and save it in the ph file;

S7. Load the optimal weights and biases saved in the ph file into the network, read the images in the test set to complete the segmentation, and save the results as a jpg format file.

Based on the above-mentioned CT image segmentation method based on Bhattacharyya coefficient active contour attention, the first five stages of the pre-trained VGG16 network are used, wherein the fifth stage does not include the maximum pooling operation, and each stage obtains feature images A1 _, A2 _, A3 _, A4 _{, and A5} respectively.

Based on the above-mentioned CT image segmentation method based on Bhattacharyya coefficient active contour attention, the decoding part includes a first Bhattacharyya coefficient active contour attention module, a first attention depth-separable convolution block, a second Bhattacharyya coefficient active contour attention module, a second attention depth-separable convolution block, a third Bhattacharyya coefficient active contour attention module, a third attention depth-separable convolution block, a fourth Bhattacharyya coefficient active contour attention module, a fourth attention depth-separable convolution block and an output module.

Based on the above-mentioned CT image segmentation method based on Bhattacharyya coefficient active contour attention, the first Bhattacharyya coefficient active contour attention module includes a first initial contour block, a first image processing block, a first Bhattacharyya coefficient active contour block and a first attention block;

(1) The implementation steps of the first initial contour block are as follows: feature map A ₅ is upsampled to obtain feature map M ₁ , feature map M ₁ is subjected to 1×1 convolution to obtain feature map B ₁ , feature map A ₄ is subjected to 1×1 convolution to obtain feature map C ₁ , feature map B ₁ and feature map C ₁ are added to obtain feature map D ₁ , feature map D ₁ is subjected to Relu activation function, 1×1 convolution, and Sigmoid activation function to obtain feature map N ₁ , feature map N ₁ is subjected to distance transformation to obtain feature map E ₁ , i.e., the initial contour;

The distance transformation is given by Find, among which is the Euclidean distance, * is the convolution multiplication, exp is the exponential function with the natural constant e as the base, is artificially given, and ;

(2) The first image processing block is implemented as follows: Feature map After 1×1 convolution, the feature map F ₁ is obtained. The input image is processed by Resize image processing function and Sigmoid activation function to obtain feature map G ₁ . Feature map F ₁ and feature map G ₁ are added to obtain feature map H ₁ . Feature map H ₁ is processed by 1×1 convolution and Sigmoid activation function to obtain feature map I ₁ .

(3) The first Bhattacharyya coefficient active contour block is implemented as follows: Use the Bhattacharyya coefficient active contour algorithm to perform a finite number of iterative segmentations on the feature map I ₁ to obtain the feature map , , where u and v are positive parameters, which are artificially given; BCV is the Bhattacharyya coefficient active contour algorithm; E ₁ is the feature map E ₁ ;

The energy function of the Bhattacharyya coefficient active contour algorithm is:

,in The following formula represents the length of the contour, The following formula represents the area within the contour. represents the Dirac function, H is the Heavyside function, is the domain, Q is the RGB color space, in and out are the initial contours For a given color q, the probability distribution function is estimated using a histogram based on a Gaussian kernel and ;

(4) The implementation steps of the first attention block are as follows: feature map A ₄ and feature map J ₁ are multiplied to obtain feature map K ₁ .

Based on the above-mentioned CT image segmentation method based on Bhattacharyya coefficient active contour attention, the first attention depth separable convolution block is implemented as follows: the feature map and feature map After splicing, the splicing results are sequentially subjected to point-by-point convolution, batch normalization, 1×7 axial depth convolution, 7×1 axial depth convolution, residual convolution, point-by-point convolution and GELU activation function to obtain the feature map. , feature map After global average pooling, 1×1 convolution, Relu activation function, 1×1 convolution, Sigmoid activation function, and feature map Multiply to get the feature map .

Based on the above-mentioned CT image segmentation method based on Bhattacharyya coefficient active contour attention, the second Bhattacharyya coefficient active contour attention module includes a second initial contour block, a second image processing block, a second Bhattacharyya coefficient active contour block, and a second attention block;

(1) The implementation steps of the second initial contour block are as follows: feature map _P1 is upsampled to obtain feature map _M2 , feature map _M2 is subjected to 1×1 convolution to obtain feature map _B2 , feature map _A3 is subjected to 1×1 convolution to obtain feature map _C2 , feature map _B2 and feature map _C2 are added to obtain feature map _D2 , feature map _D2 is subjected to Relu activation function, 1×1 convolution, and Sigmoid activation function to obtain feature map _N2 , feature map _N2 is subjected to distance transformation to obtain feature map _E2 , i.e., the initial contour;

The distance transformation is given by Find, among which is the Euclidean distance, * is the convolution multiplication, exp is the exponential function with the natural constant e as the base, is artificially given, and ;

(2) The implementation steps of the second image processing block are as follows: the feature map M ₂ is subjected to a 1×1 convolution to obtain the feature map F ₂ , the input image is subjected to the Resize image processing function and the Sigmoid activation function to obtain the feature map G ₂ , the feature map F ₂ and the feature map G ₂ are added to obtain the feature map H ₂ , the feature map H ₂ is subjected to a 1×1 convolution and the Sigmoid activation function to obtain the feature map I ₂ ;

(3) The second Bhattacharyya coefficient active contour block is implemented as follows: Use the Bhattacharyya coefficient active contour algorithm to perform a finite number of iterative segmentations on the feature map I ₂ to obtain the feature map , ,in and is a positive parameter, which is artificially given; BCV is the Bhattacharyya coefficient active contour algorithm; E ₂ is the feature map E ₂ ;

The energy function of the Bhattacharyya coefficient active contour algorithm is:

,in The following formula represents the length of the contour, The following formula represents the area within the contour. represents the Dirac function, H is the Heavyside function, is the domain, Q is the RGB color space, in and out are the initial contours For the internal and external regions, for a given color q, the probability distribution function is estimated using a histogram based on a Gaussian kernel and ;

(4) The implementation steps of the second attention block are as follows: feature map A ₃ and feature map J ₂ are multiplied to obtain feature map K ₂ .

Based on the above-mentioned CT image segmentation method based on Bhattacharyya coefficient active contour attention, the second attention depth separable convolution block is implemented as follows: the feature map and feature map After splicing, the splicing result is subjected to point-by-point convolution, batch normalization, 1×7 axial depth convolution, 7×1 axial depth convolution, residual convolution, point-by-point convolution and GELU activation function in sequence to obtain the feature map L _2. The feature map L ₂ is subjected to global average pooling, 1×1 convolution, Relu activation function, 1×1 convolution, Sigmoid activation function in sequence, and then multiplied with the feature map L ₂ to obtain the feature map P ₂ .

Based on the above-mentioned CT image segmentation method based on Bhattacharyya coefficient active contour attention, the third Bhattacharyya coefficient active contour attention module includes a third initial contour block, a third image processing block, a third Bhattacharyya coefficient active contour block, and a third attention block;

(1) The implementation steps of the third initial contour block are as follows: feature map _P2 is upsampled to obtain feature map _M3 , feature map _M3 is convolved with 1×1 to obtain feature map _B3 , feature map After 1×1 convolution, feature map C ₃ is obtained. Feature map B ₃ and feature map C ₃ are added to obtain feature map D _3. Feature map D ₃ is respectively subjected to Relu activation function, 1×1 convolution, and Sigmoid activation function to obtain feature map N _3. Feature map N ₃ is subjected to distance transformation to obtain feature map E ₃ , i.e., the initial contour.

The distance transformation is given by Find, among which is the Euclidean distance, * is the convolution multiplication, exp is the exponential function with the natural constant e as the base, is artificially given, and ;

(2) The implementation steps of the third image processing block are as follows: the feature map M ₃ is subjected to a 1×1 convolution to obtain a feature map F ₃ , the input image is subjected to a Resize image processing function and a Sigmoid activation function to obtain a feature map G ₃ , the feature map F ₃ and the feature map G ₃ are added to obtain a feature map H ₃ , and the feature map H ₃ is subjected to a 1×1 convolution and a Sigmoid activation function to obtain a feature map I ₃ ;

(3) The implementation steps of the third Bhattacharyya coefficient active contour block are as follows: Use the Bhattacharyya coefficient active contour algorithm to perform a finite number of iterative segmentations on the feature map I ₃ to obtain the feature map , ,in and is a positive parameter, which is artificially given; BCV is the Bhattacharyya coefficient active contour algorithm; E ₃ is the feature map E ₃ ;

The energy function of the Bhattacharyya coefficient active contour algorithm is: ,in The following formula represents the length of the contour, The following formula represents the area within the contour. represents the Dirac function, H is the Heavyside function, is the domain, Q is the RGB color space, in and out are the initial contours For the internal and external regions, for a given color q, the probability distribution function is estimated using a histogram based on a Gaussian kernel and ;

(4) The implementation steps of the third attention block are as follows: feature map A ₂ and feature map J ₃ are multiplied to obtain feature map K ₃ .

Based on the above-mentioned CT image segmentation method based on Bhattacharyya coefficient active contour attention, the implementation steps of the third attention deep separable convolution block are as follows: after splicing the feature map M ₃ and the feature map K ₃ , the splicing result is subjected to point-by-point convolution, batch normalization, 1×7 axial depth convolution, 7×1 axial depth convolution, residual convolution, point-by-point convolution and GELU activation function to obtain the feature map L _3. The feature map L ₃ is subjected to global average pooling, 1×1 convolution, Relu activation function, 1×1 convolution, Sigmoid activation function, and then multiplied with the feature map L ₃ to obtain the feature map P ₃ .

Based on the above-mentioned CT image segmentation method based on Bhattacharyya coefficient active contour attention, the fourth Bhattacharyya coefficient active contour attention module includes a fourth initial contour block, a fourth image processing block, a fourth Bhattacharyya coefficient active contour block and a fourth attention block.

(1) The fourth initial contour block is implemented as follows: feature map P ₃ is upsampled to obtain feature map M ₄ , feature map M ₄ is subjected to 1×1 convolution to obtain feature map B ₄ , feature map A ₁ is subjected to 1×1 convolution to obtain feature map C ₄ , feature map B ₄ and feature map C ₄ are added to obtain feature map D ₄ , feature map D ₄ is subjected to Relu activation function, 1×1 convolution, and Sigmoid activation function to obtain feature map N ₄ , feature map N ₄ is subjected to distance transformation to obtain feature map E ₄ , i.e., the initial contour;

The distance transformation is given by Find, among which is the Euclidean distance, * is the convolution multiplication, exp is the exponential function with the natural constant e as the base, is artificially given, and ;

(2) The fourth image processing block is implemented as follows: feature map M ₄ is subjected to 1×1 convolution to obtain feature map F ₄ , the input image is subjected to Resize image processing function and Sigmoid activation function to obtain feature map G ₄ , feature map F ₄ and feature map G ₄ are added to obtain feature map H ₄ , feature map H ₄ is subjected to 1×1 convolution and Sigmoid activation function to obtain feature map I ₄ ;

(3) The fourth Bhattacharyya coefficient active contour block is implemented as follows: Use the Bhattacharyya coefficient active contour algorithm to Perform a finite number of iterative segmentation to obtain the feature map , ,in and is a positive parameter, which is artificially given; BCV is the Bhattacharyya coefficient active contour algorithm; E ₄ is the feature map E ₄ ;

The energy function of the Bhattacharyya coefficient active contour algorithm is:

,

in The following formula represents the length of the contour, The following formula represents the area within the contour. represents the Dirac function, H is the Heavyside function, is the domain, Q is the RGB color space, in and out are the areas inside and outside the initial contour E ₄ , respectively. For a given color q, the probability distribution function is estimated using a histogram based on a Gaussian kernel and ;

(4) The implementation steps of the fourth attention block are as follows: Feature map and feature map Multiply to get the feature map .

Based on the above CT image segmentation method based on Bhattacharyya coefficient active contour attention,

The fourth attention depth separable convolution block is implemented as follows:

The feature map and feature map After splicing, the splicing results are sequentially subjected to point-by-point convolution, batch normalization, 1×7 axial depth convolution, 7×1 axial depth convolution, residual convolution, point-by-point convolution and GELU activation function to obtain the feature map. ;

Feature Map After global average pooling, 1×1 convolution, Relu activation function, 1×1 convolution, Sigmoid activation function, and feature map Multiply to get the feature map .

Based on the above CT image segmentation method based on Bhattacharyya coefficient active contour attention,

The output module implementation steps are as follows: Feature map After 1×1 convolution, BatchNorm accelerates the convergence of deep neural network, and Relu activation function, the segmented image is obtained. .

The effects provided in the summary of the invention are only the effects of the embodiments, rather than all the effects of the invention. The above technical solution has the following advantages or beneficial effects:

The present invention proposes a Bhattacharyya coefficient active contour attention structure, which helps our segmentation method better adapt to the grayscale differences between different regions and focus on the target structure by quantifying the similarity between the grayscale value distributions of the segmented region and the image; in the decoding stage, an attention deep separable convolution module is proposed, which has low computational complexity and can adaptively adjust the weight of each channel in the feature map, so as to capture the relationship between features more specifically and better learn important features to suppress unimportant features.

BRIEF DESCRIPTION OF THE DRAWINGS

The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

FIG1 is a flow chart of the present invention using a training set to train a network.

FIG. 2 is a diagram showing the network segmentation effect of the present invention using a test.

DETAILED DESCRIPTION

In order to make the purpose, technical solution and advantages of the present invention clearer and more understandable, the present invention is further described below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

A CT image segmentation method based on Bhattacharyya coefficient active contour attention comprises the following steps:

S1. The dataset is 3D CT images of kidneys and kidney tumors from MICCAI KiTS19. The training set of the dataset includes 210 3D-CT images, and the test set includes 90 3D-CT images.

S2. Call the nibabel library to process the volume data in the dataset to obtain 2D slices in .png format. Select 10 slices from each volume data to obtain dataset D’. Perform data enhancement on the training set in dataset D’ to obtain dataset D.

S3. The network structure includes an encoding part and a decoding part. The encoding part uses the VGG16 pre-trained network to extract features from the input image, and the decoding part uses the Bhattacharyya coefficient active contour attention structure and the attention depth separable convolution block;

S4. Use Dice loss function to calculate the loss function Loss;

S5. Use the SGD optimizer to adjust the weights and biases in the network through backpropagation;

S6. During the training process of the data set D, set the variables for saving the optimal evaluation index, and save the optimal weights and biases in the newly created file ph;

S7. Load the optimal weights and biases into the network, read the images in the test set to complete the segmentation, and save the results as a jpg file.

In this embodiment, the encoding part uses the VGG16 pre-trained network to extract features from the input image. The present invention directly uses the pre-trained VGG16 network and weight file. The pre-trained VGG16 network includes a total of six stages, and only the first five stages are used. The fifth stage does not include the Maxpooling operation, and each stage obtains feature maps A1, A2, A3, A4, and A5 respectively.

In this embodiment, the decoding part uses the Bhattacharyya coefficient active contour attention structure and the attention depth separable convolution block. The Bhattacharyya coefficient active contour attention structure reference: Chan-Vese Attention U-Net: Anattention mechanism for robust segmentation. The decoding part includes a first Bhattacharyya coefficient active contour attention module, a first attention depth separable convolution block, a second Bhattacharyya coefficient active contour attention module, a second attention depth separable convolution block, a third Bhattacharyya coefficient active contour attention module, a third attention depth separable convolution block, a fourth Bhattacharyya coefficient active contour attention module, a fourth attention depth separable convolution block, and an output module.

In this embodiment, the first Bhattacharyya coefficient active contour attention module includes a first initial contour block, a first image processing block, a first Bhattacharyya coefficient active contour block, and a first attention block;

(1) The first initial contour block is implemented as follows: feature map After upsampling, the feature map is obtained , feature map After 1×1 convolution, the feature map is obtained ,Feature map After 1×1 convolution, the feature map is obtained ,Feature map and feature map Add to get the feature map ,Feature map The feature map is obtained by Relu activation function, 1×1 convolution, and Sigmoid activation function respectively. , feature map After distance transformation, the feature map is obtained , i.e. the initial contour;

The distance transformation is given by Find, among which is the Euclidean distance, * is the convolution multiplication, exp is the exponential function with the natural constant e as the base, is artificially given, and ;

(2) The first image processing block is implemented as follows: Feature map After 1×1 convolution, the feature map is obtained The input image is processed by the Resize image processing function and the Sigmoid activation function to obtain the feature map , feature map and feature map Add to get the feature map , feature map After 1×1 convolution and Sigmoid activation function, the feature map is obtained ;

(3) The first Bhattacharyya coefficient active contour block is implemented as follows: Use the Bhattacharyya coefficient active contour algorithm to Perform a finite number of iterative segmentation to obtain the feature map , ,in and is a positive parameter, which is artificially given; BCV is the Bhattacharyya coefficient active contour algorithm, and E ₁ is the feature map E ₁ .

The energy function of the Bhattacharyya coefficient active contour algorithm is:

,

in The following formula represents the length of the contour, The following formula represents the area within the contour. represents the Dirac function, H is the Heavyside function, is the domain, Q is the RGB color space, in and out are the initial contours The inner and outer regions. For a given color q, and ;

(4) The implementation steps of the first attention block are as follows: Feature map and feature map Multiply to get the feature map .

In this embodiment, the first attention depth separable convolution block is implemented as follows: and feature map After splicing, the splicing results are sequentially subjected to point-by-point convolution, batch normalization, 1×7 axial depth convolution, 7×1 axial depth convolution, residual convolution, point-by-point convolution and GELU activation function to obtain the feature map. . Feature map After global average pooling, 1×1 convolution, Relu activation function, 1×1 convolution, Sigmoid activation function, and feature map Multiply to get the feature map .

In this embodiment, the second Bhattacharyya coefficient active contour attention module includes a second initial contour block, a second image processing block, a second Bhattacharyya coefficient active contour block, and a second attention block;

(1) The implementation steps of the second initial contour block are as follows: Feature map After upsampling, the feature map is obtained , feature map After 1×1 convolution, the feature map is obtained ,Feature map After 1×1 convolution, the feature map is obtained ,Feature map and feature map Add to get the feature map ,Feature map After Relu activation function, 1×1 convolution, and Sigmoid activation function, feature map N ₂ is obtained. After distance transformation _{, feature map E 2 is} obtained, which is the initial contour.

The distance transformation is given by Find, among which is the Euclidean distance, * is the convolution multiplication, exp is the exponential function with the natural constant e as the base, is artificially given, and ;

(2) The second image processing block is implemented as follows: Feature map After 1×1 convolution, the feature map is obtained The input image is processed by the Resize image processing function and the Sigmoid activation function to obtain the feature map ,Feature map and feature map Add to get the feature map , feature map After 1×1 convolution and Sigmoid activation function, the feature map is obtained ;

(3) The second Bhattacharyya coefficient active contour block is implemented as follows: Use the Bhattacharyya coefficient active contour algorithm to Perform a finite number of iterative segmentation to obtain the feature map , in and is a positive parameter,

is artificially given; BCV is the Bhattacharyya coefficient active contour algorithm, and E ₂ is the feature map E ₂ .

The energy function of the Bhattacharyya coefficient active contour algorithm is:

,

in The following formula represents the length of the contour, The following formula represents the area within the contour. represents the Dirac function, H is the Heavyside function, is the domain, Q is the RGB color space, in and out are the initial contours For a given color q, the probability distribution function is estimated using a histogram based on a Gaussian kernel and ;

(5) The implementation steps of the second attention block are as follows: Feature map and feature map Multiply to get the feature map .

In this embodiment, the second attention depth separable convolution block is implemented as follows: and feature map After splicing, the splicing results are sequentially subjected to point-by-point convolution, batch normalization, 1×7 axial depth convolution, 7×1 axial depth convolution, residual convolution, point-by-point convolution and GELU activation function to obtain the feature map. . Feature map After global average pooling, 1×1 convolution, Relu activation function, 1×1 convolution, Sigmoid activation function, and feature map Multiply to get the feature map .

In this embodiment, the third Bhattacharyya coefficient active contour attention module includes a third initial contour block, a third image processing block, a third Bhattacharyya coefficient active contour block, and a third attention block;

(1) The implementation steps of the third initial contour block are as follows: Feature map After upsampling, the feature map is obtained , feature map After 1×1 convolution, the feature map is obtained ,Feature map After 1×1 convolution, the feature map is obtained ,Feature map and feature map Add to get the feature map ,Feature map After Relu activation function, 1×1 convolution, and Sigmoid activation function, feature map N ₃ is obtained. After distance transformation _{, feature map E 3 is} obtained, which is the initial contour.

The distance transformation is given by Find, among which is the Euclidean distance, * is the convolution multiplication, exp is the exponential function with the natural constant e as the base, is artificially given, and ;

(2) The implementation steps of the third image processing block are as follows: Feature map After 1×1 convolution, the feature map is obtained The input image is processed by the Resize image processing function and the Sigmoid activation function to obtain the feature map ,Feature map and feature map Add to get the feature map ,Feature map After 1×1 convolution and Sigmoid activation function, the feature map is obtained ;

(3) The third Bhattacharyya coefficient active contour block is implemented as follows: Use the Bhattacharyya coefficient active contour algorithm to Perform a finite number of iterative segmentation to obtain the feature map , ,in and is a positive parameter, which is artificially given; BCV is the Bhattacharyya coefficient active contour algorithm, and E ₃ is the feature map E ₃ .

The energy function of the Bhattacharyya coefficient active contour algorithm is:

,

in The following formula represents the length of the contour, The following formula represents the area within the contour. represents the Dirac function, H is the Heavyside function, is the domain, Q is the RGB color space, in and out are the initial contours For a given color q, the probability distribution function is estimated using a histogram based on a Gaussian kernel and ;

(4) The implementation steps of the third attention block are as follows: Feature map and feature map Multiply to get the feature map .

In this embodiment, the third attention depth separable convolution block is implemented as follows: and feature map After splicing, the splicing results are sequentially subjected to point-by-point convolution, batch normalization, 1×7 axial depth convolution, 7×1 axial depth convolution, residual convolution, point-by-point convolution and GELU activation function to obtain the feature map. After global average pooling, 1×1 convolution, Relu activation function, 1×1 convolution, Sigmoid activation function, and feature map Multiply to get the feature map .

In this embodiment, the fourth Bhattacharyya coefficient active contour attention module includes a fourth initial contour block, a fourth image processing block, a fourth Bhattacharyya coefficient active contour block, and a fourth attention block.

(1) The fourth initial contour block is implemented as follows: feature map After upsampling, the feature map is obtained , feature map After 1×1 convolution, the feature map is obtained ,Feature map After 1×1 convolution, the feature map is obtained ,Feature map and feature map Add to get the feature map ,Feature map After Relu activation function, 1×1 convolution, and Sigmoid activation function, feature map N ₄ is obtained. After distance transformation _{, feature map E 4 is} obtained, which is the initial contour.

The distance transformation is given by Find, among which is the Euclidean distance, * is the convolution multiplication, exp is the exponential function with the natural constant e as the base, is artificially given, and ;

(2) The fourth image processing block is implemented as follows: Feature map After 1×1 convolution, the feature map is obtained The input image is processed by the Resize image processing function and the Sigmoid activation function to obtain the feature map ,Feature map and feature map Add to get the feature map ,Feature map After 1×1 convolution and Sigmoid activation function, the feature map is obtained .

(3) The fourth Bhattacharyya coefficient active contour block is implemented as follows: Use the Bhattacharyya coefficient active contour algorithm to Perform a finite number of iterative segmentation to obtain the feature map , ,in and is a positive parameter, which is artificially given; BCV is the Bhattacharyya coefficient active contour algorithm, and E ₄ is the feature map E ₄ .

The energy function of the Bhattacharyya coefficient active contour algorithm is:

,

in The following formula represents the length of the contour, The following formula represents the area within the contour. represents the Dirac function, H is the Heavyside function, is the domain, Q is the RGB color space, in and out are the initial contours For the internal and external regions, for a given color q, the probability distribution function is estimated using a histogram based on a Gaussian kernel and ;

(5) The implementation steps of the fourth attention block are as follows: Feature map and feature map Multiply to get the feature map .

In this embodiment, the fourth attention depth separable convolution block is implemented as follows: and feature map After splicing, the splicing results are sequentially subjected to point-by-point convolution, batch normalization, 1×7 axial depth convolution, 7×1 axial depth convolution, residual convolution, point-by-point convolution and GELU activation function to obtain the feature map. , feature map After global average pooling, 1×1 convolution, Relu activation function, 1×1 convolution, Sigmoid activation function, and feature map Multiply to get the feature map .

In this embodiment, the output module implements the following steps: feature map After 1×1 convolution, BatchNorm accelerates the convergence of deep neural network, and Relu activation function, the segmented image is obtained. .

The present invention achieved 97.97%, 95.56%, 94.04% and 95.88% results in accuracy, precision, Jaccard similarity coefficient and Dice similarity coefficient evaluation indicators respectively. Compared with advanced U-Net, UNet++, TransUNet and other models, the present invention is more able to focus on the target structure and can also obtain better segmentation results for images with uneven grayscale value distribution.

Although the above describes the specific implementation mode of the invention in conjunction with the drawings, it is not intended to limit the scope of protection of the invention. Based on the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present invention.

Claims (7)

1. A CT image segmentation method based on Pasteur coefficient active contour attention is characterized by comprising the following steps:

s1, a data set is a CT image of a 3D kidney and kidney tumor, a training set of the data set comprises 210 3D-CT images, and a test set comprises 90 3D-CT images from MICCAI KiTS 19;

s2, invoking a nibabel library to process the volume data in the data set to obtain 2D png format slices, respectively selecting 10 slices from each volume data to obtain a data set D ', and performing data enhancement on the training set in the data set D' to obtain a data set D;

s3, the network structure comprises an encoding part and a decoding part, wherein the encoding part uses a VGG16 pre-training network to extract characteristics of an input image, and uses the first five stages of the pre-training VGG16 network, wherein the fifth stage does not comprise maximum pooling operation Maxpooling operation, and each stage respectively obtains a characteristic image A ₁ 、A ₂ 、A ₃ 、A ₄ 、A ₅ The decoding part uses a Babbitt coefficient active contour attention and attention depth separable convolution block;

the decoding part comprises a first Pasteur coefficient active contour attention module, a first attention depth separable convolution block, a second Pasteur coefficient active contour attention module, a second attention depth separable convolution block, a third Pasteur coefficient active contour attention module, a third attention depth separable convolution block, a fourth Pasteur coefficient active contour attention module, a fourth attention depth separable convolution block and an output module;

the first pap coefficient active contour attention module comprises a first initial contour block, a first image processing block, a first pap coefficient active contour block and a first attention block;

(1) The first initial contour block implementation steps are as follows: feature map A ₅ Up-sampling to obtain feature map M ₁ Feature map M ₁ Obtaining a characteristic diagram B through convolution of 1 multiplied by 1 ₁ Feature map A ₄ Obtaining a characteristic diagram C through convolution of 1 multiplied by 1 ₁ Feature map B ₁ And feature map C ₁ Adding to obtain a feature map D ₁ Feature map D ₁ The feature map N is obtained through a Relu activation function, a convolution of 1 multiplied by 1 and a Sigmoid activation function respectively ₁ Feature map N ₁ Obtaining a characteristic diagram E after distance conversion ₁ I.e. the initial profile;

distance transformation is composed ofThe determination, where d (, 0) is the euclidean distance, x is the convolution product, exp is an exponential function based on a natural constant e, λ is artificially given, and 1>λ>0；

(2) The first image processing block is implemented as follows: feature map M ₁ Obtaining a characteristic diagram F through convolution of 1 multiplied by 1 ₁ The input image obtains a feature graph G through a Resize image processing function and a Sigmoid activation function ₁ Feature map F ₁ And feature map G ₁ Adding to obtain a characteristic diagram H ₁ Feature map H ₁ Obtaining a feature map I through 1X 1 convolution and Sigmoid activation function ₁ ；

(3) The first Pasteur coefficient active contour block is realized by using a Pasteur coefficient active contour algorithm to perform the following steps of ₁ Performing limited iterative segmentation to obtain a feature map J ₁ :J ₁ ＝BCV(I ₁ ,E ₁ μ, ν), wherein μ and ν are positive parameters, are artificially given; BCV is the papanicolaou coefficient active contour algorithm; e (E) ₁ Is a characteristic diagram E ₁ ；

The energy function of the papanicolaou coefficient active contour algorithm is:

wherein μ followed by formula represents the length of the contour, ν followed by formula represents the area within the contour, δ represents the dirac function, H is the heavside function, Ω is the definition domain, Q is the RGB color space, and in and out are the initial contours E, respectively ₁ Inner and outer regions; for a given color q, a gaussian kernel based histogram estimation probability distribution function P is used _in (E ₁ Q) and P _out (E ₁ ,q)；

(4) The first attention block is implemented by the following steps of a characteristic diagram A ₄ And feature map J ₁ Multiplication to obtain a feature map K ₁ ；

The first attention depth separable convolution block implementation steps are as follows: map M of features ₁ And feature map K ₁ After splicing, the splicing result is subjected to point-by-point convolution, batch Normalization batch normalization, 1×7 axial depth convolution, 7×1 axial depth convolution, residual convolution, point-by-point convolution and GELU activation function in sequence to obtain a feature map L ₁ Feature map L ₁ Sequentially performing global average pooling, 1×1 convolution, relu activation function, 1×1 convolution, sigmoid activation function, and then combining with feature map L ₁ Multiplication to obtain a feature map P ₁ ；

S4, calculating a Loss function Loss by using the Dice Loss function;

s5, using an SGD optimizer to adjust weight and bias in a network through back propagation;

s6, setting a variable for storing and evaluating an optimal index in the training process of the data set D, and storing the optimal weight and the bias in a newly built file ph;

s7, loading the optimal weight and the optimal offset into a network, reading the image in the test set to complete segmentation, and storing the result as a jpg file.

2. The CT image segmentation method based on the active contour attention of the papanicolaou coefficient according to claim 1, wherein: the second pap coefficient active contour attention module comprises a second initial contour block, a second image processing block, a second pap coefficient active contour block and a second attention block;

(1) The second initial contour block implementation steps are as follows: feature map P ₁ Up-sampling to obtain feature map M ₂ Feature map M ₂ Obtaining a characteristic diagram B through convolution of 1 multiplied by 1 ₂ Feature map A ₃ Obtaining a characteristic diagram C through convolution of 1 multiplied by 1 ₂ Feature map B ₂ And feature map C ₂ Adding to obtain a feature map D ₂ Feature map D ₂ The signature N is obtained by the Relu activation function and the 1 multiplied by 1 convolution respectively by the Sigmoid activation function ₂ Feature map N ₂ Obtaining a characteristic diagram E after distance conversion ₂ I.e. the initial profile;

distance transformation is composed ofThe determination, where d (, 0) is the euclidean distance, x is the convolution product, exp is an exponential function based on a natural constant e, λ is artificially given, and 1>λ>0；

(2) The second image processing block is implemented as follows: feature map M ₂ Obtaining a characteristic diagram F through convolution of 1 multiplied by 1 ₂ The input image obtains a feature graph G through a Resize image processing function and a Sigmoid activation function ₂ Feature map F ₂ And feature map G ₂ Adding to obtain a characteristic diagram H ₂ Feature map H ₂ Obtaining a feature map I through 1X 1 convolution and Sigmoid activation function ₂ ；

(3) The second Pasteur coefficient active contour block is realized by using the Pasteur coefficient active contour algorithm to perform the following steps of comparing the characteristic diagram I ₂ Performing limited iterative segmentation to obtain a feature map J ₂ ，J ₂ ＝BCV(I ₂ ,E ₂ μ, ν), wherein μ and ν are positive parameters, are artificially given; BCV is the papanicolaou coefficient active contour algorithm; e (E) ₂ Is a characteristic diagram E ₂ ；

(4) The energy function of the papanicolaou coefficient active contour algorithm is:

wherein μ followed by formula represents the length of the contour, ν followed by formula represents the area within the contour, δ represents the dirac function, H is the heavside function, Ω is the definition domain, Q is the RGB color space, and in and out are the initial contours E, respectively ₂ Internal and external regions, for a given color q, using a gaussian kernel based histogram estimation probability distribution function P _in (E ₂ Q) and P _out (E ₂ ,q)；

(4) The second attention block is implemented by the following steps of a characteristic diagram A ₃ And feature map J ₂ Multiplication to obtain a feature map K ₂ 。

3. The CT image segmentation method based on the active contour attention of the papanicolaou coefficient according to claim 2, wherein: the second attention depth separable convolution block implementation steps are as follows: map M of features ₂ And feature map K ₂ After splicing, the splicing result is subjected to point-by-point convolution, batch Normalization batch normalization, 1×7 axial depth convolution, 7×1 axial depth convolution, residual convolution, point-by-point convolution and GELU activation function in sequence to obtain a feature map L ₂ Feature map L ₂ Sequentially performing global average pooling, 1×1 convolution, relu activation function, 1×1 convolution, sigmoid activation function, and then combining with feature map L ₂ Multiplication to obtain a feature map P ₂ 。

4. A CT image segmentation method based on active contour attention based on papanicolaou coefficient as claimed in claim 3, wherein: the third pap coefficient active contour attention module comprises a third initial contour block, a third image processing block, a third pap coefficient active contour block and a third attention block;

(1) The third initial contour block implementation steps are as follows: feature map P ₂ Up-sampling to obtain feature map M ₃ Feature map M ₃ Is convolved by 1 x 1 to obtainTo feature map B ₃ Feature map A ₂ Obtaining a characteristic diagram C through convolution of 1 multiplied by 1 ₃ Feature map B ₃ And feature map C ₃ Adding to obtain a feature map D ₃ Feature map D ₃ The signature N is obtained by the Relu activation function and the 1 multiplied by 1 convolution respectively by the Sigmoid activation function ₃ Feature map N ₃ Obtaining a characteristic diagram E after distance conversion ₃ I.e. the initial profile;

distance transformation is composed ofThe determination, where d (, 0) is the euclidean distance, x is the convolution product, exp is an exponential function based on a natural constant e, λ is artificially given, and 1>λ>0；

(2) The third image processing block is implemented as follows: feature map M ₃ Obtaining a characteristic diagram F through convolution of 1 multiplied by 1 ₃ The input image obtains a feature graph G through a Resize image processing function and a Sigmoid activation function ₃ Feature map F ₃ And feature map G ₃ Adding to obtain a characteristic diagram H ₃ Feature map H ₃ Obtaining a feature map I through 1X 1 convolution and Sigmoid activation function ₃ ；

(3) The third Pasteur coefficient active contour block is realized by using the Pasteur coefficient active contour algorithm to perform the following steps of ₃ Performing limited iterative segmentation to obtain a feature map J ₃ :J ₃ ＝BCV(I ₃ ,E ₃ μ, ν), wherein μ and ν are positive parameters, are artificially given; BCV is the papanicolaou coefficient active contour algorithm; e (E) ₃ Is a characteristic diagram E ₃ ；

The energy function of the papanicolaou coefficient active contour algorithm is:

wherein μ followed by formula represents the length of the contour, ν followed by formula represents the area within the contour, δ represents the dirac function, H is the heaviside function, Ω is the definition domain, Q is the RGB color space, inAnd out are respectively the initial profile E ₃ Internal and external regions, for a given color q, using a gaussian kernel based histogram estimation probability distribution function P _in (E ₃ Q) and P _out (E ₃ ,q)；

(4) The third attention block is implemented by the following steps of a characteristic diagram A ₂ And feature map J ₃ Multiplication to obtain a feature map K ₃ 。

5. The CT image segmentation method based on active contour attention by a papanicolaou coefficient according to claim 4, wherein: the third attention depth separable convolution block implementation steps are as follows: map M of features ₃ And feature map K ₃ After splicing, the splicing result is subjected to point-by-point convolution, batch Normalization batch normalization, 1×7 axial depth convolution, 7×1 axial depth convolution, residual convolution, point-by-point convolution and GELU activation function in sequence to obtain a feature map L ₃ Feature map L ₃ Sequentially performing global average pooling, 1×1 convolution, relu activation function, 1×1 convolution, sigmoid activation function, and then combining with feature map L ₃ Multiplication to obtain a feature map P ₃ 。

6. The CT image segmentation method based on active contour attention by a papanicolaou coefficient according to claim 5, wherein: the fourth pap coefficient active contour attention module comprises a fourth initial contour block, a fourth image processing block, a fourth pap coefficient active contour block and a fourth attention block,

(1) The fourth initial contour block implementation steps are as follows: feature map P ₃ Up-sampling to obtain feature map M ₄ Feature map M ₄ Obtaining a characteristic diagram B through convolution of 1 multiplied by 1 ₄ Feature map A ₁ Obtaining a characteristic diagram C through convolution of 1 multiplied by 1 ₄ Feature map B ₄ And feature map C ₄ Adding to obtain a feature map D ₄ Feature map D ₄ The signature N is obtained by the Relu activation function and the 1 multiplied by 1 convolution respectively by the Sigmoid activation function ₄ Feature map N ₄ Obtaining a characteristic diagram E after distance conversion ₄ I.e. the initial profile;

distance transformation is composed ofThe determination, where d (, 0) is the euclidean distance, x is the convolution product, exp is an exponential function based on a natural constant e, λ is artificially given, and 1>λ>0；

(2) The fourth image processing block is implemented as follows: feature map M ₄ Obtaining a characteristic diagram F through convolution of 1 multiplied by 1 ₄ The input image obtains a feature graph G through a Resize image processing function and a Sigmoid activation function ₄ Feature map F ₄ And feature map G ₄ Adding to obtain a characteristic diagram H ₄ Feature map H ₄ Obtaining a feature map I through 1X 1 convolution and Sigmoid activation function ₄ ；

(3) The fourth Pasteur coefficient active contour block is realized by using the Pasteur coefficient active contour algorithm to perform the following steps of comparing the characteristic diagram I ₄ Performing limited iterative segmentation to obtain a feature map J ₄ ，J ₄ ＝BCV(I ₄ ,E ₄ μ, ν), wherein μ and ν are positive parameters, are artificially given; BCV is the papanicolaou coefficient active contour algorithm; e (E) ₄ Is a characteristic diagram E ₄ ；

The energy function of the papanicolaou coefficient active contour algorithm is:wherein μ followed by formula represents the length of the contour, ν followed by formula represents the area within the contour, δ represents the dirac function, H is the heavside function, Ω is the definition domain, Q is the RGB color space, and in and out are the initial contours E, respectively ₄ Inner and outer regions; for a given color q, a gaussian kernel based histogram estimation probability distribution function P is used _in (E ₄ Q) and P _out (E ₄ ,q)；

(4) The fourth attention block is implemented by the following steps of a characteristic diagram A ₁ And feature map J ₄ Multiplication to obtain a feature map K ₄ 。

7. The CT image segmentation method based on active contour attention by papanicolaou coefficient according to claim 6, wherein:

the fourth attention depth separable convolution block implementation steps are as follows:

map M of features ₄ And feature map K ₄ After splicing, the splicing result is subjected to point-by-point convolution, batch Normalization batch normalization, 1×7 axial depth convolution, 7×1 axial depth convolution, residual convolution, point-by-point convolution and GELU activation function in sequence to obtain a feature map L ₄ ；

Feature map L ₄ Sequentially performing global average pooling, 1×1 convolution, relu activation function, 1×1 convolution, sigmoid activation function, and then combining with feature map L ₄ Multiplication to obtain a feature map P ₄ ；

The output module comprises the following implementation steps: feature map P ₄ Obtaining a segmented image d through 1X 1 convolution, batchNorm acceleration depth neural network convergence and Relu activation function _end 。