Medical image segmentation method, system and device based on efficient focal attention

Abstract

The application discloses a medical image segmentation method, system and device based on efficient focal attention, and the method comprises the following steps: first, the image is subjected to a pretreatment stage, the image size is adjusted, and data augmentation is performed through image enhancement methods such as brightness adjustment, contrast adjustment and color adjustment; the pretreated image enters an embedding layer to generate an embedding vector, and then is subjected to feature extraction through an efficient focal attention module; three cycles are performed in this way, and the dimension of the feature vector is reduced after each cycle; the network structures at different levels are connected through a skip connection to integrate high-level semantic features into low-level features; the feature vector enters a decoder to gradually increase the dimension, and finally passes through a linear classifier to restore into a mask image; and the application can realize medical image segmentation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical image segmentation, in particular to a medical image segmentation method, system and device based on efficient focus attention. BACKGROUND

[0002] In recent years, with the continuous development of hardware technologies such as chip semiconductors and the continuous breakthroughs in the field of artificial intelligence, various artificial intelligence algorithms have been continuously innovated and developed. As an important branch of artificial intelligence, deep learning mainly abstracts and layers data and learns representation through deep neural networks, which has continuously promoted the rapid development of artificial intelligence. Artificial intelligence has been widely applied in disciplines including medical image segmentation.

[0003] Medical images such as CT, MRI, positron emission tomography (PET), ultrasound, and X-ray are important components of computer-aided medical systems (CAD) and provide important support for medical workers in the processes of disease diagnosis, treatment plan decision-making, and treatment effect evaluation. Medical image segmentation is a key step in medical image data interpretation, but traditional medical image segmentation is mainly led by radiologists and requires manual segmentation, which is affected by many factors such as human physical limits, experience levels, and medical image noise. The cost of manpower and resources is relatively high, and the accuracy of manual segmentation is often limited, resulting in that the misdiagnosis rate and missed diagnosis rate have not been significantly improved. How to accurately and efficiently segment medical images at a low cost is a difficult problem that needs to be solved.

[0004] Semi-automatic medical image segmentation based on threshold, boundary, and region assists clinicians in segmentation to some extent, but it is relatively sensitive to noise, has relatively low accuracy, and cannot distinguish the semantics of image regions, so its application prospect is limited. With the advent and rapid popularization of convolutional neural networks, deep learning has been applied to the field of medical image segmentation. Thanks to the super strong generalization ability and super high reasoning accuracy of deep neural networks, segmentation methods based on deep learning have quickly become an important means of medical image segmentation.

[0005] So far, segmentation methods based on deep learning have become the mainstream technology of medical image segmentation, and theoretical research and network models based on deep learning for medical image segmentation have been constantly emerging, but the accuracy and efficiency of model reasoning have not reached the limit and still have room for improvement. SUMMARY

[0006] The present application aims to provide a medical image segmentation method, system and device based on efficient focus attention, which aims to solve the problem of medical image segmentation.

[0007] The application provides a medical image segmentation method based on efficient focus attention, comprising:

[0008] The image preprocessing module is used for preprocessing the CT image, and the preprocessed image is input into the first embedding layer; the first embedding layer is used for dimension reduction of the preprocessed image to obtain a first dimension-reduced image, and the first dimension-reduced image is input into the first efficient focus attention module; the first efficient focus attention module is used for extracting the features of the first dimension-reduced image to obtain first features; the first features are input into the second embedding layer for dimension reduction, and the dimension-reduced features are input into the second efficient focus attention module to extract features to obtain second features; the second features are input into the third embedding layer for dimension reduction, and the dimension-reduced features are input into the third efficient focus attention module to extract features to obtain third features; the third features are input into the first upsampling layer for dimension increase to obtain first dimension-increased features; the second features and the first dimension-increased features are input into the first skip connection for fusion to obtain first fusion features; the first fusion features are input into the fourth efficient focus attention module to extract features to obtain fourth features; the fourth features are input into the second upsampling layer for dimension increase to obtain second dimension-increased features, and the first features and the second dimension-increased features are input into the second skip connection for fusion to obtain second fusion features; the second fusion features are input into the fifth efficient focus attention module to extract features to obtain fifth features; the fifth features are input into the third upsampling layer for dimension increase to obtain third dimension-increased features; and the third dimension-increased features are input into the linear classifier to restore into an image with a mask.

[0009] The application further provides a medical image segmentation system based on efficient focus attention, comprising:

[0010] The image preprocessing module is used for preprocessing the CT image, and the preprocessed image is input into the first embedding layer; the first embedding layer is used for dimension reduction of the preprocessed image to obtain a first dimension-reduced image, and the first dimension-reduced image is input into the first efficient focus attention module; the first efficient focus attention module is used for extracting the features of the first dimension-reduced image to obtain first features; the first features are input into the second embedding layer for dimension reduction, and the dimension-reduced features are input into the second efficient focus attention module to extract features to obtain second features; the second features are input into the third embedding layer for dimension reduction, and the dimension-reduced features are input into the third efficient focus attention module to extract features to obtain third features; the third features are input into the first upsampling layer for dimension increase to obtain first dimension-increased features; the second features and the first dimension-increased features are input into the first skip connection for fusion to obtain first fusion features; the first fusion features are input into the fourth efficient focus attention module to extract features to obtain fourth features; the fourth features are input into the second upsampling layer for dimension increase to obtain second dimension-increased features, and the first features and the second dimension-increased features are input into the second skip connection for fusion to obtain second fusion features; the second fusion features are input into the fifth efficient focus attention module to extract features to obtain fifth features; the fifth features are input into the third upsampling layer for dimension increase to obtain third dimension-increased features; and the third dimension-increased features are input into the linear classifier to restore into an image with a mask.

[0011] The embodiment of the present application also provides a medical image segmentation device based on efficient focus attention, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, and the computer program realizes the steps of the method when executed by the processor.

[0012] The embodiment of the present application also provides a computer readable storage medium, wherein the computer readable storage medium stores an information transmission implementation program, and the program realizes the steps of the method when executed by a processor.

[0013] The embodiment of the present application can realize medical image segmentation.

[0014] The above description is only a summary of the technical scheme of the present application, in order to more clearly understand the technical means of the present application, according to the content of the specification, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the specific embodiments of the present application or the technical scheme in the prior art, the following will briefly introduce the drawings needed to be used in the specific embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0016] Figure 1 is a schematic diagram of the medical image segmentation system based on efficient focus attention of the embodiment of the present application;

[0017] Figure 2 is a schematic diagram of the efficient focus attention module of the medical image segmentation system based on efficient focus attention of the embodiment of the present application;

[0018] Figure 3 is a schematic diagram of the granularity of focus attention calculation of the medical image segmentation system based on efficient focus attention of the embodiment of the present application;

[0019] Figure 4 is a schematic diagram of the medical image segmentation device based on efficient focus attention of the embodiment of the present application. DETAILED DESCRIPTION

[0020] The technical solutions of the present application will be described clearly and completely below in connection with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0021] System embodiments

[0022] According to the embodiments of the present application, a medical image segmentation system based on efficient focus attention is provided, Figure 1 is a schematic diagram of a medical image segmentation system based on efficient focus attention according to an embodiment of the present application, as Figure 1 shown, specifically comprising:

[0023] An image preprocessing module is configured to preprocess the CT image and input the preprocessed image into a first embedding layer. The first embedding layer is configured to reduce the dimension of the preprocessed image to obtain a first reduced dimension image and input the first reduced dimension image into a first efficient focus attention module. The first efficient focus attention module is configured to extract the features of the first reduced dimension image to obtain a first feature. The first feature is input into a second embedding layer for dimension reduction, and the dimension-reduced second feature is input into a second efficient focus attention module to extract features to obtain a second feature. The second feature is input into a third embedding layer for dimension reduction, and the dimension-reduced third feature is input into a third efficient focus attention module to extract features to obtain a third feature. The third feature is input into a first upsampling layer for dimension increase to obtain a first dimension-increased feature. The second feature and the first dimension-increased feature are input into a first skip connection for fusion to obtain a first fusion feature. The first fusion feature is input into a fourth efficient focus attention module to extract features to obtain a fourth feature. The fourth feature is input into a second upsampling layer for dimension increase to obtain a second dimension-increased feature. The first feature and the second dimension-increased feature are input into a second skip connection for fusion to obtain a second fusion feature. The second fusion feature is input into a fifth efficient focus attention module to extract features to obtain a fifth feature. The fifth feature is input into a third upsampling layer for dimension increase to obtain a third dimension-increased feature. The third dimension-increased feature is input into a linear classifier to restore into an image with a mask.

[0024] The preprocessing module is specifically configured to perform data augmentation on the CT image.

[0025] The preprocessing module is specifically configured to adjust the brightness, shape and color of the CT image to obtain a plurality of different images, thereby completing data augmentation.

[0026] The first efficient focus attention module is used to input the output X1 of the previous layer into the focus attention layer to extract local features and obtain feature vector X2. Then, X1 and X2 are added and input into the efficient attention layer to extract global features and obtain feature vector X3. Finally, X1+X2+X3 is output, which is the fusion of the original input feature vector, local features and global features. The first, second, third, fourth and fifth efficient focus attention modules have the same function.

[0027] The specific implementation method is as follows:

[0028] In this approach, a U-shaped network model based on efficient and focused attention mechanisms is constructed. The model employs an encoder-decoder structure. Trained on a labeled dataset, this deep learning model achieves accurate segmentation of different organs within an image. Figure 1 As shown, we designed an efficient focus attention module capable of extracting both global and local features, and then used this module to construct a U-shaped network with skip connections. The efficient focus attention module includes an efficient attention mechanism and a focus attention mechanism. The efficient attention mechanism has low computational time complexity and can compute global attention for the image, thereby extracting global features. The focus attention mechanism performs attention computations of different granularities on regions near a specific area of ​​the image; attention computation is fine-grained closer to the region and coarse-grained farther away. Skip connections are used for multi-scale feature fusion, which can fuse high-level semantic features with low-level features, resulting in the model's final output containing feature information at different scales, further improving model accuracy.

[0029] (2) System Workflow

[0030] like Figure 1 As shown, after the CT image enters the model, it will pass through various modules in the model along the direction of the arrow to extract features. Finally, a mask image of the same size as the original image will be output. After overlapping with the original image, different colors cover different organs, thereby achieving organ segmentation.

[0031] Image preprocessing stage: As shown in the figure, when CT images are input into the model, the model first performs preprocessing on the CT images. The main purpose of preprocessing is data augmentation. By adjusting the brightness, shape, color, etc., of a single image to obtain multiple different images, data augmentation is achieved. The benefit of increasing the amount of data input into the deep learning model is that it enhances the model's generalization ability, making the trained model more accurate.

[0032] Embedding layer: the role of the embedding layer is to reduce the dimension of the image or the feature vector passed from the previous layer, and map it to a new vector space. During model training, the vectors in the embedding layer will change constantly, and the vectors corresponding to the associated regions in the image (such as the region where an organ is located) will have a higher and higher similarity during model training.

[0033] Figure 2 is a focus attention module schematic diagram of a medical image segmentation system based on efficient focus attention according to an embodiment of the application;

[0034] Efficient focus attention module: the efficient focus attention module is the core part of the model, which is a backbone network for extracting features. The focus attention layer is used to extract local features, and the efficient focus attention layer is used to extract global features and perform feature fusion of the original features, efficient attention features and focus attention features.

[0035] Up-sampling layer: mainly increases the dimension of the feature vector passed from the upper layer to achieve the dimension size required by the next layer network.

[0036] Skip connection: used for multi-scale feature fusion, which integrates high-level semantic features into low-level features, and is the feature of the U-shaped network. Multi-scale feature fusion is very helpful for improving the accuracy of image segmentation.

[0037] The application constructs a U-shaped network model with an encoder-decoder structure. Each input batch contains multiple CT images. The images are first preprocessed to adjust the image size, and data augmentation is performed through image enhancement methods such as brightness adjustment, contrast adjustment, and color adjustment. The preprocessed images enter the embedding layer to generate embedding vectors, and then pass through the efficient focus attention module for feature extraction. This process is repeated three times, and the dimension of the feature vector decreases after each cycle. The network structures at different levels are connected through skip connection to integrate high-level semantic features into low-level features. The feature vector enters the decoder for dimension increasing, and finally passes through a linear classifier to restore it to an image with a mask. Different masks cover different types of organs. The output is compared with the label to calculate the loss value. According to the calculated gradient, the direction of parameter optimization is automatically selected, and finally the model converges in the process of parameter optimization again, realizing accurate segmentation of different organs in CT images.

[0038] Related functional details

[0039] (1) Efficient attention calculation

[0040] The formula of the traditional attention calculation method is: Wherein Q, K, V are all nxd matrices, n is the number of inputs, which is a variable, d is the length of each input, which is a constant, and n is usually much larger than d. The time complexity of the traditional attention calculation formula is O(dn 2 ), which is about the square of the input, so it is difficult to directly calculate the attention of the entire image. The efficient attention calculation formula is: E(Q, K, V) = p q (Q)(p k (K) T V), wherein p q and p k represent the softmax calculation of Q and K respectively. The difference between efficient attention calculation and traditional attention calculation is that efficient attention calculation first multiplies K^T and V, so that the time complexity becomes O(dn 2 ), since n is much larger than d and d is a constant, so such improvement makes the time complexity of attention calculation from quadratic to linear.

[0041] (2) Focus attention calculation

[0042] Figure 3 is a granularity diagram of focus attention calculation of a medical image segmentation system based on efficient focus attention according to an embodiment of the present application;

[0043] For a certain region in the image, the granularity of the focus attention calculation of the attention calculation of the nearby region depends on the distance from the region. Since the segmented object in the medical image often exists in the form of a whole, that is, the region of the same organ that needs to be segmented in the image is often connected, the focus attention performs fine-grained attention calculation on the regions with relatively short distances and coarse-grained attention calculation on the regions with relatively long distances.

[0044] How to divide the granularity: for the regions with relatively short distances, small windows are divided, and attention calculation is performed on each small window. The farther the distance, the larger the window, and no matter how large the finally divided window is, the regions in the window will be compressed into the same feature vector. Therefore, the farther the distance, the larger the window, the more information it contains, and finally compressed into the same feature vector, the more rough the information represented by the feature vector, thereby realizing the difference in granularity.

[0045] High-efficiency focal attention module. Since the segmentation objects of medical images often have structural characteristics, that is, the segmentation objects often present in the form of a whole, which requires the model to consider the extraction of local features. Different segmentation objects often have certain connections, that is, the relative positions of different segmentation objects are generally fixed, which requires the model to consider the extraction of global features. In summary, the high-efficiency focal attention module is proposed, which fuses the high-efficiency attention and focal attention features when extracting image features, realizes the comprehensive consideration of global features and local features, and can make the model more accurate in segmenting multiple organs in medical images.

[0046] The medical image segmentation model based on high-efficiency focal attention is a more efficient and accurate deep learning model, which is used for segmenting organs or lesions of patient medical CT images and plays an auxiliary role in the judgment, decision-making and treatment plan of medical workers on patient conditions.

[0047] Method embodiment

[0048] According to the embodiment of the present application, a medical image segmentation method based on high-efficiency focal attention is provided, which specifically includes:

[0049] The CT image is preprocessed by the image preprocessing module, and the preprocessed image is input into the first embedding layer. The first embedding layer is used to reduce the dimension of the preprocessed image to obtain a first reduced dimension image, and the first reduced dimension image is input into the first high-efficiency focal attention module. The first high-efficiency focal attention module is used to extract the features of the first reduced dimension image to obtain a first feature. The first feature is input into the second embedding layer for dimension reduction, and the dimension-reduced second feature is input into the second high-efficiency focal attention module for feature extraction to obtain a second feature. The second feature is input into the third embedding layer for dimension reduction, and the dimension-reduced third feature is input into the third high-efficiency focal attention module for feature extraction to obtain a third feature. The third feature is input into the first upsampling layer for dimension increase to obtain a first dimension-increased feature. The second feature and the first dimension-increased feature are input into the first jump connection for fusion to obtain a first fusion feature. The first fusion feature is input into the fourth high-efficiency focal attention module for feature extraction to obtain a fourth feature. The fourth feature is input into the second upsampling layer for dimension increase to obtain a second dimension-increased feature. The first feature and the second dimension-increased feature are input into the second jump connection for fusion to obtain a second fusion feature. The second fusion feature is input into the fifth high-efficiency focal attention module for feature extraction to obtain a fifth feature. The fifth feature is input into the third upsampling layer for dimension increase to obtain a third dimension-increased feature. The third dimension-increased feature is input into the linear classifier to restore the image with a mask.

[0050] The CT image is preprocessed by the image preprocessing module, and the preprocessed image is input into the first embedding layer. The first embedding layer is used to reduce the dimension of the preprocessed image to obtain a first reduced dimension image, and the first reduced dimension image is input into the first high-efficiency focal attention module. The first high-efficiency focal attention module is used to extract the features of the first reduced dimension image to obtain a first feature. The first feature is input into the second embedding layer for dimension reduction, and the dimension-reduced second feature is input into the second high-efficiency focal attention module for feature extraction to obtain a second feature. The second feature is input into the third embedding layer for dimension reduction, and the dimension-reduced third feature is input into the third high-efficiency focal attention module for feature extraction to obtain a third feature. The third feature is input into the first upsampling layer for dimension increase to obtain a first dimension-increased feature. The second feature and the first dimension-increased feature are input into the first jump connection for fusion to obtain a first fusion feature. The first fusion feature is input into the fourth high-efficiency focal attention module for feature extraction to obtain a fourth feature. The fourth feature is input into the second upsampling layer for dimension increase to obtain a second dimension-increased feature. The first feature and the second dimension-increased feature are input into the second jump connection for fusion to obtain a second fusion feature. The second fusion feature is input into the fifth high-efficiency focal attention module for feature extraction to obtain a fifth feature. The fifth feature is input into the third upsampling layer for dimension increase to obtain a third dimension-increased feature. The third dimension-increased feature is input into the linear classifier to restore the image with a mask.

[0051] The data augmentation of the CT image specifically comprises: adjusting brightness, shape and color of the CT image to obtain multiple different pictures, and completing data augmentation.

[0052] The first, second, third, fourth and fifth high-efficiency focal attention modules have the same function; the output X1 of the previous layer network is input to the focal attention layer through the first high-efficiency focal attention module to extract a local feature vector X2, then X1 and X2 are added and input to the high-efficiency attention layer to extract a global feature vector X3, and finally X1+X2+X3 is output, that is, the fusion of the original input feature vector, the local feature and the global feature.

[0053] The embodiment of the application is a system embodiment corresponding to the above-mentioned method embodiment, and the specific operation of each module can be understood with reference to the description of the method embodiment, which will not be repeated here.

[0054] Device embodiment one

[0055] The embodiment of the application provides a medical image segmentation device based on high-efficiency focal attention, as shown in the figure, comprising: a memory 40, a processor 42 and a computer program stored on the memory 40 and executable on the processor 42, the computer program is executed by the processor to realize the steps in the above-mentioned method embodiment. Figure 4

[0056] Device embodiment two

[0057] The embodiment of the application provides a computer readable storage medium, and the computer readable storage medium stores an implementation program of information transmission, and the program is executed by the processor 42 to realize the steps in the above-mentioned method embodiment.

[0058] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the application, and not to limit them; although the application has been described in detail with reference to the above embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the above embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the present application.​

Claims

1. A medical image segmentation system based on efficient focus attention, characterized by, The method comprises the following steps: An image preprocessing module is configured to preprocess a CT image and input the preprocessed image into a first embedding layer; The first embedding layer is configured to reduce the dimensionality of the preprocessed image to obtain a first reduced dimensionality image, and input the first reduced dimensionality image into a first efficient focal point attention module; The first efficient focal point attention module is configured to extract features of the first reduced dimensionality image to obtain first features; The first features are input into a second embedding layer for dimensionality reduction, and the dimensionally reduced features are input into a second efficient focal point attention module to extract features to obtain second features; the second features are input into a third embedding layer for dimensionality reduction, and the dimensionally reduced features are input into a third efficient focal point attention module to extract features to obtain third features; the third features are input into a first upsampling layer for dimensionality increase to obtain first dimensionally increased features; The second features and the first dimensionally increased features are input into a first skip connection for fusion to obtain first fused features; the first fused features are input into a fourth efficient focal point attention module to extract features to obtain fourth features; The fourth features are input into a second upsampling layer for dimensionality increase to obtain second dimensionally increased features; the first features and the second dimensionally increased features are input into a second skip connection for fusion to obtain second fused features; the second fused features are input into a fifth efficient focal point attention module to extract features to obtain fifth features; The fifth features are input into a third upsampling layer for dimensionality increase to obtain third dimensionally increased features; the third dimensionally increased features are input into a linear classifier to restore the features into an image with a mask; The first efficient focal point attention module is configured to input the output X1 of the previous layer network into a focal point attention layer to extract local features and obtain a feature vector X2, add X1 and X2 to obtain a feature vector X3 by inputting the sum into an efficient attention layer to extract global features, and finally output X1+X2+X3, which is the fusion of the original input feature vector, local features and global features; the first efficient focal point attention module, the second efficient focal point attention module, the third efficient focal point attention module, the fourth efficient focal point attention module and the fifth efficient focal point attention module have the same function; wherein the calculation formula of the efficient attention is wherein, and respectively represent the softmax calculation of Q and K, Q, K, and V are all nxd matrices, n is the number of inputs, and d is the length of each input; the granularity of the attention calculation of the nearby area when the focal attention is calculated depends on the distance from the area, the focal attention performs fine-grained attention calculation on the areas with shorter distance, and performs coarse-grained attention calculation on the areas with longer distance.

2. The system of claim 1, wherein, The preprocessing module is specifically configured to perform data augmentation on the CT image.

3. The system of claim 2, wherein, The preprocessing module is specifically configured to adjust the brightness, shape and color of the CT image to obtain a plurality of different images, thereby completing the data augmentation.

4. A medical image segmentation method based on efficient focus attention, characterized in that, The method comprises the following steps: An image preprocessing module is configured to preprocess a CT image and input the preprocessed image into a first embedding layer; The first embedding layer is configured to reduce the dimensionality of the preprocessed image to obtain a first reduced dimensionality image, and input the first reduced dimensionality image into a first efficient focal point attention module; The first efficient focal point attention module is configured to extract features of the first reduced dimensionality image to obtain first features; The first features are input into a second embedding layer for dimensionality reduction, and the dimensionally reduced features are input into a second efficient focal point attention module to extract features to obtain second features; the second features are input into a third embedding layer for dimensionality reduction, and the dimensionally reduced features are input into a third efficient focal point attention module to extract features to obtain third features; the third features are input into a first upsampling layer for dimensionality increase to obtain first dimensionally increased features; The second features and the first dimensionally increased features are input into a first skip connection for fusion to obtain first fused features; the first fused features are input into a fourth efficient focal point attention module to extract features to obtain fourth features; The fourth feature is input into a second upsampling layer to obtain a second dimension-increased feature, the first feature and the second dimension-increased feature are input into a second skip connection to obtain a second fusion feature after fusion; The fifth feature is input into a third upsampling layer to obtain a third dimension-increased feature; and the third dimension-increased feature is input into a linear classifier to restore an image with a mask; The first high-efficiency focal attention module, the second high-efficiency focal attention module, the third high-efficiency focal attention module, the fourth high-efficiency focal attention module and the fifth high-efficiency focal attention module have the same function; the output X1 of the last layer of network is input into a focal attention layer through the first high-efficiency focal attention module to extract a local feature and obtain a feature vector X2, then X1 and X2 are added and input into a high-efficiency attention layer to extract a global feature and obtain a feature vector X3, finally, X1+X2+X3 is output, that is, the fusion of the original input feature vector, the local feature and the global feature; wherein, the calculation formula of the high-efficiency attention is wherein, and respectively represent the softmax calculation of Q and K, Q, K and V are all nxd matrices, n is the number of inputs, and d is the length of each input; the granularity of the attention calculation of the nearby area when the focal attention is calculated depends on the distance from the area, the focal attention performs fine-grained attention calculation on the areas with shorter distance, and performs coarse-grained attention calculation on the areas with longer distance.

5. The method of claim 4, wherein, The CT image is preprocessed by the image preprocessing module, specifically, the CT image is data-augmented.

6. The method of claim 5, wherein, The CT image is data-augmented, specifically, the CT image is adjusted in brightness, shape and color to obtain multiple different pictures, and data augmentation is completed. 7.A medical image segmentation apparatus based on efficient focus attention, characterized by The method comprises the following steps: The memory, the processor and the computer program stored in the memory and executable on the processor are used to implement the steps of the medical image segmentation method based on high-efficiency focal attention as claimed in any one of claims 4 to 6.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores an information transmission implementation program, and the program is executed by the processor to implement the steps of the medical image segmentation method based on high-efficiency focal attention as claimed in any one of claims 4 to 6.

Images