Three-dimensional image segmentation method based on dual-path attention encoder-decoder network

Abstract

The present invention discloses a three-dimensional image segmentation method based on a dual-path attention codec network. The method first pre-processes the training image, then constructs a dual-path attention codec network, uses boundary loss to optimize the parameters of the network, and uses the trained model to predict and segment the three-dimensional image; finally, dense conditional random fields and maximum connected area algorithms are used to adjust the probability map, optimize the segmentation results, and save the output post-processing results. The present invention integrates the dual-path attention module into the codec network, improves the network's segmentation accuracy for three-dimensional images, reduces the time cost and labor cost of three-dimensional image processing, and promotes the progress and development of the corresponding academic and industrial circles.

Description

Technical Field

[0001] The present invention relates to the field of computer vision three-dimensional image analysis and processing. For three-dimensional image data, a three-dimensional image segmentation method based on a dual-path attention encoding and decoding network is proposed. The present invention can be applied to three-dimensional image segmentation in fields such as materials science and medicine, improve the accuracy of three-dimensional image segmentation, reduce the time cost and labor cost of three-dimensional image processing, and promote the progress and development of the corresponding academic and industrial circles. Background Art

[0002] Image semantic segmentation is a common problem in image processing and other fields. Semantic segmentation is to let the computer segment according to the content of the image. Segmentation is to separate different objects in the image at the pixel level, annotate each pixel in the original image, and classify them into different labels. The accuracy of segmentation includes the understanding of the information in the image. Three-dimensional images have the characteristics of complex imaging, high image dimensions and large information. Therefore, how to use artificial intelligence technology to quickly and accurately perform semantic segmentation on three-dimensional images and extract useful information from them is one of the research hotspots in the field of computer vision.

[0003] There are many methods for image semantic segmentation, among which image semantic segmentation based on neural network is one of the most popular research hotspots, and many research results have been achieved. FCN (Fully convolutional network) is a classic framework for image semantic segmentation. It is trained in an end-to-end method and uses the trained classification network for semantic segmentation. In order to restore the resolution of the image, FCN also uses deconvolution for upsampling. Compared with FCN, U-Net has a more symmetrical encoding and decoding structure. The jump connection from the encoding to the decoding part helps to restore the position information. However, since the basic module for building the network structure is a simple convolution block, there is a certain degree of gradient disappearance problem, which limits the increase of network depth. In addition, U-Net does not fully consider the connection between pixels and lacks the exploration of the dependency relationship between local features, which affects the accuracy of the final segmentation result. Loss function is an important tool for optimizing network parameters. Losses such as cross entropy and similarity coefficient lack the ability to optimize the network to explore the boundary features of the image, which limits the improvement of semantic segmentation accuracy. Therefore, how to build a deeper and more effective network structure and optimize the network to explore more features is the key to improving the accuracy of semantic segmentation. Summary of the invention

[0004] In order to solve the problems of the prior art, the purpose of the present invention is to overcome the shortcomings of the existing technology and design a three-dimensional image segmentation method based on a dual-path attention encoding and decoding network to enhance the network's exploration of local features and boundary features of the image and achieve high-precision segmentation of the three-dimensional image.

[0005] In order to achieve the above invention purpose, the present invention adopts the following technical solution:

[0006] A three-dimensional image segmentation method based on a dual-path attention encoding and decoding network comprises the following steps:

[0007] (1) 3D image preprocessing:

[0008] The original image used for training is randomly cropped into small image blocks, the small image blocks are preprocessed to obtain clearer images, and the preprocessed data is saved locally;

[0009] (2) Training the dual-path attention encoder-decoder network:

[0010] Build a dual-path attention encoder-decoder network, input the training set data into the network, use boundary loss to optimize the network model parameters, and save the trained network parameter file;

[0011] (3) Use the trained model to predict:

[0012] Load the trained model parameter file, input the test set data into the network, and obtain the predicted segmentation results, which are represented by probability maps;

[0013] (4) Post-processing of network prediction results:

[0014] The probability map is adjusted using dense conditional random fields to optimize the edges of the segmentation results, and then the maximum connected area algorithm is used to remove redundant false positives, and the post-processed image results are output and saved.

[0015] Preferably, a three-dimensional image segmentation method based on a dual-path attention encoding and decoding network comprises the following steps:

[0016] (1) Randomly crop the original image used for training into smaller image blocks, preprocess the small image blocks to obtain clearer images, and save the preprocessed data locally;

[0017] As a technical solution of the present invention, step 1 includes the following sub-steps:

[0018] (1-1) Cropping the 3D image data into image blocks of 12×224×244 pixels;

[0019] (1-2) Determine whether the image block is a grayscale image, and use a normalization algorithm to convert the non-grayscale image into a grayscale image;

[0020] (1-3) Use Gaussian filtering to remove noise points in the image;

[0021] (1-4) Use histogram equalization to stretch the grayscale distribution of the image and enhance the contrast of the image;

[0022] (1-5) Use the Laplace operator to sharpen the edge of the image, enhance the grayscale mutation in the image, and reduce the area where the grayscale changes slowly;

[0023] (1-6) dividing and saving the preprocessed image data;

[0024] (2) Build a dual-path attention encoder-decoder network, input the training set data into the network, use the boundary loss to optimize the network model parameters, and save the trained network parameter file;

[0025] As a technical solution of the present invention, in step 2, the dual-path attention encoding and decoding network includes three sub-network modules, namely: (a) encoder network, (b) dual-path attention network and (c) decoder network;

[0026] As a technical solution of the present invention, in step 2, a residual block, a maximum pooling block, an average pooling block and a dual-path block are used to construct an encoder, and the encoder network construction includes the following steps:

[0027] (2-1-1) Use 1 residual block to construct the first layer of the encoder to adapt to inputs of different data dimensions, and use maximum pooling to reduce the dimension of the output of the first layer;

[0028] (2-1-2) Two dual-path blocks are used in the second layer of the encoder to explore the low-level texture features of the image, and the output of the second layer is reduced in dimension using maximum pooling;

[0029] (2-1-3) Use three dual-path blocks in the third layer of the encoder to explore high-level abstract features of the image, and use average pooling to reduce the dimension of the output of the third layer;

[0030] (2-1-4) Five dual-path blocks are used in the fourth layer of the encoder to explore high-level abstract features of the image, and average pooling is used to reduce the dimension of the output of the fourth layer;

[0031] (2-1-5) Two dual-path blocks are used in the fifth layer of the encoder to integrate the explored image features;

[0032] As a technical solution of the present invention, in step 2, the attention network module includes three submodules: (a) a position attention module, (b) a channel attention module and (c) a feature fusion module; the attention network module uses a convolution block and a Softmax activation function to construct the three submodules in the attention network; the construction of the attention network module includes the following steps:

[0033] (2-2-1) Place the position attention module and the channel attention module in parallel to explore spatial features and channel features;

[0034] (2-2-2) The feature fusion module is used to integrate the explored spatial features and channel features;

[0035] As a technical solution of the present invention, in step 2, a decoder network is constructed using a residual block, a deconvolution block and a Sigmoid activation function, and the decoder construction includes the following steps:

[0036] (2-3-1) In the first to fourth layers of the decoder, two residual blocks and one deconvolution block are alternately connected to gradually restore the image feature resolution;

[0037] (2-3-2) In the fifth layer of the decoder, two residual blocks and a Sigmoid activation function are used to output the probability feature map;

[0038] As a technical solution of the present invention, in step 2, a boundary loss function is used to optimize the network to explore more contours and boundary features. The boundary loss can be expressed as:

[0039] L e =d+α·a+β·e (1)

[0040] In the above formula (1), d, a and e represent distance, area and boundary respectively. α and β are weights corresponding to area and boundary respectively. Expanding d, a and e in pixel form can be expressed as:

[0041]

[0042]

[0043] where v i，j and u i，j Represent the labeled value and predicted value respectively, x i，j and ij are the horizontal and vertical coordinates of the pixel point (i, j), N is the pixel space of the image, and Δu represents the result of subtracting the pixel values ​​of the corresponding coordinates; assuming that the true value of the image is A, B is the result of A being iteratively expanded four times, and C is the result of A being iteratively eroded four times, then the outer boundary of the true image can be obtained. and the inner boundary in represents the XOR operation, using the outer boundary of the real image O m and I m Get the inner and outer boundaries I and O of the predicted image. At this time, the boundary e of the boundary loss function can be expressed as:

[0044]

[0045] As a technical solution of the present invention, in step 2, the number of iterations epoch is set to 100 when training the network model. Usually, the network parameters can converge to the vicinity of the optimal value when the number of iterations epoch is not greater than 50. The network training includes the following steps:

[0046] (2-4-1) The Adam first-order optimization algorithm is used to optimize the network parameters and iteratively update the neural network weights based on the training data;

[0047] (2-4-2) Set the weight decay coefficient to alleviate the problem of model overfitting;

[0048] (2-4-3) In order to further obtain better network performance, set the learning rate and adopt a scheme of dynamically reducing the learning rate to further approach the optimal value of the network parameters. When the loss value no longer decreases within a certain epoch, multiply the learning rate by the attenuation factor to reduce the learning rate;

[0049] (3) Load the trained model parameter file, input the test set data into the network, and obtain the predicted segmentation results, which are represented by probability maps;

[0050] As a technical solution of the present invention, in step 3, the prediction of the three-dimensional image includes the following specific steps:

[0051] (3-1) Load the trained model parameter file;

[0052] (3-2) Input the three-dimensional image data into the network to obtain the predicted segmentation results;

[0053] (3-3) After completing the prediction of the small image blocks, the small image blocks are stitched together in order to form an image of the same size as the label image and saved locally in the form of a probability map.

[0054] Use dense conditional random fields to adjust the probability map, optimize the edges of the segmentation results, and then use the maximum connected area algorithm to remove redundant false positives, output and save the post-processed image results;

[0055] As a technical solution of the present invention, in step 4, the optimization of the network prediction result image includes the following steps:

[0056] (4-1) The probability graph output by the network is regarded as an undirected complete graph. Each pixel i can be regarded as a node v∈V in a graph G(V, E), and the node v has a fixed category label x i and the corresponding observation value y i , the dependency between nodes is regarded as edge e∈E, thus using the average approximation algorithm to construct an undirected complete graph into a dense conditional random field;

[0057] (4-2) Use dense conditional random fields to adjust the probabilities in the probability map and optimize the edge details of the segmented image;

[0058] (4-3) Use the maximum region connectivity technique to reduce false positives and false negatives in the segmentation results, further improving segmentation accuracy;

[0059] (4-4) Output and save the post-processed image results.

[0060] Compared with the prior art, the present invention has the following obvious outstanding substantive features and significant advantages:

[0061] 1. The present invention designs a dual-path attention encoder-decoder network structure, which is applied to three-dimensional image segmentation in fields such as materials science and medicine. Each layer of the encoder uses different numbers of residual blocks and dual-path blocks to more highly explore low-level texture features and high-level abstract features of the image; the dual-path attention network module is integrated into the encoder-decoder network to enhance the network's ability to express local features and explore more spatial context feature information in three-dimensional images, thereby achieving high-precision segmentation of three-dimensional images;

[0062] 2. The present invention combines dense conditional random fields and maximum region connectivity technology to post-process the prediction results of the model, which can effectively optimize the image edge details and reduce the existence of false positive examples and false negative examples in the segmentation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 The flowchart of the preferred embodiment of the present invention for segmenting a three-dimensional image is divided into the following steps: (1) inputting three-dimensional image data, randomly cropping the original image used for training and testing into smaller image blocks, preprocessing the small image blocks to obtain a clearer image, and saving the preprocessed data locally; (2) constructing a dual-path attention codec network, inputting the training set data into the network model, using the boundary loss to optimize the model parameters of the network, and saving the trained network parameter file; (3) loading the trained model parameter file, inputting the test set data into the network, obtaining the predicted segmentation result, and the segmentation result is represented by a probability map; (4) using a dense conditional random field to adjust the probability map, optimize the edge of the segmentation result, and then using the maximum connected area algorithm to remove redundant false positive examples, outputting and saving the post-processed image result.

[0064] Figure 2The flowchart of the preprocessing method of the preferred embodiment of the present invention is divided into the following steps: (1) 3D image cropping: cropping the 3D image data into an image block of 12×224×244 pixels; (2) judging whether the image block is a grayscale image, and using a normalization algorithm to convert the non-grayscale image into a grayscale image; (3) using Gaussian filtering to remove noise points in the image; (4) using histogram equalization to stretch the grayscale distribution of the image and enhance the contrast of the image; (5) using the Laplacian operator to implement edge sharpening processing of the image, enhancing the grayscale mutation in the image, that is, reducing the area where the grayscale changes slowly; (6) dividing and saving the preprocessed image data. DETAILED DESCRIPTION

[0065] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the preferred embodiment of the present invention will be reviewed and fully described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the implementation cases of the present invention, not all implementation cases. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0066] The above scheme is further described below in conjunction with specific implementation examples. The preferred embodiments of the present invention are described in detail as follows:

[0067] Embodiment 1:

[0068] See also Figure 1 In this embodiment, a three-dimensional image segmentation method based on a dual-path attention encoding and decoding network is provided. The method constructs an efficient dual-path attention encoding and decoding network structure, uses boundary loss to optimize network parameters, and improves the network's segmentation accuracy for three-dimensional image data.

[0069] The method of the present invention uses a certain three-dimensional medical image to train the model, obtains the model parameters of such data, and then obtains a high-precision prediction of similar segmentation data other than the sample. The method of the present invention comprises the following steps:

[0070] (1) Randomly crop the original image used for training into smaller image blocks, preprocess the small image blocks to obtain clearer images, and save the preprocessed data locally;

[0071] (2) Build a dual-path attention encoder-decoder network, input the training set data into the network, use the boundary loss to optimize the network model parameters, and save the trained network parameter file;

[0072] (3) Load the trained model parameter file, input the test set data into the network, and obtain the predicted segmentation results, which are represented by probability maps;

[0073] (4) Use dense conditional random fields to adjust the probability map and optimize the edges of the segmentation results. Then use the maximum connected component algorithm to remove redundant false positives and output and save the post-processed image results.

[0074] The present invention is based on a three-dimensional image segmentation method of a dual-path attention encoding and decoding network. First, the image used for training is preprocessed to obtain a clearer image, and the preprocessed data is saved locally; then the dual-path attention network is trained on the training data set using the boundary loss; then the trained model is used to predict the test data set, and the predicted probability map result is saved; finally, the probability map is adjusted using a dense conditional random field to optimize the edge of the segmentation result, and then the maximum connected area algorithm is used to remove redundant false positive examples, and the post-processed image result is output and saved.

[0075] Embodiment 2:

[0076] This embodiment is basically the same as the first embodiment, except that:

[0077] In this embodiment, if Figure 2 As shown, image preprocessing includes the following steps:

[0078] (1-1) Cropping the 3D image data into image blocks of 12×224×244 pixels;

[0079] (1-2) Determine whether the image block is a grayscale image, and use a normalization algorithm to convert the non-grayscale image into a grayscale image;

[0080] (1-3 Use Gaussian filtering to remove noise points in the image;

[0081] (1-4) Use histogram equalization to stretch the grayscale distribution of the image and enhance the contrast of the image;

[0082] (1-5) Use the Laplace operator to sharpen the edge of the image, enhance the grayscale mutation in the image, and reduce the area where the grayscale changes slowly;

[0083] (1-6) Divide and save the preprocessed image data.

[0084] This embodiment performs image preprocessing and stores the data locally, laying a good foundation for subsequent steps.

[0085] Embodiment three:

[0086] This embodiment is basically the same as the second embodiment, except that:

[0087] In this embodiment, the dual-path attention encoding and decoding network includes three sub-network modules, namely: (a) encoder network, (b) dual-path attention network and (c) decoder network; the encoder is constructed by using residual block, maximum pooling, average pooling and dual-path block, and the encoder network construction includes the following steps:

[0088] (2-1-1) Use 1 residual block to construct the first layer of the encoder to adapt to inputs of different data dimensions, and use maximum pooling to reduce the dimension of the output of the first layer;

[0089] (2-1-2) Two dual-path blocks are used in the second layer of the encoder to explore the low-level texture features of the image, and the output of the second layer is reduced in dimension using maximum pooling;

[0090] (2-1-3) Use three dual-path blocks in the third layer of the encoder to explore high-level abstract features of the image, and use average pooling to reduce the dimension of the output of the third layer;

[0091] (2-1-4) Five dual-path blocks are used in the fourth layer of the encoder to explore high-level abstract features of the image, and average pooling is used to reduce the dimension of the output of the fourth layer;

[0092] (2-1-5) Two dual-path blocks are used in the fifth layer of the encoder to integrate the explored image features;

[0093] In this embodiment, the attention network module includes three submodules: (a) a position attention module, (b) a channel attention module, and (c) a feature fusion module; the attention network module uses a convolution block and a Softmax activation function to construct the three submodules in the attention network; the construction of the attention network module includes the following steps:

[0094] (2-2-1) Place the position attention module and the channel attention module in parallel to explore spatial features and channel features;

[0095] (2-2-2) The feature fusion module is used to integrate the explored spatial features and channel features;

[0096] In this embodiment, a decoder network is constructed using a residual block, a deconvolution block, and a Sigmoid activation function. The decoder construction includes the following steps:

[0097] (2-3-1) In the first to fourth layers of the decoder, two residual blocks and one deconvolution block are alternately connected to gradually restore the image feature resolution;

[0098] (2-3-2) In the fifth layer of the decoder, two residual blocks and a Sigmoid activation function are used to output the probability feature map;

[0099] In this embodiment, the boundary loss function is used to optimize the network to explore more contours and boundary features. The boundary loss can be expressed as:

[0100] L e =d+α·a+β·e (1)

[0101] In the above formula (1), d, a and e represent distance, area and boundary respectively. α and β are the weights corresponding to area and boundary respectively. Expanding d, a and e in pixel form can be expressed as:

[0102]

[0103]

[0104] where v i，j and u i，j Represent the labeled value and predicted value respectively, x i，j and ij are the horizontal and vertical coordinates of the pixel point (i, j), N is the pixel space of the image, and Δu represents the result of subtracting the pixel values ​​of the corresponding coordinates; assuming that the true value of the image is A, B is the result of A being iteratively expanded four times, and C is the result of A being iteratively eroded four times, then the outer boundary of the true image can be obtained. and the inner boundary in represents the XOR operation, using the outer boundary of the real image O m and I m The inner and outer boundaries I and O of the predicted image can be obtained. At this time, the boundary e of the boundary loss function can be expressed as:

[0105]

[0106] In this embodiment, the number of iterations epoch is set to 100 when training the network model. Usually, the network parameters can converge to the vicinity of the optimal value when the number of iterations epoch is not greater than 50. The network training includes the following steps:

[0107] (2-4-1) The Adam first-order optimization algorithm is used to optimize the network parameters and iteratively update the neural network weights based on the training data;

[0108] (2-4-2) Set the weight decay coefficient to alleviate the problem of model overfitting;

[0109] (2-4-3) In order to further obtain better network performance, the learning rate is set, and a scheme of dynamically reducing the learning rate is adopted to further approach the optimal value of the network parameters. When the loss value no longer decreases within a certain epoch, the learning rate is multiplied by the attenuation factor to reduce the learning rate.

[0110] Embodiment 4:

[0111] This embodiment is basically the same as the third embodiment, except that:

[0112] In this embodiment, the prediction of the three-dimensional image includes the following steps:

[0113] (3-1) Load the trained model parameter file;

[0114] (3-2) Input the three-dimensional image data into the network to obtain the predicted segmentation results;

[0115] (3-3) After completing the prediction of the small image blocks, the small image blocks are stitched together in order to form an image of the same size as the label image and saved locally in the form of a probability map.

[0116] This embodiment predicts a three-dimensional image to provide prediction information for optimizing the edge of the segmentation result in subsequent steps.

[0117] Embodiment five:

[0118] This embodiment is basically the same as the third embodiment, except that:

[0119] In this embodiment, the optimization of the network prediction result image includes the following steps:

[0120] (4-1) The probability graph output by the network is regarded as an undirected complete graph. Each pixel i can be regarded as a node v∈V in a graph G(V, E), and the node v has a fixed category label x i and the corresponding observation value y i , the dependency between nodes is regarded as edge e∈E, thus using the average approximation algorithm to construct an undirected complete graph into a dense conditional random field;

[0121] (4-2) Use dense conditional random fields to adjust the probabilities in the probability map and optimize the edge details of the segmented image;

[0122] (4-3) Use the maximum region connectivity technique to reduce false positives and false negatives in the segmentation results, further improving the segmentation accuracy;

[0123] (4-4) Output and save the post-processed image results.

[0124] Based on the above embodiments, it can be seen that Figure 2 It is a flow chart of the three-dimensional image segmentation method based on the dual-path attention encoding and decoding network in the present invention, comprising the following steps:

[0125] First, the original image used for training is randomly cropped into smaller image blocks, the small image blocks are preprocessed to obtain clearer image blocks, and the preprocessed data is saved locally; the boundary loss function is designed, a dual-path attention codec network is constructed, the training set data is input into the network, the model parameters of the network are optimized using the boundary loss, and the trained network parameter file is saved; the trained model parameter file is loaded, the three-dimensional image data is input into the network, and the predicted segmentation result is obtained, and the segmentation result is represented by a probability map; the probability map is adjusted using a dense conditional random field, the edge of the segmentation result is optimized, and the maximum connected area algorithm is used to remove redundant false positive examples, and the post-processed image results are output and saved. The present invention can be applied to the three-dimensional image segmentation in the fields of materials science and medicine, and promote the progress and development of various disciplines.

[0126] The above describes the embodiments of the present invention in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments, and various changes can be made according to the purpose of the invention of the present invention. Any changes, modifications, substitutions, combinations or simplifications made according to the spirit and principle of the technical solution of the present invention should be equivalent replacement methods. As long as they meet the purpose of the invention of the present invention and do not deviate from the technical principles and inventive concepts of the three-dimensional image segmentation method based on the dual-path attention encoding and decoding network of the present invention, they all belong to the protection scope of the present invention.

Claims

1. A three-dimensional image segmentation method based on a dual-path attention encoder-decoder network, characterized in that: The steps include: (1) 3D image preprocessing: The original image used for training is randomly cropped into small image blocks, the small image blocks are preprocessed to obtain clearer images, and the preprocessed data is saved locally; (2) Training the dual-path attention encoder-decoder network: Build a dual-path attention encoder-decoder network, input the training set data into the network, use boundary loss to optimize the network model parameters, and save the trained network parameter file; (3) Use the trained model to predict: Load the trained model parameter file, input the test set data into the network, and obtain the predicted segmentation results, which are represented by probability maps; (4) Post-processing of network prediction results: Use dense conditional random fields to adjust the probability map, optimize the edges of the segmentation results, and then use the maximum connected area algorithm to remove redundant false positives, output and save the post-processed image results; In the step (2), a residual block, a maximum pooling, an average pooling and a dual-path block are used to construct an encoder network; a convolution block and a Softmax activation function are used to construct a dual-path attention network; and a residual block, a deconvolution block and a Sogmoid activation function are used to construct a decoder network; In step (2), the design and construction of the encoding network of the encoder includes the following specific steps: (2-1-1) A residual block is used in the first layer of the encoder to adapt to inputs of different data dimensions, and maximum pooling is used to reduce the dimension of the output of the first layer; (2-1-2) Two dual-path blocks are used in the second layer of the encoder to explore the low-level texture features of the image, and the output of the second layer is reduced in dimension using maximum pooling; (2-1-3) Use three dual-path blocks in the third layer of the encoder to explore high-level abstract features of the image, and use average pooling to reduce the dimension of the output of the third layer; (2-1-4) Five dual-path blocks are used in the fourth layer of the encoder to explore high-level abstract features of the image, and average pooling is used to reduce the dimension of the output of the fourth layer; (2-1-5) Two dual-path blocks are used in the fifth layer of the encoder to integrate the explored image features; The attention encoding network module consists of three submodules: (a) Position attention module; (b) Channel attention module; (c) Feature fusion module; The design and construction of the dual-path attention network module includes the following specific steps: (2-2-1) Place the position attention module and the channel attention module in parallel to explore spatial features and channel features; (2-2-2) The feature fusion module is used to integrate the explored spatial features and channel features; The design and construction of the encoder-decoder network includes the following specific steps: (2-3-1) In the first to fourth layers of the decoder, two residual blocks and one deconvolution block are alternately connected to gradually restore the image feature resolution; (2-3-2) In the fifth layer of the decoder, two residual blocks and one Sigmoid activation function are used to output the probability feature map.

2. The three-dimensional image segmentation method based on the dual-path attention encoding and decoding network according to claim 1, characterized in that: In step (1) of claim 1, the image preprocessing comprises the following specific steps: (1-1) Cropping the 3D image data into image blocks of 12×224×244 pixels; (1-2) Determine whether the image block is a grayscale image, and use a normalization algorithm to convert the non-grayscale image into a grayscale image; (1-3) Use Gaussian filtering to remove noise points in the image; (1-4) Use histogram equalization to stretch the grayscale distribution of the image and enhance the contrast of the image; (1-5) Use the Laplace operator to sharpen the edge of the image, enhance the grayscale mutation in the image, and reduce the area where the grayscale changes slowly; (1-6) Divide and save the preprocessed image data.

3. The three-dimensional image segmentation method based on the dual-path attention encoding and decoding network according to claim 1, characterized in that: In step (2), a boundary loss function is used to optimize the network model parameters, and the boundary loss is expressed as: L e =d+α·a+β·e (1) In the above formula (1), d, a and e represent distance, area and boundary respectively; α and β are weights corresponding to area and boundary respectively.

4. The image segmentation method based on the dual-path attention encoding and decoding network according to claim 1, characterized in that: In step (2), network training includes the following specific steps: (2-4-1) The Adam first-order optimization algorithm is used to optimize the network parameters and iteratively update the neural network weights based on the training data; (2-4-2) Set the weight decay coefficient to alleviate the problem of model overfitting; (2-4-3) In order to further obtain better network performance, the learning rate is set, and a scheme of dynamically reducing the learning rate is adopted to further approach the optimal value of the network parameters. When the loss value no longer decreases within a certain epoch, the learning rate lr is multiplied by the attenuation factor to reduce the learning rate.

5. The three-dimensional image segmentation method based on the dual-path attention encoding and decoding network according to claim 1, characterized in that: In the step (3), the prediction of the three-dimensional image includes the following specific steps: (3-1) Load the trained model parameter file; (3-2) Input the three-dimensional image data into the network to obtain the predicted segmentation results; (3-3) After completing the prediction of the small image blocks, the small image blocks are stitched together in order to form an image of the same size as the label image and saved locally in the form of a probability map.

6. The three-dimensional image segmentation method based on the dual-path attention encoding and decoding network according to claim 1, characterized in that: In step (4), the optimization of the network prediction result image includes the following steps: (4-1) The probability graph output by the network is regarded as an undirected complete graph, and each pixel point i is regarded as a node v∈V in a graph G(V,E), where the node v has a fixed category label x i and the corresponding observation value y i , the dependency between nodes is regarded as edge e∈E, thus using the average approximation algorithm to construct an undirected complete graph into a dense conditional random field; (4-2) Use dense conditional random fields to adjust the probabilities in the probability map and optimize the edge details of the segmented image; (4-3) Use the maximum region connectivity technique to reduce false positives and false negatives in the segmentation results, further improving the segmentation accuracy; (4-4) Output and save the post-processed image results.

Images