Intestinal Wall Vessel Segmentation Method Based on Multi-Scale Context Information and Attention Mechanism

Abstract

The present invention relates to a method for segmenting intestinal wall blood vessels based on multi-scale context information and attention mechanism. A neural network integrating multi-scale context information and attention mechanism is constructed, including a feature encoder. The multi-level outputs of the feature encoder are respectively passed through skip connections combined with channel attention modules and multi-scale context fusion modules combined with improved axial attention modules, and the obtained multi-level features are successively decoded and output through a decoder module. An intestinal wall blood vessel dataset is made and input into the neural network, and trained until the neural network is stable. After inputting the intestinal wall blood vessel image to be segmented, the stable neural network realizes the segmentation of intestinal wall blood vessels. The present invention greatly preserves the integrity of semantic information, improves the resolution ability of mucosal folds and blood vessels, pays more attention to the structural features of tiny blood vessels, improves the anti-interference ability, reduces the gap between high and low semantic information, enhances the ability to extract tiny blood vessels, better distinguishes mucosal folds from blood vessels, and has strong anti-interference ability.

Description

Technical Field

[0001] The present invention relates to image analysis, such as the technical field from bit images to non-bit images, and particularly relates to a method for segmenting intestinal wall blood vessels based on multi-scale context information and an attention mechanism in artificial intelligence image processing. Background Art

[0002] Colonoscopy is the gold standard for the diagnosis and treatment of colorectal diseases. However, due to the invasiveness of colonoscopy surgery, it may cause complications such as intestinal perforation. Intestinal perforation is usually caused by the lens coming into contact with the intestinal wall at a dangerous angle and then damaging the intestinal wall tissue, which is difficult to detect in time during the operation. If intestinal perforation occurs, intestinal contents may enter the abdominal cavity, causing diseases such as purulent peritonitis, and in severe cases, it may lead to death. Therefore, preventing intestinal perforation is an urgent problem to be solved.

[0003] According to senior endoscopists, when the colonoscope lens touches the intestinal wall, the distribution state of the blood vessels in the intestinal wall will change. Therefore, during colonoscopy surgery, doctors can prevent the occurrence of colon perforation by analyzing the distribution state of blood vessels in the image and estimating the relative position relationship between the colon lens and the intestinal wall. Intestinal wall blood vessel segmentation can provide a prerequisite for the subsequent judgment of the blood vessel distribution state. Therefore, studying an automatic segmentation method for intestinal wall blood vessels has important engineering application significance.

[0004] Currently, research work on blood vessel segmentation can be roughly classified into three main methods: The first is traditional image processing methods, such as segmentation methods based on wavelet transform and Gaussian filtering. These traditional methods are easily interfered by the subjective factors of researchers and require too much manual participation, reducing the scalability of the algorithm. The second is based on traditional machine learning methods, such as using methods like clustering and AdaBoost algorithm to first extract features and then use a classifier for pixel classification. The third is to use a method based on a convolutional neural network (CNN) for segmentation. This method effectively avoids the interference of researchers' subjective factors and can also achieve a good segmentation accuracy.

[0005] Although in the task of intestinal wall blood vessel segmentation, existing methods based on convolutional neural networks can already achieve good performance, these methods have deficiencies in the extraction ability of tiny blood vessels in the intestinal wall and the ability to distinguish between mucosal folds and blood vessels in the intestinal wall. At the same time, there are also interferences such as intestinal fluid and secretions in the intestine. Therefore, the task of intestinal wall blood vessel segmentation is extremely challenging. Summary of the Invention

[0006] The present invention solves the problems existing in the prior art and provides a method for segmenting intestinal wall blood vessels based on multi-scale context information and an attention mechanism, overcoming the problems of insufficient extraction ability of tiny blood vessels, difficulty in distinguishing mucosal folds from blood vessels, and weak anti-interference ability in segmentation.

[0007] The technical solution adopted by the present invention is an intestinal wall blood vessel segmentation method based on multi-scale context information and attention mechanism. The method constructs a neural network that fuses multi-scale context information and attention mechanism;

[0008] The neural network includes a feature encoder. The multi-level outputs of the feature encoder are respectively obtained through skip connections combined with channel attention modules and multi-scale context fusion modules combined with improved axial attention modules, and the multi-level features are successively decoded and output through a decoder module;

[0009] The intestinal wall blood vessel dataset is made and input into the neural network, trained until the neural network is stable, and after inputting the intestinal wall blood vessel image to be segmented, the intestinal wall blood vessel segmentation is realized by the stable neural network.

[0010] In the present invention, the multi-scale context fusion (MCF) module extracts multi-scale context information, which can improve the resolution ability for mucosal folds and blood vessels. At the same time, the improved axial attention (IAA) module in the module pays more attention to the structural features of tiny blood vessels, which can improve the anti-interference ability; the channel attention (CA) module is applied to the skip connection to fuse the effective low-level information in the feature encoder with the high-level information in the decoder, enhancing the ability to capture tiny blood vessels.

[0011] Preferably, the feature encoder includes 5 residual blocks. For any input image, a max-pooling layer is arranged between adjacent residual blocks; each residual block outputs corresponding features.

[0012] In the present invention, the feature encoder composed of 5 residual blocks aims to continuously extract features from the input image and solve the problem of gradient disappearance in deep networks. After the input image passes through the 5 residual blocks of the feature encoder, five levels of features are obtained respectively; when the input image enters the feature encoder, the detailed information will be lost as the number of layers deepens. Therefore, before the feature decoding module, the output feature of the last layer of the feature encoder is passed through the designed multi-scale context fusion module.

[0013] In the present invention, downsampling is performed every time the image passes through a residual block, and the downsampling operation is realized by max-pooling here.

[0014] Preferably, any one of the residual blocks includes two cascaded 3×3 convolutional layers and one 1×1 convolutional layer. The input image passes through the 3×3 convolutional layer and the 1×1 convolutional layer at the same time, and the two results are added pixel by pixel, and the corresponding features are output after passing through the ReLU layer of the activation function.

[0015] Preferably, the features output by the first four residual blocks are passed through a skip connection that combines a channel attention module, and the features output by the last residual block are passed through a multi-scale context fusion module that combines an improved axial attention module.

[0016] Preferably, the multi-scale context fusion module that combines the improved axial attention module includes two branches;

[0017] The first branch consists of a parallel structure of a 3×3 standard convolutional layer and a 3×3 dilated convolutional layer. The obtained feature information passes through the improved axial attention module to obtain local context information;

[0018] The second branch consists of a parallel structure of a 5×5 standard convolutional layer and a 5×5 dilated convolutional layer. The obtained feature information passes through the improved axial attention module to obtain global context information;

[0019] It outputs the sum of the local context information and the global context information.

[0020] Preferably, any one of the improved axial attention modules includes a 1×1 convolutional layer. The output of the 1×1 convolutional layer is respectively input into a vertical direction attention block and a horizontal direction attention module, and the outputs of the vertical direction attention block and the horizontal direction attention block are added together.

[0021] In the present invention, the standard convolution mainly extracts the local information of the target, and the dilated convolution adjusts the dilation rate and increases the receptive field size to obtain the multi-scale information of the target. These two branches can effectively integrate the local and global context information and improve the resolution ability for mucosal folds and blood vessels.

[0022] In the present invention, the feature information obtained from the two branches passes through the improved axial attention module (IAA) to construct the correlation between the local and global information. In order to make full use of the parameters of the convolutional kernel, the input feature map first passes through a 1×1 convolution to preserve the continuity and integrity of the relevant information, and then enters the vertical direction and horizontal direction attention structures respectively. This structure reshapes, transposes, and multiplies the shapes through 1×1 convolutions to aggregate different information. At the same time, the use of the attention mechanism can pay more attention to the structural features of tiny blood vessels, effectively reducing the interference of factors such as intestinal fluid and secretions. Finally, the vertical direction attention and horizontal direction attention results are effectively aggregated through the summation method for the feature map.

[0023] Preferably, the skip connection that combines the channel attention module includes 2 branches arranged in parallel;

[0024] One branch is used to output the image texture features, including a global max pooling layer, a fully connected layer, and a Sigmoid function layer arranged in sequence;

[0025] Another branch is used to output the overall data features, including a globally average pooling layer, a fully connected layer, and a Sigmoid function layer that are sequentially arranged;

[0026] The outputs of the two branches are added together.

[0027] In the present invention, the channel attention module is combined with skip connections, which can enhance the model's ability to capture tiny blood vessels while also significantly reducing the gap between high-level and low-level semantic information. Since max pooling can preserve image texture features and global average pooling can preserve overall data features, global max pooling and global average pooling are used to integrate different feature information respectively; the first four features obtained thereby have more tiny blood vessel features and the gap between high-level and low-level semantic information is reduced.

[0028] Preferably, the output of the multi-scale context fusion module combined with the improved axial attention module is restored to a feature map through four decoder modules, and the output channels of the result after restoring the feature map for each layer of decoding are concatenated with the outputs of the skip connections corresponding to each layer that combine the channel attention module until the final output result.

[0029] In the present invention, transposed convolution can restore more detailed feature information through adaptive mapping. Therefore, transposed convolution is used to restore the feature map. The feature map is restored through four decoder modules. After restoring the feature map for each layer of decoding, the output is concatenated with the output of the skip connection module corresponding to each layer, so that the effective low-level information in the feature encoder is fused with the high-level information in the decoder, enhancing the ability to capture tiny blood vessels; finally, the feature undergoes a 1×1 convolution for channel transformation to obtain the final output result, and the output result is then deeply supervised using GroundTruth to obtain the final model.

[0030] Preferably, in the method, intestinal wall images are collected, and the collected images are processed to remove the light spots in the images and remove sensitive information. At the same time, the blood vessels (true images) in the intestinal wall images are manually labeled to produce a corresponding intestinal wall blood vessel dataset.

[0031] Preferably, in the method, the intestinal wall blood vessel dataset is randomly divided into a training set and a test set, a gradient descent algorithm and a loss function are selected, and after adjusting the learning rate, training is carried out.

[0032] In the present invention, the designed model is trained using intestinal wall blood vessel images as input to obtain parameters, and then the trained parameters are used to input test set images to predict the final segmentation result.

[0033] The present invention relates to a method for segmenting intestinal wall blood vessels based on multi-scale context information and attention mechanism. A neural network integrating multi-scale context information and attention mechanism is constructed, including a feature encoder. The multi-level outputs of the feature encoder are respectively obtained through skip connections combined with channel attention modules and multi-scale context fusion modules combined with improved axial attention modules, and the multi-level features obtained are successively decoded and output through a decoder module. An intestinal wall blood vessel dataset is made and input into the neural network, trained until the neural network is stable, and after inputting an image of the intestinal wall blood vessel to be segmented, the stable neural network realizes the segmentation of the intestinal wall blood vessel.

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

[0035] (1) Based on the encoder-decoder structure, a feature encoder composed of residual blocks is used to prevent the problem of gradient disappearance in deep networks and greatly preserve the integrity of semantic information;

[0036] (2) The MCF module is used to aggregate context information, adaptively learn multi-scale features, improve the resolution ability for mucosal folds and blood vessels. At the same time, the improved axial attention module in the MCF module can pay more attention to the structural features of tiny blood vessels and improve the anti-interference ability;

[0037] (3) It is proposed to combine the CA module with skip connections, and through the attention mechanism, pay more attention to the features of tiny blood vessels and narrow the gap between high and low semantic information;

[0038] (4) Finally, the ability to extract tiny blood vessels is enhanced in the segmentation, better distinguish mucosal folds and blood vessels, and has strong anti-interference ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is the overall network structure diagram of the present invention, where Figure 1 (a) is the residual block in the feature encoder, Figure 1 (b) is the decoder module in the feature decoder;

[0040] Figure 2 is the structure diagram of the multi-scale context fusion module in the present invention;

[0041] Figure 3 is the structure diagram of the improved axial attention module in the multi-scale context fusion module of the present invention;

[0042] Figure 4 is the structure diagram of the channel attention module in the skip connection of the present invention;

[0043] Figure 5 is the comparison schematic diagram of the final segmentation results between the present invention and the comparative method. DETAILED DESCRIPTION OF THE INVENTION

[0044] The present invention will be further described in detail below in conjunction with embodiments, but the protection scope of the present invention is not limited thereto.

[0045] The present invention relates to a method for segmenting intestinal wall blood vessels based on multi-scale context information and attention mechanism, and the steps are as follows:

[0046] Step 1: Collect intestinal wall blood vessel images, and perform preprocessing operations such as removing light spots and sensitive information on the collected images. Professional physicians manually annotate the real blood vessel images to make an intestinal wall blood vessel dataset.

[0047] In the present invention, the real image here and the image segmented by the network model can be compared to obtain corresponding indicators, and at the same time, the loss of both can be obtained to optimize the network model.

[0048] Step 2: Input the intestinal wall blood vessel dataset X = {x1, x2,..., x n}, where X represents the input samples in the dataset, n represents the number of samples, represents the input image of the RGB three channels, and input the dataset into the proposed neural network.

[0049] Step 3: Design a feature encoder composed of 5 residual blocks to continuously extract features from the input image in Step 2, and at the same time solve the problem of gradient disappearance in deep networks. After the input image passes through the 5 residual blocks of the feature encoder, five levels of features {W i (x): i = 1, 2,..., 5} are obtained respectively.

[0050] As Figure 1 (a) shows, each residual block is composed of two cascaded 3×3 convolutions and a 1×1 convolution. The input image passes through the 3×3 convolution layer and the 1×1 convolution at the same time, then the results are added pixel by pixel, and finally passed through a ReLU activation function layer to ensure non-linearity. The image is downsampled after passing through each residual block. Here, the downsampling operation is implemented by max pooling. The input feature map x passes through a residual block (as Figure 1 (a) shows) and the downsampling is defined as follows:

[0051] W(x) = Down(Φ(Conv 3×3 (Conv 3×3 (x)) + Conv 1×1 (x))) (1)

[0052] In formula (1), Φ represents the ReLU activation function, Conv 3×3 , Conv 1×1 respectively represent 3×3 and 1×1 convolutions, Down represents downsampling, and x is the input feature map.

[0053] Step 4: When the input image enters the feature encoder, detailed information is lost as the number of layers deepens. Therefore, before the feature decoding module, the final output W5(x) of the feature encoder is passed through the designed MCF module;

[0054] As Figure 2 shown, this module has two branches. The first branch consists of a parallel structure of a 3×3 conventional convolution and a 3×3 dilated convolution (dilation rate of 3), and the second branch consists of a parallel structure of a 5×5 conventional convolution and a 5×5 dilated convolution (dilation rate of 5); among them, the conventional convolution mainly extracts the local information of the target, and the dilated convolution adjusts the dilation rate to increase the receptive field size and obtain the multi-scale information of the target. These two branches can effectively integrate local and global context information and improve the resolution ability of mucosal folds and blood vessels.

[0055] After that, the feature information obtained by the two branches will pass through the improved axial attention module to build the correlation between local and global information. In the improved axial attention module, in order to make full use of the parameters of the convolution kernel, the input feature map first passes through a 1×1 convolution to preserve the continuity and integrity of relevant information; then it enters the vertical and horizontal attention structures respectively. The vertical attention structure reshapes, transposes, and multiplies the shape in the vertical direction through three 1×1 convolutions, and the horizontal attention structure reshapes, transposes, and multiplies the shape in the horizontal direction through three 1×1 convolutions to aggregate different information; at the same time, the use of the attention mechanism can pay more attention to the structural features of tiny blood vessels and effectively reduce the interference of factors such as intestinal fluid and secretions; finally, the vertical attention and horizontal attention results are effectively aggregated into the feature map by summation.

[0056] Step 5: It is proposed to combine the channel attention module with skip connections to enhance the model's ability to capture tiny blood vessels and also greatly reduce the gap between high and low semantic information. Since max pooling can retain the image texture features and global average pooling can retain the overall data features, global max pooling and global average pooling are used to integrate different feature information respectively, and then through the fully connected layer FC(k), and finally through the Sigmoid function to obtain the channel attention weights; the fully connected layer FC(k) is implemented using a 1×1 one-dimensional convolution, and let the input image of the channel attention module Then there is the following definition:

[0057]

[0058]

[0059] S(x) = S GAP (x) + S GMP (x) (4)

[0060] In the formula, represents a one-dimensional convolution with a convolution kernel size of 1×1, σ is the Sigmoid activation function, respectively represent the output feature maps of the two branches, ⊙ is the matrix dot product, GAP represents the global average pooling operation, GMP represents the global maximum pooling operation, and S(x) represents the final output of the channel attention module.

[0061] For the first four hierarchical features {W i (x): i = 1, 2, 3, 4} obtained by the feature encoder in step 1, they respectively pass through the channel attention module to obtain the features {S i (x): i = 1, 2, 3, 4}. There are more microvascular features in S i (x), and at the same time, the gap between high and low semantic information can be greatly reduced.

[0062] Step 6: The transposed convolution can restore more detailed feature information through adaptive mapping. Therefore, the transposed convolution is used to restore the feature map, and a decoder module is designed based on the transposed convolution, as Figure 1 shown. The output F(x) of the MCF module in step 2 needs to pass through four decoder modules to restore the feature map. The features obtained at each layer are {F j (x): j = 1, 2, 3, 4}. At the same time, after completing the restoration of the feature map by each layer of decoding, F j (x) will be concatenated with the output S i (x) of the corresponding skip connection module in each layer, so that the effective low-level information in the feature encoder is fused with the high-level information in the decoder, enhancing the ability to capture microvessels.

[0063] The specific process can be defined as C i,j (x) = [S i (x), F j (x)], where i = 1, 2, 3, 4; j = 4, 3, 2, 1, S i (x) is the output of the channel attention module, F j (x) is the output of the decoder module, and C i,j represents the feature after channel concatenation.

[0064] Finally, the feature C 1,4 (x) undergoes a 1×1 convolution for channel transformation to obtain the final output result, and then the output result is deeply supervised using the GroundTruth to obtain the final model. The designed model is trained using the intestinal wall blood vessel images as input to obtain parameters, and then the trained parameters are used to input the test set images to predict the final segmentation result.

[0065] Step 7: The training strategy of the network model is as follows:

[0066] 7.1 First, randomly divide the intestinal wall blood vessel dataset into a training set and a test set. The resolution of the intestinal wall blood vessel images is 560×560;

[0067] 7.2 In the training process, the Adam algorithm is selected for the gradient descent algorithm;

[0068] 7.3 The selected loss function is a combined loss function composed of cross-entropy loss (Binary Cross Entropy, BCE) and Dice loss

[0069] Cross-entropy loss The calculation formula is as follows:

[0070]

[0071] In the formula, y i ∈{0,1}, p i ∈[0,1] represent the true label value and the pixel prediction probability value respectively.

[0072] Dice loss The calculation formula is as follows:

[0073]

[0074] In the formula, N is the total number of pixels, y i ∈{0,1}, p i ∈[0,1] are the true label value and the pixel prediction probability respectively.

[0075] Finally, the combined loss function can be expressed as

[0076] 7.4 Use the StepLR mechanism to adjust the learning rate. The batch size of the training samples fed into the network each time is 4, and the number of training times is 200 times.

[0077] The present invention provides an experimental example:

[0078] (1) Experimental conditions

[0079] The experiment used a workstation configured with an Intel(R) Xeon(R) Gold 6161 CPU @ 2.20GHz 2.2GHz (2 processors), 64GB of memory, a Windows operating system, and 2 Nvidia GTX 3080 Ti graphics cards. The model was implemented based on the PyTorch deep learning framework, with a PyTorch version of 1.8.0 and a Python version of 3.7. The ADAM algorithm was used to optimize the overall parameters, and the learning rate was set to 1e-4. The method of this patent was compared with four medical image segmentation methods: U-Net, CTF-Net, CE-Net, and CS-Net, respectively.

[0080] (2) Experimental results

[0081] The method proposed in this invention was compared with the U-Net, CTF-Net, CE-Net, and CS-Net networks on the same dataset. For the intestinal wall blood vessel segmentation task with an unbalanced foreground and background, in order to optimize the model, this patent used a combined loss function composed of cross-entropy loss and Dice loss. In addition, commonly used metrics in medical image segmentation, such as Accuracy, Sensitivity (also called Recall), Specificity, and F1-Score, were used to evaluate the model. The four metrics are expressed as equations (7), (8), (9), and (11) respectively:

[0082]

[0083]

[0084]

[0085]

[0086]

[0087] Among them, TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative respectively. True positive means that the intestinal wall blood vessel pixels are correctly segmented, true negative means that the background pixels are correctly segmented, false positive means that the background pixels are wrongly segmented as blood vessel pixels, and false negative means that the blood vessel pixels are wrongly segmented as background pixels. As shown in Table 1, it is a comparison table of the metrics of other neural networks and the network constructed in this invention;

[0088] Table 1 Comparison table of the metrics of other neural networks and the network constructed in this invention

[0089]

[0090]

[0091] As can be seen from Table 1, the model proposed in this patent obtains the highest F1 score of 73.57%, indicating that the model can accurately segment the background and blood vessels. At the same time, it also obtains the highest accuracy, specificity, and AUC, which are 94.76%, 97.76%, and 96.33% respectively, indicating that this patent can counteract and eliminate interferences such as mucosal folds and intestinal fluid accumulation. Comparing with the experimental indicators of U-Net, the F1 score of the model in this patent increases by 3.86%, the accuracy increases by 1.0%, the specificity increases by 0.65%, the sensitivity increases by 4.86%, and the AUC value increases by 2.2%. Generally speaking, this data proves that the method proposed in this patent is superior to other methods in the performance of intestinal wall blood vessel segmentation and has better practical engineering application value.

[0092] To achieve the above content, the present invention also stores an intestinal wall blood vessel segmentation program based on multi-scale context information and attention mechanism on a medium, and adopts an intestinal wall blood vessel segmentation method based on multi-scale context information and attention mechanism during device operation. Then, by inputting the intestinal wall blood vessel image to be segmented into the device, the segmented result is output through a neural network.

[0093] Those skilled in the art should understand that the embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer-usable program code.

[0094] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to generate a machine, so that the instructions executed by the processors of the computer or other programmable data processing devices generate means for realizing the functions specified in Figure 1 one or more processes or multiple processes and / or blocks Figure 1 one or more blocks or multiple blocks.

[0095] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured product including instruction means, and the instruction means realizes the functions in the processFigure 1 one process or multiple processes and / or blocks Figure 1 the functions specified in one block or multiple blocks.

[0096] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide for implementing in the process Figure 1 one process or multiple processes and / or blocks Figure 1 the steps of the functions specified in one block or multiple blocks.

[0097] Although the preferred embodiments of the present invention have been described, those skilled in the art can make additional changes and modifications to these embodiments once they learn the basic creative concept. Therefore, the appended claims are intended to be construed to include the preferred embodiments as well as all changes and modifications falling within the scope of the present invention.

[0098] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these changes and modifications.

Claims

1. An intestinal wall blood vessel segmentation method based on multi-scale context information and attention mechanism, characterized in that: The method constructs a neural network that fuses multi-scale context information and an attention mechanism; The neural network includes a feature encoder, and the feature encoder includes 5 residual blocks. For any input image, a max-pooling layer is arranged between adjacent residual blocks; each residual block outputs corresponding features; the features output by the first 4 residual blocks are passed through a skip connection that combines a channel attention module, and the features output by the last residual block are passed through a multi-scale context fusion module that combines an improved axial attention module; The multi-scale context fusion module that combines the improved axial attention module includes two branches; the first branch consists of a parallel structure of a 3×3 standard convolutional layer and a 3×3 dilated convolutional layer, and the obtained feature information passes through the improved axial attention module to obtain local context information; the second branch consists of a parallel structure of a 5×5 standard convolutional layer and a 5×5 dilated convolutional layer, and the obtained feature information passes through the improved axial attention module to obtain global context information; Output with the sum of the local context information and the global context information; Any one of the improved axial attention modules first passes through a 1×1 convolutional layer, and the output of the 1×1 convolutional layer is respectively input into a vertical direction attention block and a horizontal direction attention structure. The vertical direction attention structure reshapes, transposes, and multiplies the shape in the vertical direction through three convolutions respectively, and the horizontal direction attention structure reshapes, transposes, and multiplies the shape in the horizontal direction through three convolutions respectively to aggregate different information; finally, the outputs of the vertical direction attention block and the horizontal direction attention block are added; The obtained multi-level features are decoded and output successively through a decoder module; Make an intestinal wall blood vessel dataset and input it into the neural network, train until the neural network is stable, and after inputting the image of the intestinal wall blood vessel to be segmented, the stable neural network realizes the segmentation of the intestinal wall blood vessel.

2. The intestinal wall blood vessel segmentation method based on multi-scale context information and attention mechanism according to claim 1, characterized in that: Any one of the residual blocks includes two cascaded 3×3 convolutional layers and a 1×1 convolutional layer. The input image passes through the 3×3 convolutional layer and the 1×1 convolutional layer at the same time, and the two results are added pixel by pixel, and the corresponding features are output after passing through the ReLU activation function layer.

3. A method for intestinal wall blood vessel segmentation based on multi-scale context information and attention mechanism according to claim 1, characterized in that: The skip connection that combines the channel attention module includes two branches arranged in parallel; One branch is used to output image texture features, including a global max-pooling layer, a fully connected layer, and a Sigmoid function layer arranged in sequence; The other branch is used to output overall data features, including a global average-pooling layer, a fully connected layer, and a Sigmoid function layer arranged in sequence; The outputs of the above two branches are added.

4. A method for intestinal wall blood vessel segmentation based on multi-scale context information and attention mechanism according to claim 1, characterized in that: The output of the multi-scale context fusion module that combines the improved axial attention module restores the feature map through four decoder modules, and the result after restoring the feature map for each layer is concatenated with the output channels of the skip connections of the corresponding channel attention modules for each layer until the final output result.

5. A method for segmenting intestinal wall blood vessels based on multi-scale context information and attention mechanism according to claim 1, characterized in that: In the method, intestinal wall blood vessel images are collected, the collected images are processed to remove the light spots in the images and remove sensitive information, and the blood vessels in the intestinal wall blood vessel images are manually annotated to make the corresponding intestinal wall blood vessel dataset.

6. A method for intestinal wall blood vessel segmentation based on multi-scale context information and attention mechanism according to claim 1, characterized in that: In the method, the intestinal wall blood vessel dataset is randomly divided into a training set and a test set, the gradient descent algorithm and the loss function are selected, and training is carried out after adjusting the learning rate.

Images