Medical image segmentation system and method based on enhanced multi-scale convolution aggregation

Abstract

The present application relates to the technical field of image segmentation, and the existing model based on convolutional neural network still has limitations such as limited receptive field and difficulty in extracting deep features in the image, and important feature channel information is ignored when using a skip connection for multi-scale information fusion. The present application provides a medical image segmentation system and method based on enhanced multi-scale convolution aggregation, which utilizes a multi-receptive field aggregation module between an encoder and a decoder to reduce the semantic gap between the encoding and decoding subnetworks, and after the encoding features output by the multi-scale large kernel attention module of the encoder are fused, the fused features are transmitted to the multi-scale large kernel attention module of the decoder, and the encoding outputs of different stages are cooperatively utilized for feature fusion, so that the ability of the network to capture boundary features is improved, and more accurate segmentation of the target region is realized.

Description

TECHNICAL FIELD

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

[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 architecture, the efficiency and accuracy of image segmentation have been significantly improved, and have been 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. Through end-to-end segmentation training, the accuracy of segmentation is further improved. However, medical images usually have characteristics such as high complexity, irregular and fuzzy boundaries, small feature regions, overlapping tissue structures, and low image contrast. Therefore, it is particularly important to construct a segmentation model suitable for medical images. The UNet model has a milestone significance in the field of medical image segmentation, and its performance in various medical image segmentation tasks is better than traditional image processing methods. The UNet model adopts a fully symmetric encoder-decoder structure with skip connections, which can effectively combine low-level and high-level semantic information to generate the final mask. However, when dealing with 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 skip connections in the network, thereby improving the segmentation performance. DoubleU-Net proposes to combine two Unet structures. The first Unet is used to learn the context information and pass the features to the second Unet to achieve more efficient information extraction. LCU-Net enhances segmentation performance through concatenation operations and introduces Inception modules to expand the receptive field, so as to be able to extract deeper features. 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 field is relatively limited, making it difficult to fully extract deep features in the image. At the same time, when using skip connections for multi-scale information fusion, these models often ignore important feature channel information. SUMMARY

[0004] In order to solve the above technical problems, the application provides a medical image segmentation system and method based on enhanced multi-scale convolution aggregation, which expands the receiving domain by using a deep feature preservation module in the processing of an input image, effectively reduces the loss of key detail features, realizes cross-channel interaction of multi-scale information, and reduces information loss in the severe down-sampling process.

[0005] The application is mainly realized by the following technical solutions:

[0006] A medical image segmentation system based on enhanced multi-scale convolution aggregation, the network structure of the medical image segmentation system is a convolutional neural network with a U-shaped structure, including an encoder, 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 preservation module and a plurality of multi-scale large kernel attention modules, the decoder includes a plurality of multi-scale large kernel attention modules, each module of the encoder corresponds to a multi-scale large kernel attention module in the decoder, and each module of the encoder is connected to the multi-scale large kernel attention module in the corresponding decoder by jumping connection, wherein:

[0007] The encoder inputs the medical image to be segmented into the deep feature preservation module and performs maximum pooling by the deep feature preservation module, realizes multi-scale feature fusion at different network depths by a plurality of multi-scale large kernel attention modules, and serves as a plurality of input features of the multi-receptive field aggregation module;

[0008] The decoder decodes the encoded features of each module in the encoder after splicing by a plurality of multi-scale large kernel attention modules to obtain a segmentation image;

[0009] The input end of the multi-receptive field aggregation module is connected to the encoder, and the output end is connected to the decoder, thereby reducing the semantic gap between the encoding and decoding sub-networks, fusing the encoded features output by the multi-scale large kernel attention modules of the encoder, and transmitting the features to the multi-scale large kernel attention modules of the decoder, so as to cooperatively utilize the encoding outputs at different stages for feature fusion and improve the ability of the network to capture boundary features.

[0010] Further, the deep feature preserving module replaces the traditional stacked standard convolution with a depth separable convolution, the deep feature preserving module adopts a residual structure, an input medical image to be segmented passes through a 3x3 convolution layer, increases the feature channel, expands the receptive field, and obtains an input feature map; the input feature map passes through a 7x7 depth convolution and a 1x1 point convolution, the point convolution increases the number of channels by four times, and an intermediate feature map is obtained; after the intermediate feature map is processed by a ReLU activation function, a 1x1 point convolution reduces the number of channels to one fourth of the original size, and a feature map with global features is obtained; the input feature map and the feature map with global features are summed to obtain the output result of the deep feature preserving module.

[0011] Further, the multi-scale large kernel attention module combines a large kernel convolution with cross-channel information interaction, splits the feature map into two parts in the channel dimension after 1x1 convolution, and then interacts the two parts of the split feature map after different operations to enhance the multi-scale feature extraction capability and expand the receptive field to extract 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 1x1 convolution to obtain an initial feature map, the initial feature map splits the feature map along the channel dimension into feature map M c and feature map M l , feature map M l , applies 3x3 convolution to capture more channel information, and obtains feature map M k ; after feature map M k passes through 5x5 depth convolution and 7x7 depth convolution respectively, feature maps and feature maps and feature maps are obtained respectively, and then a 1x1 point convolution is performed to obtain a summed feature map The summed feature map is applied with max pooling and average pooling to obtain feature maps S max and S avg :

[0012]

[0013] where P max represents a max pooling function, P avg represents an average pooling function, and the feature map M is obtained through the following formula:

[0014]

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

[0016]

[0017] M = M k × S

[0018] wherein F 7×7 represents a large kernel convolution with a 7x7 convolution kernel, [] represents concatenation in the channel dimension, σ represents a sigmoid activation function, F 1×1 represents a point convolution with a 1x1 convolution kernel; represents a feature map after a large kernel convolution with a 7x7 convolution kernel; S j represents a feature map after a sigmoid activation function; S represents a feature map after a point convolution with a 1x1 convolution kernel;

[0019] a feature map M c is obtained by 3x3 convolution, activated by a ReLU activation function, summed with a feature map M r :

[0020] M r = M + ReLU (F 3×3 (M c ))

[0021] wherein F 3×3 represents 3x3 convolution, and a feature map M r fuses information of different channels and then aggregates the feature maps by using a channel-by-channel weighted combination to obtain an output Y of the multi-scale large kernel attention module.

[0022] Further, the multi-receptive field aggregation module fuses features of outputs of a plurality of multi-scale large kernel attention modules as follows:

[0023] The feature maps X i of different encoding stages are applied with maximum pooling and then concatenated in the channel dimension to obtain a feature map S c , which is expressed by the formula:

[0024]

[0025] The feature maps X i of different encoding stages are applied with maximum pooling and then concatenated in the channel dimension to obtain a feature map S

[0026]

[0027] The feature maps X iSix different outputs O1, O2, O3, O4, O5 and O6 are obtained through the cross-channel separation module, and the six outputs are fused by weighting to obtain a feature map Q c , and the formula is:

[0028]

[0029] wherein x i (c) is a soft attention weight value, and the formula is:

[0030]

[0031] wherein P represents global spatial information obtained by globally pooling the six different outputs, represents a weight of the c-th channel, and the feature map S c and the feature map Q c are added to output as the output of the multi-receptive field aggregation module.

[0032] Further, the cross-channel separation module expands the receptive field and captures more abundant channel information through a 1x1 convolution and a 3x3 convolution, obtains a cross-channel separation module initial feature map, divides the cross-channel separation module initial feature map into two feature maps M1 and M2 in the channel dimension; the feature map M1 is taken as an output L1 of the cross-channel separation module; the feature map M1 and the feature map M2 are summed, and then cross-channel information fusion is realized through a 3x3 convolution 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 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 realizes multi-scale feature fusion at different network depths by using multiple multi-scale large kernel attention modules, and the decoder decodes the encoded features of each module in the encoder after splicing to obtain a segmentation image, and specifically includes the following steps:

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

[0035] S2. sequentially pass the first feature map obtained in step S1 through a plurality of multi-scale large kernel attention modules of the encoder to obtain a first encoded feature, a second encoded feature,..., and a tth encoded feature, wherein t represents the number of multi-scale large kernel attention modules in the encoder;

[0036] S3. input all the encoded features into a multi-acceptance domain aggregation module to obtain a second feature map;

[0037] S4. upsample the second feature map obtained in step S3, splice it with the tth encoded feature, and pass it through the first multi-scale large kernel attention module of the decoder to obtain the first decoded feature, upsample the first decoded feature, splice it with the t-1th encoded feature, and pass it through the second multi-scale large kernel attention module of the decoder to obtain the second decoded feature, and so on, until the t-1th encoded feature is upsampled, spliced with the first encoded feature, and passed through the tth scale large kernel attention module to obtain the tth decoded feature; the tth decoded feature is upsampled, spliced with the first feature map, and passed through the t+1th multi-scale large kernel attention module of the decoder to obtain the t+1th decoded feature;

[0038] S5. pass the t+1th decoded feature obtained in step S4 through a sigmoid layer and a 1x1 convolution layer to obtain the final image segmentation result.

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

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

[0041] 2. The deep feature preservation module is used in the processing of the input medical image to expand the acceptance domain, effectively reducing the loss of key detail features in the medical image, and the deep convolution and point convolution are used in the deep feature preservation module instead of ordinary convolution to improve the calculation efficiency, the application provides a medical image segmentation system and method capable of more comprehensively capturing and fusing deep features and multi-scale information in images, realizing efficient segmentation of medical images, and can be applied to the field of medical image segmentation.

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

[0043] 4. The multi-acceptance domain aggregation module is used between the encoding and decoding stages to reduce the semantic gap between the encoding and decoding sub-networks, and the feature fusion is performed using the encoding outputs of different stages, thereby realizing more accurate segmentation of the target region. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 It is a network structure diagram of the image segmentation system of the application;

[0045] Figure 2 It is a schematic diagram of a deep feature preservation module;

[0046] Figure 3 schematic diagram of a multi-scale large kernel attention module;

[0047] Figure 4 schematic diagram of a multi-acceptance domain aggregation module;

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

[0049] Figure 6 qualitative comparison of the present application and the rest of the methods on the CVC-ColonD dataset. DETAILED DESCRIPTION

[0050] Embodiments of the present application are described below in detail, examples of which are shown in the accompanying drawings, in which the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the accompanying drawings are exemplary and are intended to explain the embodiments of the present application, and cannot be understood as limiting the present application.

[0051] As shown in Figures 1-5 The present application discloses a medical image segmentation system based on enhanced multi-scale convolution aggregation. The network structure of the medical image segmentation system is a convolutional neural network with a U-shaped structure, which includes an encoder, a decoder, and a multi-acceptance domain aggregation module embedded in the network structure of the medical image segmentation system as a conversion link between the encoder and the decoder, for encoding and decoding the medical image to be segmented. The encoder includes a deep feature preservation module and a plurality of multi-scale large kernel attention modules, and the decoder includes a plurality of multi-scale large kernel attention modules. Each module of the encoder corresponds to a multi-scale large kernel attention module in the decoder, and each module of the encoder is connected to the corresponding multi-scale large kernel attention module in the decoder by a skip connection. In the Figure 1 In the embodiment shown, three multi-scale large kernel attention modules are provided in the encoder, and four multi-scale large kernel attention modules are provided in the decoder. The deep feature preservation module and the three multi-scale large kernel attention modules in the encoder each correspond to a multi-scale large kernel attention module.

[0052] The encoder inputs the medical image to be segmented into the deep feature preservation module and performs maximum pooling through the deep feature preservation module. The multi-scale large kernel attention modules are used to realize multi-scale feature fusion at different network depths and serve as multiple input features of the multi-acceptance domain aggregation module. The decoder decodes the encoded features of each module in the encoder after splicing them through the multi-scale large kernel attention modules to obtain a segmentation 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 and cross-channel information interaction. After 1x1 convolution, the feature map is split into two parts in the channel dimension. The first part is processed by ordinary convolution to retain the original features, and the other part is processed by different large kernel convolution operations to extract deeper features. The feature maps are interacted after different operations. In this embodiment, 7x7 and 5x5 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-acceptance domain 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-acceptance domain aggregation module is connected to the encoder, and the output end is connected to the decoder. This reduces the semantic gap between the encoding and decoding sub-networks, effectively transmits the encoding results of multiple stages to the decoding stage, realizes the fusion of multi-scale features at different network depths, and cooperatively utilizes the encoding outputs of different stages for feature fusion, as shown in Figure 1 The three input ends of the multi-acceptance domain aggregation module are respectively connected to the outputs of the three multi-scale large kernel attention modules. The multi-acceptance domain 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 ability of the network 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. The input medical image to be segmented is processed through a 3x3 convolution layer to increase the feature channels and expand the receptive field, obtaining the input feature map.

[0056] The input feature map is processed through a 7x7 depth convolution and a 1x1 point convolution. The depth convolution does not change the number of channels, while the point convolution increases the number of channels by four times, obtaining the intermediate feature map.

[0057] After the intermediate feature map is processed through a ReLU activation function, a 1x1 point convolution reduces the number of channels to one quarter of the original size, obtaining a feature map with global features. By changing the number of channels through point convolution, the number of channels can be increased or decreased according to specific needs. 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 1x1 convolution to obtain an initial feature map.

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

[0061] The feature maps M l apply a 3x3 convolution to capture more channel information, resulting in feature maps M k .

[0062] The feature maps M k respectively through a 5x5 deep convolution and a 7x7 deep convolution, resulting in feature maps and feature maps Feature maps and feature maps are summed through a 1x1 point convolution respectively to obtain the summed feature maps.

[0063] Maximum pooling and average pooling are applied to the summed feature maps to obtain the maximum pooled feature maps S max and the average pooled feature maps S avg :

[0064]

[0065] where P max represents the maximum pooling function, P avg represents the average pooling function, and to fuse more rich spatial information, the features are connected through a concat concatenation function to obtain feature maps The feature maps M are obtained through the following formula:

[0066]

[0067]

[0068] M = M k x S

[0069] where F 7×7 represents a large kernel convolution with a 7x7 convolution kernel, [] represents concatenation in the channel dimension, σ represents a sigmoid activation function, F 1×1 represents a point convolution with a 1x1 convolution kernel, S i represents a feature map after sigmoid activation, and S represents a feature map after point convolution with a 1x1 convolution kernel.

[0070] The feature maps M c are activated by a 3x3 convolution and a ReLU activation function, and then summed with the feature maps M to obtain the feature maps M r :

[0071] M r= M + ReLU(F 3×3 (M c ))

[0072] where ReLU denotes a ReLU activation function, F 3×3 represents a 3x3 convolution. The feature map M r After fusing the information of different channels, the feature maps are aggregated by using a channel-wise weighted combination to obtain the output Y of the multi-scale large kernel attention module.

[0073] The information of different channels is fused to obtain more rich semantic information, and the fusion of the feature maps is defined as:

[0074]

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

[0076]

[0077] The feature maps are aggregated by using a channel-wise soft attention weighted combination, P c represents the feature maps after applying global pooling, and the channel of each feature map is generated by splitting and weighted combination, and the calculation of the c-th channel is:

[0078]

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

[0080]

[0081] where represents the weight of the global spatial information P on the c-th channel. Finally, the feature map W c is fused with the initial feature map through a skip connection to obtain the output Y of the multi-scale large kernel attention module.

[0082] The multi-receptive field aggregation module integrates the feature map outputs of different encoding stages into the decoding subnetwork as a connection bridge between the encoding and decoding parts.

[0083] The feature maps X iThe feature maps S are spliced in the channel dimension after maximum pooling to obtain a feature map S c , which is expressed by the formula:

[0084]

[0085] where P max represents maximum pooling, represents a feature map after maximum pooling, and [] represents splicing in the channel dimension, represents a feature map X i obtained by applying maximum pooling to different encoding stages, which are feature maps , respectively, in this embodiment.

[0086] Meanwhile, the feature maps X i of different encoding stages are respectively obtained by the cross-channel separation module to obtain six different outputs O1, O2, O3, O4, O5, and O6.

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

[0088]

[0089] where x i (c) is a soft attention weight, which is expressed by the formula:

[0090]

[0091] where P represents global spatial information obtained by performing global pooling on the six different outputs, represents a weight of the global spatial information P on the c-th channel, and the feature map S c and the feature map Q c are added to generate an output of the multi-acceptance field integration module.

[0092] The cross-channel separation module first performs a 1×1 convolution and a 3×3 convolution to expand the acceptance field and capture richer channel information to obtain an 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 taken 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 realize cross-channel information fusion to obtain another output L2 of the cross-channel separation module.​

[0094] As shown in Figure 1 and Figure 5 The application also discloses 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, 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 realizes multi-scale feature fusion at different network depths by using a plurality of multi-scale large kernel attention modules, and the decoder decodes the encoded features of each module in the encoder after splicing to obtain a segmentation image by a plurality of multi-scale large kernel attention modules, and the method comprises the following steps:

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

[0096] S2: sequentially passing the first feature map through three layers of multi-scale large kernel attention modules to obtain a first encoded feature, a second encoded feature, and a tth encoded feature, wherein t represents the number of multi-scale large kernel attention modules in the encoder; in this embodiment, t is 3;

[0097] S3: inputting all the encoded features into a multi-acceptance domain aggregation module to obtain a second feature map; in this embodiment, the first, second and third encoded features are inputted into the multi-acceptance domain aggregation module to obtain the second feature map;

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

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

[0100] To prove the effectiveness of the present application, the present application is compared with other nine most advanced models of the CVC-ColonDB dataset using five evaluation indexes. The five evaluation indexes are: F1 score, accuracy, precision, recall, and average intersection over union. The nine most advanced models are: UNet, AttentionUnet, UNet++, ResUNet++, DCSAUNet, LW-IRSTNet, PDFUnet, AMSUnet, and MSDANet. Table 1 shows the comparison between the present application and the other methods on the CVC-ColonD dataset. EMCANet represents the present application. As can be seen from the data in Table 1, the present application is obviously superior to the other compared methods in the five evaluation indexes.

[0101] Table 1

[0102]

[0103] The above only describes the preferred embodiments of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solution falling within the concept of the present application shall fall within the protection scope of the present application. It should be noted that, for ordinary skilled persons in the art, some improvements and refinements without departing from the principles of the present application shall also be considered as falling within the protection scope of the present application.

Claims

1. A medical image segmentation system based on enhanced multi-scale convolution aggregation, characterized by, The network structure of the medical image segmentation system is a convolutional neural network with a U-shaped structure, including an encoder for encoding the medical image to be segmented, a decoder, 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 preservation module and a plurality of multi-scale large kernel attention modules, the decoder includes a plurality of multi-scale large kernel attention modules, each module of the encoder corresponds to a multi-scale large kernel attention module in the decoder, and each module of the encoder is connected to the corresponding multi-scale large kernel attention module in the decoder by jumping connection, wherein: The encoder inputs the medical image to be segmented into the deep feature preservation module and performs maximum pooling through the deep feature preservation module, uses a plurality of multi-scale large kernel attention modules to realize multi-scale feature fusion at different network depths, and uses the plurality of multi-scale large kernel attention modules as a plurality of input features of the multi-receptive field aggregation module; The decoder decodes the encoded features of each module in the encoder through a plurality of multi-scale large kernel attention modules to obtain a segmentation image; The multi-scale large kernel attention module combines large kernel convolution and cross-channel information interaction, splits the feature map into two parts after 1x1 convolution in the channel dimension, and then interacts the two parts after different operations to enhance the multi-scale feature extraction capability and expand the receptive field to extract deep representation features from the medical image to be segmented; The input end of the multi-receptive field aggregation module is connected to the encoder, and the output end is connected to the decoder, thereby reducing the semantic gap between the encoding and decoding sub-networks, fusing the encoded features output by the multi-scale large kernel attention modules of the encoder, and transmitting the fused features to the multi-scale large kernel attention modules of the decoder, so as to cooperatively utilize the encoded outputs at different stages for feature fusion and improve the network's ability to capture boundary features; The multi-receptive field aggregation module fuses the outputs of the plurality of multi-scale large kernel attention modules as follows: Feature maps of different encoded stages X i After applying max pooling, concatenation is performed on the channel dimension to obtain feature maps S c The formula is: feature maps representing different encoding stages X i respectively applying max pooling to the feature maps, respectively , ; Feature maps of different encoded stages X i Six different outputs are obtained respectively through the cross-channel separation module O 1, output O 2, output O 3, output O 4, output O 5 and output O 6, the six outputs are weighted and fused to obtain a feature map Q c The formula is: wherein, x i c is a soft attention weight, the formula is:​ wherein, P represents global spatial information obtained after global pooling on six different outputs, represents global spatial information P a weight of the c-th channel, and the feature map S c and the feature map Q c are added and output as the output of the multi-receptive field aggregation module; The cross-channel separation module accepts the domain and captures richer channel information through a 1x1 convolution and a 3x3 convolution expansion, and obtains a cross-channel separation module initial feature map, and the cross-channel separation module initial feature map is divided into M 1 and M 2 two feature maps in the channel dimension; said feature map M 1 as an output of said cross-channel separation module L 1; The feature map M 1 and the feature map M 2 are summed up, and cross-channel information fusion is realized through a 3x3 convolution to obtain another output L 2 of the cross-channel separation module.

2. The medical image segmentation system based on enhanced multi-scale convolution aggregation according to claim 1, wherein, The deep feature preservation module uses a depth separable convolution to replace a traditional stacked standard convolution, and adopts a residual structure, wherein an input medical image to be segmented is input into a 3x3 convolution layer to increase the feature channel and expand the receptive field to obtain an input feature map; The input feature map is input into a 7x7 depth convolution and a 1x1 point convolution, the point convolution increases the number of channels by four times to obtain an intermediate feature map; The intermediate feature map is processed by a ReLU activation function, and the 1x1 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.

3. The medical image segmentation system based on enhanced multi-scale convolution aggregation according to claim 1, wherein, The multi-scale large kernel attention module adjusts the channel number of the feature map through 1x1 convolution to obtain an initial feature map, and the initial feature map splits the feature mapping along the channel dimension into feature maps M C and feature maps M l , feature maps M l , and applies 3x3 convolution to capture more channel information to obtain feature maps M k ; The feature maps M k After 5x5 depth convolution and 7x7 depth convolution respectively, the feature maps and , the feature maps and the feature maps are summed up by a 1x1 point convolution respectively to obtain summed feature maps ; maximum pooling and average pooling are applied on the summed feature maps to obtain feature maps S max and feature maps S avg : wherein P max represents a max-pooling function, P avg represents an average-pooling function, and the feature map is obtained by the following formula M : Will S j Decomposed into channels S 1 and S 2 wherein, F 7×7 represents a large kernel convolution with a convolution kernel of 7x7, [ ] represents concatenation in the channel dimension, By encoding and decoding the medical image to be segmented through the medical image segmentation system based on the enhanced multi-scale convolution aggregation, the encoder uses a plurality of multi-scale large kernel attention modules to realize multi-scale feature fusion at different network depths, and the decoder decodes the encoded features of each module in the encoder through a plurality of multi-scale large kernel attention modules to obtain a segmentation image, which specifically includes the following steps: represents a sigmoid activation function, F 1×1 represents a point convolution with a convolution kernel of 1x1; represents a feature map after a large kernel convolution with a convolution kernel of 7x7; S j represents a feature map a feature map after a sigmoid activation function; S represents a feature map after a point convolution with a convolution kernel of 1x1; the feature map M C by a 3x3 convolution, activated by a ReLU activation function, with the feature map M summed to obtain a feature map M r : wherein, F 3×3 representing a 3x3 convolution, a feature map M r fusing information of different channels and then adopting a channel-by-channel weighting combination to aggregate the feature maps to obtain an output of the multi-scale large kernel attention module Y .

4. The medical image segmentation method based on enhanced multi-scale convolution aggregation according to any one of claims 1-3, based on the medical image segmentation system based on enhanced multi-scale convolution aggregation, characterized in that, ​ S1. input a medical image to be segmented into a deep feature preserving module of an encoder to obtain a first feature map; S2. sequentially pass the first feature map obtained in step S1 through a plurality of multi-scale large kernel attention modules of the encoder to obtain a first encoding feature, a second encoding feature, …, and a tth encoding feature, wherein t represents the number of the multi-scale large kernel attention modules in the encoder; S3. input all the encoding features into a multi-acceptance domain aggregation module to obtain a second feature map; S4. up-sample the second feature map obtained in step S3, splice the tth encoding feature, and pass the splice through a first multi-scale large kernel attention module of a decoder to obtain a first decoding feature, up-sample the first decoding feature, splice the t-1th encoding feature, and pass the splice through a second multi-scale large kernel attention module of the decoder to obtain a second decoding feature, and so on and so forth until the t-1th encoding feature is up-sampled, spliced with the first encoding feature, and passed through a tth multi-scale large kernel attention module to obtain a tth decoding feature; up-sample the tth decoding feature, splice the first feature map, and pass the splice through a (t+1)th multi-scale large kernel attention module of the decoder to obtain a (t+1)th decoding feature; S5. pass the (t+1)th decoding feature obtained in step S4 through a sigmoid layer and a 1×1 convolution layer to obtain a final image segmentation result.

Images