A medical image segmentation method integrating multi-scale features and attention mechanism

By introducing multi-scale feature extraction and attention mechanisms into the medical image segmentation model, the detailed information loss problem caused by downsampling is solved, the segmentation accuracy and the model's attention to detailed information are improved, and it is suitable for a variety of medical image segmentation tasks.

CN114119638BActive Publication Date: 2024-11-01UNIV OF SHANGHAI FOR SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111458212.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-02
Publication Date
2024-11-01
Estimated Expiration
2041-12-02

AI Technical Summary

Technical Problem

The existing medical image segmentation model will lose detailed information during the downsampling process, and the detailed information of the low-level feature map is not fully utilized, resulting in difficulty in target boundary segmentation and small-target capture.

Method used

Using a medical image segmentation method that integrates multi-scale feature extraction and attention mechanisms, the encoder-decoder structure is constructed through the U-Net backbone network, combined with the multi-scale feature extraction module and attention module, and fused image information using the multi-scale pooling layer , and guide advanced feature extraction through attention weights, pay attention to spatial locations with rich details.

Benefits of technology

It improves the accuracy of medical image segmentation, enhances the model's modeling ability of long-distance dependencies, makes up for the lack of detailed information of advanced feature maps, and is suitable for a variety of medical image segmentation tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114119638B_ABST
    Figure CN114119638B_ABST
Patent Text Reader

Abstract

The present invention provides a medical image segmentation method integrating multi-scale features and attention mechanism. Firstly, a medical image segmentation model including an encoder-decoder, a multi-scale feature extraction module and an attention module is constructed based on U-Net, and the model takes the image to be segmented as input; then, a high-level feature map obtained based on the encoder and a low-level feature map integrated with multi-scale information based on the multi-scale feature extraction module are input into the attention module to obtain a feature map integrating the relationship between local information and global information; finally, the feature map is input into the decoder for upsampling processing to restore the image resolution, and a corresponding segmentation result is obtained; the medical image segmentation method utilizes the multi-scale feature extraction module to integrate the multi-scale information of the image, and at the same time adopts the attention mechanism to make the medical image segmentation model pay attention to some of the most relevant information, thereby improving the modeling ability of the model for long-distance dependencies, thereby improving the segmentation accuracy of medical images.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of computer vision, and specifically relates to a method for performing medical image segmentation based on deep learning in scene analysis. Background Art

[0002] Medical image segmentation is a very important task in computer-aided diagnosis and treatment planning. The task is to achieve pixel-level segmentation of the region of interest on the medical image. Early medical image segmentation systems were mainly based on improvements to traditional image segmentation algorithms, such as edge detection-based segmentation algorithms, threshold-based segmentation algorithms, and region-based segmentation algorithms. However, these methods rely heavily on the professional knowledge of doctors, have poor generalization capabilities, and cannot be migrated to new task scenarios. With the development of deep learning technology, it has gradually become the mainstream method for medical image segmentation with its powerful modeling capabilities. Compared with traditional algorithms, deep learning algorithms based on convolutional neural networks have made breakthroughs in many visual recognition tasks, such as image classification, semantic segmentation, etc., and have therefore been introduced into medical image segmentation.

[0003] With the development of deep learning technology, natural image segmentation algorithms have gradually been applied to the medical field. The most representative research work is the fully convolutional neural network. In the fully convolutional neural network, the convolutional layer is used to replace the fully connected layer, which improves the network's ability to extract image features. In addition, in order to improve the accuracy of segmentation, jump connections are introduced to fuse features at different levels. Without changing the image size, the image is segmented at the pixel level, which is a significant improvement compared to traditional methods. With the success of FCN, researchers began to focus on how to improve the segmentation model based on the characteristics of medical images. Considering that medical images have rich spatial information and the network downsampling process is prone to lose spatial information, the encoder-decoder based network structure began to emerge.

[0004] In the encoder-decoder network, the encoder is composed of a series of convolutional layers. High-level features extract semantic information of the image, and low-level features extract detail information of the image. The decoder then upsamples the feature map, restores the image resolution, and outputs the segmented result. Although this type of model has strong feature extraction capabilities, there are still some shortcomings: 1) Due to the existence of downsampling operations, detailed information will inevitably be lost, which will be a great challenge for segmenting at the boundary of the target or capturing small targets. Although the image resolution will be restored by upsampling in the decoder part, the valuable information lost during the downsampling process cannot be restored; 2) Low-level feature maps often have more detailed information, but they are not fully utilized. Most models operate on high-level feature maps with more semantic information, and do not pay enough attention to low-level feature maps. Therefore, a more accurate segmentation method is needed to solve the above problems. Summary of the invention

[0005] In order to solve the above problems, a medical image intelligent segmentation method integrating multi-scale feature extraction and attention mechanism is provided. The present invention adopts the following technical solutions:

[0006] The present invention provides a medical image segmentation method integrating multi-scale features and an attention mechanism, characterized in that the method comprises the following steps: step S1, using U-Net as a backbone network, constructing a medical image segmentation model based on an encoder-decoder structure, wherein the medical image segmentation model comprises an encoder, a decoder, a multi-scale feature extraction module and an attention module; step S2, inputting an image to be segmented into the medical image segmentation model, encoding the image to be segmented based on the encoder to obtain a feature map E; step S3, extracting features from the image to be segmented based on the multi-scale feature extraction module to obtain a low-level feature map M of information of multiple scales; step S4, inputting the low-level feature map M and the feature map E as low-level features and high-level features, respectively, into the attention module to obtain a feature map A integrating the relationship between local information and global information; step S5, inputting the feature map A into the decoder for upsampling processing, and obtaining a segmentation result corresponding to the image to be segmented.

[0007] The medical image segmentation method integrating multi-scale features and attention mechanism provided by the present invention may also have such a technical feature, wherein the encoder-decoder structure includes 4 encoding layers and 4 decoding layers, the encoding layer is composed of a convolution layer, a pooling layer and an activation function, and the process of encoding the image to be segmented by the encoding layer is expressed as:

[0008] Ei=Sigmoid(Pooling(Conv(Input)))

[0009] In the formula, Conv, Pooling, and Sigmoid represent the processing of convolution operation, pooling operation, and activation function respectively, Input represents the input of the encoding layer, and Ei represents the output feature map of each encoding layer, where i = 1, 2, 3, 4.

[0010] The medical image segmentation method integrating multi-scale features and attention mechanism provided by the present invention may also have such a technical feature, wherein step S3 includes the following sub-steps: step S3-1, inputting the image to be segmented into the multi-scale feature extraction module for downsampling to obtain four feature maps of different sizes; step S3-2, reducing the feature dimensions of the four feature maps of different sizes through a convolutional layer and upsampling, and obtaining a feature map P restored to the size of the image to be segmented:

[0011] P=Upsample(Conv(Pooling(Input)))

[0012] In the formula, Upsample represents the upsampling operation; step S3-3, the feature map P is fused and spliced ​​with the image to be segmented to obtain the feature map C:

[0013] C=Concat(P,Input)

[0014] In the formula, Concat represents the concatenation operation; in step S3-4, the feature map C is convolved once to obtain the low-level feature map M:

[0015] M = Conv(C).

[0016] The medical image segmentation method integrating multi-scale features and attention mechanism provided by the present invention may also have such a technical feature, wherein the low-level feature is recorded as Xl, the high-level feature is recorded as Xh, and step S4 includes the following sub-steps: step S4-1, performing a convolution on the high-level feature Xh to obtain feature Xh′:

[0017] Xh′=Conv(Xh);

[0018] Step S4-2: perform matrix multiplication on the low-level feature Xl and feature Xh′ and perform normalization to obtain the attention weight map X am :

[0019] X am =softmax(Xl(Xh′)T)

[0020] Where softmax represents the normalization operation; step S4-3, based on the attention weight map X am Guide feature extraction for feature Xh′ to obtain feature map Xh″:

[0021] Xh″=Xam Xh′;

[0022] Step S4-4, concatenate the feature map Xh″ and the feature map Xh′ and perform a convolution to obtain the feature map A:

[0023] A=Conv(Concat(Xh″,Xh′)).

[0024] The medical image segmentation method integrating multi-scale features and attention mechanism provided by the present invention may also have such a technical feature, wherein step S5 is: the input of each decoding layer is composed of the output of the previous decoding layer and the output of the corresponding encoding layer through jump connection, wherein each decoding layer includes a Dropout operation, a convolution operation and an upsampling operation on the input:

[0025] Di=Upsample(Conv(Dropout(Concat(Input,Ei))))

[0026] In the formula, Di represents the output feature map of each decoding layer, where i = 1, 2, 3, 4.

[0027] Function and Effect of the Invention

[0028] According to a medical image segmentation method integrating multi-scale features and attention mechanism of the present invention, a medical image segmentation model with a multi-scale feature extraction module and an attention module is constructed based on an encoder-decoder, and the model is used to intelligently segment a variety of medical images. Among them, due to the use of a multi-scale feature extraction module, multi-scale information of the image to be segmented from the underlying features is integrated by using a pooling layer of multiple scales, and the multi-scale features with more details are output as the input of the low-level features of the attention module, so as to guide the feature extraction of the high-level feature map, so that the medical image segmentation model can pay more attention to the spatial position with rich detail information, effectively making up for the deficiency that the high-level feature map has relatively rich semantic information but lacks detail information.

[0029] The medical image segmentation method of the present invention that integrates multi-scale features and attention mechanism can utilize pooling layers of multiple scales to fuse the multi-scale information of the image, and at the same time adopts the attention mechanism to allow the model to focus on some of the most relevant information, thereby improving the model's ability to model long-distance dependencies, thereby improving the segmentation accuracy of medical images. The method has good versatility and is suitable for the segmentation of various medical images. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a flow chart of a medical image segmentation method integrating multi-scale features and attention mechanism in an embodiment of the present invention;

[0031] Figure 2 is a schematic diagram of the structure of a medical image segmentation model in an embodiment of the present invention;

[0032] Figure 3 is a schematic diagram of the structure of a multi-scale feature extraction module in an embodiment of the present invention;

[0033] Figure 4 is a schematic diagram of the structure of the attention module in an embodiment of the present invention;

[0034] Figure 5 is a schematic diagram comparing the segmentation effect of the embodiment of the present invention on the DRIVE retinal blood vessel dataset with other mainstream segmentation algorithms;

[0035] Figure 6 is a schematic diagram of experimental comparison results of an embodiment of the present invention on a DRIVE retinal vessel dataset with other mainstream segmentation algorithms;

[0036] Figure 7 It is a schematic diagram of experimental comparison results of an embodiment of the present invention on the CHASE_DB1 retinal vascular dataset with other mainstream segmentation algorithms. DETAILED DESCRIPTION

[0037] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the following is a detailed description of a medical image segmentation method that integrates multi-scale features and attention mechanism of the present invention in combination with embodiments and drawings.

[0038] <Example>

[0039] Figure 1 is a flowchart of a medical image segmentation method integrating multi-scale features and attention mechanism in an embodiment of the present invention. Figure 2 It is a schematic diagram of the structure of the medical image segmentation model in an embodiment of the present invention.

[0040] like Figure 1 and Figure 2 As shown in FIG, the medical image segmentation method integrating multi-scale features and attention mechanism includes the following steps:

[0041] Step S1, using U-Net as the backbone network, constructing a medical image segmentation model based on an encoder-decoder structure, the medical image segmentation model having an encoder, a decoder, a multi-scale feature extraction module, and an attention module.

[0042] Step S2: input the image to be segmented into the medical image segmentation model, and encode the image to be segmented based on the encoder to obtain a feature map E.

[0043] In this embodiment, the encoder is used to extract image features, and the decoder is used to restore image resolution to achieve pixel-level segmentation. The encoder-decoder structure includes 4 encoding layers and 4 decoding layers, each of which is composed of a convolutional layer, a pooling layer, and an activation function. The process of encoding the image to be segmented in each encoding layer is expressed as:

[0044] Ei=Sigmoid(Pooling(Conv(Input)))

[0045] In the formula, Conv, Pooling, and Sigmoid represent the processing of convolution operation, pooling operation, and activation function respectively, Input represents the input of the encoding layer, and Ei represents the output feature map of each encoding layer, where i = 1, 2, 3, 4.

[0046] Among them, the sizes of the output feature maps E1, E2, E3 and E4 of each encoding layer are 1 / 2, 1 / 4, 1 / 8 and 1 / 16 of the size of the image to be segmented, respectively.

[0047] Step S3: extract features from the image to be segmented based on a multi-scale feature extraction module to obtain a low-level feature map M of information of multiple scales.

[0048] Figure 3 Schematic diagram of the structure of a multi-scale feature extraction module in an embodiment of the present invention.

[0049] Since the downsampling operation of the pooling layer on the image to be segmented will cause the feature map to lose some detail information, in order to obtain richer spatial information, in this embodiment, the original image to be segmented is input into the multi-scale feature extraction module, and the multi-scale feature information (such as Figure 3 shown). Specifically:

[0050] Step S3-1: Input the image to be segmented into a multi-scale feature extraction module for downsampling to obtain four feature maps of different sizes.

[0051] Since four pooling layers with different pooling kernels are used in parallel to downsample the segmented image, four feature maps with sizes of 1×1, 2×2, 3×3, and 6×6 are obtained.

[0052] Step S3-2, reduce the feature dimensions of the four feature maps of different sizes to 1 / 4 of the original through a 1×1 convolution layer, and perform upsampling on the pyramid pooling structure to obtain a feature map P restored to the size of the image to be segmented:

[0053] P=Upsample(Conv(Pooling(Input)))

[0054] Where Upsample represents the upsampling operation.

[0055] In the semantic segmentation task, if the segmentation target is a single object and the size of the object does not vary much, more feature information can be extracted using traditional convolution, and the encoder model using traditional convolution can also achieve good segmentation results. However, when the target to be segmented on a picture is multiple types of objects and the sizes of the objects are also different, it will be difficult to achieve accurate segmentation of various objects using the traditional convolution encoder model, because the convolution kernel size, step size, and padding value of the traditional convolution are fixed during the training process, so it can only effectively extract feature information of objects of one size, but it cannot perform as well for objects of other sizes, and there may be many mis-segmentation and misclassification results.

[0056] In this embodiment, in order to better obtain feature information of multiple scales, a pyramid pooling structure is used to form the main part of the multi-scale feature extraction module, and the detail features and global features of the target are integrated to obtain feature information of multiple scales.

[0057] Step S3-3, fusing the feature map P with the image to be segmented to obtain the feature map C:

[0058] C=Concat(P,Input)

[0059] In the formula, Concat represents the concatenation operation.

[0060] Since the image to be segmented can be used as a global feature and has rich feature information, in this embodiment, splicing the feature map P with the image to be segmented is to fuse the detailed features and global features of the target.

[0061] Step S3-4, perform a convolution on the feature map C to ensure that the number of channels of the final output result is the same as that of the image to be segmented, and obtain a low-level feature map M:

[0062] M = Conv(C).

[0063] In step S4, the low-level feature map M and the feature map E are respectively input into the attention module as low-level features and high-level features to obtain a feature map A that integrates the relationship between local information and global information.

[0064] The low-level feature map M is obtained by directly parallel pooling the original image to be segmented by the multi-scale feature extraction module. It contains features of multiple scales and is rich in detail information. The feature map E comes from the last layer output of the encoder. Since it has undergone multiple levels of encoding, the feature map at this time carries rich semantic information.

[0065] Figure 4Schematic diagram of the structure of the attention module in an embodiment of the present invention.

[0066] In this embodiment, low-level features and high-level features are integrated to perform local and global information relationships. Figure 4 As shown, the attention module contains two inputs, namely the low-level features Xl of the low-level feature map M and the high-level features Xh of the feature map E obtained by the encoder. Specifically:

[0067] Step S4-1, convolve the high-level feature Xh to obtain feature Xh′:

[0068] Xh′=Conv(Xh);

[0069] Step S4-2: perform matrix multiplication on the low-level feature Xl and feature Xh′ and perform normalization to obtain the attention weight map X am :

[0070] X am =softmax(Xl(Xh′)T)

[0071] In the formula, softmax represents the normalization operation;

[0072] Step S4-3: In order to pay more attention to the locations with rich details on the high-level feature map E, the attention weight map X is used. am Guide feature extraction for feature Xh′ to obtain feature map Xh″:

[0073] Xh″=X am Xh′;

[0074] Step S4-4, concatenate the feature map Xh″ and the feature Xh′ to further enhance the features and perform a convolution to obtain the feature map A:

[0075] A=Conv(Concat(Xh″,Xh′)).

[0076] In deep neural networks, capturing long-distance dependencies is crucial. In sequence data processing, loop operations are usually used to achieve this, and in image data processing, stacking of convolution operations is usually used to increase the receptive field. The attention mechanism reweights features by calculating attention weights to achieve the purpose of strengthening effective features and suppressing invalid features. Unlike multi-scale feature fusion, which is only fused into a static representation, the attention mechanism allows the model to focus on some of the most relevant information, effectively integrating the relationship between local information and global information, and improving the model's ability to model long-distance dependencies.

[0077] Therefore, in this embodiment, a module is designed to use low-level features with richer detail information to guide high-level coding layer feature extraction (i.e., step S4-3), so that the medical image segmentation model can pay more attention to spatial locations rich in detail information, effectively making up for the deficiency of high-level feature maps that have relatively rich semantic information but lack detail information.

[0078] Step S5, input the feature map A to the decoder for upsampling to restore the image resolution and obtain a segmentation result corresponding to the image to be segmented.

[0079] In this embodiment, since the encoder downsamples the original image to be segmented 4 times, the size of the obtained feature map is 1 / 16 of the original image. The size of the output feature map A after the attention module does not change and is still 1 / 16 of the original image. Therefore, the decoder is used to perform 4 upsampling to restore the feature map size to the same size as the original image. Specifically:

[0080] The outputs of the four coding layers are denoted as D1, D2, D3 and D4 respectively, and the input of each decoding layer is formed by jump connection between the output of the previous decoding layer and the output of the corresponding coding layer, thereby improving the information loss in the upsampling process.

[0081] Each decoding layer includes a Dropout operation, a convolution operation, and an upsampling operation on the input:

[0082] Di=Upsample(Conv(Dropout(Concat(Input,Ei))))

[0083] In the formula, Di represents the output feature map of each decoding layer, where i = 1, 2, 3, 4.

[0084] After 4 decoding layers, the final output is the segmentation result of the image to be segmented.

[0085] Figure 5 It is a schematic diagram comparing the segmentation effect of the embodiment of the present invention on the DRIVE retinal blood vessel dataset with other mainstream segmentation algorithms.

[0086] In this embodiment, a comparative experiment was conducted on the DRIVE retinal vascular dataset using the medical image segmentation method integrating multi-scale features and the attention mechanism of this embodiment and other existing segmentation methods.

[0087] like Figure 5As shown, from left to right, each column is the original image, the gold standard image, the U-Net segmentation result, the LCP-Net segmentation result and the segmentation result of the present invention. In the first three rows, the rectangular box is the segmentation effect of the details. It can be seen that compared with the segmentation results of the U-Net segmentation method and the LCP-Net segmentation method, the segmentation of the details of the present invention is more obvious, which shows that the proposed attention module makes full use of the detail information rich in low-level features, so that a more satisfactory effect can be achieved in the details. Among them, in the fourth row, the rectangular box marks the situation of mis-segmentation. Compared with the gold standard image, it can be clearly seen that the U-Net segmentation method and the LCP-Net segmentation method are wrong segmentations, while the medical image segmentation model of this embodiment performs very well and correctly segments the image.

[0088] Figure 6 is a schematic diagram of experimental comparison results of an embodiment of the present invention on the DRIVE retinal blood vessel dataset with other mainstream segmentation algorithms, Figure 7 It is a schematic diagram of experimental comparison results of an embodiment of the present invention on the CHASE_DB1 retinal vascular dataset with other mainstream segmentation algorithms.

[0089] In this embodiment, four indicators, namely, accuracy (Acc), sensitivity (Sen), specificity (Spe) and area under the curve (AUC), are used to evaluate different segmentation methods. Specifically:

[0090]

[0091]

[0092]

[0093] Where TP, TN, FP and FN represent the number of true positives, true negatives, false positives and false negatives, respectively.

[0094] The AUC metric is used to evaluate the classifier's ability to distinguish between positive and negative samples. The closer the AUC is to 1, the more authentic the detection method is.

[0095] AUC is the area enclosed by the ROC curve and the coordinate axis. The ROC curve is drawn with Sen as the horizontal axis and Spe as the vertical axis. The closer the ROC curve is to the upper left corner, the better the performance of the model.

[0096] like Figure 6 and Figure 7 As shown, the medical image segmentation method integrating multi-scale features and attention mechanism of the present invention performs best in all three indicators, which shows that the present invention is effective in the task of medical image segmentation.

[0097] Example Function and Effect

[0098] According to a medical image segmentation method that integrates multi-scale features and an attention mechanism provided in this embodiment, a medical image segmentation model with a multi-scale feature extraction module and an attention module is constructed based on an encoder-decoder, and the model is used to intelligently segment a variety of medical images. Among them, due to the use of a multi-scale feature extraction module, multi-scale information of the image to be segmented from the underlying features is integrated using pooling layers of multiple scales, and the multi-scale features that output more details are aggregated as the input of the low-level features of the attention module, which is used to guide the feature extraction of the high-level feature map, so that the medical image segmentation model can pay more attention to the spatial position with rich detail information, effectively making up for the deficiency that the high-level feature map has relatively rich semantic information but lacks detail information.

[0099] A medical image segmentation method that integrates multi-scale features and an attention mechanism in this embodiment can utilize pooling layers of multiple scales to fuse multi-scale information of an image, and at the same time use an attention mechanism to allow the model to focus on some of the most relevant information, thereby improving the model's ability to model long-distance dependencies, thereby improving the segmentation accuracy of medical images. This method has good versatility and is suitable for the segmentation of various medical images.

[0100] In the embodiment, a pyramid redization structure is used to form the main part of the multi-scale feature extraction module, and the detail features and global features of the target are integrated, thereby improving the accuracy of model segmentation and avoiding the problems of mis-segmentation and misclassification.

[0101] The above embodiments are only used to illustrate specific implementation modes of the present invention, and the present invention is not limited to the description scope of the above embodiments.

Claims

1. A medical image segmentation method integrating multi-scale features and attention mechanism, characterized in that: The following steps are involved: Step S1, using U-Net as a backbone network, constructing a medical image segmentation model based on an encoder-decoder structure, wherein the medical image segmentation model has an encoder, a decoder, a multi-scale feature extraction module, and an attention module; Step S2, inputting the image to be segmented into the medical image segmentation model, and encoding the image to be segmented based on the encoder to obtain a feature map E; Step S3, extracting features from the image to be segmented based on the multi-scale feature extraction module to obtain a low-level feature map M of information of multiple scales; Step S4, inputting the low-level feature map M and the feature map E as low-level features and high-level features into the attention module respectively to obtain a feature map A that integrates the relationship between local information and global information; Step S5, inputting the feature map A into the decoder for upsampling processing to obtain a segmentation result corresponding to the image to be segmented, Wherein, the step S3 includes the following sub-steps: Step S3-1, inputting the image to be segmented into the multi-scale feature extraction module for downsampling to obtain four feature maps of different sizes; Step S3-2, reducing the feature dimensions of the four feature maps of different sizes through a convolutional layer and performing upsampling to obtain a feature map P restored to the size of the image to be segmented: P=Upsample(Conv(Pooling(Input))) In the formula, Upsample represents the upsampling operation; Step S3-3, fusing and splicing the feature map P with the image to be segmented to obtain a feature map C: C=Concat(P,Input) In the formula, Concat represents the concatenation operation; Step S3-4, convolve the feature map C once to obtain the low-level feature map M: M=Conv(C), The low-level feature is recorded as Xl, and the high-level feature is recorded as Xh. The step S4 comprises the following sub-steps: Step S4-1, convolve the high-level feature Xh to obtain feature Xh′: Xh′=Conv(Xh); Step S4-2: perform matrix multiplication on the low-level feature Xl and the feature Xh′ and perform normalization operation to obtain the attention weight map X am : X am =softmax(Xl(Xh′) T ) In the formula, softmax represents the normalization operation; Step S4-3, based on the attention weight map X am The feature Xh′ is guided to extract the feature, and the feature graph Xh″ is obtained: Xh″=X am Xh′; Step S4-4, concatenate the feature map Xh″ and the feature map Xh′ and perform a convolution to obtain the feature map A: A=Conv(Concat(Xh″,Xh′)).

2. The medical image segmentation method integrating multi-scale features and attention mechanism according to claim 1, characterized in that: in, The encoder-decoder structure includes 4 encoding layers and 4 decoding layers. The encoding layer consists of a convolutional layer, a pooling layer, and an activation function. The process of encoding the image to be segmented by the encoding layer is expressed as: Ei=Sigmoid(Pooling(Conv(Input))) In the formula, Conv, Pooling, and Sigmoid represent the processing of convolution operation, pooling operation, and activation function respectively, Input represents the input of the encoding layer of this layer, and Ei represents the output feature map of the encoding layer of each layer, where i = 1, 2, 3, 4.

3. The medical image segmentation method integrating multi-scale features and attention mechanism according to claim 2, characterized in that: in, The step S5 is: The input of each decoding layer is composed of the output of the previous decoding layer and the output of the corresponding encoding layer through jump connection, wherein each decoding layer includes Dropout operation, convolution operation and upsampling operation on the input: Di=Upsample(Conv(Dropout(Concat(Input,Ei)))) In the formula, Di represents the output feature map of each decoding layer, where i = 1, 2, 3, 4.