Medical image automatic segmentation method based on dual-aggregation hybrid neural network

Abstract

The invention discloses a medical image automatic segmentation method based on a dual-aggregation hybrid neural network, and the method comprises the steps: inputting a to-be-segmented medical image into a pre-trained medical image segmentation network, and outputting a medical image segmentation result; according to the medical image segmentation network, a dual-polymerization mixed U-shaped network is adopted, the dual-polymerization mixed U-shaped network comprises two parallel encoders and a single decoder, and the two parallel encoders are a convolutional neural network and a Transformer respectively; local features and global features of the medical image are extracted at the same time through a convolutional neural network and a Transform; performing multi-scale feature fusion on the extracted local features and global features through a collaborative feature aggregation module CFA and a non-collaborative feature aggregation module NCFA; and inputting the features subjected to multi-scale feature fusion into a decoder, and outputting a medical image segmentation result. According to the invention, multi-scale feature fusion can be realized, the segmentation precision of a medical image segmentation result is enhanced, and the performance of a medical image segmentation network is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image segmentation, and in particular relates to an automatic medical image segmentation method based on a dual-aggregation hybrid neural network. Background Art

[0002] The purpose of medical image segmentation is to extract regions of interest in medical images for computer-aided diagnosis. It can also help healthcare professionals with injury visualization, disease monitoring, and treatment planning.

[0003] Convolutional neural networks (CNNs) have powerful representation extraction capabilities and have achieved great success in medical image analysis. Among them, the U-net (UNet) is the most widely used. It consists of a symmetrical encoder-decoder network, and high-resolution features of different scales are aggregated using skip connections to enhance detail retention. Its derivative network models (such as UNet++, Attention UNet, Res-UNet) have successively achieved state-of-the-art performance in various segmentation tasks, but CNN-based methods have poor global modeling capabilities due to inherent inductive bias. In particular, the segmentation performance is poor in terms of texture, shape, and size, which involve global vision.

[0004] Transformer, a brand new architecture that only requires an attention mechanism to model global dependencies without any convolution operations, has achieved great success in the field of natural language processing (NLP). Researchers have tried to introduce Transformer into the field of computer vision (CV), and the results show that under the condition of pre-training on large-scale datasets, Transformer provides better performance than CNN in computer vision CV and has become an alternative architecture. Among them, Vision Transformer (ViT) divides the image into a series of tokens to adapt to the input requirements and models their global dependencies by stacking Transformer blocks. ViT achieves impressive performance on image classification tasks, but requires large datasets for pre-training. ViT's results in image classification are outstanding, but due to its low-resolution feature maps and the quadratic growth of complexity with the image size, its structure is not suitable as a general backbone network other than image classification, such as dense vision tasks or high-resolution input images. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a medical image automatic segmentation method based on a dual-aggregation hybrid neural network, which includes: inputting a medical image to be segmented into a pre-trained medical image segmentation network, and outputting a medical image segmentation result;

[0006] The medical image segmentation network adopts a dual-aggregation hybrid U-type network, which includes two parallel encoders and a single decoder, and the two parallel encoders are convolutional neural networks CNN and Transformer respectively;

[0007] The local features and global features of medical images are extracted simultaneously through convolutional neural networks (CNN) and Transformer.

[0008] The extracted local features and global features are fused at multiple scales through the collaborative feature aggregation module CFA and the non-collaborative feature aggregation module NCFA;

[0009] The multi-scale feature fusion features are input into the decoder and the medical image segmentation results are output.

[0010] The training process of the medical image segmentation network includes:

[0011] S1: Collect medical image sets, annotate and preprocess them, and save them in a medical image database;

[0012] S2: Screen images from the medical image database and obtain training and test sets after data augmentation;

[0013] S3: Medical images and their segmented images are selected from the training set and input into the medical image segmentation network, respectively, and the global and local features of the medical images are extracted after 4 downsamplings through CNN and Transformer, respectively;

[0014] S4: Feature fusion of the extracted global features and local features through the collaborative aggregation module CFA and the non-collaborative aggregation module NCFA;

[0015] S5: Input the fused features into a decoder to obtain predicted segmentation results of the medical image and its segmented image respectively;

[0016] S6: Calculate the loss between the predicted segmentation results of the medical image and its segmented image, and optimize the medical image segmentation network through back propagation of the loss.

[0017] Beneficial effects of the present invention:

[0018] First, the present invention adopts a dual encoder-single decoder U-shaped network architecture for medical image segmentation, wherein the dual encoder includes CNN and Transformer; this architecture can combine the edge local semantic information of CNN with the global context interaction of Transformer, thereby enhancing the integration of complementary features and obtaining better segmentation performance. In medical image segmentation, this method performs better in dealing with irregular and challenging boundaries.

[0019] Secondly, the present invention proposes a collaborative fusion and non-collaborative fusion method of heterogeneous hierarchical downsampling features, that is, the hierarchical features collected by CNN and Transformer are fused and processed together in a collaborative and non-collaborative manner based on a multi-attention mechanism. It effectively integrates coarse-grained and fine-grained feature representations and significantly improves the accuracy of medical image segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A schematic diagram of the structure of a medical image segmentation network in an embodiment of the present invention;

[0021] Figure 2 This is a schematic diagram of the network structure of a CNN in an embodiment of the present invention;

[0022] Figure 3 This is a schematic diagram of the network structure of the collaborative feature aggregation module CFA in an embodiment of the present invention;

[0023] Figure 4 Schematic diagram of the network structure of the non-collaborative feature aggregation module NCFA in an embodiment of the present invention. DETAILED DESCRIPTION

[0024] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] The embodiment of the present invention proposes a medical image automatic segmentation method based on a dual-aggregation hybrid neural network. Figure 1 As shown, the method includes: inputting the medical image to be segmented into a pre-trained medical image segmentation network, and outputting the medical image segmentation result;

[0026] The medical image segmentation network adopts a dual-aggregation hybrid U-type network, which includes two parallel encoders and a single decoder, and the two parallel encoders are convolutional neural networks CNN and Transformer respectively;

[0027] The local features and global features of medical images are extracted simultaneously through convolutional neural networks (CNN) and Transformer.

[0028] The extracted local features and global features are fused at multiple scales through the collaborative feature aggregation module CFA and the non-collaborative feature aggregation module NCFA;

[0029] The multi-scale feature fusion features are input into the decoder and the medical image segmentation results are output.

[0030] Figure 1 Schematic diagram of the structure of the medical image segmentation network in an embodiment of the present invention. Figure 1 In the invention, the medical image segmentation network adopts a dual-aggregation hybrid U-type network, which includes two parallel encoders and a single decoder. The image input end is connected to two parallel encoders, which are a convolutional neural network CNN and a Transformer. The input image is input into the Transformer after embedding. In the embodiment of the present invention, the Transformer includes SwinTransformer and Patch merging to extract the local features and global features of the input image data. The input image is input into the CNN to extract the local features and global features of the input image data. After the first downsampling, the sampled image data is respectively input into the first layer of Swin Transformer and the first layer of CNN. The first layer of Swin Transformer and the first layer of CNN output global features and local features respectively, which are input into the first collaborative feature aggregation module (Collaborative Feature Aggregation, CFA) for feature fusion. The outputs of the second layer of Swin Transformer, the second layer of CNN and the first CFA are respectively input into the second CFA, and so on. After at least 4 downsampling processes, the fused features are obtained. The outputs of each layer of SwinTransformer and each layer of CNN are used as non-collaborative feature aggregation modules (NCFA), and after at least 4 downsampling processes, fused features are obtained. The fused features output by CFA are reconstructed, and the reconstructed features and the fused features output by at least 4 NCFAs are input into the decoder to obtain the image segmentation result.

[0031] Reference Figure 1As shown, the backbone network Swin Transformer (ST) of the Transformer in an embodiment of the present invention includes at least 4 layers of Swin Transformer Layer (STL). The first three layers of STL are composed of two Swin TransformerBlock (STB) and Patch Merging, and the last layer of STL is composed of only two STBs. The Swin Transformer (ST) block consists of a window-based multi-head self attention (W-MSA) module or a shifted window-based multi-head self attention (SW-MSA) module and a 2-layer MLP. A LN (Layer Norm) layer is applied before each multi-headself attention (MSA) module and each MLP module, and a residual connection is applied after each module. In two consecutive SBT blocks, W-MSA modules or SW-MSA modules are used alternately to achieve cross-window connection while maintaining efficient calculation of non-overlapping windows. The calculation formula is:

[0032]

[0033] in, represents the output of the l-th layer (S)W-MSA module, (S)W-MSA(.) represents the (moving) window multi-head self-attention mechanism, LN(.) represents layer normalization, MLP(.) represents multi-layer perceptron, x l represents the output of the l-th layer MLP module, and x l+1 Represents the above operation of the l+1th layer.

[0034] The calculation of the self-attention mechanism is as follows:

[0035]

[0036] Among them, Attention(Q,K,V) represents the calculation of the multi-head self-attention mechanism for Q,K,V, Q,K,V represent the query matrix, key matrix and value matrix respectively. p represents the number of patches in the window, d represents the dimension of the query or key, and the values ​​in B are taken from the bias matrix

[0037] The patch merging block downsamples the input feature map of dimension [N, C, H, W] with 0 and 1 as the starting point and 2 as the step size in the H and W dimensions to obtain 4 feature maps of dimension [N, C, H / 2, W / 2]. These 4 feature maps are channel-concatenated to obtain feature maps of dimension [N, C*4, H / 2, W / 2]. In order to align with the feature channels corresponding to the output of the CNN module, a linear layer is applied after channel concatenation to unify the output feature dimension to twice the input dimension, obtaining a feature map of dimension [N, C*2, H / 2, W / 2].

[0038] Figure 2 Schematic diagram of the network structure of CNN in an embodiment of the present invention.

[0039] As some preferred embodiments, refer to Figure 2 As shown, the convolutional neural network CNN includes a root layer and three body layers, the root layer includes a 7×7 convolution kernel conv, a group normalization layer (Group Normalization, GN) and a Relu activation function layer; the body layer includes N bottleneck blocks (bottleneck); each bottleneck block (bottleneck) includes 3 layers of cascaded depthwise separable convolution kernels (DWConv), a group normalization layer GN and a Relu activation function.

[0040] Figure 2 In the figure, CNN includes a root layer and three body layers. The root layer uses standard convolution with a kernel of 7, a stride of 2, and a padding of 3 for feature representation learning and downsampling. The three body layers first use DWConv with a stride of 2 for feature representation learning and downsampling, and then use multiple DWConv for deeper feature representation learning. All convolution operations are followed by a group normalization layer (GN), and each layer is finally processed with an activation function ReLU. Finally, the output features are linearly embedded and positionally encoded for subsequent feature fusion.

[0041] Figure 3 Schematic diagram of the network structure of the collaborative feature aggregation module CFA in an embodiment of the present invention.

[0042] As some preferred embodiments, refer to Figure 3As shown, the collaborative feature aggregation module CFA includes N layers of cascaded layer normalization layers LN and multi-head channel attention (MCA), and then connected to a layer normalization layer (LN) and a multilayer perceptron (MLP).

[0043] Figure 3 In the figure, the features output by ST (Swin Transformer), CNN and the previous CFA are embedded into the same dimension after LN, and the three embedded features are input into MCA after channel concatenation and output after N calculations. Finally, the global and local features extracted by collaborative learning are processed by LN and MLP. E1 to E3 represent three different embedded features.

[0044] Figure 4 Schematic diagram of the network structure of the non-collaborative feature aggregation module NCFA in an embodiment of the present invention.

[0045] The non-collaborative feature aggregation module NCFA includes a concatenation layer contact, N cascaded layer normalization layers LN, multi-head self-attention (MSA), layer normalization layer LN and a multi-layer perceptron MLP, and then connected to a reconstruction recovery layer Reconstruct, S1~S4 represent 4 non-collaborative output features.

[0046] Figure 4 In the figure, the hierarchical down-sampling features collected by the Transformer branch and the CNN branch are processed by the embedding layer and then channel-joined. They are then sent to the N-layer feature aggregation module composed of LN, MSA, residual connection and MLP to restore the original shape and perform residual connection with the input to keep the features stable. The final output is the non-collaborative feature.

[0047] In some preferred embodiments, the training process of the medical image segmentation network includes:

[0048] S1: Collect medical image sets, annotate and preprocess them, and save them in a medical image database.

[0049] Collaborate with hospitals or medical research institutions to obtain medical image datasets of patients.

[0050] In some preferred embodiments, the medical image set specifically includes X-ray images, CT images, MRI images and ultrasound images of lesions.

[0051] Specifically, medical image data includes: X-ray (such as chest X-ray), CT (computed tomography) (such as brain or lung CT), MRI (magnetic resonance imaging) (such as brain or spine MRI images), ultrasound (such as fetal ultrasound, cardiac ultrasound, etc.). Due to the diversity and complexity of medical images, the collected images need to be preprocessed to ensure the consistency of data quality.

[0052] In some preferred embodiments, the collected medical images are manually annotated, which specifically includes:

[0053] Set the pixel category ID of the medical image background to 0;

[0054] The medical image is segmented into segmented images of different categories, and the pixels of different categories are sequentially set to category IDs: 1, 2, ..., N, where N is the number of segmentation categories.

[0055] Specifically, various algorithms are used to remove noise from the image, such as artifacts, interference introduced by imaging devices, etc., to improve the clarity of the image. Due to different imaging devices and imaging parameters, the resolution and contrast of medical images may vary, so the images need to be standardized so that the images in the dataset have consistent resolution and grayscale range. Crop or resize the images as needed to adapt to the subsequent model input format. For example, all images are resized to 224×224 pixel resolution, and the pixel category ID of the background is set to 0 value; the pixels of different categories of segmentation are set to category IDs: 1, 2, ..., N in turn. Where N is the number of segmentation categories (excluding background).

[0056] S2: Filter images from the medical image database and obtain training and test sets after data augmentation.

[0057] In some preferred embodiments, the collected medical images are preprocessed, specifically:

[0058] Remove noise from acquired medical images;

[0059] The medical image after noise removal is standardized to adjust the resolution and grayscale range of the medical image.

[0060] Specifically, some categories with obvious characteristics are selected from the medical image dataset, similar, repeated and single-background categories are removed, and a series of data enhancement operations such as horizontal flipping, vertical flipping, 45-degree rotation, contrast enhancement, and noise addition are performed on the selected images to form the medical image dataset required for training. The medical image dataset is divided into a training set and a test set in a ratio of 9:1.

[0061] S3: Medical images and their segmented images are selected from the training set and input into the medical image segmentation network respectively, and the local features and global features of the medical images are extracted after 4 times of downsampling through CNN and Transformer respectively.

[0062] The medical image is specifically: a lesion medical image of a patient containing a lesion. The segmented image of the medical image is: a segmentation mask image of the lesion area marked by a doctor.

[0063] Specifically, the Root layer uses a standard convolution with a convolution kernel of 7, a stride of 2, and a padding of 3 for feature representation learning and downsampling. Each convolution generates a new feature map by sliding on the image (i.e., moving the stride), calculating the weighted sum of the local pixel area, and detecting local patterns of the image (such as edges, textures, color changes, etc.) to extract low-level features. As the network deepens, the convolution kernel will gradually extract higher-level features (such as shapes, object parts, etc.) from low-level features (such as edges). For multi-channel input images (such as RGB images) and the corresponding convolution kernels, the formula needs to sum the convolution results for each channel. Assume that the input image has C channels and the convolution kernel size of each channel is k. h ×k w , then the calculation formula for the pixel value F(i,j) at position (i,j) in the feature map is

[0064]

[0065] Where: I c is the cth channel of the input image. K c is the convolution kernel corresponding to the cth channel. c is the number of channels in the input image.

[0066] Each convolution kernel generates a corresponding feature map, and the feature maps of multiple convolution kernels are stacked into a three-dimensional tensor.

[0067] The three Body layers first use DWConv with a stride of 2 for feature representation learning and downsampling, and then use multiple DWConv for deeper feature representation learning. GN is used after all convolution operations, and Relu is performed at the end of each layer. Finally, the output features are linearly embedded and positionally encoded for subsequent feature fusion. Compared with standard convolution, DWConv significantly reduces the amount of calculation and parameters. Deep convolution can extract spatial information of each channel separately. In addition, point-by-point convolution achieves cross-channel information fusion through linear combinations between channels. This separation operation can extract features more efficiently. Compared with Batch Normalization (BN), GN needs to calculate the mean and variance through the entire mini-batch, which does not work well in small batches. GN's normalization operation is based on the channel grouping of each sample and does not depend on the batch size, so it is more stable in small batches or single sample reasoning.

[0068] Specifically, unlike the convolution kernel of traditional CNN, Swin Transformer first divides the input image into non-overlapping patches of fixed size and flattens each patch into a vector.

[0069] Input image size: H×W×C (height, width, number of channels).

[0070] The image is divided into non-overlapping blocks of size P×P, each block is of size P 2 ×C.

[0071] These blocks are flattened and mapped to a fixed dimension through a linear transformation to obtain the initial Patch Embedding.

[0072] The final output is an embedding vector in the form of a sequence with a size of:

[0073]

[0074] Where D is the dimension of the embedding.

[0075] In each layer, Swin Transformer uses a fixed-size window of 7×7 to perform local self-attention calculations on the image. The pixels in each window will obtain dependencies through the self-attention mechanism to achieve local feature extraction, without requiring global calculations like traditional Transformers. The feature extraction of Swin Transformer is performed layer by layer. Each layer consists of multiple Transformer blocks, and downsampling operations (similar to the pooling operation of CNN) are introduced in each stage to gradually reduce the resolution and increase the number of channels. From high resolution to low channel number, the resolution is halved after downsampling, the number of channels is increased, and further downsampling, the resolution is further halved, and finally the resolution becomes one sixteenth of the original global feature. Downsampling is achieved through the Patch Merging operation. The adjacent blocks are merged and the number of channels is adjusted through linear transformation. The advantage of Swin Transformer over traditional CNN and standard Vision Transformer (ViT) is that it uses local window self-attention to avoid the high computational cost of global self-attention. Secondly, the sliding window mechanism gradually captures long-distance dependencies.

[0076] S4: The extracted global features and local features are fused through the collaborative aggregation module CFA and the non-collaborative aggregation module NCFA.

[0077] Specifically, the Swin Transformer module, CNN module and the previous CFA module output 3 features as the current CFA input. The shapes of the 3 collaborative features of the 4 inputs are [B, 2*C, H / 2, W / 2], [B, 4*C, H / 4, W / 4], [B, 8*C, H / 8, W / 8], [B, 16*C, H / 16, W / 16]. The 3 collaborative features of the 4 inputs will be processed by LN and linearly embedded into [B, L, 2*C], [B, L, 4*C], [B, L, 8*C], [B, L, 16*C] of the same height and width. Each time, these 3 embedded features are channel-concatenated into 1 feature [B, L, 30*C]. Then, these 4 embedded features are aggregated through N layers of multi-head channel attention layers, and then the aggregated features are input into LN and MLP as the final collaborative features. The calculation of the channel attention mechanism is as follows:

[0078]

[0079] in, represents the query matrix, key matrix and value matrix. d is the sequence length and C i (i=1,2,3,4) four skip connection channel dimensions, specifically C1=64, C2=128, C3=256, C4=512.

[0080] The collaborative feature shapes of the 4 outputs are [B, 2*C, H / 2, W / 2], [B, 4*C, H / 4, W / 4], [B, 8*C, H / 8, W / 8], [B, 16*C, H / 16, W / 16]. Compared with the traditional skip connection feature fusion method, CFA can simultaneously utilize the global and local feature information extracted by Transformer and CNN, avoiding the low decomposition accuracy caused by using a single feature, thereby improving the segmentation accuracy.

[0081] Specifically, the Swin Transformer module and the CNN module output four downsampled features at different stages as NCFA inputs. The four non-collaborative feature input shapes are [B, 2*C, H / 2, W / 2], [B, 4*C, H / 4, W / 4], [B, 8*C, H / 8, W / 8], [B, 16*C, H / 16, W / 16]. The four input features will be linearly embedded into [B, L, 2*C], [B, L, 4*C], [B, L, 8*C], [B, L, 16*C] with unchanged height and width. These four embedded features are then channel-concatenated into one feature [B, L, 30*C]. After N layers of Transformer layers for feature aggregation, the aggregated features are reconstructed into the input four downsampled feature shapes of different scales and are connected to the input features with corresponding residuals as four output features. The calculation of the self-attention mechanism in Transformer is as follows:

[0082]

[0083] Among them, Attention(Q,K,V) represents the calculation of the multi-head self-attention mechanism for Q,K,V, Q,K,V represent the query matrix, key matrix and value matrix respectively. p represents the number of patches in the window, d represents the dimension of the query or key, and the values ​​in B are taken from the bias matrix

[0084] The output shapes of the four non-collaborative features are [B, 2*C, H / 2, W / 2], [B, 4*C, H / 4, W / 4], [B, 8*C, H / 8, W / 8], and [B, 16*C, H / 16, W / 16]. Compared with the traditional upsampling interpolation method, NCFA independently uses Transformer and CNN to extract hierarchical downsampling features. In the subsequent upsampling recovery process, its global and local features can be used to assist in restoring the image without relying entirely on interpolation upsampling, which can further improve the segmentation accuracy.

[0085] S5: Input the fused features into the decoder to obtain the predicted segmentation results of the medical image and its segmented image respectively.

[0086] The collaborative features output by the collaborative aggregation module CFA are reconstructed, and then together with the 4 sets of fusion features output by the non-collaborative aggregation module NCFA, they are processed by 4 upsampling decoding in the decoder, and the predicted segmentation results of the medical image and its segmented image are output respectively.

[0087] S6: Calculate the loss between the predicted segmentation results of the medical image and its segmented image, and optimize the medical image segmentation network by back propagating the loss, which is a mixed cross entropy and Dice loss.

[0088] A person skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium, which can include: ROM, RAM, disk or CD, etc.

[0089] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for automatic segmentation of medical images based on dual-aggregation hybrid neural network, characterized in that: include: Inputting the medical image to be segmented into a pre-trained medical image segmentation network and outputting the medical image segmentation result; The medical image segmentation network adopts a dual-aggregation hybrid U-type network, which includes two parallel encoders and a single decoder, and the two parallel encoders are convolutional neural networks CNN and Transformer respectively; The local features and global features of medical images are extracted simultaneously through convolutional neural networks (CNN) and Transformer. The extracted local features and global features are fused at multiple scales through the collaborative feature aggregation module CFA and the non-collaborative feature aggregation module NCFA; The multi-scale feature fusion features are input into the decoder and the medical image segmentation results are output.

2. The method for automatic segmentation of medical images based on dual-aggregation hybrid neural network according to claim 1, characterized in that: The training process of the medical image segmentation network includes: Collect medical image sets, annotate and pre-process them, and save them in a medical image database; Images are selected from the medical image database, and training and test sets are obtained after data enhancement; Medical images and their segmented images were selected from the training set and input into the medical image segmentation network, respectively, and the global and local features of the medical images were extracted after 4 downsamplings through CNN and Transformer, respectively; The extracted global features and local features are fused through the collaborative aggregation module CFA and the non-collaborative aggregation module NCFA; The fused features are input into the decoder to obtain the predicted segmentation results of the medical image and its segmented image respectively; The loss between the predicted segmentation results of the medical image and its segmented image is calculated, and the medical image segmentation network is optimized by back-propagation through the loss.

3. The method for automatic segmentation of medical images based on a dual-aggregation hybrid neural network according to claim 1 or 2, characterized in that: The convolutional neural network CNN includes a root layer and three body layers. The root layer includes a 7×7 convolution kernel conv, a group normalization layer GN and a Relu activation function layer; the body layer includes N bottleneck blocks bottleneck; each bottleneck block includes a 3-layer cascaded depth-separable convolution kernel DWConv, a group normalization layer GN and a Relu activation function.

4. The method for automatic segmentation of medical images based on dual-aggregation hybrid neural network according to claim 1 or 2, characterized in that: The collaborative feature aggregation module CFA includes N layers of cascaded layer normalization layers LN and multi-head channel attention MCA, and then the layer normalization layers LN and the multi-layer perceptron MLP are connected respectively.

5. The method for automatic segmentation of medical images based on dual-aggregation hybrid neural network according to claim 1 or 2, characterized in that: The non-collaborative feature aggregation module NCFA includes a concatenation layer contact, N cascaded layer normalization layers LN, multi-head self-attention MSA, layer normalization layer LN and multi-layer perceptron MLP, and finally a reconstruction recovery layer Reconstruct restores the original scale.

6. The method for automatic segmentation of medical images based on dual-aggregation hybrid neural network according to claim 2, characterized in that: The medical image set specifically includes X-ray images, CT images, MRI images and ultrasonic images of lesions.

7. The method for automatic segmentation of medical images based on dual-aggregation hybrid neural network according to claim 2, characterized in that: Manual annotation of the collected medical images is performed as follows: Set the pixel category ID of the medical image background to 0; The medical image is segmented into segmented images of different categories, and the pixels of different categories are sequentially set to category IDs: 1, 2, ..., N, where N is the number of segmentation categories.

8. The method for automatic segmentation of medical images based on dual-aggregation hybrid neural network according to claim 2, characterized in that: The collected medical images are preprocessed as follows: Remove noise from acquired medical images; The medical image after noise removal is standardized to adjust the resolution and grayscale range of the medical image.

9. The method for automatic segmentation of medical images based on dual-aggregation hybrid neural network according to claim 2, characterized in that: Images are selected from the medical image database and data enhancement is performed on them, which specifically includes: removing similar, repeated and single-background categories in the medical images, and performing horizontal flipping, vertical flipping, 45-degree rotation, contrast enhancement or noise addition on the medical images.

Images