Image segmentation method based on global perception U-Net model

Abstract

The invention discloses an image segmentation method based on a global perception U-Net model, particularly relates to the technical field of image segmentation, and is technically characterized in that the global perception U-Net model comprises an encoder, a decoder, a CFGC module and a CSCP module, a first layer of the encoder adopts a first convolution module and a maximum pooling module which are connected in sequence, and a second layer of the encoder adopts a second convolution module and a maximum pooling module which are connected in sequence; the first layer of the encoder adopts a first convolution module, the other four layers of the encoder adopt CFGC modules and maximum pooling modules which are connected in sequence, the last layer of the decoder adopts a second convolution module, the other four layers of the decoder adopt transpose convolution modules, CSCP modules and CFGC modules which are connected in sequence, and the first four layers of the encoder are in jump connection with the first four layers of the decoder respectively; and the CFGC module comprises an H-dimensional feature selection unit and a W-dimensional feature selection unit, and is used for selecting feature information under H and W dimensions under different scales, and performing feature enhancement on global features in a matrix multiplication manner based on the H-dimensional feature selection unit and the W-dimensional feature selection unit.

Description

Technical Field

[0001] The present invention relates to the technical field of image segmentation, and particularly to an image segmentation method based on a global perception U-Net model. Background Art

[0002] In the field of medical image segmentation, most organs, tissues, or lesion sites need to be extracted for medical analysis and research. Although organs or tissues of the same type, there are sometimes problems of inconsistent scales among different individuals, and some have a high degree of overlap with the background area, which leads to the high requirement for the segmentation accuracy of the network in the medical field.

[0003] To solve the above problems, most of the existing technologies focus on improving the network's analysis ability for objects of different scales. For example, by stacking multiple convolutional layers to obtain features of different scales and more fully realizing feature fusion. However, stacking too many convolutional layers, especially large kernels, will not only lead to a sudden increase in the number of parameters and computational complexity, but even lead to a decrease in the stability of the network; moreover, simply stacking convolutional layers cannot make the segmentation network get rid of the disadvantage that convolution pays more attention to local area information. On the contrary, Transformer gives full play to its advantage of focusing on global information in the field of medical image segmentation by virtue of its advantage of modeling long-range dependencies, but it brings a relatively high computational complexity, which will lead to a high-cost model deployment in practical applications.

[0004] In view of this, the present invention aims to provide an image segmentation method based on a global perception U-Net model to solve the above-mentioned related problems. Summary of the Invention

[0005] The technical problem to be solved by the present invention is that the prior art in the field of medical image segmentation is difficult to achieve the technical problem of simultaneously paying attention to local features and global features while reducing the computational complexity. The purpose is to provide an image segmentation method based on a global perception U-Net model. By setting CFGC modules with global perception collaborative fusion in each layer of the encoder and decoder, it can not only have the feature expression ability at different scales, but also strengthen the feature representation based on the global features at different scales; by setting CSCP modules and CFGC modules in the decoder, and making the encoder layers jump-connect to the CSCP modules of the decoder layers one by one, it can use the CFGC modules and CSCP modules to recover features layer by layer, and achieve effective fusion during the feature splicing process of each layer of encoder and decoder; at the same time, through the CSCP module that gradually fuses cross-channels, it can directly fuse the features in the encoder and decoder stages, optimize the semantic gap between the encoder and decoder without changing the jump connection structure, and will not cause additional semantic ambiguity; at the same time, using different divided channels of the CSCP module to respectively perform feature division on the feature images output from the encoder and the feature images output from the transposed convolution module, and gradually cross-splicing and fusing the divided deep features and shallow features, and performing global average pooling processing on different feature images through different global average pooling units to obtain different feature weight vectors, and combining the feature weight vectors with the spliced and fused feature images can simultaneously pay attention to the important semantic information in the encoder stage and the decoder channels.

[0006] The present invention is realized through the following technical solutions: An image segmentation method based on a global perception U-Net model, the method includes: Obtain the medical image to be segmented, input the medical image to be segmented into the global perception U-Net model for image segmentation processing, and obtain the segmented image segmentation result; Among them, the global perception U-Net model includes an encoder, a decoder, CFGC modules and CSCP modules. Both the encoder and the decoder adopt a five-layer structure connected in sequence. The first layer of the encoder adopts a first convolution module and a max pooling module connected in sequence. The remaining four layers of the encoder both adopt a CFGC module and a max pooling module connected in sequence. The last layer of the decoder adopts a second convolution module. The remaining four layers of the decoder both adopt a transposed convolution module, a CSCP module and a CFGC module connected in sequence. The first four layers of the encoder are respectively jump-connected to the first four layers of the decoder; the CFGC module includes H a dimensional feature selection unit and W a dimensional feature selection unit, which is used to select the feature information in the H and W dimensions at different scales, and based on H a dimensional feature selection unit and WThe dimension feature selection unit enhances the global features in the form of matrix multiplication; the CSCP module includes a feature image division channel, a feature splicing and fusion unit, and a global average pooling unit, which are used to perform channel-wise average division on the feature images in the encoder and decoder stages, splice the divided features in the encoder and decoder stages respectively, and complete step-by-step feature fusion based on the channel dependence relationship between the feature images in the encoder and decoder stages.

[0007] Furthermore, the first convolutional module in the first layer of the encoder is jump-connected to the CSCP module in the fourth layer of the decoder, the CFGC module in the second layer of the encoder is jump-connected to the CSCP module in the third layer of the decoder, the CFGC module in the third layer of the encoder is jump-connected to the CSCP module in the second layer of the decoder, and the CFGC module in the fourth layer of the encoder is jump-connected to the CSCP module in the first layer of the decoder.

[0008] Furthermore, the CFGC module in the last layer of the encoder is connected to the transposed convolutional module in the first layer of the decoder.

[0009] Furthermore, the CFGC module includes a first depth convolutional unit, a second depth convolutional unit, a first feature splicing and fusion unit, a first H dimension feature selection unit, a first W dimension feature selection unit, a second H dimension feature selection unit, a second W dimension feature selection unit, a second feature splicing and fusion unit, a third feature splicing and fusion unit, a third standard convolutional unit, a fourth standard convolutional unit, a fifth standard convolutional unit, a first feature shape reshaping unit, a second feature shape reshaping unit, a first global feature calculation unit, a first activation function, a second feature calculation unit, and a sixth standard convolutional unit; The input ends of the first depth convolutional unit and the second depth convolutional unit jointly serve as the input end of the CFGC module; The output end of the first depth convolutional unit is respectively connected to the input ends of the first feature splicing and fusion unit, the first H dimension feature selection unit, and the first W dimension feature selection unit. The output end of the second depth convolutional unit is respectively connected to the input ends of the first feature splicing and fusion unit, the second H dimension feature selection unit, and the second W dimension feature selection unit; The first H dimension feature selection unit and the second H dimension feature selection unit are both connected to the input end of the second feature splicing and fusion unit. The first W dimension feature selection unit and the second WThe output ends of the dimensional feature selection units are all connected to the input ends of the third feature splicing and fusion unit; The output end of the second feature splicing and fusion unit is connected to the input end of the third standard convolution unit, the output end of the first feature splicing and fusion unit is connected to the input end of the fourth standard convolution unit, and the output end of the third feature splicing and fusion unit is connected to the input end of the fifth standard convolution unit; The output end of the third standard convolution unit is connected to the input end of the first feature shape reshaping unit, and the output end of the fifth standard convolution unit is connected to the input end of the second feature shape reshaping unit; The output ends of the first feature shape reshaping unit and the second feature shape reshaping unit are both connected to the input end of the first global feature calculation unit, the output end of the first global feature calculation unit is connected to the first activation function, the output ends of the first activation function and the fourth standard convolution unit are both connected to the input end of the second feature calculation unit, the output end of the second feature calculation unit is connected to the input end of the sixth standard convolution unit, and the output end of the sixth standard convolution unit serves as the output end of the CFGC module.

[0010] Furthermore, the CSCP module includes a first feature image division channel, a second feature image division channel, a fourth feature splicing and fusion unit, a fifth feature splicing and fusion unit, a seventh standard convolution unit, an eighth standard convolution unit, a third feature product calculation unit, a fourth feature product calculation unit, a sixth feature splicing and fusion unit, a ninth standard convolution unit, a first global average pooling unit, and a second global average pooling unit; The input end of the first feature image division channel and the input end of the first global average pooling unit jointly serve as the first input end of the CSCP, the input end of the second feature image division channel and the input end of the second global average pooling unit jointly serve as the second input end of the CSCP, the output end of the first feature image division channel is respectively connected to the input ends of the fourth feature splicing and fusion unit and the fifth feature splicing and fusion unit, the output end of the second feature image division channel is respectively connected to the input ends of the fourth feature splicing and fusion unit and the fifth feature splicing and fusion unit, the output end of the fourth feature splicing and fusion unit is connected to the input end of the seventh standard convolution unit, and the output end of the fifth feature splicing and fusion unit is connected to the input end of the eighth standard convolution unit; The output ends of the seventh standard convolution unit and the first global average pooling unit are both connected to the input end of the third feature product calculation unit, and the output ends of the eighth standard convolution unit and the second global average pooling unit are both connected to the input end of the fourth feature product calculation unit; The output ends of the third feature product calculation unit and the fourth feature product calculation unit are both connected to the input end of the sixth feature splicing and fusion unit. The output end of the sixth feature splicing and fusion unit is connected to the input end of the ninth standard convolution unit. The output end of the ninth standard convolution unit serves as the output end of the CSCP module.

[0011] Further, the first depth convolution unit and the second depth convolution unit are both used to receive the first feature image input to the CFGC module, and respectively perform depth convolution processing on the first feature image at different scales to obtain the second feature image and the third feature image. The first feature splicing and fusion unit is used to splice the second feature image and the third feature image, and fuse the spliced features through the fourth standard convolution unit to obtain the fourth feature image. First H dimensional feature selection unit is used to select and splice the maximum feature and the average feature of the second feature image in the H dimension to obtain the first H dimensional feature; the first W dimensional feature selection unit is used to select and splice the maximum feature and the average feature of the second feature image in the W dimension to obtain the first W dimensional feature; the second H dimensional feature selection unit is used to select and splice the maximum feature and the average feature of the third feature image in the H dimension to obtain the second H dimensional feature, and the second W dimensional feature selection unit is used to select and splice the maximum feature and the average feature of the third feature image in the W dimension to obtain the second W dimensional feature; The second feature splicing and fusion unit is used to splice the first H dimensional feature and the second H dimensional feature to obtain the fifth feature image; the third feature splicing and fusion unit is used to splice the first W dimensional feature and the second W dimensional feature to obtain the sixth feature image; The third standard convolution unit is used to perform standard convolution processing on the fifth feature image. The first feature shape reshaping unit is used to reshape the feature shape of the fifth feature image after the convolution processing by the third standard convolution unit to obtain the first reshaped feature image. The fifth standard convolution unit is used to perform standard convolution processing on the sixth feature image. The second feature shape reshaping unit is used to reshape the feature shape of the sixth feature image after the convolution processing by the fifth standard convolution unit to obtain the second reshaped feature image. The first global feature calculation unit is used to perform matrix calculations on the first reshaped feature image and the second reshaped feature image to obtain a global feature representation matrix; the first activation function is used to activate the global feature representation matrix to obtain a global response matrix; The second feature calculation unit is used to perform a Hadamard product calculation on the fourth feature image and the global response matrix to obtain a seventh feature image; The seventh feature image is input into the sixth standard convolution unit for convolution fusion to obtain the output feature image of the CFGC module.

[0012] Further, the first feature image channel division and the first global average pooling unit are used to receive the eighth feature image input to the CSCP module, the second feature image channel division and the second global average pooling unit are used to receive the ninth feature image input to the CSCP module, the first feature image channel division is used to perform an average division of the eighth feature image on the channels to obtain a tenth feature image and an eleventh feature image, the first global average pooling unit is used to perform global average pooling on the eighth feature image to obtain a first encoder global weight vector, the second feature image channel division is used to perform an average division of the ninth feature image on the channels to obtain a twelfth feature image and a thirteenth feature image, and the second global average pooling unit is used to perform global average pooling on the ninth feature image to obtain a second decoder global weight vector; The fourth feature splicing and fusion unit is used to splice the tenth feature image and the twelfth feature image, and perform convolution fusion on the spliced feature image through the seventh standard convolution unit to obtain a fourteenth feature image; The fifth feature splicing and fusion unit is used to splice the eleventh feature image and the thirteenth feature image, and perform convolution fusion on the spliced feature image through the eighth standard convolution unit to obtain a fifteenth feature image; The third feature product calculation unit is used to perform a product calculation on the first encoder global weight vector and the fourteenth feature image to obtain a sixteenth feature image, and the fourth feature product calculation unit is used to perform a product calculation on the second decoder global weight vector and the fifteenth feature image to obtain a seventeenth feature image; The sixth feature splicing and fusion unit is used to splice the fifteenth feature image and the sixteenth feature image, and perform feature fusion on the spliced feature image using the ninth standard convolution unit to obtain the output feature image of the CSCP module.

[0013] Further, the convolution kernels of the first depth convolution unit, the third standard convolution unit, the fifth standard convolution unit, and the sixth standard convolution unit are all 3×3, the convolution kernel of the second depth convolution unit is 7×7, and the convolution kernel of the fourth standard convolution unit is 1×1.

[0014] Further, the convolutional kernels of the seventh standard convolutional unit, the eighth standard convolutional unit, and the ninth standard convolutional unit are all 1×1.

[0015] Further, the first activation function uses an activation function.

[0016] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1. In the present invention, by setting CFGC modules with global perception collaborative fusion in each layer of the encoder and decoder, it is possible to not only have the ability to express features at different scales, but also strengthen feature representation based on global features at different scales; by setting CSCP modules and CFGC modules in the decoder, and making the encoder layers jump-connect with the CSCP modules of the decoder layers one by one, it is possible to use the CFGC modules and CSCP modules to recover features layer by layer, and achieve effective fusion during the feature splicing process of each layer of the encoder and decoder; at the same time, through the CSCP module that gradually fuses cross-channels, it is possible to directly fuse the features in the encoder and decoder stages, optimize the semantic gap between the encoder and decoder without changing the jump-connection structure, and will not cause additional semantic ambiguity; at the same time, using different divided channels of the CSCP module to respectively divide the feature images output from the encoder and the feature images output from the transposed convolution module, and gradually cross-splicing and fusing the divided deep features and shallow features, and performing global average pooling processing on different feature images through different global average pooling units to obtain different feature weight vectors, and combining the feature weight vectors with the spliced and fused feature images, it is possible to simultaneously focus on the important semantic information in the encoder stage and the decoder channels.

[0017] 2. In the present invention, by using two different depth convolutional units in the CFGC module with global perception collaborative fusion, it is possible to capture local features from different scales and select these two scale features at the same time, thus making up for the limitation that convolution pays more attention to local region features. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following will briefly introduce the drawings required for the embodiments. It should be understood that the following drawings only show some embodiments of the present invention, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts. In the drawings: Figure 1 is a schematic structural diagram of the X-UNet model in an image segmentation method based on a global perception U-Net model in this embodiment; Figure 2It is a schematic structural diagram of the CFGC module in an image segmentation method based on a global perception U-Net model in this embodiment; Figure 3 It is a schematic diagram of the operation process of the CSCP module in an image segmentation method based on a global perception U-Net model in this embodiment; Figure 4 It is a schematic structural diagram of the CSCP module in an image segmentation method based on a global perception U-Net model in this embodiment; Figure 5 It is a schematic diagram of the operation process of the CSCP module in an image segmentation method based on a global perception U-Net model in this embodiment; Figure 6 It is a schematic diagram of model optimization in an image segmentation method based on a global perception U-Net model in this embodiment; Figure 7 It is a schematic diagram of comparison results in an image segmentation method based on a global perception U-Net model in this embodiment. Specific embodiments

[0019] The following describes exemplary embodiments of the present disclosure with reference to the accompanying drawings. Various details of the embodiments of the present disclosure are included to assist understanding, and they should be considered merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of the present disclosure. Similarly, for clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0020] In the present disclosure, unless otherwise specified, the terms "first", "second", etc. are used to describe various elements and are not intended to limit the positional relationship, temporal relationship, or importance relationship of these elements. Such terms are only used to distinguish one element from another. In some examples, the first element and the second element may refer to the same instance of the element, and in certain cases, based on the context description, they may also refer to different instances.

[0021] The terms used in the description of various examples in the present disclosure are only for the purpose of describing specific examples and are not intended to be limiting. Unless the context clearly indicates otherwise, if the number of elements is not specifically limited, the element may be one or more. In addition, the term " / and" used in the present disclosure covers any one of the listed items and all possible combinations.

[0022] Embodiment 1

[0023] See Figure 1 As shown, in this embodiment, an image segmentation method based on a global perception U-Net model is provided. The method includes: Obtain the medical image to be segmented, input the medical image to be segmented into the global perception U-Net model for image segmentation processing, and obtain the segmented image segmentation result; Among them, the global perception U-Net model (X-UNet model) includes an encoder, a decoder, a CFGC module, and a CSCP module. Both the encoder and the decoder adopt a five-layer structure connected in sequence. The first layer of the encoder adopts a first convolution module and a max pooling module connected in sequence. The remaining four layers of the encoder all adopt a CFGC module and a max pooling module connected in sequence. The last layer of the decoder adopts a second convolution module. The remaining four layers of the decoder all adopt a transposed convolution module, a CSCP module, and a CFGC module connected in sequence. The first four layers of the encoder are respectively skip-connected to the first four layers of the decoder. Specifically: the first convolution module of the first layer of the encoder is skip-connected to the CSCP module of the fourth layer of the decoder, the CFGC module of the second layer of the encoder is skip-connected to the CSCP module of the third layer of the decoder, the CFGC module of the third layer of the encoder is skip-connected to the CSCP module of the second layer of the decoder, and the CFGC module of the fourth layer of the encoder is skip-connected to the CSCP module of the first layer of the decoder.

[0024] It should be noted that in this embodiment, the global perception U-Net model is similar to the existing U-Net model architecture and is obtained by improving the existing U-Net model. The global perception U-Net model is a U-shaped encoder-decoder architecture; at the same time, it should be noted that in this embodiment, the number of filters in each layer of the encoder and the decoder is 32, 64, 128, 256, and 512 in sequence from the first layer to the fifth layer; and in order to improve the stability during training and introduce non-linearity, a BN unit and a ReLU activation function are set after all convolutional units of the global perception U-Net model in this embodiment. At the same time, the max pooling module and the transposed convolution module both adopt a max pooling layer with a convolution kernel of 2×2 and a transposed convolution layer with a convolution kernel of 2×2 in the conventional U-Net model. This technical solution is a conventional technical means in the art and will not be elaborated here.

[0025] At the same time, it should be noted that since the channel dimension of the first layer of the encoder is relatively low, the performance of depth convolution will decline in a relatively low channel dimension. Therefore, the first layer of the encoder in this embodiment adopts a first convolution module with a convolution kernel of 3×3 to extract primary semantic information; and then a max pooling module with a kernel size of 2×2 is used to complete the first downsampling operation.

[0026] The working process of the global perception U-Net model is specifically as follows: First, a medical image to be segmented with a size of 256×256×3 (H×W×C) is input into the first convolutional module of the first layer of the encoder, and a conventional convolution operation with a convolution kernel of 3×3, a step size of 1, and a padding of 3 is performed, and an image with a size of 256×256×32 is output; then the output image is respectively input into the max-pooling module of the first layer of the encoder and the CSCP module of the fourth layer of the decoder. The max-pooling module of the first layer of the encoder performs a max-pooling operation with a kernel of 2×2, a step size of 2, and a padding of 1 on it, and outputs an image with a size of 128×128×32 to the CFGC module of the second layer of the encoder; After the CFGC module of the second layer of the encoder processes the received image, an image with a size of 128×128×64 is output; then the output image is respectively input into the max-pooling module of the second layer of the encoder and the CSCP module of the third layer of the decoder. The max-pooling module of the second layer of the encoder performs a max-pooling operation with a kernel of 2×2, a step size of 2, and a padding of 1 on it, and outputs an image with a size of 64×64×64 to the CFGC module of the third layer of the encoder; After the CFGC module of the third layer of the encoder processes the received image, an image with a size of 64×64×128 is output; then the output image is respectively input into the max-pooling module of the third layer of the encoder and the CSCP module of the second layer of the decoder. The max-pooling module of the third layer of the encoder performs a max-pooling operation with a kernel of 2×2, a step size of 2, and a padding of 1 on it, and outputs an image with a size of 32×32×128 to the CFGC module of the fourth layer of the encoder; After the CFGC module of the fourth layer of the encoder processes the received image, an image with a size of 32×32×256 is output; then the output image is respectively input into the max-pooling module of the fourth layer of the encoder and the CSCP module of the first layer of the decoder. The max-pooling module of the fourth layer of the encoder performs a max-pooling operation with a kernel of 2×2, a step size of 2, and a padding of 1 on it, and outputs an image with a size of 16×16×256 to the CFGC module of the fifth layer of the encoder; After the CFGC module of the fifth layer of the encoder processes the received image, an image with a size of 16×16×512 is output, and then the output image is input into the transposed convolution module of the first layer of the decoder; The transposed convolution module in the first layer of the decoder performs deconvolution on the received image with a convolution kernel of 2×2 and a stride of 2, and outputs an image with a size of 32×32×256. Then, the output image is input into the CSCP module in the first layer of the decoder; the CSCP module in the first layer of the decoder processes the image input by the transposed convolution module and the image output by the fourth layer of the encoder, and outputs an image with a size of 32×32×512. Then, the output image is input into the CFGC module in the first layer of the decoder; the CFGC module in the first layer of the decoder processes the input image and outputs an image with a size of 32×32×256. Then, the output image is input into the transposed convolution module in the second layer of the decoder; The transposed convolution module in the second layer of the decoder performs deconvolution on the received image with a convolution kernel of 2×2 and a stride of 2, and outputs an image with a size of 64×64×128. Then, the output image is input into the CSCP module in the second layer of the decoder; the CSCP module in the second layer of the decoder processes the image input by the transposed convolution module and the image output by the third layer of the encoder, and outputs an image with a size of 64×64×256. Then, the output image is input into the CFGC module in the second layer of the decoder; the CFGC module in the second layer of the decoder processes the input image and outputs an image with a size of 64×64×128. Then, the output image is input into the transposed convolution module in the third layer of the decoder; The transposed convolution module in the third layer of the decoder performs deconvolution on the received image with a convolution kernel of 2×2 and a stride of 2, and outputs an image with a size of 128×128×64. Then, the output image is input into the CSCP module in the third layer of the decoder; the CSCP module in the third layer of the decoder processes the image input by the transposed convolution module and the image output by the second layer of the encoder, and outputs an image with a size of 128×128×128. Then, the output image is input into the CFGC module in the third layer of the decoder; the CFGC module in the third layer of the decoder processes the input image and outputs an image with a size of 128×128×64. Then, the output image is input into the transposed convolution module in the fourth layer of the decoder; The transposed convolution module in the fourth layer of the decoder performs deconvolution on the received image with a convolution kernel of 2×2 and a stride of 2, and outputs an image with a size of 256×256×32. Then, the output image is input into the CSCP module in the fourth layer of the decoder; the CSCP module in the fourth layer of the decoder processes the image input by the transposed convolution module and the image output by the first layer of the encoder, and outputs an image with a size of 256×256×64. Then, the output image is input into the CFGC module in the fourth layer of the decoder; the CFGC module in the fourth layer of the decoder processes the input image and outputs an image with a size of 256×256×32. Then, the output image is input into the second convolution module in the fifth layer of the decoder; The second convolutional module of the fifth layer of the decoder performs a 1×1 convolution operation on the received image and outputs an image with a size of 256×256×1.

[0027] Specifically, in this embodiment, by setting the CFGC module with global perception collaborative fusion in each layer of the encoder and decoder, it is possible to not only have the feature expression ability at different scales, but also strengthen the feature representation based on the global features at different scales; by setting the CSCP module and CFGC module in the decoder and making the CSCP modules of each layer of the encoder jump-connect to the CSCP modules of each layer of the decoder one by one, it is possible to use the CFGC module and CSCP module to gradually recover features layer by layer, realizing effective fusion in the process of feature splicing of each layer of the encoder and decoder; at the same time, through the CSCP module that gradually fuses cross-channels, it is possible to directly fuse the features in the encoder and decoder stages, optimize the semantic gap between the encoder and decoder without changing the jump-connection structure, and will not cause additional semantic ambiguity; at the same time, using different divided channels of the CSCP module to respectively perform feature division on the feature image output from the encoder and the feature image output from the transposed convolutional module, and gradually cross-splicing and fusing the divided deep features and shallow features, and performing global average pooling processing on different feature images through different global average pooling units to obtain different feature weight vectors, and combining the feature weight vectors with the spliced and fused feature images, it is possible to simultaneously pay attention to the important semantic information in the encoder stage and the decoder channels.

[0028] Further, referring to Figure 2 As shown, the CFGC module includes a first depth convolution unit, a second depth convolution unit, a first feature splicing and fusion unit, a first H dimensional feature selection unit, a first W dimensional feature selection unit, a second H dimensional feature selection unit, a second W dimensional feature selection unit, a second feature splicing and fusion unit, a third feature splicing and fusion unit, a third standard convolution unit, a fourth standard convolution unit, a fifth standard convolution unit, a first feature shape reshaping unit, a second feature shape reshaping unit, a first global feature calculation unit, a first activation function, a second feature calculation unit, and a sixth standard convolution unit; The input ends of the first depth convolution unit and the second depth convolution unit together serve as the input end of the CFGC module; The output end of the first depth convolution unit is respectively connected to the input ends of the first feature splicing and fusion unit, the first H dimensional feature selection unit, and the first W dimensional feature selection unit. The output end of the second depth convolution unit is respectively connected to the input ends of the first feature splicing and fusion unit, the second HThe dimensional feature selection unit and the second W are connected to the input end of the dimensional feature selection unit; The first H dimensional feature selection unit and the second H are both connected to the input end of the second feature splicing and fusion unit. The output ends of the first W dimensional feature selection unit and the second W dimensional feature selection unit are both connected to the input end of the third feature splicing and fusion unit; The output end of the second feature splicing and fusion unit is connected to the input end of the third standard convolution unit. The output end of the first feature splicing and fusion unit is connected to the input end of the fourth standard convolution unit. The output end of the third feature splicing and fusion unit is connected to the input end of the fifth standard convolution unit; The output end of the third standard convolution unit is connected to the input end of the first feature shape reshaping unit. The output end of the fifth standard convolution unit is connected to the input end of the second feature shape reshaping unit; The output ends of the first feature shape reshaping unit and the second feature shape reshaping unit are both connected to the input end of the first global feature calculation unit. The output end of the first global feature calculation unit is connected to the first activation function. The output ends of the first activation function and the fourth standard convolution unit are both connected to the input end of the second feature calculation unit. The output end of the second feature calculation unit is connected to the input end of the sixth standard convolution unit. The output end of the sixth standard convolution unit is used as the output end of the CFGC module.

[0029] It should be noted that in this embodiment, the convolution kernels of the first depth convolution unit, the third standard convolution unit, the fifth standard convolution unit, and the sixth standard convolution unit are all 3×3. The convolution kernel of the second depth convolution unit is 7×7. The convolution kernel of the fourth standard convolution unit is 1×1; at the same time, the first depth convolution unit and the second depth convolution unit are both depth convolutions. The third standard convolution unit, the fourth standard convolution unit, the fifth standard convolution unit, and the sixth standard convolution unit are all standard convolutions. It should be noted that standard convolution and depth convolution are both well-known common knowledge in this field and will not be elaborated here; the first activation function uses the activation function.

[0030] Specifically, in this embodiment, the CFGC module not only has the ability to extract multi-scale features, but also perceives global feature information in a lightweight manner, successfully realizing effective attention to both local and global information; and through H the dimensional feature selection unit and W the dimensional feature selection unit, global feature enhancement is effectively realized in the dimensions of H and W.

[0031] Furthermore, referring toFigure 4 As shown, the CSCP module includes a first feature image division channel, a second feature image division channel, a fourth feature splicing and fusion unit, a fifth feature splicing and fusion unit, a seventh standard convolution unit, an eighth standard convolution unit, a third feature product calculation unit, a fourth feature product calculation unit, a sixth feature splicing and fusion unit, a ninth standard convolution unit, a first global average pooling unit, and a second global average pooling unit; The input end of the first feature image division channel and the input end of the first global average pooling unit together serve as the first input end of the CSCP. The input end of the second feature image division channel and the input end of the second global average pooling unit together serve as the second input end of the CSCP. The output end of the first feature image division channel is respectively connected to the input ends of the fourth feature splicing and fusion unit and the fifth feature splicing and fusion unit. The output end of the second feature image division channel is respectively connected to the input ends of the fourth feature splicing and fusion unit and the fifth feature splicing and fusion unit. The output end of the fourth feature splicing and fusion unit is connected to the input end of the seventh standard convolution unit. The output end of the fifth feature splicing and fusion unit is connected to the input end of the eighth standard convolution unit; The output end of the seventh standard convolution unit and the output end of the first global average pooling unit are both connected to the input end of the third feature product calculation unit. The output end of the eighth standard convolution unit and the output end of the second global average pooling unit are both connected to the input end of the fourth feature product calculation unit; The output end of the third feature product calculation unit and the output end of the fourth feature product calculation unit are both connected to the input end of the sixth feature splicing and fusion unit. The output end of the sixth feature splicing and fusion unit is connected to the input end of the ninth standard convolution unit. The output end of the ninth standard convolution unit serves as the output end of the CSCP module.

[0032] It should be noted that in this embodiment, the convolution kernels of the seventh standard convolution unit, the eighth standard convolution unit, and the ninth standard convolution unit are all 1×1; at the same time, the seventh standard convolution unit, the eighth standard convolution unit, and the ninth standard convolution unit are all standard convolutions.

[0033] Meanwhile, it should be noted that the excellent performance of U-Net largely owes to the design of the skip connections between the encoder and the decoder, which enables the network to preserve semantic information at different stages lost during the pooling process, while allowing semantic interaction between the shallow features in the encoder stage and the deep semantic information in the decoder stage, thus better restoring the details and boundary features of the target. However, most of the recent existing technologies seem to overemphasize optimizing the features in the encoder stage by improving the structure of the skip connections. They often neglect the subsequent fusion process after splicing with the decoder stage. U-Net and most of its variants still use a simple splicing operation to fuse the features of the encoder-decoder stages. As far as we know, there still exists a semantic gap between the features in the encoder stage optimized by various methods and the features in the decoder stage. A simple splicing operation may not meet the requirements of high segmentation accuracy. Therefore, we provide the CSCP module to directly fuse the features of the encoder-decoder stages, optimize the semantic gap between the encoder and the decoder from another perspective, and reduce semantic ambiguity.

[0034] Further, referring to Figure 3 shown, both the first depth convolution unit and the second depth convolution unit are used to receive the first feature image input to the CFGC module, and perform depth convolution processing on the first feature image at different scales respectively to obtain the second feature image and the third feature image. Specifically: ; , where 、 represent the second feature image and the third feature image respectively, 、 ; represents a depth convolution with a convolution kernel size of ; represents a depth convolution with a convolution kernel size of ; represents the input first feature image, .

[0035] The first feature splicing and fusion unit is used to splice the second feature image and the third feature image, and fuse the spliced features through the fourth standard convolution unit to obtain the fourth feature image. Specifically: , where represents the fourth feature image, ; represents a standard convolution with a convolution kernel size of 1×1; represents splicing the feature maps along the channel dimension.

[0036] The first H dimensional feature selection unit is used to select in HSelect and splice the maximum feature and the average feature of the second feature image in the dimension to obtain the first H dimensional feature; the first W dimensional feature selection unit is used to select and splice the maximum feature and the average feature of the second feature image in the W dimension to obtain the first W dimensional feature; the second H dimensional feature selection unit is used to select and splice the maximum feature and the average feature of the third feature image in the H dimension to obtain the second H dimensional feature, the second W dimensional feature selection unit is used to select and splice the maximum feature and the average feature of the third feature image in the W dimension to obtain the second W dimensional feature, specifically: ; ; ; , where , respectively represent the first H dimensional (height dimension) feature and the second H dimensional (height dimension) feature output, , ; respectively represent the first W dimensional (width dimension) feature and the second W dimensional (height dimension) feature output, , ; represents the average selection of the layer-by-layer features in the dimension; represents the maximum selection of the layer-by-layer features in the dimension, represents W or H dimension.

[0037] The second feature splicing and fusion unit is used to splice the first H dimensional feature and the second H dimensional feature to obtain the fifth feature image; the third feature splicing and fusion unit is used to splice the first W dimensional feature and the second W dimensional feature to obtain the sixth feature image, specifically: ; , where represents the fifth feature image, ; represents the sixth feature image, ; The third standard convolution unit is used to perform standard convolution processing on the fifth feature image, and the first feature shape reshaping unit is used to reshape the feature shape of the fifth feature image after the convolution processing of the third standard convolution unit to obtain the first reshaped feature image; the fifth standard convolution unit is used to perform standard convolution processing on the sixth feature image, and the second feature shape reshaping unit is used to reshape the feature shape of the sixth feature image after the convolution processing of the fifth standard convolution unit to obtain the second reshaped feature image, specifically: ; , where represents feature shape reshaping; represents standard convolution with a convolution kernel size of 3×3; represents the first reshaped feature image; represents the second reshaped feature image; The first global feature calculation unit is used to perform matrix calculation on the first reshaped feature image and the second reshaped feature image to obtain the global feature representation matrix, specifically: , where represents the influence of the th position on the th position, represents the total amount of pixels, ; respectively represent the pixel points at position of the first reshaped feature image and the pixel points at position of the second reshaped feature image; The first activation function is used to activate the global feature representation matrix to obtain the global response matrix, specifically: , where represents the global response matrix, ; represents the global response matrix.

[0038] The second feature calculation unit is used to perform Hadamard product calculation on the fourth feature image and the global response matrix to obtain the seventh feature image, specifically: , where represents the fourth feature image; represents the global response matrix; represents the seventh feature image, .

[0039] The seventh feature image is input into the sixth standard convolution unit for convolution fusion to obtain the output feature image of the CFGC module, specifically: , where represents the output feature image of the CFGC module, .

[0040] Furthermore, referring toFigure 5 As shown, the first feature image division channel and the first global average pooling unit are used to receive the eighth feature image input to the CSCP module , ; The second feature image division channel and the second global average pooling unit are used to receive the ninth feature image input to the CSCP module , ; The first feature image division channel is used to perform an average division of the eighth feature image on the channels to obtain the tenth feature image and the eleventh feature image , , ; The first global average pooling unit is used to perform global average pooling on the eighth feature image to obtain the first encoder global weight vector, specifically: , where represents the first encoder global weight vector; represents global average pooling; The second feature image division channel is used to perform an average division of the ninth feature image on the channels to obtain the twelfth feature image and the thirteenth feature image , , ; The second global average pooling unit is used to perform global average pooling on the ninth feature image to obtain the second decoder global weight vector, specifically: , where represents the first encoder global weight vector; The fourth feature splicing and fusion unit is used to splice the tenth feature image and the twelfth feature image, and perform convolutional fusion on the spliced feature image through the seventh standard convolutional unit to obtain the fourteenth feature image. The fifth feature splicing and fusion unit is used to splice the eleventh feature image and the thirteenth feature image, and perform convolutional fusion on the spliced feature image through the eighth standard convolutional unit to obtain the fifteenth feature image, specifically: ; , where represents a standard convolution with a kernel size of 1×1; represent the fourteenth feature image and the fifteenth feature image respectively.

[0041] The third feature product calculation unit is used to calculate the product of the first encoder global weight vector and the fourteenth feature image to obtain the sixteenth feature image. The fourth feature product calculation unit is used to calculate the product of the second decoder global weight vector and the fifteenth feature image to obtain the seventeenth feature image, specifically: ; , where represents the sixteenth feature image; Represents the seventeenth feature image; The sixth feature splicing and fusion unit is used to splice the features of the fifteenth feature image and the sixteenth feature image, and use the ninth standard convolution unit to perform feature fusion on the spliced feature image to obtain the output feature image of the CSCP module. Specifically: , where represents the output feature image of the CSCP module.

[0042] Specifically, see Figure 6 As shown, in this embodiment, the model parameters of the X-UNet model are also updated by using the binary cross-entropy loss function in combination with the Adam optimizer. This technical solution is a conventional technical means and will not be elaborated here.

[0043] It should be noted that in this embodiment, in order to verify the effect of the X-UNet model provided by this method, we have completed all comparative experiments and ablation studies. And, we use a random division strategy for all the data sets in this embodiment, where 80% is used for training and the remaining 20% is used for testing. At the same time, in order to ensure the rigor and fairness of the comparative experiments and ensure the repeatability of data segmentation, 3 random number seeds are used by us to divide the data sets. During the training process, we use the cosine annealing decay learning rate. In the CVC-ClinicDB, Kvasir-SEG, and PH2 data sets, the initial learning rate value is 0.0001, and in BUSI it is 0.0002, and the minimum value of the cosine annealing decay learning rate in the four data sets is 0.00001. In the CVC-ClinicDB, Kvasir-SEG, and BUSI data sets, the initial image size is set to 256×256, and in the PH2 data set, the image size is set to 320×320, so as to explore the segmentation performance of the network under different image resolutions. The Batch size and epoch are uniformly set to 4 and 100 respectively, and the BCEDiceLoss loss function and the Adam optimizer are used by us during the training process. Most importantly, all the algorithm models used for comparison in this embodiment (including this technical solution) have not undergone any pre-training.

[0044] IoU (Intersection over Union), Dice (Dice coefficient, DICE), and ACC (Accuracy) are respectively common classical evaluation indicators in medical image segmentation. IoU is the ratio of the intersection to the union between the prediction result and the ground truth result. Dice is always used to capture the measure of the overlap between two samples. The formulas are as follows: ; , where A and B represent two sets of sampling points. Both IoU and Dice are used to measure the similarity between two samples. Obviously, the higher the value, the better the performance of the network. Acc is the proportion of the number of correct samples in the predicted results to the total number of predictions. The higher the value of Acc, the stronger the ability of the network to correctly predict true positive targets. The formula is as follows: , where, 、 、 and represent true positive, false positive, true negative, and false negative respectively.

[0045] The results of the comparative experiments show: 1) The detailed results on the Kvasir-SEG dataset can be seen in Table 1. Compared with other advanced baseline methods, our X-UNet (78.65%) is in the leading level among the three basic evaluation indicators of IoU, Dice, and Acc while having less computational complexity and fewer parameters. Compared with the most classical UNet (75.55%), the IoU has increased by 3.1%, proving the superiority of our model. At the same time, the TransAttUnet (78.45%) based on Transfomer achieved sub-optimal performance with the most computational complexity, and its IoU is 0.2% lower than ours. And the IoU of the model FDFUNet (78.37%) designed based on CNN is 0.28% lower than that of X-UNet. Under the condition of comparable computational complexity, the number of parameters of FDFUNet is more than three times that of ours, which is due to the advantage of the depth convolution we adopted by focusing on lightweight design. Obviously, X-UNet has achieved a good balance among computational complexity, number of parameters, and performance.

[0046] Table 1: Experimental results of X-UNet and other baseline methods on the Kvasir-SEG dataset

[0047] Among them, GFLOPS is the abbreviation of Giga Floating-point Operations Per Second, which refers to the index of one billion floating-point operations per second; Params represents the variable-length parameter index.

[0048] 2) All experimental settings are consistent with the Kvasir-SEG dataset. The detailed experimental results on the CVC-ClinicDB dataset can be referred to Table 2 below. Obviously, our X-UNet (87.90%) also achieved a leading level in the segmentation experiments on another polyp dataset, with the IoU improved by 3.02% compared to UNet (84.88%). It is worth noting that FDFUNet (87.81%) and TransAttUNet (87.66%) have relatively small performance differences with our X-UNet in the CVC-ClinicDB dataset, being 0.09% and 0.24% lower than ours respectively. However, their parameter counts are significantly more than ours. Especially for TransAttUnet, due to the use of the multi-head attention mechanism to establish long-range dependencies, the computational cost is even significantly leading. Compared with them, X-UNet can save more computational resources while globally modeling long-range dependencies through its CFGC module, and has higher practical application value.

[0049] Table 2: Experimental results of X-UNet and other baseline methods on the CVC-ClinicDB dataset

[0050] 3) Different from the above two datasets, in the BUSI dataset, we explored the network performance at different learning rates, and the images in the BUSI dataset are ultrasound images with noise. More detailed experimental results can be referred to Table 3 below. Even under the condition of ultrasound images, our X-UNet (69.08%) still maintained the best segmentation performance. Compared with UNet (63.25%), the IoU is 5.83% higher. CANet (68.52%), which performed mediocrely in the previous two polyp datasets, finally achieved the second-best performance in the BUSI dataset, with the IoU 0.56% lower than ours. Different from the previous polyp datasets, FDFUNet (68.12%) showed a significant performance decline in the environment of ultrasound images, with the IoU 0.96% lower than ours. This indicates that our network model can also achieve good performance in the ultrasound image dataset, demonstrating the generality and adaptability of X-UNet to a certain extent.

[0051] Table 3: Experimental results of X-UNet and other baseline methods on the BUSI dataset

[0052] 4) Other experimental settings are the same as those in the Kvasir-SEG and CVC-ClinicDB datasets. However, the resolution of the input images in the PH2 dataset is set to 320×320. More detailed experimental data can be referred to Table 4 below. It can be clearly seen that although the performance of our X-UNet (91.49%) does not significantly exceed that of other baseline networks, with the IoU being 1.34% higher than that of the original UNet (90.15%), X-UNet still leads in the three metrics of IoU, Dice, and Acc with fewer parameters and computational costs. TransAttUnet (90.27%), which achieved good performance in previous datasets, showed a significant decline in the PH2 dataset, probably because the training images in the PH2 dataset are too scarce, making it difficult to learn the features of the data. This shows that our X-UNet can still have good performance when the training images are scarce, and at the same time verifies the relatively strong generality of X-UNet on different types of datasets.

[0053] Table 4: Experimental results of X-UNet and other baseline methods on the PH2 dataset

[0054] In Figure 7 , we completed some visualization experimental results of X-UNet and other standard baseline networks. Obviously, compared with other advanced baseline networks, our X-UNet has a more accurate perception ability of the boundary information of the lesion area, resulting in the final prediction result of X-UNet being the most consistent with the ground truth. This is probably due to the feature enhancement based on global response in the CFGC module to establish long-range dependencies, leading to the sensitivity to the boundaries of the lesion area.

[0055] In summary, in the present invention, a lightweight and efficient X-UNet for medical image segmentation is proposed based on the Global Perception Collaborative Fusion Module (CFGC module) and the Cross-Split Channel Step-by-Step Fusion Module (CSCO module). To model long-range dependencies globally, we designed the CFGC module, which not only preserves feature information at different scales but also effectively enhances global features in the H and W dimensions. To better reduce semantic ambiguity between the encoder and decoder, we proposed the CSCP module to directly act on the features in the encoder-decoder fusion stage to promote more sufficient fusion. Compared with traditional CNNs, X-UNet is not limited by the drawback that convolution overly focuses on local regions; compared with most Transformer-based networks, X-UNet models long-range dependencies with less computational cost and fewer parameters while protecting the 2D structure of the image, captures precious spatial information that cannot be attended to by multi-head attention through convolution, and pays attention to enhancing global features while realizing local semantic interaction. Finally, our experiments prove that X-UNet achieves leading-level performance with lower computational cost and fewer parameters compared with other advanced algorithms on multiple datasets of different scales.

[0056] The specific embodiments described above further elaborate on the purpose, technical solutions, and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An image segmentation method based on a global perception U-Net model, characterized in that, The method includes: Obtain the medical image to be segmented, input the medical image to be segmented into the global perception U-Net model for image segmentation processing, and obtain the segmented image segmentation result; Among them, the global perception U-Net model includes an encoder, a decoder, a CFGC module, and a CSCP module. Both the encoder and the decoder adopt a five-layer structure connected in sequence. The first layer of the encoder adopts a first convolutional module and a max pooling module connected in sequence. The remaining four layers of the encoder all adopt a CFGC module and a max pooling module connected in sequence. The last layer of the decoder adopts a second convolutional module. The remaining four layers of the decoder all adopt a transposed convolutional module, a CSCP module, and a CFGC module connected in sequence. The first four layers of the encoder are respectively skip-connected to the first four layers of the decoder; the CFGC module includes H a dimensional feature selection unit and W a dimensional feature selection unit, which is used to select the feature information in the H and W dimensions at different scales, and based on H the dimensional feature selection unit and W the dimensional feature selection unit, the global feature is enhanced by matrix multiplication; The CSCP module includes a feature image division channel, a feature splicing and fusion unit, and a global average pooling unit, which are used to perform channel-wise feature averaging on the feature images in the encoder and decoder stages, splice the divided features in the encoder and decoder stages respectively, and complete step-by-step feature fusion based on the channel dependence relationship between the feature images in the encoder and decoder stages.

2. The image segmentation method based on a global perception U-Net model according to claim 1, characterized in that, The first convolutional module in the first layer of the encoder is skip-connected to the CSCP module in the fourth layer of the decoder, the CFGC module in the second layer of the encoder is skip-connected to the CSCP module in the third layer of the decoder, the CFGC module in the third layer of the encoder is skip-connected to the CSCP module in the second layer of the decoder, and the CFGC module in the fourth layer of the encoder is skip-connected to the CSCP module in the first layer of the decoder.

3. The image segmentation method based on a global perception U-Net model according to claim 2, characterized in that, The CFGC module in the last layer of the encoder is connected to the transposed convolutional module in the first layer of the decoder.

4. The image segmentation method based on a global perception U-Net model according to claim 1, characterized in that, The CFGC module includes a first depth convolution unit, a second depth convolution unit, a first feature splicing and fusion unit, a first H dimensional feature selection unit, a first W dimensional feature selection unit, a second H dimensional feature selection unit, a second W dimensional feature selection unit, a second feature splicing and fusion unit, a third feature splicing and fusion unit, a third standard convolution unit, a fourth standard convolution unit, a fifth standard convolution unit, a first feature shape reshaping unit, a second feature shape reshaping unit, a first global feature calculation unit, a first activation function, a second feature calculation unit, and a sixth standard convolution unit; The input ends of the first depth convolutional unit and the second depth convolutional unit together serve as the input end of the CFGC module; The output ends of the first depth convolution unit are respectively connected to the input ends of the first feature splicing and fusion unit, the first H dimensional feature selection unit and the first W dimensional feature selection unit. The output ends of the second depth convolution unit are respectively connected to the input ends of the first feature splicing and fusion unit, the second H dimensional feature selection unit and the second W dimensional feature selection unit; First H The output ends of the first-dimensional feature selection unit and the second H -dimensional feature selection unit are both connected to the input end of the second feature splicing and fusion unit. The output ends of the first W -dimensional feature selection unit and the second W -dimensional feature selection unit are both connected to the input end of the third feature splicing and fusion unit; The output end of the second feature splicing and fusion unit is connected to the input end of the third standard convolutional unit, the output end of the first feature splicing and fusion unit is connected to the input end of the fourth standard convolutional unit, and the output end of the third feature splicing and fusion unit is connected to the input end of the fifth standard convolutional unit; The output end of the third standard convolutional unit is connected to the input end of the first feature reshaping unit, and the output end of the fifth standard convolutional unit is connected to the input end of the second feature reshaping unit; The output ends of the first feature reshaping unit and the second feature reshaping unit are both connected to the input end of the first global feature calculation unit, the output end of the first global feature calculation unit is connected to the first activation function, the output ends of the first activation function and the fourth standard convolutional unit are both connected to the input end of the second feature calculation unit, the output end of the second feature calculation unit is connected to the input end of the sixth standard convolutional unit, and the output end of the sixth standard convolutional unit serves as the output end of the CFGC module.

5. The image segmentation method based on a global perception U-Net model according to claim 1, characterized in that, The CSCP module includes a first feature image division channel, a second feature image division channel, a fourth feature splicing and fusion unit, a fifth feature splicing and fusion unit, a seventh standard convolutional unit, an eighth standard convolutional unit, a third feature product calculation unit, a fourth feature product calculation unit, a sixth feature splicing and fusion unit, a ninth standard convolutional unit, a first global average pooling unit, and a second global average pooling unit; The input end of the first feature image division channel and the input end of the first global average pooling unit together serve as the first input end of the CSCP. The input end of the second feature image division channel and the input end of the second global average pooling unit together serve as the second input end of the CSCP. The output end of the first feature image division channel is respectively connected to the input ends of the fourth feature splicing and fusion unit and the fifth feature splicing and fusion unit. The output end of the second feature image division channel is respectively connected to the input ends of the fourth feature splicing and fusion unit and the fifth feature splicing and fusion unit. The output end of the fourth feature splicing and fusion unit is connected to the input end of the seventh standard convolution unit. The output end of the fifth feature splicing and fusion unit is connected to the input end of the eighth standard convolution unit; The output end of the seventh standard convolution unit and the output end of the first global average pooling unit are both connected to the input end of the third feature product calculation unit. The output end of the eighth standard convolution unit and the output end of the second global average pooling unit are both connected to the input end of the fourth feature product calculation unit; The output end of the third feature product calculation unit and the output end of the fourth feature product calculation unit are both connected to the input end of the sixth feature splicing and fusion unit. The output end of the sixth feature splicing and fusion unit is connected to the input end of the ninth standard convolution unit. The output end of the ninth standard convolution unit serves as the output end of the CSCP module.

6. The image segmentation method based on a global perception U-Net model according to claim 4, characterized in that, The first depth convolution unit and the second depth convolution unit are both used to receive the first feature image input to the CFGC module and respectively perform depth convolution processing on the first feature image at different scales to obtain the second feature image and the third feature image; The first feature splicing and fusion unit is used to splice the second feature image and the third feature image and fuse the spliced features through the fourth standard convolution unit to obtain the fourth feature image; First H The dimensional feature selection unit is used to select and splice the maximum feature and the average feature of the second feature image on the H dimension to obtain the first H dimensional feature; First W The dimensional feature selection unit is used to select and splice the maximum feature and the average feature of the second feature image in the W dimension to obtain the first W dimensional feature; Second H The second - dimension feature selection unit is used to select and splice the maximum feature and the average feature of the third - feature image on the H dimension to obtain the second H dimension feature. The second W dimension feature selection unit is used to select and splice the maximum feature and the average feature of the third - feature image on the W dimension to obtain the second W dimension feature; The second feature splicing and fusion unit is used to splice the first H dimensional feature and the second H dimensional feature to obtain a fifth feature image; The third feature splicing and fusion unit is used to splice the first W dimensional feature and the second W dimensional feature to obtain a sixth feature image; The third standard convolution unit is used to perform standard convolution processing on the fifth feature image. The first feature shape reshaping unit is used to reshape the feature shape of the fifth feature image after convolution processing by the third standard convolution unit to obtain the first reshaped feature image; The fifth standard convolution unit is used to perform standard convolution processing on the sixth feature image. The second feature shape reshaping unit is used to reshape the feature shape of the sixth feature image after convolution processing by the fifth standard convolution unit to obtain the second reshaped feature image; The first global feature calculation unit is used to perform matrix calculation on the first reshaped feature image and the second reshaped feature image to obtain the global feature representation matrix. The first activation function is used to activate the global feature representation matrix to obtain the global response matrix; The second feature calculation unit is used to perform Hadamard product calculation on the fourth feature image and the global response matrix to obtain the seventh feature image; The seventh feature image is input to the sixth standard convolution unit for convolution fusion to obtain the output feature image of the CFGC module.

7. The image segmentation method based on a global perception U-Net model according to claim 5, characterized in that, The first feature image partitioning channel and the first global average pooling unit are used to receive the eighth feature image input to the CSCP module. The second feature image partitioning channel and the second global average pooling unit are used to receive the ninth feature image input to the CSCP module. The first feature image partitioning channel is used to perform an average partitioning of the eighth feature image on the channels to obtain a tenth feature image and an eleventh feature image. The first global average pooling unit is used to perform global average pooling on the eighth feature image to obtain a first encoder global weight vector. The second feature image partitioning channel is used to perform an average partitioning of the ninth feature image on the channels to obtain a twelfth feature image and a thirteenth feature image. The second global average pooling unit is used to perform global average pooling on the ninth feature image to obtain a second decoder global weight vector; The fourth feature splicing and fusion unit is used to splice the tenth feature image and the twelfth feature image, and perform convolutional fusion on the spliced feature image through the seventh standard convolutional unit to obtain a fourteenth feature image; The fifth feature splicing and fusion unit is used to splice the eleventh feature image and the thirteenth feature image, and perform convolutional fusion on the spliced feature image through the eighth standard convolutional unit to obtain a fifteenth feature image; The third feature product calculation unit is used to perform product calculation on the first encoder global weight vector and the fourteenth feature image to obtain a sixteenth feature image. The fourth feature product calculation unit is used to perform product calculation on the second decoder global weight vector and the fifteenth feature image to obtain a seventeenth feature image; The sixth feature splicing and fusion unit is used to splice the fifteenth feature image and the sixteenth feature image, and perform feature fusion on the spliced feature image using the ninth standard convolutional unit to obtain the output feature image of the CSCP module.

8. A method for image segmentation based on a global perception U-Net model according to claim 4, characterized in that, The convolution kernels of the first depth convolutional unit, the third standard convolutional unit, the fifth standard convolutional unit, and the sixth standard convolutional unit are all 3×3. The convolution kernel of the second depth convolutional unit is 7×7. The convolution kernel of the fourth standard convolutional unit is 1×1.

9. A method for image segmentation based on a global perception U-Net model according to claim 5, characterized in that, The convolution kernels of the seventh standard convolutional unit, the eighth standard convolutional unit, and the ninth standard convolutional unit are all 1×1.

10. A method for image segmentation based on a global perception U-Net model according to claim 4, characterized in that, The first activation function adopts activation function.

Images