System for medical image processing, medium and computer equipment

Abstract

The invention provides a system for medical image processing, a medium and computer equipment. The system comprises an image preprocessing module, an image convolutional network module, a convolutional neural network module and a multi-modal feature fusion module. The image preprocessing module performs morphological operation on the tumor core area of the first image to obtain a second image; the graph convolutional network module outputs a first feature vector integrating information of a tumor core region and information of a peritumor neighborhood region; the convolutional neural network module performs convolution to extract deep features of the image to obtain a second feature vector; performing feature fusion on the first feature vector and the second feature vector to obtain a first output and a second output, and determining a final tumor identification result by the first output and the second output through a multi-layer perceptron. According to the system disclosed by the invention, the input breast ultrasonic image is subjected to image processing, and judgment is carried out based on related characteristics of a tumor core region and a peritumor neighborhood region, so that benign and malignant types of breast tumors are divided.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and in particular to a system, medium and computer equipment for medical image processing. Background Art

[0002] According to statistics, breast cancer has become the most common malignant tumor in women and one of the most common causes of cancer death in women. Early detection and diagnosis can provide the best guidance for tumor treatment plans, are the key to reducing mortality, and have important clinical research value. X-rays, ultrasound and magnetic resonance imaging are the main screening methods for breast cancer detection, especially ultrasound, which is widely used because of its convenience, non-invasiveness, no radiation, real-time and good tolerance. However, to ensure the diagnostic accuracy of breast ultrasound, radiologists need to have rich experience and proficiency. In addition, medical imaging experts are required to interpret the results that may be highly subjective. However, for the huge amount of ultrasound image data generated in clinical practice, medical imaging experts need to read the films one by one and give corresponding judgment results. The huge workload makes it easy to miss small features on the image, and it is easy to be misjudged by subjective factors.

[0003] In view of the above problems, a computer-aided diagnosis system is needed to assist in identification based on medical images. Summary of the invention

[0004] The purpose of the present invention is to provide a system, medium and computer equipment for medical image processing, which can realize automatic identification of benign and malignant breast tumors based on ultrasound images.

[0005] To achieve the above-mentioned object, the present invention provides a system for medical image processing, which comprises an image preprocessing module, a graph convolutional network module, a convolutional neural network module and a multimodal feature fusion module;

[0006] The image preprocessing module is used to receive a plurality of first images, wherein the tumor core area of ​​the first images carries a mask mark image, and obtain a plurality of second images including an expansion area by performing morphological operations on the tumor core areas of the plurality of first images, wherein a difference set between the expansion area and the tumor core area is used as a peritumoral area, and the peritumoral area is equally divided into eight peritumoral neighborhood areas along a circumferential direction, and a plurality of high-throughput imaging omics features are extracted from the tumor core area and each of the peritumoral neighborhood areas;

[0007] The graph convolution network module integrates the relevant information of the images of each of the peritumoral neighborhood areas and the tumor core area to obtain the relevant information between each of the peritumoral neighborhood areas and the tumor core area, and outputs a first feature vector that integrates the information of the tumor core area and the peritumoral neighborhood area;

[0008] The convolutional neural network module is used to perform normalization processing on the plurality of second images to obtain a plurality of third images with uniform resolution, the convolutional neural network module comprises a convolution block with a three-layer bottleneck structure, and the convolutional neural network module performs convolution on the third image to extract deep features of the image to obtain a second feature vector;

[0009] The multimodal feature fusion module performs feature fusion on the first feature vector and the second feature vector to obtain a first output and a second output. The first output and the second output are respectively added after passing through a multi-layer perceptron to determine a final tumor identification result.

[0010] Optionally, the high-throughput imaging features extracted by the image preprocessing module from the tumor core area and each of the tumor surrounding neighboring areas include 16 grayscale features, 53 texture features and 276 wavelet transform features.

[0011] Optionally, the 53 texture features include 22 grayscale co-occurrence matrix features, 13 grayscale run-length matrix features, 13 grayscale area matrix features and 5 neighborhood grayscale difference matrix features.

[0012] Optionally, the convolutional neural network module includes an input layer, a multi-layer convolution block including a three-layer bottleneck structure, an average pooling layer and an output layer.

[0013] Optionally, the three-layer bottleneck structure convolution block includes a 1×1 convolution layer, a 3×3 convolution layer and a 1×1 convolution layer connected in sequence.

[0014] Optionally, four layers of convolutional blocks including three bottleneck structures are included.

[0015] Optionally, the multimodal feature fusion module performs feature fusion on the first feature vector and the second feature vector to obtain a first output and a second output, specifically including:

[0016] Map the first feature vector to the intermediate representation space as Query1, map the second feature vector to the intermediate representation space as Key1 and Value1, and calculate their attention scores:

[0017]

[0018] In the formula, Q1 represents Query1, K1 represents Key1, and d represents the dimension of Q1 and K1;

[0019] Combine the attention score with Value1 to guide the feature representation of Value1, as shown in the following formula:

[0020]

[0021] Where α represents an optional scaling factor used to control the influence of the attention score; V1 represents Value1; output1 represents the first output;

[0022] Map the second feature vector to the intermediate representation space as Query2, map the first feature vector to the intermediate representation space as Key2 and Value2, and calculate their attention scores:

[0023]

[0024] In the formula, Q2 represents Query2, K2 represents Key2, and d represents the dimension of Q2 and K2;

[0025] Combine the attention score with Value2 to guide the feature representation of Value2, as shown in the following formula:

[0026]

[0027] Where α represents an optional scaling factor used to control the influence of the attention score; V2 represents Value2; output2 represents the second output.

[0028] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores the system for medical image processing as described above.

[0029] The present invention also provides a computer device, which includes the storage medium as described above.

[0030] In summary, compared with the prior art, the system, medium and computer device for medical image processing provided by the present invention have the following beneficial effects:

[0031] The system, medium and computer equipment for medical image processing of the present invention perform image processing on the input breast ultrasound image, make judgments based on the relevant features of the tumor core area and the surrounding area, and classify the benign and malignant types of breast tumors. The process is automatically completed by the algorithm without the need for manual intervention to extract identification parameters, and has strong robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic diagram of the structure of the system for medical image processing of the present invention.

[0033] Figure 2 This is a flow chart of a method for distinguishing benign and malignant breast cancer based on a combination of internal and external texture structure features of a tumor and spatial position features in a system for medical image processing of the present invention.

[0034] Figure 3It is a process schematic diagram of the system identification method for medical image processing of the present invention.

[0035] Figure 4 This is the receiver operating characteristic curve. DETAILED DESCRIPTION

[0036] The following will be combined with the attached embodiment of the present invention Figure 1 ~Attached Figure 4 , the technical solutions, structural features, objectives achieved and effects in the embodiments of the present invention are described in detail.

[0037] It should be noted that the drawings are in a very simplified form and use non-precise proportions. They are only used to conveniently and clearly assist in explaining the embodiments of the present invention, and are not used to limit the conditions for the implementation of the present invention. Therefore, they have no substantive technical significance. Any structural modification, change in proportional relationship or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical content disclosed by the present invention.

[0038] It should be noted that, in the present invention, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only the elements explicitly listed, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0039] like Figure 1-Figure 3 As shown, the present invention provides a system for medical image processing, which includes an image preprocessing module, a graph convolutional network module, a convolutional neural network module and a multimodal feature fusion module.

[0040] The division of the tumor core area and the surrounding area is the premise and basis for subsequent feature extraction and classification recognition. First, the original image is manually drawn or the tumor core area mask is obtained through the tumor segmentation system. The mask is used to highlight or isolate the core part of the tumor for further analysis or processing. The tumor core area refers to the most representative part of the tumor that may contain the most malignant cells. This is usually the most active growth area or the part with the richest blood flow in the tumor.

[0041] The image preprocessing module receives multiple first images, where the first image is an image with a mask mark on the core area of ​​the tumor. Multiple second images including the expansion area are obtained by performing morphological operations on the core areas of the tumor of the multiple first images. Specifically, the morphological operation includes taking the mask area as the intratumoral area, using the cv2.dilate function, setting the parameters kernel to 4×4, and iteration to 10, and performing morphological operations on the mask area to obtain the expansion area. The difference between the expansion area and the core area of ​​the tumor is taken as the peritumoral area. A rectangular coordinate system is established with the center of the tumor as the origin, and each quadrant is divided into two equal parts, so that 8 sector-shaped areas can be obtained. The intersection of the peritumoral area and one of the sector-shaped areas is recorded as a peritumoral neighborhood area, and finally a total of 8 peritumoral neighborhood areas are obtained.

[0042] In order to effectively utilize the spatial position correlation between each peritumoral neighborhood area, the graph convolutional network module integrates the relevant information of the images of each peritumoral neighborhood area and the tumor core area to obtain the relevant information between each peritumoral neighborhood area and the tumor core area, and outputs the first feature vector that combines the tumor core area and the peritumoral neighborhood area.

[0043] Specifically, the graph convolutional network module extracts multiple high-throughput imaging features for the tumor core area and each peritumoral neighborhood area. A total of 345 high-throughput imaging features are extracted, including 16 grayscale features, 53 texture features and 276 wavelet transform features. In this way, the information of each image can be represented by a 9×345 feature vector.

[0044] The 53 texture features include 22 gray-level co-occurrence matrix features, 13 gray-level run-length matrix features, 13 gray-level region matrix features and 5 neighborhood gray-level difference matrix features.

[0045] The 16 grayscale features are: 1. Energy; 2. Entropy of histogram; 3. Peak; 4. Mean absolute error; 5. Average; 6. Median; 7. Grayscale range; 8. Root mean square; 9. Skew; 10. Standard deviation; 11. Histogram uniformity; 12. Variance; 13. Histogram mean; 14. Histogram variance; 15. Histogram skew; 16. Histogram peak.

[0046] The 22 gray-level co-occurrence matrix features are: 1. Energy; 2. Entropy; 3. Difference; 4. Contrast; 5. Contrast; 6. Correlation; 7. Homogeneity; 8. Autocorrelation; 9. Cluster shading; 10. Cluster prominence; 11. Maximum probability; 12. Sum of squares; 13. Average sum; 14. Total variance; 15. Sum of entropy; 16. Difference variance; 17. Difference entropy; 18. Co-correlation normalization; 19. Co-correlation normalization; 20. Maximum correlation coefficient; 21. Contrast normalization; 22. Contrast moment normalization.

[0047] The 13 grayscale run matrix features are specifically as follows: 23. Short-term emphasis; 24. Long-term emphasis; 25. Grayscale non-uniformity; 26. Long-line non-uniformity; 27. Long-line percentage; 28. Line measure of low grayscale value; 29. ​​Line length of high grayscale value; 30. Line measure of low grayscale value of short line; 31. Line measure of high grayscale value of short line; 32. Line measure of low grayscale value of long line; 33. Line measure of high grayscale value of long line; 34. Grayscale value variance; 35. Long-line variance.

[0048] The 13 grayscale area matrix features are specifically as follows: 36, small block measurement; 37, large block measurement; 38, grayscale non-uniformity; 39, block non-uniformity; 40, block percentage; 41, block measurement of low grayscale values; 42, block measurement of high grayscale values; 43, block measurement of low grayscale values ​​in small areas; 44, block measurement of high grayscale values ​​in small areas; 45, block measurement of low grayscale values ​​in large areas; 46, block measurement of high grayscale values ​​in large areas; 47, grayscale value variance; 48, block size variance.

[0049] The five neighborhood grayscale difference matrix features are: 49, roughness; 50, contrast; 51, busyness; 52, complexity; 53, intensity.

[0050] The 276 wavelet transform features are components obtained by wavelet decomposition and analyzed separately in 4 sub-bands. Wavelet transform features are used to represent information that cannot be directly reflected by grayscale and texture features.

[0051] The graph convolutional network module includes a graph convolutional network. The core idea of ​​the graph convolutional network is that each node will receive messages from its neighboring nodes, fuse their features and update its representation. The graph can be represented as G = (V, E), where V and E represent the nodes and adjacent edges of the graph convolutional network respectively.

[0052] Each node in the figure represents an area obtained by the above segmentation, and the relationship between the nodes will be dynamically updated as the network is trained. Since the 9 areas obtained by segmentation have an intuitive connection relationship, they are used as the basis for constructing the adjacency matrix. That is, if two areas are connected, the adjacent edge value between the nodes of the two areas is 1, and if they are not connected, it is 0. The 9×345 feature vector obtained by the above steps is first normalized to the range of 0-1, and then updated through the graph convolution network. Finally, a feature vector of dimension 9×512 is obtained. The graph convolution process on the figure can be described as:

[0053]

[0054] In the formula, the matrix H l represents the node features of the lth layer, H 0 =X input σ represents the nonlinear activation function, Wl-1 is the learnable weight matrix, represents the normalized adjacency matrix, which can be calculated as:

[0055]

[0056] Where I is the unit matrix, yes The whole process can be regarded as the propagation and aggregation of information on the graph. By learning the complex interactions between nodes, the graph convolutional network can finally obtain a feature vector that integrates the information of the tumor core area and the surrounding area.

[0057] In the present invention, a convolutional neural network is constructed as a parallel branch to extract local structural texture features of the tumor. Since a large part of the information in the breast ultrasound image is irrelevant glands and tissues, in order to reduce the impact of redundant information on classification accuracy, the convolutional neural network module normalizes the multiple second images to obtain multiple third images with uniform resolution. To facilitate subsequent unified processing, the third image uses a resolution of 512×512.

[0058] The convolutional neural network module includes a convolution block with a three-layer bottleneck structure. The convolutional neural network module performs convolution on the third image to extract deep image features to obtain a second feature vector FC, wherein the output size of the second feature vector FC is 8×8×1024.

[0059] The convolutional neural network module includes an input layer, four convolution blocks including three bottleneck structures, an average pooling layer, and an output layer. In the first convolution stage, the input image is processed preliminarily to extract some basic features, and a certain degree of downsampling is performed to reduce the resolution of the image. For example, the size of the input image is reduced by half, while the number of channels is moderately increased; the second convolution stage further extracts features based on the first layer, the resolution continues to decrease, and the number of channels is further increased. At this stage, the features begin to become more abstract and representative; the third convolution stage continues to deepen the feature extraction, the resolution is further reduced, the number of channels continues to increase, and the information in the feature map is more focused on the key features and patterns of the image; the fourth convolution stage finally extracts the features in depth, the resolution reaches the minimum, and the number of feature channels reaches a relatively high value. At this time, the extracted features are deep features of the image, containing rich semantic information and structural information. In general, the convolutional neural network extracts deep features of the image by gradually reducing the resolution and increasing the number of feature channels through four convolution blocks, and finally obtains a feature vector containing local structural texture information through the average pooling layer.

[0060] The stages of the three-layer bottleneck structure convolution block include connecting 1×1 convolution layer, 3×3 convolution layer and 1×1 convolution layer in sequence. Among them, the first 1×1 convolution layer is responsible for reducing the dimension of the input feature map, reducing the number of channels, and reducing the complexity and parameter amount of subsequent calculations; the second 3×3 convolution layer is the main feature extraction layer, which is used to capture local features of the image, such as edges, textures and other information. Due to the dimensionality reduction processing of the first layer, the amount of calculation of this layer is relatively small; the third layer is also a 1×1 convolution layer, which is used to restore the number of channels of the feature map to an appropriate number in order to connect with the subsequent layers. Such a structure can ensure that the number of filters in the 3×3 convolution layer is independent of the upper and lower layers when reducing or expanding the dimension of the feature map, which not only maintains the accuracy of the model, but also reduces the number of parameters. At the same time, the residual structure is added to accelerate the convergence of the network and avoid gradient disappearance.

[0061] The multimodal feature fusion module performs feature fusion on the first feature vector FG (9×512) and the second feature vector FC (8×8×1024) to obtain the first output output1 and the second output output2. The first output output1 and the second output output2 are respectively added through the multi-layer perceptron to determine the final tumor identification result.

[0062] The multimodal feature fusion module performs feature fusion on the first feature vector FG and the second feature vector FC to obtain the first output output1 and the second output output2, specifically including:

[0063] Map the first feature vector FG to the intermediate representation space as Query1, map the second feature vector FC to the intermediate representation space as Key1 and Value1, and calculate their attention scores:

[0064]

[0065] Where Q1 represents the vector Query1, K1 represents the vector Key1, and d represents the dimension of Q1 and K1.

[0066] Combining the attention score with Value1, specifically, multiplying the attention score by Value1 and adding the weighted learnable parameter α to Value1 can improve the stability of the module to guide the feature representation of Value1. The process is shown in the following formula:

[0067]

[0068] Where α represents an optional scaling factor used to control the influence of the attention score; V1 represents Value1; output1 represents the first output;

[0069] Map the second feature vector FC to the intermediate representation space as Query2, map the first feature vector FG to the intermediate representation space as Key2 and Value2, and calculate their attention scores:

[0070]

[0071] In the formula, Q2 represents Query2, K2 represents Key2, and d represents the dimension of Q2 and K2;

[0072] Combine the attention score with Value2 to guide the feature representation of Value2, as shown in the following formula:

[0073]

[0074] Where α represents an optional scaling factor used to control the influence of the attention score; V2 represents Value2; output2 represents the second output.

[0075] Finally, ACC (accuracy), SEN (sensitivity), SPE (specificity), AUC (Area Under the Curve), PPV (Positive Predictive Value) and NPV (Negative Predictive Value) were calculated to quantitatively evaluate the identification performance.

[0076] The network parameters of the system for medical image processing of the present invention are set as follows: the network is trained using the PyTorch deep learning framework on Nvidia GTX Titan X, the batch size is set to 8, the Adam optimization algorithm with an initial learning rate of 1e-4, and the binary cross entropy loss function is used, which is defined as follows:

[0077] L=∑[Y·logP+(1-Y)·log(1-P)]

[0078] In the formula, P is the sample probability value identified by the model, and the overall loss function of the network is designed as:

[0079] L total =λ1·L QKV +λ2·L G +λ3·L C

[0080] Where L G and L C Represent the output of graph convolutional network and convolutional neural network respectively, L QKVrepresents the output of the feature fusion module, and λ1, λ2, and λ3 represent the weights of each loss respectively.

[0081] The medical image processing system of the present invention is experimented on an ultrasound breast cancer dataset and the effect is confirmed by a five-fold cross validation. A public dataset is used, consisting of 437 benign cases and 210 malignant cases, with an average size of 500×500 pixels, and corresponding segmentation masks from 600 female patients aged between 25 and 75 years old. These data are collected by the LOGIQ E9 US system and the LOGIQ E9 Agile US system, and the sensor frequency of the linear array probe is 1–5MHz.

[0082] like Figure 4 As shown, the present invention conducted comparative experiments to prove the effectiveness of the proposed method. The method of the present invention was compared with four existing methods, including Logistic Regression, Resnest18, Vision Transformer, and a certain literature method. Table 1 below shows the classification and identification of benign and malignant breast cancer by different methods on data set A. It can be seen that the method of the present invention achieved the best performance. Compared with the other four methods, the highest accuracy rate of 96.66% was achieved. In addition, the AUC was 0.9839, the SEN was 92.58%, the SPE was 98.86%, the PPV was 97.71%, and the NPV was 96.14%. It can be seen that compared with the existing methods, the method of the present invention has been improved overall, has better feature extraction capabilities, and illustrates the role of information interaction in the tumor and surrounding areas.

[0083] Table 1. Results of different methods for the classification and identification of benign and malignant breast cancer

[0084]

[0085] The present invention uses ultrasonic grayscale images to distinguish benign and malignant breast cancer and achieves satisfactory results, which reflects the practicality and universal applicability of the method of the present invention. In addition, since the entire identification process of the method of the present invention is automatically completed by the algorithm, there is no need for manual intervention to extract identification parameters, so it has strong robustness in practical applications.

[0086] The system for medical image processing of the present invention first constructs a graph convolutional network module to extract the potential connection between the tumor core area and the surrounding neighboring area as the global feature of the image; secondly, a convolutional neural network module is constructed to extract the structural texture features of the tumor area as the local feature of the image; finally, a multimodal feature fusion module is established for the interactive learning of the above-mentioned global features and local features, so as to further improve the accuracy of distinguishing benign and malignant tumors.

[0087] The difference between benign and malignant tumors is not only reflected in the core area of ​​the tumor, but also in the surrounding neighborhood areas. Therefore, the present invention first divides the core area of ​​the tumor and the surrounding multi-neighborhood areas, extracts high-throughput features of the images of each area, and effectively converts tumor images of different volumes and shapes into high-dimensional data that depict tumor information. A graph convolutional network model is constructed to screen and learn features based on graph node features and graph structure, which not only enhances the correlation of features in different regions, but also helps to improve the accuracy of identification; a convolutional neural network is established to extract deep detail features of the tumor area as another modal feature complement, and a feature fusion framework is established based on the potential connection between multimodal features, which can further improve the performance of classification and identification. The entire process is completed automatically without the need for manual extraction of parameters, and has strong robustness.

[0088] The present invention also provides a computer-readable storage medium, which stores the system for medical image processing as described above.

[0089] The present invention also provides a computer device, which includes the storage medium as described above.

[0090] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be appreciated that the above description should not be considered as a limitation of the present invention. After reading the above content, it will be apparent to those skilled in the art that various modifications and substitutions of the present invention will occur. Therefore, the protection scope of the present invention should be limited by the appended claims.

Claims

1. A system for medical image processing, characterized in that: The system includes an image preprocessing module, a graph convolutional network module, a convolutional neural network module and a multimodal feature fusion module; The image preprocessing module is used to receive a plurality of first images, wherein the tumor core area of ​​the first images carries a mask mark image, and obtain a plurality of second images including an expansion area by performing morphological operations on the tumor core areas of the plurality of first images, wherein a difference set between the expansion area and the tumor core area is used as a peritumoral area, and the peritumoral area is equally divided into eight peritumoral neighborhood areas along a circumferential direction, and a plurality of high-throughput imaging omics features are extracted from the tumor core area and each of the peritumoral neighborhood areas; The graph convolution network module integrates the relevant information of the images of each of the peritumoral neighborhood areas and the tumor core area to obtain the relevant information between each of the peritumoral neighborhood areas and the tumor core area, and outputs a first feature vector that integrates the information of the tumor core area and the peritumoral neighborhood area; The convolutional neural network module is used to perform normalization processing on the plurality of second images to obtain a plurality of third images with uniform resolution, the convolutional neural network module comprises a convolution block with a three-layer bottleneck structure, and the convolutional neural network module performs convolution on the third image to extract deep features of the image to obtain a second feature vector; The multimodal feature fusion module performs feature fusion on the first feature vector and the second feature vector to obtain a first output and a second output. The first output and the second output are respectively added after passing through a multi-layer perceptron to determine a final tumor identification result.

2. The system for medical image processing according to claim 1, characterized in that: The high-throughput imaging features extracted by the image preprocessing module from the tumor core area and each of the tumor surrounding neighboring areas include 16 grayscale features, 53 texture features and 276 wavelet transform features.

3. The system for medical image processing according to claim 2, characterized in that: The 53 texture features include 22 grayscale co-occurrence matrix features, 13 grayscale run matrix features, 13 grayscale region matrix features and 5 neighborhood grayscale difference matrix features.

4. The system for medical image processing according to claim 1, characterized in that: The convolutional neural network module includes an input layer, a multi-layer convolution block including a three-layer bottleneck structure, an average pooling layer and an output layer.

5. The system for medical image processing according to claim 4, characterized in that: The three-layer bottleneck structure convolution block includes a 1×1 convolution layer, a 3×3 convolution layer and a 1×1 convolution layer connected in sequence.

6. The system for medical image processing according to claim 4, characterized in that: It includes four layers of convolutional blocks with three bottleneck structures.

7. The system for medical image processing according to claim 1, characterized in that: The multimodal feature fusion module performs feature fusion on the first feature vector and the second feature vector to obtain a first output and a second output, specifically comprising: Map the first feature vector to the intermediate representation space as Query1, map the second feature vector to the intermediate representation space as Key1 and Value1, and calculate their attention scores: In the formula, Q1 represents Query1, K1 represents Key1, and d represents the dimension of Q1 and K1; Combine the attention score with Value1 to guide the feature representation of Value1, as shown in the following formula: Where α represents an optional scaling factor used to control the influence of the attention score; V1 represents Value1; output1 represents the first output; Map the second feature vector to the intermediate representation space as Query2, map the first feature vector to the intermediate representation space as Key2 and Value2, and calculate their attention scores: In the formula, Q2 represents Query2, K2 represents Key2, and d represents the dimension of Q2 and K2; Combine the attention score with Value2 to guide the feature representation of Value2, as shown in the following formula: Where α represents an optional scaling factor used to control the influence of the attention score; V2 represents Value2; output2 represents the second output.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores the system for medical image processing according to any one of claims 1 to 7.

9. A computer device, characterized in that: The computer device comprises the storage medium of claim 8.

Images