Medical image segmentation system and method based on enhanced multi-scale volume accumulation

Abstract

The invention relates to the technical field of image segmentation, and aims to solve the problems that an existing model based on a convolutional neural network still has the limitations of limited receptive field, difficulty in extracting deep features in an image and the like, and important feature channel information is ignored when multi-scale information fusion is carried out by using jump connection. The invention provides a medical image segmentation system and method based on enhanced multi-scale volume integration, a multi-acceptance domain aggregation module is utilized between an encoder and a decoder, the semantic gap between encoding and decoding sub-networks is reduced, and after encoding features output by a multi-scale large kernel attention module of the encoder are fused, the fusion is carried out to obtain a fusion image. And a multi-scale large-kernel attention module of the decoder carries out feature fusion by cooperatively utilizing coding output at different stages, so that the capability of capturing boundary features by a network is improved, and more accurate segmentation of a target region is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of image segmentation, and in particular to a medical image segmentation system and method based on enhanced multi-scale convolution aggregation. Background Art

[0002] In recent years, with the gradual application of deep learning in the field of image segmentation, especially the continuous development of convolutional neural networks and Transformer architectures, the efficiency and accuracy of image segmentation have been significantly improved and widely used in the field of medical image segmentation. At the same time, a large number of convolutional neural network variants have also emerged, such as fully convolutional networks. The accuracy of segmentation has been further improved through end-to-end segmentation training. However, medical images are usually characterized by high complexity, irregular and blurred boundaries, small feature areas, overlapping tissue structures, and low image contrast. Therefore, it is particularly important to build a segmentation model suitable for medical images. The UNet model is a milestone in the field of medical image segmentation, and its performance in various medical image segmentation tasks is better than that of traditional image processing methods. The UNet model adopts a fully symmetric encoder-decoder structure with jump connections, which can effectively combine low-level and high-level semantic information to generate the final mask. However, when processing more complex medical images, the performance of UNet is not perfect, and improved models have been widely used in the field of medical image segmentation. Unet++ overcomes the semantic information gap in feature fusion by adding a series of nested jump connections in the network, thereby improving the segmentation performance. DoubleU-Net proposes to combine two Unet structures. The first Unet is used to learn contextual information and pass features to the second Unet for more efficient information extraction. LCU-Net enhances segmentation performance through connection operations and introduces the Inception module to expand the receptive field, so that deeper features can be extracted. RefineNet extracts features of different resolutions through a multi-path network to achieve better segmentation performance.

[0003] UNet and its improved models have shown excellent performance in fast medical image segmentation. However, these convolutional neural network-based models still have certain limitations. For example, their receptive fields are relatively limited, making it difficult to fully extract deep features in images. At the same time, when using skip connections for multi-scale information fusion, these models often ignore important feature channel information. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a medical image segmentation system and method based on enhanced multi-scale convolutional aggregation, which uses a deep feature preservation module to expand the receptive field in the processing of the input image, effectively reduces the loss of key detail features, realizes cross-channel interaction of multi-scale information, and reduces information loss during drastic downsampling.

[0005] The present invention is mainly achieved through the following technical solutions:

[0006] A medical image segmentation system based on enhanced multi-scale convolutional aggregation, wherein the network structure of the medical image segmentation system is a convolutional neural network with a U-shaped structure, including an encoder and a decoder for encoding and decoding a medical image to be segmented, and a multi-receptive field aggregation module embedded in the network structure of the medical image segmentation system as a conversion link between the encoder and the decoder, wherein the encoder includes a deep feature retention module and a plurality of multi-scale large core attention modules, and the decoder includes a plurality of multi-scale large core attention modules, each module of the encoder corresponds to a multi-scale large core attention module in the decoder, and each module of the encoder is jump-connected with the corresponding multi-scale large core attention module in the decoder, wherein:

[0007] The encoder uses the deep feature preservation module to input the medical image to be segmented into the deep feature preservation module and performs maximum pooling. It uses multiple multi-scale large-core attention modules to achieve multi-scale feature fusion under different network depths and serves as multiple input features of the multi-receptive field aggregation module.

[0008] The decoder concatenates the coded features of each module in the encoder through multiple multi-scale large-core attention modules and decodes them to obtain the segmented image;

[0009] The input of the multi-receptive field aggregation module is connected to the encoder, and the output is connected to the decoder, which reduces the semantic gap between the encoding and decoding sub-networks. The encoded features output by the encoder's multi-scale large-core attention module are fused and transmitted to the decoder's multi-scale large-core attention module, which synergistically utilizes the encoding outputs of different stages for feature fusion, thereby improving the network's ability to capture boundary features.

[0010] Furthermore, the deep feature preservation module replaces the traditional stacked standard convolution with a deep separable convolution. The deep feature preservation module adopts a residual structure. The medical image to be segmented is input through a 3×3 convolution layer to increase the feature channels and expand the receptive field to obtain an input feature map; the input feature map is passed through a 7×7 deep convolution and a 1×1 point convolution. The point convolution will increase the number of channels by four times to obtain an intermediate feature map; after the intermediate feature map is processed by the ReLU activation function, the 1×1 point convolution reduces the number of channels to one quarter of the original size to obtain a feature map with global features; the input feature map and the feature map with global features are summed to obtain the output result of the deep feature preservation module.

[0011] Furthermore, the multi-scale large kernel attention module combines large kernel convolution with cross-channel information interaction, and splits the feature map into two parts in the channel dimension after 1×1 convolution. The two parts of the split feature map interact with each other after different operations, which enhances the multi-scale feature extraction capability and expands the receptive field, and extracts deep representation features from the medical image to be segmented; the multi-scale large kernel attention module adjusts the number of channels of the feature map through 1×1 convolution to obtain the initial feature map, and the initial feature map splits the feature map into feature maps M along the channel dimension. c and feature map M l , feature map M l , apply 3×3 convolution to capture more channel information and obtain the feature map M k ; The feature map M k After 5×5 depth convolution and 7×7 depth convolution respectively, the feature map is obtained and Feature Map and feature map Then, we sum them up through a 1×1 point convolution to get the summed feature map. The summed feature map Apply maximum pooling and average pooling to obtain the feature map S max and feature map S avg :

[0012]

[0013] Where P max represents the maximum pooling function, P avg Represents the average pooling function, and then the feature map M is obtained by the following formula:

[0014]

[0015] S j Split into S1 and S2 in the channel dimension:

[0016]

[0017] M=M k ×S

[0018] Among them, F 7×7 represents a large convolution kernel of 7×7, [] represents concatenation in the channel dimension, σ represents the sigmoid activation function, F 1×1 Represents a point convolution with a convolution kernel of 1×1; represents the feature map after convolution with a large kernel of 7×7; S j Representative feature map The feature map after the sigmoid activation function; S represents the feature map after the point convolution with a convolution kernel of 1×1;

[0019] Feature map M c After 3×3 convolution, it is activated by the ReLU activation function and summed with the feature map M to obtain the feature map M r :

[0020] M r =M+ReLU(F 3×3 (M c ))

[0021] Among them, F 3×3 Represents 3×3 convolution, feature map M r After fusing the information of different channels, the feature maps are aggregated using channel-by-channel weighted combination to obtain the output Y of the multi-scale large-core attention module.

[0022] Furthermore, the multi-receptive field aggregation module performs feature fusion on the outputs of multiple multi-scale large-kernel attention modules as follows:

[0023] Feature maps X at different encoding stages i After applying the maximum pooling, concatenate them in the channel dimension to obtain the feature map S c , the formula is:

[0024]

[0025] Represents the feature map X at different encoding stages i The feature maps obtained after applying the maximum pooling are

[0026]

[0027] Feature maps X at different encoding stages iAfter the cross-channel separation module, six different outputs O1, output O2, output O3, output O4, output O5 and output O6 are obtained respectively. The six outputs are weighted fused to obtain the feature map Q c , the formula is:

[0028]

[0029] Among them, x i (c) is the soft attention weight, the formula is:

[0030]

[0031] Among them, P represents the global spatial information obtained after global pooling of six different outputs, Represents the global spatial information {weight for the c-th channel, the feature map S c and feature map Q c The added output is used as the output of the multi-receptive field aggregation module.

[0032] Furthermore, the cross-channel separation module expands the receptive field and captures richer channel information through a 1×1 convolution and a 3×3 convolution to obtain an initial feature map of the cross-channel separation module, and divides the initial feature map of the cross-channel separation module into two feature maps M1 and M2 in the channel dimension; the feature map M1 is used as an output L1 of the cross-channel separation module; the feature map M1 and the feature map M2 are summed, and then a 3×3 convolution is used to realize cross-channel information fusion to obtain another output L2 of the cross-channel separation module.

[0033] A medical image segmentation method based on enhanced multi-scale convolution aggregation, based on the above-mentioned medical image segmentation system based on enhanced multi-scale convolution aggregation, encodes and decodes the medical image to be segmented through the medical image segmentation system based on enhanced multi-scale convolution aggregation, the encoder uses multiple multi-scale large-core attention modules to realize multi-scale feature fusion under different network depths, and the decoder uses multiple multi-scale large-core attention modules to splice the coded features of each module in the encoder and decode to obtain the segmented image, which specifically includes the following steps:

[0034] S1. Inputting the medical image to be segmented into the deep feature preservation module of the encoder to obtain a first feature map;

[0035] S2. The first feature map obtained in step S1 is sequentially passed through a plurality of multi-scale large-core attention modules of the encoder to obtain a first encoding feature, a second encoding feature, and a t-th encoding feature, respectively, where t represents the number of multi-scale large-core attention modules in the encoder;

[0036] S3. Input all the encoded features into the multi-receptive field aggregation module to obtain the second feature map;

[0037] S4. Upsampling the second feature map obtained in step S3 and splicing it with the t-th encoding feature through the first multi-scale large core attention module of the decoder to obtain the first decoding feature, upsampling the first decoding feature and splicing it with the t-1-th encoding feature through the second multi-scale large core attention module of the decoder to obtain the second decoding feature, and so on, until the t-1-th encoding feature is upsampled and spliced ​​with the first encoding feature through the t-th scale large core attention module to obtain the t-th decoding feature; upsampling the t-th decoding feature and splicing it with the first feature map through the t+1-th multi-scale large core attention module of the decoder to obtain the t+1-th decoding feature;

[0038] S5. Pass the t+1th decoding feature obtained in step S4 through a sigmoid layer and a 1×1 convolution layer to obtain the final image segmentation result.

[0039] In summary, the invention has the following beneficial effects:

[0040] 1. The network structure of the present invention realizes cross-channel interaction of multi-scale information, reduces information loss in the process of drastic downsampling, and improves the spatial feature extraction process, especially in capturing boundary features, thereby realizing efficient segmentation of medical images.

[0041] 2. In the processing of the input medical image, the deep feature preservation module is used to expand the receptive field, effectively reducing the loss of key detail features in the medical image. The deep convolution and point convolution are used instead of ordinary convolution in the deep feature preservation module to improve the computational efficiency. The present invention provides a medical image segmentation system and method that can more comprehensively capture and fuse the deep features and multi-scale information in the image and achieve efficient segmentation of medical images, which can be applied to the field of medical image segmentation.

[0042] 3. In the feature extraction process, the use of large kernel convolution combined with cross-channel interaction of multi-scale information can more comprehensively capture and fuse the deep features and multi-scale information in the image.

[0043] 4. The multi-receptive field aggregation module is used between the encoding and decoding stages to reduce the semantic gap between the encoding and decoding sub-networks, and the encoding outputs of different stages are used for feature fusion to achieve more accurate segmentation of the target area. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A schematic diagram of the network structure of the image segmentation system of the present invention;

[0045] Figure 2 Schematic diagram of the deep feature preservation module;

[0046] Figure 3 Schematic diagram of the multi-scale large kernel attention module;

[0047] Figure 4 It is a schematic diagram of the multi-receptive field aggregation module;

[0048] Figure 5 This is a flow chart of a medical image segmentation method based on enhanced multi-scale convolution aggregation;

[0049] Figure 6 A qualitative comparison is performed between the present invention and other methods on the CVC-ColonD dataset. DETAILED DESCRIPTION

[0050] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the embodiments of the present invention, and should not be construed as limiting the present invention.

[0051] like Figures 1 to 5 As shown, the present invention discloses a medical image segmentation system based on enhanced multi-scale convolutional aggregation. The network structure of the medical image segmentation system is a convolutional neural network with a U-shaped structure, including an encoder and a decoder for encoding and decoding the medical image to be segmented, and a multi-receptive field aggregation module embedded in the network structure of the medical image segmentation system as a conversion link between the encoder and the decoder. The encoder includes a deep feature retention module and a plurality of multi-scale large core attention modules, and the decoder includes a plurality of multi-scale large core attention modules. Each module of the encoder corresponds to a multi-scale large core attention module in the decoder. Each module of the encoder is jump-connected with the corresponding multi-scale large core attention module in the decoder. Figure 1 In the illustrated embodiment, three multi-scale large core attention modules are arranged in the encoder, and four multi-scale large core attention modules are arranged in the decoder. The deep feature preservation module in the encoder and the three multi-scale large core attention modules each correspond to one multi-scale large core attention module.

[0052] The encoder inputs the medical image to be segmented into the deep feature preservation module through the deep feature preservation module and performs maximum pooling, and uses multiple multi-scale large kernel attention modules to realize multi-scale feature fusion under different network depths, and serves as multiple input features of the multi-receptive field aggregation module; the decoder splices the encoded features of each module in the encoder through multiple multi-scale large kernel attention modules and then decodes to obtain the segmented image.

[0053] The multi-scale large kernel attention module divides the feature map into two processing paths in the channel dimension. The multi-scale large kernel attention module combines large kernel convolution with cross-channel information interaction. After 1×1 convolution, the feature map is split into two parts in the channel dimension. The first part retains the original features after ordinary convolution, and the other part extracts deeper features after different large kernel convolution operations. The feature map interacts after different operations. In this embodiment, 7×7 and 5×5 large kernel convolutions are used to enhance the multi-scale feature extraction capability and expand the receptive field. This module can extract deep representation features from medical images.

[0054] The multi-receptive field aggregation module is embedded in the network structure of the medical image segmentation system as a transition link between the encoder and the decoder. The input end of the multi-receptive field aggregation module is connected to the encoder, and the output end is connected to the decoder, which reduces the semantic gap between the encoding and decoding sub-networks. It can effectively transmit the encoding results of multiple stages to the decoding stage, realize the fusion of multi-scale features under different network depths, and collaboratively use the encoding outputs of different stages for feature fusion, such as Figure 1 As shown in the figure, the three input ends of the multi-receptive field aggregation module are respectively connected to the outputs of the three multi-scale large-core attention modules. The multi-receptive field aggregation module can effectively transmit the encoding results of multiple stages to the decoding stage, realize the fusion of multi-scale features at different network depths, and improve the network's ability to capture boundary features.

[0055] The deep feature preservation module replaces the traditional stacked standard convolution with a depth-separable convolution. The deep feature preservation module adopts a residual structure, inputs the medical image to be segmented through a 3×3 convolution layer, increases the feature channel, expands the receptive field, and obtains the input feature map.

[0056] The input feature map passes through a 7×7 depth convolution and a 1×1 point convolution. The depth convolution does not change the number of channels, while the point convolution increases the number of channels by four times to obtain the intermediate feature map.

[0057] After the intermediate feature map is processed by the ReLU activation function, the 1×1 point convolution reduces the number of channels to one-fourth of the original size, and obtains a feature map with global features. The number of channels can be increased or decreased according to specific needs by changing the number of channels through point convolution. The input feature map and the feature map with global features are summed to obtain the output result of the deep feature preservation module.

[0058] The multi-scale large kernel attention module divides the feature map into two processing paths in the channel dimension.

[0059] The multi-scale large-kernel attention module first adjusts the number of channels of the feature map through a 1×1 convolution to obtain the initial feature map.

[0060] The initial feature map splits the feature map into feature maps M along the channel dimension c and feature map M l .

[0061] The feature map M l Apply 3×3 convolution to capture more channel information to obtain the feature map M k .

[0062] The feature map M k After 5×5 depth convolution and 7×7 depth convolution respectively, the feature map is obtained and feature map Feature Map and feature map Then, they are summed up through a 1×1 point convolution to obtain the summed feature map.

[0063] Apply maximum pooling and average pooling to the summed feature map to obtain the maximum pooled feature map S max And the feature map S after average pooling avg :

[0064]

[0065] Where P max represents the maximum pooling function, P avg Represents the average pooling function. In order to integrate richer spatial information, the features are connected through the concat function to obtain the feature map Then the feature map M is obtained by the following formula:

[0066]

[0067]

[0068] M=M k ×S

[0069] Among them, F 7×7 represents a large convolution kernel of 7×7, [] represents concatenation in the channel dimension, σ represents the sigmoid activation function, F 1×1 represents a point convolution with a convolution kernel of 1×1, S i represents the feature map after the sigmoid activation function, and S represents the feature map after the point convolution with a convolution kernel of 1×1.

[0070] Feature map M c After 3×3 convolution, it is activated by the ReLU activation function and then summed with the feature map M to obtain the feature map M r :

[0071] M r=M+ReLU(F 3×3 (M c ))

[0072] Among them, ReLU represents the ReLU activation function, F 3×3 Represents 3×3 convolution. Feature map M r After fusing the information of different channels, the feature maps are aggregated using channel-by-channel weighted combination to obtain the output Y of the multi-scale large-core attention module.

[0073] The information from different channels is fused to obtain richer semantic information. The fusion of feature maps is defined as:

[0074]

[0075] Where H, W, and C represent the height, width, and number of channels of the feature map, respectively, and V represents the fused feature map. j Represents different convolutions, M j represents the feature map to be fused, and then global pooling is applied to the global spatial information. The calculation of the c-th component is expressed as:

[0076]

[0077] The feature maps are aggregated using a channel-by-channel soft attention weighted combination, P c Represents the feature map after global pooling is applied, split weighted combination generates the channels of each feature map, and the c-th channel is calculated as:

[0078]

[0079] Among them, W c is the feature map after weighted combination, F i Represents different convolutions, M i represents the feature map to be weighted combined, x i (c) is the soft attention weight, the formula is:

[0080]

[0081] in represents the weight of the global spatial information P on the c-th channel. Finally, the feature map W is skipped through the connection c Fused with the initial feature map, the output Y of the multi-scale large-core attention module is obtained.

[0082] The multi-receptive field aggregation module serves as a bridge between the encoding and decoding parts and integrates the feature maps of different encoding stages to output to the decoding subnet.

[0083] Feature maps X at different encoding stages iAfter applying the maximum pooling, concatenate them in the channel dimension to obtain the feature map S c , the formula is:

[0084]

[0085] Where P max represents the maximum pooling, represents the feature map after maximum pooling, [] represents concatenation in the channel dimension, Represents the feature map X at different encoding stages i The feature maps obtained after applying the maximum pooling are respectively, in this embodiment, the feature maps are Feature Map and feature map

[0086] At the same time, the feature maps X at different encoding stages i After the cross-channel separation module, six different outputs O1, output O2, output O3, output O4, output O5 and output O6 are obtained respectively.

[0087] The output O1, output O2, output O3, output O4, output O5, and output O6 are weighted fused to obtain a feature map Q c , the formula is:

[0088]

[0089] Among them, x i (c) is the soft attention weight, the formula is:

[0090]

[0091] Where P represents the global spatial information obtained by global pooling of six different outputs, Represents the weight of the global spatial information P on the c-th channel, and the feature map S c and the feature map Q c The summation generates the output of the multi-receptive field integration module.

[0092] The cross-channel separation module first undergoes a 1×1 convolution and a 3×3 convolution to expand the receptive field and capture richer channel information to obtain the initial feature map of the cross-channel separation module.

[0093] The initial feature map of the cross-channel separation module is divided into two feature maps, M1 and M2, in the channel dimension; the feature map M1 is used as an output L1 of the cross-channel separation module; the feature map M1 and the feature map M2 are summed, and then a 3×3 convolution is used to realize cross-channel information fusion to obtain another output L2 of the cross-channel separation module.

[0094] like Figure 1 and Figure 5 As shown, the present invention also discloses a medical image segmentation method based on enhanced multi-scale convolution aggregation. Based on the above-mentioned medical image segmentation system based on enhanced multi-scale convolution aggregation, the medical image to be segmented is encoded and decoded by the medical image segmentation system based on enhanced multi-scale convolution aggregation. The encoder uses multiple multi-scale large-core attention modules to realize multi-scale feature fusion under different network depths. The decoder uses multiple multi-scale large-core attention modules to splice the encoded features of each module in the encoder and decode to obtain a segmented image, which specifically includes the following steps:

[0095] S1: Inputting the medical image to be segmented into the deep feature preservation module to obtain a first feature map;

[0096] S2: passing the first feature map through three layers of multi-scale large-core attention modules in sequence to obtain a first encoding feature, a second encoding feature, and a t-th encoding feature, respectively, where t represents the number of multi-scale large-core attention modules in the encoder; in this embodiment, t is 3;

[0097] S3: input all the encoding features into a multi-receptive field aggregation module to obtain a second feature map; in this embodiment, the first, second, and third encoding features are input into a multi-receptive field aggregation module to obtain a second feature map;

[0098] S4: upsampling the second feature map obtained in step S3 and splicing it with the t-th encoding feature to obtain the first decoding feature after passing through the first multi-scale large core attention module of the decoder; upsampling the first decoding feature and splicing it with the t-1-th encoding feature to obtain the second decoding feature after passing through the second multi-scale large core attention module of the decoder; and so on, until the t-1-th encoding feature is upsampled and spliced ​​with the first encoding feature to obtain the t-th decoding feature after passing through the t-th scale large core attention module; the t-th decoding feature is upsampled and spliced ​​with the first feature map to obtain the t+1-th multi-scale large core attention module of the decoder. +1 decoding feature; in this embodiment, the second feature map is upsampled and then spliced ​​with the third encoding feature through the first multi-scale large core attention module of the decoder to obtain a first decoding feature, the first decoding feature is upsampled and then spliced ​​with the second encoding feature through the second multi-scale large core attention module of the decoder to obtain a second decoding feature, the second decoding feature is upsampled and then spliced ​​with the first encoding feature through the third multi-scale large core attention module of the decoder to obtain a third decoding feature; the third decoding feature is upsampled and then spliced ​​with the first feature map through the fourth multi-scale large core attention module of the decoder to obtain a fourth decoding feature;

[0099] S5: The t+1th decoding feature obtained in step S4 is passed through a sigmoid layer and a 1×1 convolution layer to obtain the final image segmentation result. In this embodiment, the fourth decoding feature is passed through a sigmoid layer and a 1×1 convolution layer to obtain the final image segmentation result.

[0100] In order to demonstrate the effectiveness of the present invention, five evaluation indicators are used to compare the present invention with nine other state-of-the-art models on the CVC-ColonDB dataset. The five evaluation indicators are: F1 score, accuracy, precision, recall, and mean intersection and union. Nine state-of-the-art models: UNet, AttentionUnet, UNet++, ResUNet++, DCSAUNet, LW-IRSTNet, PDFUnet, AMSUnet, and MSDANet. Table 1 compares the present invention with other methods on the CVC-ColonD dataset. EMCANet represents the present invention. It can be seen from the data in Table 1 that the present invention is significantly superior to the other compared methods in terms of the five evaluation indicators.

[0101] Table 1

[0102]

[0103] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A medical image segmentation system based on enhanced multi-scale convolution aggregation, characterized in that: The network structure of the medical image segmentation system is a convolutional neural network with a U-shaped structure, including an encoder and a decoder for encoding and decoding the medical image to be segmented, and a multi-receptive field aggregation module embedded in the network structure of the medical image segmentation system as a conversion link between the encoder and the decoder, the encoder includes a deep feature retention module and a plurality of multi-scale large core attention modules, the decoder includes a plurality of multi-scale large core attention modules, each module of the encoder corresponds to a multi-scale large core attention module in the decoder, and each module of the encoder is jump-connected with the corresponding multi-scale large core attention module in the decoder, wherein: The encoder inputs the medical image to be segmented into the deep feature preservation module through the deep feature preservation module and performs maximum pooling, uses multiple multi-scale large-core attention modules to achieve multi-scale feature fusion under different network depths, and serves as multiple input features of the multi-receptive field aggregation module; The decoder concatenates the coded features of each module in the encoder through multiple multi-scale large-core attention modules and then decodes to obtain a segmented image; The input end of the multi-receptive field aggregation module is connected to the encoder, and the output end is connected to the decoder, so as to reduce the semantic gap between the encoding and decoding sub-networks, fuse the encoding features output by the multi-scale large-core attention module of the encoder, and transmit them to the multi-scale large-core attention module of the decoder, so as to collaboratively utilize the encoding outputs of different stages for feature fusion, thereby improving the network's ability to capture boundary features.

2. The medical image segmentation system based on enhanced multi-scale convolution aggregation according to claim 1, characterized in that: The deep feature preservation module replaces the traditional stacked standard convolution with a deep separable convolution. The deep feature preservation module adopts a residual structure, inputs the medical image to be segmented through a 3×3 convolution layer, increases the feature channel, expands the receptive field, and obtains the input feature map; The input feature map is passed through a 7×7 depth convolution and a 1×1 point convolution. The point convolution increases the number of channels by four times to obtain an intermediate feature map. After the intermediate feature map is processed by the ReLU activation function, the 1×1 point convolution reduces the number of channels to one-fourth of the original size, thereby obtaining a feature map with global features; The input feature map and the feature map with global features are summed to obtain an output result of the deep feature preservation module.

3. The medical image segmentation system based on enhanced multi-scale convolution aggregation according to claim 1, characterized in that: The multi-scale large kernel attention module combines large kernel convolution with cross-channel information interaction, splits the feature map into two parts in the channel dimension after 1×1 convolution, and then interacts with the two parts of the split feature map after different operations, thereby enhancing the multi-scale feature extraction capability and expanding the receptive field, and extracting deep representation features from the medical image to be segmented; The multi-scale large-core attention module adjusts the number of channels of the feature map through 1×1 convolution to obtain the initial feature map, which splits the feature map into feature maps M along the channel dimension. c and feature map M l , feature map M l , apply 3×3 convolution to capture more channel information and obtain the feature map M k ; The feature map M k After 5×5 depth convolution and 7×7 depth convolution respectively, the feature map is obtained and Feature Map and feature map Then, we sum them up through a 1×1 point convolution to get the summed feature map. The summed feature map Apply maximum pooling and average pooling to obtain the feature map S max and feature map S avg : Where P max represents the maximum pooling function, P avg Represents the average pooling function, and then the feature map M is obtained by the following formula: S j Split into S1 and S2 in the channel dimension M=M k ×S Among them, F 7×7 represents a large convolution kernel of 7×7, [] represents concatenation in the channel dimension, σ represents the sigmoid activation function, F 1×1 Represents a point convolution with a convolution kernel of 1×1; represents the feature map after convolution with a large kernel of 7×7; S j Representative feature map The feature map after the sigmoid activation function; S represents the feature map after the point convolution with a convolution kernel of 1×1; The characteristic map M c After 3×3 convolution, it is activated by the ReLU activation function and summed with the feature map M to obtain the feature map M r : M r =M+ReLU(F 3×3 (M c )) Among them, F 3×3 Represents 3×3 convolution, feature map M r After fusing the information of different channels, the feature maps are aggregated using channel-by-channel weighted combination to obtain the output Y of the multi-scale large-core attention module.

4. The medical image segmentation system based on enhanced multi-scale convolution aggregation according to claim 1, characterized in that: The multi-receptive field aggregation module performs feature fusion on the outputs of multiple multi-scale large-core attention modules as follows: Feature maps X at different encoding stages i After applying the maximum pooling, concatenate them in the channel dimension to obtain the feature map S c , the formula is: Represents the feature map X at different encoding stages i The feature maps obtained after applying the maximum pooling are Feature maps X at different encoding stages i After the cross-channel separation module, six different outputs O1, output O2, output O3, output O4, output O5 and output O6 are obtained respectively. The six outputs are weighted fused to obtain the feature map Q c , the formula is: Among them, x i (c) is the soft attention weight, the formula is: Among them, P represents the global spatial information obtained after global pooling of six different outputs, Represents the weight of the global spatial information P on the c-th channel, and the feature map S c and feature map Q c The added output is used as the output of the multi-receptive field aggregation module.

5. The medical image segmentation system based on enhanced multi-scale convolution aggregation according to claim 4, characterized in that: The cross-channel separation module expands the receptive field through a 1×1 convolution and a 3×3 convolution and captures richer channel information to obtain an initial feature map of the cross-channel separation module, and divides the initial feature map of the cross-channel separation module into two feature maps, M1 and M2, in the channel dimension; Using the feature map M1 as an output L1 of the cross-channel separation module; The feature map M1 and the feature map M2 are summed, and then a 3×3 convolution is performed to achieve cross-channel information fusion to obtain another output L2 of the cross-channel separation module.

6. A medical image segmentation method based on enhanced multi-scale convolution aggregation, based on the medical image segmentation system based on enhanced multi-scale convolution aggregation according to any one of claims 1 to 5, characterized in that: The medical image to be segmented is encoded and decoded by a medical image segmentation system based on enhanced multi-scale convolution aggregation. The encoder uses multiple multi-scale large-core attention modules to realize multi-scale feature fusion under different network depths. The decoder uses multiple multi-scale large-core attention modules to splice the encoded features of each module in the encoder and decode to obtain the segmented image. Specifically, the following steps are included: S1. Inputting the medical image to be segmented into the deep feature preservation module of the encoder to obtain a first feature map; S2. The first feature map obtained in step S1 is sequentially passed through a plurality of multi-scale large-core attention modules of the encoder to obtain a first encoding feature, a second encoding feature, and a t-th encoding feature, respectively, where t represents the number of multi-scale large-core attention modules in the encoder; S3. Input all the encoded features into the multi-receptive field aggregation module to obtain the second feature map; S4. Upsampling the second feature map obtained in step S3 and splicing it with the t-th encoding feature through the first multi-scale large core attention module of the decoder to obtain the first decoding feature, upsampling the first decoding feature and splicing it with the t-1-th encoding feature through the second multi-scale large core attention module of the decoder to obtain the second decoding feature, and so on, until the t-1-th encoding feature is upsampled and spliced ​​with the first encoding feature through the t-th scale large core attention module to obtain the t-th decoding feature; upsampling the t-th decoding feature and splicing it with the first feature map through the t+1-th multi-scale large core attention module of the decoder to obtain the t+1-th decoding feature; S5. Pass the t+1th decoding feature obtained in step S4 through a sigmoid layer and a 1×1 convolution layer to obtain the final image segmentation result.

Images