Medical image segmentation method, system, medium and device based on collaborative extraction of heterogeneous features

Abstract

The present invention discloses a medical image segmentation method, system, medium and device based on the collaborative extraction of heterogeneous features, which belongs to the technical field of image segmentation. A medical image is acquired and preprocessed; the preprocessed medical image is input into a first encoder and a second encoder respectively, the first encoder extracts the edge features of the medical image based on the Sobel operator, and the second encoder extracts the global features of the medical image based on the channel attention mechanism. Based on the iterative attention feature fusion mechanism, the initial edge features and global features are fused, and the features are fused with the upsampled features of the corresponding layer of the decoder through jump connections. The fused features are decoded with the global features to obtain a segmentation map. The present invention can automatically learn the importance of features from different edge and global positions and scales, so as to better capture the complex information in the image and improve the accuracy of medical image segmentation.

Description

Technical Field

[0001] The present invention belongs to the technical field of image segmentation, and in particular relates to a medical image segmentation method, system, medium and equipment based on collaborative extraction of heterogeneous features. Background Art

[0002] Medical image segmentation technology divides medical images into several regions with distinct characteristics, maintaining certain similarities within regions and certain differences between them, thereby extracting the portion of interest. Medical image segmentation technology can accurately divide different parts of medical images into different regions, helping doctors accurately diagnose diseases, better plan surgeries, detect lesions, and perform quantitative analysis of tissues and organs. For example, by segmenting lesion regions, doctors can more easily determine their size, shape, and location, thereby formulating more effective treatment plans. Therefore, improving the accuracy of image segmentation has significant practical significance.

[0003] Medical image segmentation technology evolved from automatic image segmentation techniques. Traditional image segmentation methods include edge-based, threshold-based, region-based, and morphology-based methods. These methods each have their own advantages and provide strong support for image segmentation in various application scenarios. However, they struggle to achieve ideal performance when faced with complex medical images. The development of deep learning has brought new opportunities for medical image segmentation.

[0004] In recent years, deep learning-based image segmentation technology has made groundbreaking progress. Since Ronenberg et al. proposed the U-Net in 2015, this method has greatly advanced research in medical image segmentation. The U-Net is an encoder-decoder network structure that effectively preserves image information and achieves accurate segmentation of image structures. Since then, researchers have proposed many excellent segmentation models based on the U-Net structure. For example, U-Net++ introduces a nested structure that better integrates information between layers. Huang et al. proposed U-Net3+, which uses all-scale skip connections to effectively integrate features at different scales. DenseU-Net, which incorporates the concept of an attention mechanism, can effectively handle complex and variable medical image data. However, these single encoder structures focus more on local features when processing images and struggle to fully capture contextual information. The convolution operation of convolutional neural networks (CNNs) still faces challenges such as limited boundary perception and the loss of spatial information, often resulting in inaccurate object contour segmentation. Summary of the Invention

[0005] The purpose of the present invention is to provide a medical image segmentation method, system, medium and device based on collaborative extraction of heterogeneous features, aiming to solve the above-mentioned problems.

[0006] The present invention is mainly achieved through the following technical solutions:

[0007] A medical image segmentation method based on collaborative extraction of heterogeneous features includes the following steps:

[0008] Step S1: Obtain a medical image dataset and preprocess the dataset;

[0009] Step S2: segmenting the medical image using the trained image segmentation network model;

[0010] Step S21: inputting the preprocessed medical image into a first encoder, wherein the first encoder extracts edge features of the medical image based on a Sobel operator;

[0011] Step S22: inputting the preprocessed medical image into a second encoder, which extracts global features of the medical image based on a channel attention mechanism;

[0012] Step S23: Based on the iterative attention feature fusion mechanism, the initial edge features and global features are fused, and then fused with the upsampled features of the corresponding layer of the decoder through skip connections;

[0013] Step S24: Decode the features fused in step S23 with the global features to obtain a segmentation map.

[0014] In order to better implement the present invention, further, step S21 includes the following steps:

[0015] Step A1: First, input the preprocessed medical image, use the Sobel operator to calculate the horizontal and vertical gradients respectively, and process them to obtain the edge intensity map to obtain the initial edge features, and then perform nonlinear enhancement through the ReLU activation function;

[0016] Step A2: Then, the gradient-enhanced features are input into several convolution branches respectively to capture multi-scale features; the output of the convolution branches is spliced ​​with the initial edge features and the features are recombined through 1×1 convolution; the multi-scale features after the recombined features are fused;

[0017] Step A3: Finally, the fused multi-scale features are deeply optimized through the feature refinement layer to enhance the consistency of edge responses.

[0018] In order to better implement the present invention, further, step S22 includes the following steps:

[0019] Step B1: The preprocessed medical image is sequentially input into a 3x3 convolutional layer, a feature calibration unit, a 5x5 convolutional layer, a feature calibration unit, and a 7x7 convolutional layer to obtain a first feature map; the feature calibration unit processes features based on a channel attention mechanism;

[0020] Step B2: Input the preprocessed medical image into the two-dimensional convolutional layer to obtain the second feature map;

[0021] Step B3: Concatenate the first feature map and the second feature map to obtain global features.

[0022] The present invention is mainly achieved through the following technical solutions:

[0023] A medical image segmentation system based on collaborative extraction of heterogeneous features, used to implement the above-mentioned medical image segmentation method based on collaborative extraction of heterogeneous features, comprising an image acquisition module and an image segmentation network module. The image segmentation network module comprises a first encoder and a second encoder arranged in parallel, and a decoder. The first encoder and the second encoder respectively comprise a hierarchical gradient fusion module and a multi-receptive field feature refinement module. The decoder comprises a feature adaptive fusion module.

[0024] The image acquisition module is used to acquire medical images;

[0025] The hierarchical gradient fusion module and the multi-receptive field feature refinement module are used to extract edge features and global features of the image respectively;

[0026] The hierarchical gradient fusion module and the multi-receptive field feature refinement module are respectively connected to the feature adaptive fusion module; the feature adaptive fusion module is used to fuse edge features and global features based on an iterative attention feature fusion mechanism, and fuse them with upsampled features through jump connections to decode and obtain a segmentation map.

[0027] Specifically, the image segmentation network model for collaborative extraction of heterogeneous features proposed in the present invention follows the classic encoder-decoder architecture. The encoder consists of convolutional and pooling layers, designed to extract features from images. The decoder is composed of transposed convolutions and skip connections. The transposed convolutions increase the spatial resolution of the feature map to match the size of the original input image, gradually recovering detailed features. Skip connections connect corresponding layers between the encoder and decoder, helping the network better recover image details.

[0028] In order to better implement the present invention, further, the hierarchical gradient fusion module includes a bidirectional gradient perception unit and a first convolution branch and a second convolution branch arranged in parallel, the bidirectional gradient perception unit is connected to the first convolution branch and the second convolution branch respectively, and the output ends of the first convolution branch and the second convolution branch are connected to the 3×3 convolution layer through a feature splicing layer;

[0029] The bidirectional gradient perception unit includes a Sobel operator and a ReLU activation function arranged sequentially from front to back;

[0030] The first convolution branch includes a 3×3 convolution layer, a first feature splicing layer, and a 1×1 convolution layer, which are arranged in sequence from front to back, and the input end of the first feature splicing layer is connected to the 3×3 convolution layer and the bidirectional gradient perception unit respectively;

[0031] The second convolution branch includes a 5×5 convolution layer, a second feature splicing layer, and a 1×1 convolution layer, which are arranged in sequence from front to back, and the input end of the second feature splicing layer is connected to the 5×5 convolution layer and the bidirectional gradient perception unit respectively.

[0032] In order to better realize the present invention, further, the multi-receptive field feature refinement module includes a first feature branch and a second feature branch arranged in parallel, and the output ends of the first feature branch and the second feature branch are respectively connected to the feature splicing layer; the first feature branch includes a 3x3 convolution layer, a feature calibration unit, a 5x5 convolution layer, a feature calibration unit and a 7x7 convolution layer arranged in sequence from front to back; the second feature branch includes a two-dimensional convolution layer; the feature calibration unit includes several stacked channel attention modules and a two-dimensional convolution layer arranged in sequence from front to back.

[0033] In order to better implement the present invention, further, the feature adaptive fusion module includes several iterative attention feature fusion modules to fuse the fused features with the upsampled features through jump connections; the iterative attention feature fusion module is used to fuse edge features and global features based on the iterative attention feature fusion mechanism, and includes two multi-scale convolutional attention modules;

[0034] The edge features and global features are spliced ​​and input into one of the multi-scale convolutional attention modules, and the output is multiplied by the edge features and the global features respectively, and input into another multi-scale convolutional attention module after feature splicing; the output of the other multi-scale convolutional attention module is multiplied by the edge features and the global features respectively, and finally the fused features are output after feature splicing.

[0035] In order to better realize the present invention, further, the multi-scale convolutional attention module includes a first convolutional attention branch and a second convolutional attention branch, the output ends of the first convolutional attention branch and the second convolutional attention branch are respectively connected to the feature splicing layer, and are spliced ​​with the input feature through the Sigmoid activation function layer; the first convolutional attention branch includes a global average pooling layer, a point-to-point convolution layer, a ReLU activation function and a point-to-point convolution layer arranged in sequence from top to bottom; the second convolutional attention branch includes a point-to-point convolution layer, a ReLU activation function and a point-to-point convolution layer arranged in sequence from top to bottom.

[0036] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned medical image segmentation method based on collaborative extraction of heterogeneous features.

[0037] An electronic device comprises a memory and a processor; the memory stores a computer program; the processor is used to execute the computer program in the memory to implement the above-mentioned medical image segmentation method based on collaborative extraction of heterogeneous features.

[0038] The beneficial effects of the present invention are as follows:

[0039] This paper addresses the difficulty in fully capturing image context and blurred details during image processing by collaboratively extracting heterogeneous features. This paper proposes an image segmentation network that automatically learns the importance of features from different locations and scales, both edge and global, to better capture complex information in images and improve the accuracy of medical image segmentation.

[0040] This invention uses a hierarchical gradient fusion module (HGM) to extract edge features, improving the accuracy of medical image contour segmentation. Specifically, this invention uses a Sobel edge detection filter combined with convolutions of different scales to design a hierarchical gradient fusion module (HGM) to extract edge features, addressing the problem of insufficient contour segmentation accuracy in medical image segmentation. This invention also uses a multi-receptive field feature refinement module (MFM) to extract global features, preserving more detailed information and improving segmentation accuracy. Specifically, this invention uses a combination of channel attention and spatial attention mechanisms to design a multi-receptive field feature refinement module (MFM), preserving image details to extract global features, addressing the detail loss problem of feature extraction in conventional convolution. This invention also incorporates an adaptive feature fusion module (AFM) to enhance skip links, addressing the difficulty of linear fusion in fusing features from different feature streams. Specifically, this invention uses a multi-scale attention feature fusion mechanism to design an adaptive feature fusion module (AFM) to fuse edge features with global features, addressing the inability of linear fusion to effectively handle the complex relationships between features of different scales. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Flowchart of the medical image segmentation method based on collaborative extraction of heterogeneous features of the present invention;

[0042] Figure 2 Schematic diagram of the overall structure of the image segmentation network model of the present invention;

[0043] Figure 3 It is a structural diagram of the hierarchical gradient fusion module;

[0044] Figure 4 This is a schematic diagram of the structure of the multi-receptive field feature refinement module;

[0045] Figure 5 It is a structural diagram of the feature adaptive fusion module;

[0046] Figure 6 Schematic diagram of the structure of the iterative attention feature fusion module (IAFF);

[0047] Figure 7 Schematic diagram of the structure of the multi-scale convolutional attention module (MS-CAM);

[0048] Figure 8 This is a diagram showing the segmentation performance of the model of the present invention and other models in Example 1 on the PCL dataset. DETAILED DESCRIPTION

[0049] Example 1:

[0050] A medical image segmentation method based on the collaborative extraction of heterogeneous features uses an image segmentation network model with an encoding-decoding structure to segment medical images. Specifically, Figure 2 As shown, the present invention uses two parallel first and second encoders with different structures to extract image features. The first encoder includes a hierarchical gradient fusion module (HGM) composed of a Sobel edge operator and convolution kernels of different sizes; the second encoder includes a multi-receptive field feature refinement module (MFM) composed of a channel attention module (CA) and a residual mechanism, as well as convolution kernels of different sizes. The hierarchical gradient fusion module (HGM) and the multi-receptive field feature refinement module (MFM) are used as the convolutional layers for feature extraction to extract image features.

[0051] Secondly, the present invention uses a feature adaptive fusion module (AFM) to perform jump connection during upsampling to participate in the entire upsampling process. When performing jump connection, the present invention passes the features obtained from the two feature streams through the feature adaptive fusion module (AFM), and then connects them to the feature module of the corresponding layer of the decoder as a jump connection. After obtaining the fused features, they are connected with the transposed convolution features passed in from the bottleneck layer. The features obtained by the connection are the final features of the upsampling process, thereby gradually restoring the image detail features.

[0052] Preferably, if Figure 1 As shown, the specific steps include:

[0053] Step S1: Obtain a medical image dataset and preprocess the dataset;

[0054] Step S2: segmenting the medical image using the trained image segmentation network model;

[0055] Step S21: inputting the preprocessed medical image into a first encoder, wherein the first encoder extracts edge features of the medical image based on a Sobel operator;

[0056] The image after data preprocessing is input into the Hierarchical Gradient Fusion Module (HGM) to extract edge features of the image. Figure 3 As shown, the hierarchical gradient fusion module (HGM) includes a bidirectional gradient sensing unit and a first convolution branch and a second convolution branch arranged in parallel. The bidirectional gradient sensing unit is connected to the first convolution branch and the second convolution branch respectively, and the output ends of the first convolution branch and the second convolution branch are connected to the 3×3 convolution layer through the feature splicing layer. Specifically, it includes the following steps:

[0057] (1) First, the input image is processed by a bidirectional gradient perception unit: the Sobel operator is used to calculate the horizontal and vertical gradients respectively, and the edge intensity map is obtained by square and square root operations, and then nonlinear enhancement is performed by the ReLU activation function. This step retains the physical interpretability advantage of traditional edge detection methods.

[0058] (2) Subsequently, the multi-scale feature learning phase begins: the gradient-enhanced features are fed into the 3×3 and 5×5 convolution branches, respectively, to capture multi-scale features through convolution operations with different receptive fields. Before feature fusion, the present invention designs a gradient-preserving connection mechanism: the initial edge features are concatenated with the convolution-processed features, and then feature reorganization is achieved through 1×1 convolution. Combined with batch normalization and ReLU activation, this effectively avoids the attenuation of edge information in deep networks.

[0059] (3) Finally, the fused multi-scale features are deeply optimized through the feature refinement layer (3×3 convolution layer) to further enhance the consistency of edge responses. This hierarchical design enables HGM to maintain the accuracy advantage of traditional edge detection while leveraging the powerful capabilities of deep learning in feature representation, thereby demonstrating excellent generalization performance in various complex medical image segmentation tasks. Preferably, the specific operation of the Sobel operator is as follows:

[0060] Sobel_x is used to detect edges in the horizontal direction, and sobel_y is used to detect edges in the vertical direction.

[0061] sobel_x: [[-1,0,1],[-2,0,2],[-1,0,1]];

[0062] sobel_y: [[-1,-2,-1],[0,0,0],[1,2,1]].

[0063] Step S22: inputting the preprocessed medical image into a second encoder, which extracts global features of the medical image based on a channel attention mechanism;

[0064] The image after data preprocessing is input into the multi-receptive field feature refinement module (MFM), and then passes through four multi-receptive field feature refinement modules with the same structure in succession. The features obtained from each module are passed into the feature adaptive fusion module (AFM) through jump connections.

[0065] like Figure 4As shown, the multi-receptive field feature refinement module (MFM) includes a first feature branch and a second feature branch arranged in parallel, and the output ends of the first feature branch and the second feature branch are respectively connected to the feature splicing layer; the first feature branch includes a 3x3 convolutional layer, a feature calibration unit, a 5x5 convolutional layer, a feature calibration unit, and a 7x7 convolutional layer arranged in sequence from front to back; the second feature branch includes a two-dimensional convolutional layer. Specifically, three channel attention modules (CA) are stacked and residually connected to form a feature calibration unit (FCU), and the feature calibration unit (FCU) is then cross-connected and residually connected with the 3x3 convolutional layer, the 5x5 convolutional layer, and the 7x7 convolutional layer to form an adaptive multi-receptive field feature refinement module.

[0066] The channel attention module (CA) can dynamically recalibrate the feature map by calculating the weight of each channel, emphasizing important features and suppressing unimportant features. This helps the network focus more on information that is useful for the task, thereby improving the performance of the model. Specifically, Figure 4 As shown in the figure, the feature map is first fed into the max pooling layer and the average pooling layer, respectively, to obtain the maximum pooling value and the average pooling value; this is equivalent to having a global receptive field. The max pooling value and the average pooling value are then fed into a shared multi-layer perceptron (MLP) (linear layer, ReLU activation function, and linear layer) for learning, resulting in two 1×1×C feature maps. These two maps are then element-wise added to generate a rich feature channel. Finally, the two 1×1×C feature maps are element-wise added and then mapped using a sigmoid activation function to obtain a channel attention weight matrix. The channel attention weight matrix is ​​then multiplied by the input feature map to obtain a channel attention feature map, now of size (C, H, W). Similarly, using this module as the convolutional layer of the encoder during downsampling to extract image detail can reduce the loss of detail typically associated with feature extraction using conventional convolution.

[0067] Step S23: Based on the iterative attention feature fusion mechanism, the initial edge features and global features are fused, and then fused with the upsampled features of the corresponding layer of the decoder through skip connections;

[0068] Step S24: Decode the features fused in step S23 with the global features to obtain a segmentation map.

[0069] After obtaining the edge and global heterogeneous features, the feature adaptive fusion module (AFM) is used to fuse the features. After fusion, the features are combined with the upsampled features and transmitted forward to finally output the segmentation map.

[0070] like Figure 5As shown, the adaptive feature fusion module (AFM) includes several iterative attention feature fusion modules (IAFF) to fuse the fused features with the upsampled features via skip connections. The iterative attention feature fusion module (IAFF) is used to fuse edge features and global features based on the iterative attention feature fusion mechanism and includes two multi-scale convolutional attention modules (MS-CAM). Specifically, the present invention introduces an iterative attention feature fusion mechanism and sets a layer-by-layer progressive adaptive feature fusion module (AFM) on the skip connection branch. This module fuses different features from two heterogeneous feature streams via skip connections. After passing through this module, the obtained features are then fused with the upsampled features via skip connections to complement features whose details disappeared during downsampling and contour features that were not accurately extracted.

[0071] like Figure 6 As shown in , the Iterative Attention Feature Fusion Module (IAFF) includes two Multi-Scale Convolutional Attention Modules (MS-CAM), where x is the global feature, y is the edge feature, and z is the fused feature. Figure 7 As shown, the multi-scale convolutional attention module (MS-CAM) includes a first convolutional attention branch and a second convolutional attention branch. The output ends of the first convolutional attention branch and the second convolutional attention branch are respectively connected to the feature splicing layer and spliced ​​with the input features through the Sigmoid activation function layer; the first convolutional attention branch includes a global average pooling layer, a point-to-point convolution layer, a ReLU activation function, and a point-to-point convolution layer arranged in sequence from top to bottom; the second convolutional attention branch includes a point-to-point convolution layer, a ReLU activation function, and a point-to-point convolution layer arranged in sequence from top to bottom. The specific operations are:

[0072] The input features are divided into two branches. One branch performs global average pooling and then two point-to-point convolutions (Point-wise Conv) and a ReLU activation function. The other branch only performs two point-to-point convolutions (Point-wise Conv) and a ReLU activation function. The two are then added together and the Sigmoid activation function is used to obtain the weight, and finally the input features are added.

[0073] like Figure 8 As shown, the network segmentation result of the present invention is very close to the real segmentation, performs well in capturing the shape and boundary of the target area and has strong generalization ability while retaining more details of the image, which shows that the model of the present invention has good segmentation ability and good robustness.

[0074] As shown in Table 1, Acc(%) represents the proportion of pixels correctly predicted by the model, Dice(%) measures the similarity between the prediction and the true data, and Jaccard(%), also known as the intersection-over-union ratio, is the ratio of the intersection of the prediction and the true data to their union. Pre(%) represents the proportion of pixels predicted to be positive that are actually positive. Recall(%) represents the proportion of pixels correctly predicted to be positive among all pixels that are actually positive. Spec(%) represents the proportion of pixels correctly predicted to be negative among all pixels that are actually negative. On the digital retinal image (DRIVE) dataset for vascular extraction, the present invention shows that the important evaluation indicators, the Jaccard coefficient, increased by 4.051%, the Acc coefficient increased by 6.590%, and the Specificity coefficient was also the best. It can be seen that HCE-Net has good segmentation and generalization capabilities.

[0075] As shown in Table 2, the settings of the ablation experiment are as follows:

[0076] (1) MAS: The present invention retains the hierarchical gradient fusion module based on the encoding-decoding architecture and uses it for downsampling feature extraction, aiming to reveal the role of the hierarchical gradient fusion module proposed in the present invention.

[0077] (2) MA: This paper only retains the feature adaptive fusion module, aiming to verify whether the performance of the network model can be improved by introducing the multi-scale attention mechanism into the corresponding layer of the network's skip connection.

[0078] (3) MASE: retains the multi-receptive field feature refinement module proposed in this invention, aiming to verify that this module can retain more detailed features and improve segmentation accuracy when used for feature extraction.

[0079] As shown in Table 2, the analysis of the indicators indicates that each module significantly improves the network model's segmentation capabilities. This improvement is particularly pronounced in retinal vascular images, where edge segmentation is challenging and detailed feature extraction is difficult. MAS and MA achieve significant improvements on the DRIVE dataset, fully demonstrating the effectiveness and feasibility of introducing skip connections in the hierarchical gradient fusion module and feature adaptive fusion module proposed in this paper. Ablation experiments fully demonstrate that each module proposed in this paper plays a targeted role, improving the network's segmentation capabilities.

[0080] Table 1 Indicators of the proposed model and other different network models on the DRIVE dataset

[0081]

[0082] Table 2 Ablation test indicators of the proposed model on PCL and DRIVE datasets

[0083]

[0084] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification or equivalent change made to the above embodiment based on the technical essence of the present invention shall fall within the scope of protection of the present invention.

Claims

1. A medical image segmentation method based on collaborative extraction of heterogeneous features, characterized in that: The following steps are involved: Step S1: Obtain a medical image dataset and preprocess the dataset; Step S2: segmenting the medical image using the trained image segmentation network model; Step S21: inputting the preprocessed medical image into a first encoder, wherein the first encoder extracts edge features of the medical image based on a Sobel operator; Step S22: inputting the preprocessed medical image into a second encoder, which extracts global features of the medical image based on a channel attention mechanism; Step S23: Based on the iterative attention feature fusion mechanism, the initial edge features and global features are fused, and then fused with the upsampled features of the corresponding layer of the decoder through skip connections; Step S24: Decode the features fused in step S23 with the global features to obtain a segmentation map; The step S21 includes the following steps: Step A1: First, input the preprocessed medical image, use the Sobel operator to calculate the horizontal and vertical gradients respectively, and process them to obtain the edge intensity map to obtain the initial edge features, and then perform nonlinear enhancement through the ReLU activation function; Step A2: Then, the gradient-enhanced features are input into several convolution branches respectively to capture multi-scale features; the output of the convolution branches is spliced ​​with the initial edge features and the features are recombined through 1×1 convolution; the multi-scale features after the recombined features are fused; Step A3: Finally, the fused multi-scale features are deeply optimized through the feature refinement layer to enhance the consistency of edge responses; The step S22 includes the following steps: Step B1: The preprocessed medical image is sequentially input into a 3x3 convolutional layer, a feature calibration unit, a 5x5 convolutional layer, a feature calibration unit, and a 7x7 convolutional layer to obtain a first feature map; the feature calibration unit processes features based on a channel attention mechanism; Step B2: Input the preprocessed medical image into the two-dimensional convolutional layer to obtain the second feature map; Step B3: Concatenate the first feature map and the second feature map to obtain global features.

2. A medical image segmentation system based on collaborative extraction of heterogeneous features, used to implement the medical image segmentation method based on collaborative extraction of heterogeneous features according to claim 1, characterized in that: It includes an image acquisition module and an image segmentation network module. The image segmentation network module includes a first encoder and a second encoder and a decoder arranged in parallel. The first encoder and the second encoder respectively include a hierarchical gradient fusion module and a multi-receptive field feature refinement module. The decoder includes a feature adaptive fusion module. The image acquisition module is used to acquire medical images; The hierarchical gradient fusion module and the multi-receptive field feature refinement module are used to extract edge features and global features of the image respectively; The hierarchical gradient fusion module and the multi-receptive field feature refinement module are respectively connected to the feature adaptive fusion module; the feature adaptive fusion module is used to fuse edge features and global features based on an iterative attention feature fusion mechanism, and fuse them with upsampled features through jump connections to decode and obtain a segmentation map.

3. The medical image segmentation system based on collaborative extraction of heterogeneous features according to claim 2, characterized in that: The hierarchical gradient fusion module includes a bidirectional gradient sensing unit and a first convolution branch and a second convolution branch arranged in parallel. The bidirectional gradient sensing unit is connected to the first convolution branch and the second convolution branch respectively, and the output ends of the first convolution branch and the second convolution branch are connected to the 3×3 convolution layer through a feature splicing layer. The bidirectional gradient perception unit includes a Sobel operator and a ReLU activation function arranged sequentially from front to back; The first convolution branch includes a 3×3 convolution layer, a first feature splicing layer, and a 1×1 convolution layer, which are arranged in sequence from front to back, and the input end of the first feature splicing layer is connected to the 3×3 convolution layer and the bidirectional gradient perception unit respectively; The second convolution branch includes a 5×5 convolution layer, a second feature splicing layer, and a 1×1 convolution layer, which are arranged in sequence from front to back, and the input end of the second feature splicing layer is connected to the 5×5 convolution layer and the bidirectional gradient perception unit respectively.

4. The medical image segmentation system based on collaborative extraction of heterogeneous features according to claim 2, characterized in that: The multi-receptive field feature refinement module includes a first feature branch and a second feature branch arranged in parallel, and the output ends of the first feature branch and the second feature branch are respectively connected to the feature splicing layer; the first feature branch includes a 3x3 convolution layer, a feature calibration unit, a 5x5 convolution layer, a feature calibration unit and a 7x7 convolution layer arranged in sequence from front to back; the second feature branch includes a two-dimensional convolution layer; the feature calibration unit includes several stacked channel attention modules and a two-dimensional convolution layer arranged in sequence from front to back.

5. The medical image segmentation system based on collaborative extraction of heterogeneous features according to claim 2, characterized in that: The feature adaptive fusion module includes several iterative attention feature fusion modules to fuse the fused features with the upsampled features through skip connections; the iterative attention feature fusion module is used to fuse edge features and global features based on the iterative attention feature fusion mechanism, and includes two multi-scale convolutional attention modules; The edge features and global features are spliced ​​and input into one of the multi-scale convolutional attention modules, and the output is multiplied by the edge features and the global features respectively, and input into another multi-scale convolutional attention module after feature splicing; the output of the other multi-scale convolutional attention module is multiplied by the edge features and the global features respectively, and finally the fused features are output after feature splicing.

6. The medical image segmentation system based on collaborative extraction of heterogeneous features according to claim 5, characterized in that: The multi-scale convolutional attention module includes a first convolutional attention branch and a second convolutional attention branch, wherein the output ends of the first convolutional attention branch and the second convolutional attention branch are respectively connected to the feature splicing layer and spliced ​​with the input feature through the Sigmoid activation function layer; The first convolutional attention branch includes a global average pooling layer, a point-to-point convolution layer, a ReLU activation function and a point-to-point convolution layer arranged in sequence from top to bottom; the second convolutional attention branch includes a point-to-point convolution layer, a ReLU activation function and a point-to-point convolution layer arranged in sequence from top to bottom.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the medical image segmentation method based on collaborative extraction of heterogeneous features described in claim 1 is implemented.

8. An electronic device, characterized in that: It comprises a memory and a processor; the memory stores a computer program; the processor is used to execute the computer program in the memory to implement a medical image segmentation method based on collaborative extraction of heterogeneous features as described in any one of claim 1.

Images