A lightweight segmentation method for colorectal cancer lesions based on deep learning

Abstract

The present invention proposes a lightweight segmentation method for colorectal cancer lesion areas based on deep learning, comprising the following steps: step S1, data processing: performing data preprocessing and data enhancement operations on the obtained abdominal image; step S2, network construction and training: the network is a neural network constructed based on a context-guided network, an efficient pyramid channel attention module is added to the context-guided block, and the residual idea is combined with the local features extracted by the context-guided block to perform neural network training; step S3, predictive segmentation: inputting test data into the neural network, reading the model weights saved during training and performing predictive segmentation, and comparing and evaluating the output image lesion area prediction segmentation result with the neural network label; the present invention can achieve relatively excellent segmentation effect, and can achieve a good balance between model lightweighting and performance.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to a lightweight segmentation method for colorectal cancer lesion areas based on deep learning. Background Art

[0002] The diagnosis and treatment of colorectal cancer has become an important research issue. The purpose of medical image segmentation is to segment the target area of interest in the medical image to provide a more reliable basis for disease diagnosis and treatment. In recent years, relying on the powerful feature learning ability of deep neural networks, deep learning has been successfully applied to a variety of medical image segmentation problems. The U-Net proposed by Ronneberger et al. is one of the most commonly used medical segmentation network models. It has an encoder-decoder structure, in which the encoder aims to capture higher-level semantic features and gradually reduce the spatial dimension of the feature map, and the decoder is used to restore object details and spatial dimensions. In order to further improve the performance of U-Net, many scholars have integrated network structures such as VGGNet proposed by Simonyan et al., ResNet proposed by He et al., and DenseNet proposed by Huang et al., which have stronger learning capabilities, into the U-Net framework. For example, Guo et al. used the VGG module as the encoder for a U-Net, achieving high segmentation accuracy in colorectal polyp segmentation. Abedalla et al. used a ResNet pre-trained on ImageNet as the encoder, achieving excellent results in pneumothorax segmentation in chest X-ray images. Li et al. designed an encoder based on the DenseNet model, improving the accuracy of liver and liver tumor segmentation. Furthermore, many researchers have incorporated attention mechanisms into U-Net networks to select features that are more critical to the task at hand. For example, Oktay et al. added a gated attention module to the U-Net framework, successfully applying it to the multi-label segmentation problem of abdominal CT.

[0003] The above methods designed different network structures for medical image segmentation. However, these network algorithms have some limitations: 1. While achieving good segmentation effects, the number of model parameters also increases significantly, consuming more hardware resources and time; 2. The network has some deficiencies in the detailed segmentation of lesions and cannot fully learn the global and local features of the image, resulting in poor segmentation of the lesion area. The detailed segmentation of the lesion area is prone to missed detection, and the number of network model parameters cannot be well controlled. Summary of the Invention

[0004] The present invention proposes a lightweight segmentation method for colorectal cancer lesion areas based on deep learning, which can achieve relatively excellent segmentation effect and achieve a good balance between model lightweight and performance.

[0005] The present invention adopts the following technical solutions.

[0006] A light-weight segmentation method for colorectal cancer lesion regions based on deep learning, comprising the following steps:

[0007] Step S1, data processing: performing data preprocessing and data enhancement operations on the obtained abdominal images;

[0008] Step S2, network construction and training: The network is a neural network built based on the context-guided network. An efficient pyramid channel attention module is added to the context-guided block. The residual idea is combined with the local features extracted by the context-guided block to perform neural network training;

[0009] Step S3, predictive segmentation: input test data into the neural network, read the model weights saved during training and perform predictive segmentation, and compare and evaluate the output image lesion area prediction segmentation result with the neural network label.

[0010] In step S1, the abdominal image is an abdominal MRI data set, and the data set is divided into a training set and a test set according to the required ratio. The data preprocessing is used to improve the lesion segmentation accuracy output in step S3. The method is as follows: after normalizing the image, using a Gaussian Laplacian filter to enhance the edge texture expression of the abdominal image, then combining the original image with the filtered image, and at the same time expanding the Gaussian smoothing component of the operator to reduce the noise component after superposition; wherein the formula of the Gaussian Laplacian filter is as follows:

[0011]

[0012] Where σ is the Gaussian standard deviation, LoG(x,y) is the pixel point after the Gaussian Laplace filter, x,y are the horizontal pixel and vertical pixel values respectively, and e is a natural constant.

[0013] The dataset is divided into training set and test set in a ratio of 3:2.

[0014] In step S2, the network model is an efficient pyramid context-guided network model, that is, a lightweight context-guided network model. On this basis, depthwise separable convolution is introduced to improve the model as a whole and strike a balance between lightness and high performance. An efficient pyramid channel attention module is added to the encoding module of the neural network to learn multi-scale global features in the image, and the residual idea is combined with the local features extracted by the context-guided block to learn and enrich the high-level and low-level semantic features in the image.

[0015] The network model is an EPCG-Net model with three downsampling parts. In the first downsampling part, only three standard convolutional layers are stacked for preliminary feature extraction. In the second and third parts, M and N EPCG blocks are stacked to downsample the feature map to the input image respectively.

[0016] For the second and third parts, the input of the first layer is obtained by combining the first and last blocks of the previous part, so that the image features are reused and the possibility of gradient disappearance is reduced;

[0017] Finally, a 1×1 convolutional layer and linear interpolation upsampling are used to generate segmentation predictions. Specifically, the hole convolution of the EPCG block for feature extraction adopts depth-wise separable convolution to further reduce the number of parameters, eliminating the cross-channel computational cost and avoiding excessive degradation of network performance. The attention mechanism adopts grouped convolution to reduce the number of parameters.

[0018] In the EPCG module, a 1×1 convolution operation is first performed, and then the vector after the convolution operation is fed into a 3×3 convolution and a 3×3 dilated depth-separable convolution respectively. Then, the two tensors of the ordinary convolution and the dilated depth-separable convolution are channel-superimposed, and then the combined tensor is batch-normalized. The PReLU activation function is added after the normalized output;

[0019] The 3×3 ordinary convolution is used to extract local features, and the 3×3 dilated depthwise separable convolution is used to extract local contextual features. By increasing the dilation rate of the dilated depthwise separable convolution, the range of extracted contextual features is gradually increased, so as to utilize the depthwise separable convolution to reduce the performance degradation caused by the depthwise convolution, while avoiding an excessive increase in the number of parameters, thereby achieving a balance between the number of parameters and performance;

[0020] The network is given nonlinear mapping learning capabilities through PReLU function activation. Specifically, the activated tensor is passed through the efficient pyramid channel attention module EPCA to alleviate gradient disappearance and enhance network performance. The processed tensor is added to the initial input tensor using the residual method.

[0021] The EPCA module uses the channel attention mechanism to allow the network to selectively weight the importance of each channel to help the network better learn the required features. Specifically, let X∈R C×H×WRepresent the input feature map, where C, H, and W represent its number of channels, height, and width, respectively. Set a SE block consisting of compression and excitation, which are used to encode global information and adaptively redistribute channel relationship weights, respectively. Generate a channel-level description by using global average pooling, which embeds global spatial information into the channel description. The global average pooling operator can be expressed by the following formula:

[0022]

[0023] The attention weight of the c-th channel in the SE module can be expressed as:

[0024] w c =σ(W1δ(W0(g c ))) Formula 3;

[0025] Among them, the symbol δ represents a specific linear unit ReLU operation, and Represents two fully connected (FC) layers; two fully connected layers are used to more effectively combine linear information between channels and facilitate the interaction of high- and low-channel dimensional information; the symbol σ represents the activation function, using the Sigmoid function; by using the activation function, weights are assigned to channels after channel interaction to more effectively extract information; the process of generating channel attention described above is called the SEWeight module;

[0026] In the EPCA module structure, in the efficient pyramid split attention module EPSA, when inputting image features, the channels are first divided evenly, and then the segmented feature tensors are fed into the group convolution;

[0027] In order to prevent the extraction of features of different scales from containing insufficient information and failing to give full play to the attention mechanism, the EPCA attention module is introduced. Specifically, firstly, the group convolution module GConv is used to perform group convolution operations of different scales on the input tensor in parallel to obtain multi-scale features F. i (i=0,1,2,3), which aims to extract features of different scales. The number of convolution groups at each scale increases according to the size of the convolution kernel to reduce the number of parameters brought by convolution of different scales and improve network efficiency. Then, the multi-scale features are reconnected on the channel to obtain the reconstructed feature tensor F. The SEWeight module extracts the channel attention weights of the multi-scale feature tensor to construct cross-dimensional interaction. The Softmax operation is used to recalibrate the attention weights of the corresponding channels to obtain AttWeight, thereby establishing channel dependency. The element-by-element product operation is applied to the recalibrated weights and the corresponding feature maps. Finally, a reconstructed feature map with richer multi-scale feature information is obtained as the output out. F iThe relationship between F, AttWeight, and out is shown in the following formula:

[0028] F=Cat([F1,…,F i ])i=1,2,3,4 Formula 4;

[0029]

[0030] AttWeightCat=([AttWeight1,…,AttWeight i ])i=0,1,2,3 Formula 6;

[0031]

[0032] In the GConv module structure, when inputting image features, four different convolution kernels of 3×3, 5×5, 7×7, and 9×9 are used for convolution, and the feature channel after each convolution is reduced to Channel, each convolution kernel produces a different spatial receptive field, and parallel convolution is used to simultaneously process multi-scale tensor information; grouped convolution is introduced to reduce the number of convolution parameters;

[0033] After the grouped convolution, the multi-scale features are channel-superimposed to form a recombined multi-scale feature tensor with the same size as the module input. The relationship between the number of groups corresponding to the grouped convolution and the convolution kernel size is:

[0034]

[0035] Among them, G is the number of groups, and K is the corresponding convolution kernel size.

[0036] The method proposed in this paper can achieve relatively excellent segmentation effect and strike a good balance between model lightweight and performance. Specifically:

[0037] Advantage 1:

[0038] The introduction of the multi-scale attention mechanism can pay more attention to multi-scale features, thereby avoiding missing lesion details.

[0039] Advantage 2:

[0040] By using the idea of lightweight networks, the number of network parameters can be reduced. While reducing the number of network parameters, some measures are taken to avoid excessive performance degradation caused by network lightweighting.

[0041] Advantage 3:

[0042] A good balance is achieved between network lightweight and performance, which has certain guiding significance for medical clinical diagnosis segmentation applications and future deployment on mobile devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0044] Attachment Figure 1 It is a schematic flow chart of the method of the present invention;

[0045] Attachment Figure 2 This is a schematic diagram of the before and after comparison of the images in the image dataset after the pre-processing filtering process ((a) is before processing, and (b) is after processing);

[0046] Attachment Figure 3 It is a schematic diagram of the overall framework for training the neural network model;

[0047] Attachment Figure 4 It is a structural diagram of the EPCG module;

[0048] Attachment Figure 5 It is a structural diagram of the EPCA module;

[0049] Attachment Figure 6 It is a structural diagram of the GConv module;

[0050] Attachment Figure 7 It is a comparative diagram of different technical solutions when testing image datasets (the present invention is the EPCG-Net solution);

[0051] Attachment Figure 8 is with Figure 7 Schematic diagram of the evaluation index results of the corresponding test experiment (the present invention is the EPCG-Net solution). DETAILED DESCRIPTION

[0052] As shown in the figure, a lightweight segmentation method for colorectal cancer lesion area based on deep learning includes the following steps:

[0053] Step S1, data processing: performing data preprocessing and data enhancement operations on the obtained abdominal images;

[0054] Step S2, network construction and training: The network is a neural network built based on the context-guided network. An efficient pyramid channel attention module is added to the context-guided block. The residual idea is combined with the local features extracted by the context-guided block to perform neural network training;

[0055] Step S3, predictive segmentation: input test data into the neural network, read the model weights saved during training and perform predictive segmentation, and compare and evaluate the output image lesion area prediction segmentation result with the neural network label.

[0056] like Figure 2As shown, in step S1, the abdominal image is an abdominal MRI data image dataset, and the dataset is divided into a training set and a test set according to a required ratio. The data preprocessing is used to improve the lesion segmentation accuracy output in step S3. The method is as follows: after normalizing the image, using a Gaussian Laplacian filter to enhance the edge texture expression of the abdominal image, then combining the original image with the filtered image, and at the same time expanding the Gaussian smoothing component of the operator to reduce the noise component after superposition; wherein the formula of the Gaussian Laplacian filter is as follows:

[0057]

[0058] Where σ is the Gaussian standard deviation, LoG(x,y) is the pixel point after the Gaussian Laplace filter, x,y are the horizontal pixel and vertical pixel values respectively, and e is a natural constant.

[0059] The dataset is divided into training set and test set in a ratio of 3:2.

[0060] In step S2, the network model is an efficient pyramid context-guided network model, that is, a lightweight context-guided network model. On this basis, depthwise separable convolution is introduced to improve the model as a whole and strike a balance between lightness and high performance. An efficient pyramid channel attention module is added to the encoding module of the neural network to learn multi-scale global features in the image, and the residual idea is combined with the local features extracted by the context-guided block to learn and enrich the high-level and low-level semantic features in the image.

[0061] like Figure 3 As shown in the figure, the network model is an EPCG-Net model with three downsampling parts. In the first downsampling part, only three standard convolutional layers are stacked for preliminary feature extraction. In the second and third parts, M and N EPCG blocks are stacked to downsample the feature map to the input image respectively. For the second and third parts, the input of the first layer is obtained by combining the first and last blocks of the previous part, so that the image features are reused and the possibility of gradient disappearance is reduced.

[0062] Finally, a 1×1 convolutional layer and linear interpolation upsampling are used to generate segmentation predictions. Specifically, the hole convolution of the EPCG block for feature extraction adopts depth-wise separable convolution to further reduce the number of parameters, eliminating the cross-channel computational cost and avoiding excessive degradation of network performance. The attention mechanism adopts grouped convolution to reduce the number of parameters.

[0063] like Figure 4As shown in the figure, in the EPCG module, a 1×1 convolution operation is first performed, and then the vector after the convolution operation is sent to a 3×3 convolution and a 3×3 dilated depth-separable convolution respectively. Then, the two tensors of the ordinary convolution and the dilated depth-separable convolution are channel-superimposed, and then the combined tensor is batch-normalized. After the normalized output, the PReLU activation function is added;

[0064] The 3×3 ordinary convolution is used to extract local features, and the 3×3 dilated depthwise separable convolution is used to extract local contextual features. By increasing the dilation rate of the dilated depthwise separable convolution, the range of extracted contextual features is gradually increased, so as to utilize the depthwise separable convolution to reduce the performance degradation caused by the depthwise convolution, while avoiding an excessive increase in the number of parameters, thereby achieving a balance between the number of parameters and performance;

[0065] The network is given nonlinear mapping learning capabilities through PReLU function activation. Specifically, the activated tensor is passed through the efficient pyramid channel attention module EPCA to alleviate gradient disappearance and enhance network performance. The processed tensor is added to the initial input tensor using the residual method.

[0066] like Figure 5 As shown, the EPCA module uses a channel attention mechanism to allow the network to selectively weight the importance of each channel to help the network better learn the required features. Specifically, let X∈R C×H×W Represent the input feature map, where C, H, and W represent its number of channels, height, and width, respectively. Set a SE block consisting of compression and excitation, which are used to encode global information and adaptively redistribute channel relationship weights, respectively. Generate a channel-level description by using global average pooling, which embeds global spatial information into the channel description. The global average pooling operator can be expressed by the following formula:

[0067]

[0068] The attention weight of the c-th channel in the SE module can be expressed as:

[0069] w c =σ(W1δ(W0(g c ))) Formula 3;

[0070] Among them, the symbol δ represents a specific linear unit ReLU operation, and Represents two fully connected (FC) layers; two fully connected layers are used to more effectively combine linear information between channels and facilitate the interaction of high- and low-channel dimensional information; the symbol σ represents the activation function, using the Sigmoid function; by using the activation function, weights are assigned to channels after channel interaction to more effectively extract information; the process of generating channel attention described above is called the SEWeight module;

[0071] In the EPCA module structure, in the efficient pyramid split attention module EPSA, when inputting image features, the channels are first divided evenly, and then the segmented feature tensors are fed into the group convolution;

[0072] In order to prevent the extraction of features of different scales from containing insufficient information and failing to give full play to the attention mechanism, the EPCA attention module is introduced. Specifically, firstly, the group convolution module GConv is used to perform group convolution operations of different scales on the input tensor in parallel to obtain multi-scale features F. i (i=0,1,2,3), which aims to extract features of different scales. The number of convolution groups at each scale increases according to the size of the convolution kernel to reduce the number of parameters brought by convolution of different scales and improve network efficiency. Then, the multi-scale features are reconnected on the channel to obtain the reconstructed feature tensor F. The SEWeight module extracts the channel attention weights of the multi-scale feature tensor to construct cross-dimensional interaction. The Softmax operation is used to recalibrate the attention weights of the corresponding channels to obtain AttWeight, thereby establishing channel dependency. The element-by-element product operation is applied to the recalibrated weights and the corresponding feature maps. Finally, a reconstructed feature map with richer multi-scale feature information is obtained as the output out. F i The relationship between F, AttWeight, and out is shown in the following formula:

[0073] F=Cat([F1,…,F i ])i=1,2,3,4 Formula 4;

[0074]

[0075] AttWeightCat=([AttWeight1,…,AttWeight i ])i=0,1,2,3 Formula 6;

[0076]

[0077] like Figure 6 As shown in the GConv module structure, when inputting image features, four different convolution kernels of 3×3, 5×5, 7×7, and 9×9 are used for convolution, and the feature channel after each convolution is reduced to Channel, each convolution kernel produces a different spatial receptive field, and parallel convolution is used to simultaneously process multi-scale tensor information; grouped convolution is introduced to reduce the number of convolution parameters;

[0078] After the grouped convolution, the multi-scale features are channel-superimposed to form a recombined multi-scale feature tensor with the same size as the module input. The relationship between the number of groups corresponding to the grouped convolution and the convolution kernel size is:

[0079]

[0080] Among them, G is the number of groups, and K is the corresponding convolution kernel size.

[0081] like Figure 7 、 Figure 8 To further verify the effectiveness of EPCG-Net, we compared its performance with the current classic segmentation network U-Net and its variants. This was also done on a dataset of abdominal images. For ease of comparison, both quantitative and qualitative analyses were performed on the network.

[0082] Figure 8 The table shows the experimental evaluation index results. It can be seen that EPCG-Net achieved performance comparable to AttU-Net, but the number of model parameters and floating-point size were much smaller than AttU-Net; compared with the basic U-Net, the number of model parameters and floating-point size of EPCG-Net were also significantly reduced. Compared with U-Net, ResU-Net, and DenseU-Net, Dice improved by 6.9%, 3.9%, and 0.3%, respectively, and mIoU improved by 7.8%, 4.2%, and 0.6%, respectively. In terms of the three indicators of Acc, Sen, and Spe, EPCG-Net also has a significant advantage over the other networks. At the same time, it can be seen that when EPCG-Net takes M3N21, the network performance decreases due to the increased network complexity and excessive depth.

[0083] The qualitative segmentation results on the dataset are as follows Figure 7 As shown in the figure, CG-Net suffers from some missed detections due to its lack of structures like the multi-scale pyramid attention mechanism, while U-Net and its variants exhibit a certain degree of misjudgment. Similarly, EPCG-Net exhibits a decline in visual segmentation performance when using M3N21 due to factors such as excessive network complexity. The qualitative segmentation results are consistent with the quantitative metrics, indicating that EPCG-Net's segmentation results closely represent the true gold standard.

[0084] In summary, the method proposed in this invention can achieve relatively excellent segmentation effect and achieve a good balance between model lightweight and performance.

Claims

1. A lightweight segmentation method for colorectal cancer lesions based on deep learning, characterized by: The following steps are included: Step S1, data processing: performing data preprocessing and data enhancement operations on the obtained abdominal images; Step S2, network construction and training: The network is a neural network built based on the context-guided network. An efficient pyramid channel attention module is added to the context-guided block. The residual idea is combined with the local features extracted by the context-guided block to perform neural network training; Step S3, predictive segmentation: input test data into the neural network, read the model weights saved during training and perform predictive segmentation, and compare and evaluate the output image lesion area prediction segmentation result with the neural network label; In step S2, the network model is an efficient pyramid context-guided network model, that is, a lightweight context-guided network model. On this basis, depthwise separable convolution is introduced to improve the overall model. An efficient pyramid channel attention module is added to the encoding module of the neural network to learn multi-scale global features in the image. The residual idea is combined with the local features extracted by the context-guided block to learn and enrich the high-level and low-level semantic features in the image. The network model is an EPCG-Net model with three downsampling parts. In the first downsampling part, only three standard convolutional layers are stacked for preliminary feature extraction. In the second and third parts, M and N EPCG blocks are stacked to downsample the feature map to the input image respectively. For the second and third parts, the input of the first layer is obtained by combining the first and last blocks of the previous part, so that the image features are reused; Finally, a 1×1 convolutional layer and linear interpolation upsampling are used to produce segmentation predictions. Specifically, the hole convolution of the EPCG block for feature extraction adopts depth-wise separable convolution; and the attention module adopts grouped convolution.

2. The method for lightweight segmentation of colorectal cancer lesions based on deep learning according to claim 1, characterized in that: In step S1, the abdominal image is an abdominal MRI data set, and the data set is divided into a training set and a test set according to the required ratio. The data preprocessing is used to improve the lesion segmentation accuracy output in step S3. The method is as follows: after normalizing the image, using a Gaussian Laplacian filter to enhance the edge texture expression of the abdominal image, then combining the original image with the filtered image, and at the same time expanding the Gaussian smoothing component of the operator; wherein the formula of the Gaussian Laplacian filter is as follows: Where σ is the Gaussian standard deviation, LoG(x,y) is the pixel point after the Gaussian Laplace filter, x,y are the horizontal pixel and vertical pixel values respectively, and e is a natural constant.

3. The method for lightweight segmentation of colorectal cancer lesions based on deep learning according to claim 2, characterized in that: The dataset is divided into training set and test set in a ratio of 3:

2.

4. The method for lightweight segmentation of colorectal cancer lesions based on deep learning according to claim 1, characterized in that: In the EPCG block, a 1×1 convolution operation is first performed, and then the vector after the convolution operation is fed into a 3×3 convolution and a 3×3 dilated depthwise separable convolution respectively. Then, the two tensors of the ordinary convolution and the dilated depthwise separable convolution are channel-wise superimposed, and then the combined tensor is batch-normalized. The PReLU activation function is added after the normalized output; The 3×3 ordinary convolution is used to extract local features, and the 3×3 hole depth-separable convolution is used to extract local context features; The network is activated by the PReLU function to enable nonlinear mapping learning capabilities. Specifically, the activated tensor is passed through the efficient pyramid channel attention module EPCA, and the processed tensor is added to the initial input tensor using the residual method.

5. The method for lightweight segmentation of colorectal cancer lesions based on deep learning according to claim 4, characterized in that: The EPCA module uses the channel attention mechanism to allow the network to selectively weight the importance of each channel to help the network better learn the required features. Specifically: Let X∈R C×H×W Represent the input feature map, where C, H, and W represent its number of channels, height, and width, respectively. Set a SE block consisting of compression and excitation, which are used to encode global information and adaptively redistribute channel relationship weights, respectively. Generate a channel-level description by using global average pooling, which embeds global spatial information into the channel description. The global average pooling operator is expressed by the following formula: The attention weight of the c-th channel in the SE module is expressed as: w c =σ(W1δ(W0(g c ))) Formula 3; Among them, the symbol δ represents a specific linear unit ReLU operation, and Represents two fully connected layers; the symbol σ represents the activation function, using the Sigmoid function; the above process of generating channel attention is called the SEWeight module; In the EPCA module structure, in the efficient pyramid split attention module EPSA, when inputting image features, the channels are first divided evenly, and then the segmented feature tensors are fed into the group convolution; The EPCA attention module is introduced. Specifically, first enter the group convolution module GConv and perform group convolution operations of different scales on the input tensor in parallel to obtain multi-scale features F i (i=0,1,2,3), and then reconnect the multi-scale features on the channel to obtain the reconstructed feature tensor F; extract the channel attention weights of the multi-scale feature tensor through the SEWeight module to construct cross-dimensional interaction, use the Softmax operation to recalibrate the attention weights of the corresponding channels to obtain AttWeight, thereby establishing channel dependency, and apply the element-by-element product operation to the recalibrated weights and the corresponding feature map; finally, obtain the reconstructed feature map with richer multi-scale feature information as the output out; F i The relationship between F, AttWeight, and out is shown in the following formula: F=Cat([F1,…,F i ])i=1,2,3,4 Formula 4; AttWeight = Cat([AttWeight0,…,AttWeight i ) i = 0, 1, 2, 3 Formula VI; 6. The method for lightweight segmentation of colorectal cancer lesions based on deep learning according to claim 5, characterized in that: In the GConv module structure, when inputting image features, four different convolution kernels of 3×3, 5×5, 7×7, and 9×9 are used for convolution. The feature channel after each convolution is reduced to Channel, each convolution kernel produces a different spatial receptive field, and parallel convolution is used to simultaneously process multi-scale tensor information; grouped convolution is introduced to reduce the number of convolution parameters; After the grouped convolution, the multi-scale features are channel-superimposed to form a recombined multi-scale feature tensor with the same size as the module input. The relationship between the number of groups corresponding to the grouped convolution and the convolution kernel size is: Among them, G is the number of groups, and K is the corresponding convolution kernel size.

Images