A colonoscopy polyp image segmentation method based on hybrid model

Abstract

The present invention discloses a colonoscopic polyp image segmentation method based on a hybrid model, the method comprising: step S1: constructing a data set; step S2: constructing a polyp image segmentation network, importing the colonoscopic polyp image of the data set in step S1 into the Res2Net backbone, and obtaining the multi-level features of the image; step S3: importing the multi-level features of the image in step S2 into a feature extraction module; step S4: performing channel dimension splicing on the features in step S3 respectively through an expert routing module, obtaining a corresponding weight map, and further obtaining a combined feature map; step S5: a feature fusion module performs pooling convolution on the combined feature map of step S4 to obtain a segmented image. The feature extraction module designed by the present invention can capture different features in the polyp image, and the effective fusion of these features successfully improves the performance of the model; based on this, the polyp image is segmented and decoupled, and the interference between features can be reduced during the reasoning process to obtain a more accurate segmented image.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing, and in particular relates to a colonoscopic polyp image segmentation method based on a hybrid model. Background Art

[0002] Colorectal cancer is a common malignant tumor in the gastrointestinal tract. A wide range of deep learning methods have been developed to assist in colorectal cancer screening, early diagnosis, and treatment, thereby effectively improving patient survival rates.

[0003] Early traditional polyp image segmentation methods mainly rely on hand-crafted low-level features such as color, texture, and shape. Recent research has focused on deep learning-based methods and has achieved promising results. For example, PraNet uses parallel partial decoders to aggregate high-level features and adopts a reverse attention module to identify boundary cues. LDNet proposes a lesion-aware dynamic kernel for polyp segmentation, which enhances the flexibility of learning. H2Former introduces a multi-scale channel attention module to capture multi-scale dependencies.

[0004] However, despite the progress made by these methods, the performance degradation for small polyps remains a significant challenge. Small polyps are often similar to background normal tissues in texture and color, making them difficult to accurately identify. To address this problem, existing methods mainly focus on designing advanced attention mechanisms. For example, CaraNet introduces an axial reverse attention module to analyze target position information and multi-scale features to enhance the sensitivity to key information. SANet designs a shallow attention mechanism to filter out background noise from shallow features. However, these methods are usually based on the U-Net with a U-shaped structure, which has shown excellent versatility in many medical image segmentation tasks. Specifically, the decoder of the U-shaped network operates in a cascade manner, gradually recovering the decoded features from the deep layer to the shallow layer to generate the final output. However, in this process, features at different levels are concatenated, which may lose valuable information when continuously propagating in the decoder, thereby affecting the segmentation performance. In addition, the gradual concatenation of features also leads to increased redundancy and computational overhead. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art, the implementation of the present invention provides a colonoscopic polyp image segmentation method based on a hybrid model to solve the problems raised in the above-mentioned background technology.

[0006] The method comprises:

[0007] Step S1: constructing a data set, the data set including a number of colonoscopic polyp images;

[0008] Step S2: Construct a polyp image segmentation network, which includes a Res2Net backbone, a feature extraction module, an expert routing module, and a feature fusion module. Import the colonoscopy polyp image of the data set in step S1 into the Res2Net backbone to obtain the multi-level features of the image, which are the first image multi-level features , the second image multi-level features , the third image multi-level features and the fourth image multi-level features ;

[0009] Step S3: importing the multi-level features of the image in step S2 into the feature extraction module, respectively obtaining the first spatial image features for boundary perception, the second spatial image features for retaining rich internal features of the polyp, the third spatial image features for retaining external spatial details, and the multi-scale global spatial image features;

[0010] Step S4: using an expert routing module to perform channel dimension splicing on the first spatial image feature, the second spatial image feature, the third spatial image feature and the global spatial image feature in step S3, respectively, to obtain a corresponding weight map, and further obtain a combined feature map;

[0011] Step S5: The feature fusion module performs pooling convolution on the combined feature map of step S4 to obtain the final segmented image;

[0012] Step S6: construct a loss function and minimize the loss function to optimize the parameters of the polyp image segmentation network.

[0013] Further, the Res2Net backbone in step S2 includes a plurality of encoder blocks connected in series; the encoder blocks are respectively a first, a second, a third, a fourth and a fifth encoder block;

[0014] The process of obtaining the multi-level features of the image in step S2 is specifically as follows:

[0015] Import the images from the dataset in step S1 ,Through the encoder block of the Res2Net backbone, the multi-level features of the image are obtained ,in, , represents the set of real numbers, Indicates the height of the image. Indicates the width of the image. Indicates the number of channels corresponding to each layer of encoder block;

[0016] Eliminate multi-level features with large spatial resolution , get the first image multi-level features , the second image multi-level features , the third image multi-level features and the fourth image multi-level features .

[0017] Further, the feature extraction module in step S3 includes a first spatial expert component, a second spatial expert component, a third spatial expert component and a global spatial expert component;

[0018] Step S3 is specifically as follows:

[0019] Step S31:

[0020] The first spatial expert component performs an operation on the first image multi-level features in step S2. Perform feature extraction to obtain high-frequency first spatial image features for boundary perception ;

[0021] Step S31 is specifically as follows:

[0022] Step S311: The first image multi-level features generated in step S2 Separation into high frequency features and low frequency characteristics ; expressed as:

[0023] ;

[0024] ;

[0025] in, represents a convolution operation with a convolution kernel size of 3. The kernel size is , average pooling operation with a step size of 2;

[0026] Step S312: high-frequency features with rich boundary and contour information are transformed into With low-frequency features with global structure Perform feature interaction to obtain high-frequency boundary information and low frequency boundary information , expressed as:

[0027] ;

[0028] ;

[0029] in, represents a convolution operation with learnable parameters, Represents high frequency features The kernel size is The average pooling operation, represents the convolution kernel used to go from high-frequency input to high-frequency output, represents the convolution kernel used to go from low-frequency input to high-frequency output, represents the convolution kernel used to transform from high-frequency input to low-frequency output, represents the convolution kernel used to go from low-frequency input to low-frequency output, Represents the nearest neighbor interpolation obtained by convolution conduct times upsampling operation;

[0030] Step S313: discard low-frequency boundary information , only retain high-frequency boundary information First spatial image features for edge perception in images , expressed as:

[0031] ;

[0032] The second spatial expert component and the third spatial expert component receive the second image multi-level features respectively and the third image multi-level features As input, obtain the second space image features that retain the rich internal features of the polyp and third space image features that retain exterior space details ;

[0033] Step S32 is specifically as follows:

[0034] Step S321: compress the multi-level features of the second image through a convolution operation with a convolution kernel size of 1 and the third image multi-level features The number of channels can reduce information redundancy and computational complexity, which can be expressed as:

[0035] ;

[0036] ;

[0037] in, represents a convolution operation with a convolution kernel size of 1, represents the second compressed image multi-level features, represents the third compressed image multi-level feature;

[0038] Step S322: horizontal pooling and vertical pooling are then used to capture the context information of the second compressed image multi-level features and the third compressed image multi-level features in the horizontal direction and the vertical direction respectively, and broadcast addition is used to model the rectangular region of interest;

[0039] The process of obtaining context information in the horizontal and vertical directions is expressed as:

[0040] ;

[0041] ;

[0042] ;

[0043] in, and Represent horizontal pooling and vertical pooling operations respectively, and Respectively represent the context information in the horizontal and vertical directions, and together constitute a coarse region;

[0044] Step S323: The obtained rough area is passed through Convolution for refinement and obtain spatial attention map , that is, the rectangular region of interest, is expressed as:

[0045] ;

[0046] in, represents a convolution operation with a convolution kernel size of 3. Represents a broadcast addition operation;

[0047] Step S324: Obtain the spatial attention map Multi-level features with the second image and the third image multi-level features Multiply to obtain the second spatial image feature and third space image features ; expressed as:

[0048] ;

[0049] in, , Represents element-wise multiplication;

[0050] Step S33: Obtain multi-scale global space image features through the global space expert component ;

[0051] Step S33 is specifically as follows:

[0052] Step S331: The global spatial expert component extracts the fourth image multi-level features through different pooling layers and convolution layers, and uses different convolution kernel sizes and step sizes. The multi-scale features are respectively the first multi-scale features , the second multi-scale feature , the third multi-scale feature and the fourth multi-scale feature , expressed as:

[0053] ;

[0054] ;

[0055] ;

[0056] ;

[0057] in, represents global average pooling, represents an average pooling operation with a kernel size of 5 and a stride of 2. represents an average pooling operation with a kernel size of 3 and a stride of 1. Represents a convolution operation with a convolution kernel size of 1;

[0058] Step S332: Using the fourth multi-scale feature Further guide the first multi-scale features , the second multi-scale feature and the third multi-scale feature , and further obtain the first Interactive multi-scale features at different scales , expressed as:

[0059] ;

[0060] ;

[0061] in, Represents an upsampling operation;

[0062] Step S333: Interactive multi-scale features at different scales Connect with each other and conduct Convolution is used to reduce the number of channels; finally, the fourth image multi-level features of the input Add together to obtain global spatial image features , expressed as:

[0063] ;

[0064] in, Represents concatenation in the channel dimension.

[0065] Furthermore, step S4 is specifically as follows:

[0066] Step S41: The first spatial image feature , Second space image features , Third Space Image Features and global spatial image features splicing along the channel dimension and passing through a multi-layer perceptron and Activation function layer, processing the concatenated features into a weight map , expressed as:

[0067] ;

[0068] in, represents a multi-layer perceptron, represents the activation function layer;

[0069] Get four weight graphs , the size of each weight graph is , Indicates the height of the image. Indicates the width of the image;

[0070] Weight Graph The constraints are expressed as:

[0071] ;

[0072] Step S42: Upsample the first spatial image features , Second space image features , Third Space Image Features and global spatial image features According to the weight graph Re-weight and generate a combined feature map , expressed as:

[0073] ;

[0074] ;

[0075] in, Represents pixel-by-pixel multiplication.

[0076] Furthermore, step S5 is specifically as follows:

[0077] Combine feature maps The final segmented image is generated by further multi-layer perceptron fusion and convolution with a convolution kernel of size 1. , expressed as:

[0078] .

[0079] Furthermore, the loss function in step S6 It is expressed as:

[0080] ;

[0081] ;

[0082] in, Represents the true mask map of the image, Represents the polyp region segmentation and The combined loss of and They represent weighted binary cross entropy loss and weighted intersection-over-union loss respectively.

[0083] Compared with the prior art, the present invention has the following beneficial effects:

[0084] 1) The feature extraction module designed in the present invention can capture different features in the polyp image, including high-frequency first-space image features for boundary perception, second-space image features and third-space image features that respectively retain rich internal features and external spatial details of the polyp, and multi-scale global spatial image features; the effective fusion of the above features successfully improves the performance of the model; based on this, the polyp image can be segmented and decoupled, reducing interference between them during the reasoning process, thereby obtaining a more accurate segmented image.

[0085] 2) Furthermore, compared with the traditional U-shaped structure that supplements the encoder features to the parallel decoder with skip connections, which generates a huge computational burden, the present invention designs an expert routing module to dynamically select appropriate parts of different features to achieve efficient feature aggregation and generate the final segmented image. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 The present invention is a flow chart of the steps of a colonoscopic polyp image segmentation method based on a hybrid model. DETAILED DESCRIPTION

[0087] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0088] Reference Figure 1 , a colonoscopic polyp image segmentation method based on a hybrid model, in one example, the method includes:

[0089] Step S1: constructing a data set, the data set including a number of colonoscopic polyp images;

[0090] Step S2: Construct a polyp image segmentation network, which includes a Res2Net backbone, a feature extraction module, an expert routing module, and a feature fusion module. Import the colonoscopy polyp image of the data set in step S1 into the Res2Net backbone to obtain the multi-level features of the image, which are the first image multi-level features , the second image multi-level features , the third image multi-level features and the fourth image multi-level features ;

[0091] Step S3: importing the multi-level features of the image in step S2 into the feature extraction module, respectively obtaining the first spatial image features for boundary perception, the second spatial image features for retaining rich internal features of the polyp, the third spatial image features for retaining external spatial details, and the multi-scale global spatial image features;

[0092] Step S4: using an expert routing module to perform channel dimension splicing on the first spatial image feature, the second spatial image feature, the third spatial image feature and the global spatial image feature in step S3, respectively, to obtain a corresponding weight map, and further obtain a combined feature map;

[0093] Step S5: The feature fusion module performs pooling convolution on the combined feature map of step S4 to obtain the final segmented image;

[0094] Step S6: construct a loss function and minimize the loss function to optimize the parameters of the polyp image segmentation network.

[0095] Further, in one example, the Res2Net backbone in step S2 includes a plurality of encoder blocks connected in series; the encoder blocks are respectively a first, a second, a third, a fourth and a fifth encoder block;

[0096] The process of obtaining the multi-level features of the image in step S2 is specifically as follows:

[0097] Import the images from the dataset in step S1 ,Through the encoder block of the Res2Net backbone, the multi-level features of the image are obtained ,in, , represents the set of real numbers, Indicates the height of the image. Indicates the width of the image. Indicates the number of channels corresponding to each layer of encoder block;

[0098] Eliminate multi-level features with large spatial resolution , get the first image multi-level features , the second image multi-level features , the third image multi-level features and the fourth image multi-level features .

[0099] Further, in one example, the feature extraction module in step S3 includes a first spatial expert component, a second spatial expert component, a third spatial expert component and a global spatial expert component;

[0100] Step S3 is specifically as follows:

[0101] Step S31:

[0102] The first spatial expert component performs an operation on the first image multi-level features in step S2. Perform feature extraction to obtain high-frequency first spatial image features for boundary perception ;

[0103] Step S31 is specifically as follows:

[0104] Step S311: The first image multi-level features generated in step S2 Separation into high frequency features and low frequency characteristics ; expressed as:

[0105] ;

[0106] ;

[0107] in, represents a convolution operation with a convolution kernel size of 3. The kernel size is , average pooling operation with a step size of 2;

[0108] Step S312: high-frequency features with rich boundary and contour information are transformed into With low-frequency features with global structure Perform feature interaction to obtain high-frequency boundary information and low frequency boundary information , expressed as:

[0109] ;

[0110] ;

[0111] in, represents a convolution operation with learnable parameters, Represents high frequency features The kernel size is The average pooling operation, represents the convolution kernel used to go from high-frequency input to high-frequency output, represents the convolution kernel used to go from low-frequency input to high-frequency output, represents the convolution kernel used to transform from high-frequency input to low-frequency output, represents the convolution kernel used to go from low-frequency input to low-frequency output, Represents the nearest neighbor interpolation obtained by convolution conduct times upsampling operation;

[0112] Step S313: discard low-frequency boundary information , only retain high-frequency boundary information First spatial image features for edge perception in images , expressed as:

[0113] ;

[0114] Step S32: The second spatial expert component and the third spatial expert component receive the second image multi-level features respectively and the third image multi-level features As input, obtain the second space image features that retain the rich internal features of the polyp and third space image features that retain exterior space details ;

[0115] Step S32 is specifically as follows:

[0116] Step S321: compress the multi-level features of the second image through a convolution operation with a convolution kernel size of 1 and the third image multi-level features The number of channels can reduce information redundancy and computational complexity, which can be expressed as:

[0117] ;

[0118] ;

[0119] in, represents a convolution operation with a convolution kernel size of 1, represents the second compressed image multi-level features, represents the third compressed image multi-level feature;

[0120] Step S322: horizontal pooling and vertical pooling are then used to capture the context information of the second compressed image multi-level features and the third compressed image multi-level features in the horizontal direction and the vertical direction respectively, and broadcast addition is used to model the rectangular region of interest;

[0121] The process of obtaining context information in the horizontal and vertical directions is expressed as:

[0122] ;

[0123] ;

[0124] ;

[0125] in, and Represent horizontal pooling and vertical pooling operations respectively, and Respectively represent the context information in the horizontal and vertical directions, and together constitute a coarse region;

[0126] Step S323: The obtained rough area is passed through Convolution for refinement and obtain spatial attention map , that is, the rectangular region of interest, is expressed as:

[0127] ;

[0128] in, represents a convolution operation with a convolution kernel size of 3. Represents a broadcast addition operation;

[0129] Step S324: Obtain the spatial attention map Multi-level features with the second image and the third image multi-level features Multiply to obtain the second spatial image feature and third space image features ; expressed as:

[0130] ;

[0131] in, , Represents element-wise multiplication;

[0132] Step S33: Obtain multi-scale global space image features through the global space expert component ;

[0133] Step S33 is specifically as follows:

[0134] Step S331: The global spatial expert component extracts the fourth image multi-level features through different pooling layers and convolution layers, and uses different convolution kernel sizes and step sizes. The multi-scale features are respectively the first multi-scale features , the second multi-scale feature , the third multi-scale feature and the fourth multi-scale feature , expressed as:

[0135] ;

[0136] ;

[0137] ;

[0138] ;

[0139] in, represents global average pooling, represents an average pooling operation with a kernel size of 5 and a stride of 2. represents an average pooling operation with a kernel size of 3 and a stride of 1. Represents a convolution operation with a convolution kernel size of 1;

[0140] Step S332: Using the fourth multi-scale feature Further guide the first multi-scale features , the second multi-scale feature and the third multi-scale feature , and further obtain the first Interactive multi-scale features at different scales , expressed as:

[0141] ;

[0142] ;

[0143] in, Represents an upsampling operation;

[0144] Step S333: Interactive multi-scale features at different scales Connect with each other and conduct Convolution is used to reduce the number of channels; finally, the fourth image multi-level features of the input Add together to obtain global spatial image features , expressed as:

[0145] ;

[0146] in, Represents concatenation in the channel dimension.

[0147] Furthermore, in one example, step S4 is specifically as follows:

[0148] Step S41: The first spatial image feature , Second space image features , Third Space Image Features and global spatial image features splicing along the channel dimension and passing through a multi-layer perceptron and Activation function layer, processing the concatenated features into a weight map , expressed as:

[0149] ;

[0150] in, represents a multi-layer perceptron, represents the activation function layer;

[0151] Get four weight graphs , the size of each weight graph is , Indicates the height of the image. Indicates the width of the image;

[0152] Weight Graph The constraints are expressed as:

[0153] ;

[0154] Step S42: Upsample the first spatial image features , Second space image features , Third Space Image Features and global spatial image features According to the weight graph Re-weight and generate a combined feature map , expressed as:

[0155] ;

[0156] ;

[0157] in, Represents pixel-by-pixel multiplication.

[0158] Furthermore, in one example, step S5 is specifically as follows:

[0159] Combine feature maps The final segmented image is generated by further multi-layer perceptron fusion and convolution with a convolution kernel of size 1. , expressed as:

[0160] .

[0161] Furthermore, in one example, the loss function in step S6 is It is expressed as:

[0162] ;

[0163] ;

[0164] in, Represents the true mask map of the image, Represents the polyp region segmentation and The combined loss of and They represent weighted binary cross entropy loss and weighted intersection-over-union loss respectively.

[0165] Finally, it should be noted that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A colonoscopic polyp image segmentation method based on a hybrid model, characterized in that: The method comprises: Step S1: constructing a data set, the data set including a number of colonoscopic polyp images; Step S2: Construct a polyp image segmentation network, which includes a Res2Net backbone, a feature extraction module, an expert routing module, and a feature fusion module. Import the colonoscopy polyp image of the data set in step S1 into the Res2Net backbone to obtain the multi-level features of the image, which are the first image multi-level features , the second image multi-level features , the third image multi-level features and the fourth image multi-level features ; Step S3: importing the multi-level features of the image in step S2 into the feature extraction module, respectively obtaining the first spatial image features for boundary perception, the second spatial image features for retaining rich internal features of the polyp, the third spatial image features for retaining external spatial details, and the multi-scale global spatial image features; Step S4: using an expert routing module to perform channel dimension splicing on the first spatial image feature, the second spatial image feature, the third spatial image feature and the global spatial image feature in step S3, respectively, to obtain a corresponding weight map, and further obtain a combined feature map; Step S5: The feature fusion module performs pooling convolution on the combined feature map of step S4 to obtain the final segmented image; Step S6: constructing a loss function, minimizing the loss function to optimize the parameters of the polyp image segmentation network; The Res2Net backbone in step S2 includes a number of encoder blocks connected in series; the encoder blocks are the first, second, third, fourth and fifth encoder blocks; The process of obtaining the multi-level features of the image in step S2 is specifically as follows: Import the images from the dataset in step S1 ,Through the encoder block of the Res2Net backbone, the multi-level features of the image are obtained ,in, , represents the set of real numbers, Indicates the height of the image. Indicates the width of the image. Indicates the number of channels corresponding to each layer of encoder block; Eliminate multi-level features with large spatial resolution , get the first image multi-level features , the second image multi-level features , the third image multi-level features and the fourth image multi-level features ; The feature extraction module in step S3 includes a first spatial expert component, a second spatial expert component, a third spatial expert component and a global spatial expert component; Step S3 is specifically as follows: Step S31: The first spatial expert component performs an operation on the first image multi-level features in step S2. Perform feature extraction to obtain high-frequency first spatial image features for boundary perception ; Step S31 is specifically as follows: Step S311: The first image multi-level features generated in step S2 Separation into high frequency features and low frequency characteristics ; expressed as: ； ； in, represents a convolution operation with a convolution kernel size of 3. The kernel size is , average pooling operation with a step size of 2; Step S312: high-frequency features with rich boundary and contour information are transformed into With low-frequency features with global structure Perform feature interaction to obtain high-frequency boundary information and low frequency boundary information , expressed as: ； ； in, represents a convolution operation with learnable parameters, Represents high frequency features The kernel size is The average pooling operation, represents the convolution kernel used to go from high-frequency input to high-frequency output, represents the convolution kernel used to go from low-frequency input to high-frequency output, represents the convolution kernel used to transform from high-frequency input to low-frequency output, represents the convolution kernel used to go from low-frequency input to low-frequency output, Represents the nearest neighbor interpolation obtained by convolution conduct times upsampling operation; Step S313: discard low-frequency boundary information , only retain high-frequency boundary information First spatial image features for edge perception in images , expressed as: ； Step S32: The second spatial expert component and the third spatial expert component receive the second image multi-level features respectively and the third image multi-level features As input, obtain the second space image features that retain the rich internal features of the polyp and third space image features that retain exterior space details ; Step S32 is specifically as follows: Step S321: compress the multi-level features of the second image through a convolution operation with a convolution kernel size of 1 and the third image multi-level features The number of channels can reduce information redundancy and computational complexity, which can be expressed as: ； ； in, represents a convolution operation with a convolution kernel size of 1, represents the second compressed image multi-level features, represents the multi-level features of the third compressed image; Step S322: horizontal pooling and vertical pooling are then used to capture the context information of the second compressed image multi-level features and the third compressed image multi-level features in the horizontal direction and the vertical direction respectively, and broadcast addition is used to model the rectangular region of interest; The process of obtaining context information in the horizontal and vertical directions is expressed as: ； ； ； in, and Represent horizontal pooling and vertical pooling operations respectively, and Respectively represent the context information in the horizontal and vertical directions, and together constitute a coarse region; Step S323: The obtained rough area is passed through Convolution for refinement and obtain spatial attention map , that is, the rectangular region of interest, is expressed as: ； in, represents a convolution operation with a convolution kernel size of 3. Represents a broadcast addition operation; Step S324: Obtain the spatial attention map Multi-level features with the second image and the third image multi-level features Multiply to obtain the second spatial image feature and third space image features ; expressed as: ； in, , Represents element-wise multiplication; Step S33: Obtain multi-scale global space image features through the global space expert component ; Step S33 is specifically as follows: Step S331: The global spatial expert component extracts the fourth image multi-level features through different pooling layers and convolution layers, and uses different convolution kernel sizes and step sizes. The multi-scale features are respectively the first multi-scale features , the second multi-scale feature , the third multi-scale feature and the fourth multi-scale feature , expressed as: ； ； ； ； in, represents global average pooling, represents an average pooling operation with a kernel size of 5 and a stride of 2. represents an average pooling operation with a kernel size of 3 and a stride of 1. Represents a convolution operation with a convolution kernel size of 1; Step S332: Using the fourth multi-scale feature Further guide the first multi-scale features , the second multi-scale feature and the third multi-scale feature , and further obtain the first Interactive multi-scale features at different scales , expressed as: ； ； in, Represents an upsampling operation; Step S333: Interactive multi-scale features at different scales Connect with each other and conduct Convolution is used to reduce the number of channels; finally, the fourth image multi-level features of the input Add together to obtain global spatial image features , expressed as: ； in, Indicates splicing in the channel dimension; Step S4 is specifically as follows: Step S41: The first spatial image feature , Second space image features , Third Space Image Features and global spatial image features splicing along the channel dimension and passing through a multi-layer perceptron and Activation function layer, processing the concatenated features into a weight map , expressed as: ； in, represents a multi-layer perceptron, represents the activation function layer; Get four weight graphs , the size of each weight graph is , Indicates the height of the image. Indicates the width of the image; Weight Graph The constraints are expressed as: ； Step S42: Upsample the first spatial image features , Second space image features , Third Space Image Features and global spatial image features According to the weight graph Re-weight and generate a combined feature map , expressed as: ； ； in, Represents pixel-by-pixel multiplication.

2. The colonoscopic polyp image segmentation method based on a hybrid model as claimed in claim 1, characterized in that: Step S5 is specifically as follows: Combine feature maps The final segmented image is generated by further multi-layer perceptron fusion and convolution with a convolution kernel of size 1. , expressed as: 。 3. The colonoscopic polyp image segmentation method based on hybrid model as claimed in claim 2, characterized in that: Loss function in step S6 It is expressed as: ； ； in, Represents the true mask map of the image, Represents the polyp region segmentation and The combined loss of and They represent weighted binary cross entropy loss and weighted intersection-over-union loss respectively.

Images