A few-shot learning tumor segmentation method driven by graph attention model

Abstract

The present invention discloses a small sample learning tumor segmentation method driven by a graph attention model, wherein data is divided into a support set and a query set, the support set and the query set are input into an encoder for encoding, and support features and query features are obtained; feature A of a tumor region of interest is obtained according to the support feature, and feature A is de-pooled so that the support feature and the query feature have the same resolution; the de-pooled support feature and the query feature are superimposed on a channel basis, the superimposed features are put into a graph attention module, and the spatial information between feature slices is learned through the graph attention module; the output of the graph attention module is put into a decoder, and a segmentation result is obtained in combination with a jump connection method; based on the graph attention model, the present invention improves the accuracy of tumor image data segmentation in small sample conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision, and in particular to a small sample learning tumor segmentation method driven by a graph attention model. Background Art

[0002] Semantic segmentation plays a vital role in medical imaging in recent years. In the past few decades, with the extensive research on automated segmentation techniques, deep learning-based segmentation models are driving progress in this problem. In particular, convolutional neural networks (CNNs) can achieve state-of-the-art performance when trained with sufficient labeled data. However, due to cost and time considerations, it is impractical to train such high-performance models in the field of medical imaging due to the lack of large amounts of manually annotated data.

[0003] In order to reduce the burden of annotating a large number of samples, few-shot learning has emerged. In the context of few-shot learning, the few-shot learning model is trained to learn previously unseen categories using only a small number of samples, which are usually represented as support images, and the corresponding segmentation masks to guide the segmentation of unlabeled samples, usually query images represented by features derived from the support images. In this case, when it comes to breast cancer datasets, a model trained on several cancer molecular subtypes is able to accurately segment other molecular subtypes with a small number of samples. However, most few-shot learning methods based on 2D natural images cannot be directly applied to 3D images because they lose the spatial context information between 3D image slices and may cause overfitting, resulting in low cutting accuracy under small samples and inability to process in 3D. Summary of the invention

[0004] The purpose of this section is to summarize some aspects of embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the specification abstract and the invention title of this application to avoid blurring the purpose of this section, the specification abstract and the invention title, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0005] In view of the above existing problems, the present invention is proposed.

[0006] Therefore, the present invention provides a small sample learning tumor segmentation method driven by a graph attention model, which can (1) avoid the problem that 3D images cannot be operated;

[0007] (2) Avoid the problem of inaccurate segmentation under small samples.

[0008] To solve the above technical problems, the present invention provides the following technical solutions: including, dividing the data into a support set and a query set, inputting the support set and the query set into an encoder for encoding, and obtaining support features and query features; obtaining feature A of the tumor region of interest according to the support features, and de-pooling the feature A so that the support features and the query features have the same resolution; superimposing the de-pooled support features and the query features by channel, putting the superimposed features into a graph attention module, and learning the spatial information between feature slices through the graph attention module; putting the output of the graph attention module into a decoder, and obtaining the segmentation result in combination with a jump connection method.

[0009] This is a preferred solution for the small sample learning tumor segmentation method driven by a graph attention model described in the present invention, wherein: the data includes obtaining a sagittal MRI image of a 3D breast cancer tumor; obtaining a binary image corresponding to the 3D tumor image; obtaining a molecular typing report corresponding to each 3D tumor, and the molecular typing includes: normal type, cavitary type, HER2 overexpression type and triple-negative breast cancer.

[0010] This is a preferred solution for the small sample learning tumor segmentation method driven by a graph attention model described in the present invention, wherein: the data is divided into a support set and a query set, including classifying the tumor data using the molecular typing report; each type of tumor data is divided into a training set and a test set, and a strategy training network is trained using three types of tumors to test the remaining type of tumor; the support set and query set are randomly selected from the other three types of training data.

[0011] This is a preferred solution of a small sample learning tumor segmentation method driven by a graph attention model according to the present invention, wherein: the obtaining of the supporting features and the query features includes that the supporting features and the query features obtained after encoding the t-th slice of the support and query images are:

[0012] X q，t =E(I q，t )

[0013] Among them, E is the encoder, K is the K samples in the support set, is the t-th slice image of the k-th sample in the support set, I q,t is a slice of the query image, X s,t To support the feature, X q,t For query features.

[0014] This is a preferred solution of a small sample learning tumor segmentation method driven by a graph attention model described in the present invention, wherein: obtaining a feature A of a tumor region of interest from the support features, and performing unpooling on the feature A includes scaling the size of the support set segmentation label using a bilinear interpolation technique; performing a Hadamard product between the segmentation label of the support set and the encoded support features to calculate the features of the tumor region of interest; and pooling the features of the tumor region of interest using a global average pooling technique:

[0015]

[0016] in, is the support feature of the t-th slice of the k-th sample after encoding, for The corresponding segmentation label, ⊙ is the Hadamard product, is the supporting feature of the t-th slice of the representative k-th sample after pooling, and GAP() is the global average pooling technique.

[0017] This is a preferred solution of a small sample learning tumor segmentation method driven by a graph attention model described in the present invention, wherein: the channel-by-channel superposition of the de-pooled support features and the query features includes:

[0018]

[0019] in, is the feature of the tumor region of interest of the support set of the kth sample and the tth slice after processing, For channel superposition operation, h t Slice feature.

[0020] This is a preferred solution of a small sample learning tumor segmentation method driven by a graph attention model described in the present invention, wherein: the graph attention module learns the spatial information between feature slices, including treating a slice of a 3D tumor image as a vertex on the graph, and the feature of the slice As vertex features, a series of new slice features are needed Using the new slice features, a shared attention mechanism is constructed: Calculate the adjacent slice h i 、h j The attention weights are:

[0021] a ij = att(Wh i ,Wh j )

[0022] Among them, a ijis the attention weight of vertex i and vertex j, att() is the attention mechanism, Wh i ,Wh j Represents the matrix multiplication with the feature; the slice h is subjected to the normalized exponential function i 、h j The attention weight a ij To standardize:

[0023]

[0024] Among them, σ refers to the activation function, || is the concatenation operation, and a T represents the transpose of the attention vector a, and exp() is the exponential function e x ,Wh i The randomly generated matrix is ​​multiplied by the feature of the i-th slice, Wh j The randomly generated matrix is ​​multiplied by the feature of the jth slice, Wh k is a randomly generated matrix multiplied by the feature of an adjacent slice of the i-th slice, N i are all adjacent slices of slice i, k is the sequence number of a slice among all adjacent slices of slice i, α ij The slice h is normalized by the normalized exponential function. i and all its adjacent slices N i The attention value, is a collection of slices of dimension F, is a set of slices of D;

[0025] Slice h i The features of can be obtained by summing the features of adjacent slices with corresponding attention weights:

[0026]

[0027] Among them, h′ i h i The slice feature after a graph attention layer, j is the slice number of all adjacent slices of slice i.

[0028] This is a preferred solution for a small sample learning tumor segmentation method driven by a graph attention model described in the present invention, which also includes extending the attention layer to multi-head attention to increase the stability of training:

[0029]

[0030] Where r is the number of times the graph attention layer is reused.

[0031] This is a preferred solution of a small sample learning tumor segmentation method driven by a graph attention model described in the present invention, wherein: it also includes, when extended to a k-shot segmentation task, given K supporting data: The graph attention module will act on each support and query image, and the features obtained are:

[0032]

[0033] This is a preferred solution for the small sample learning tumor segmentation method driven by a graph attention model described in the present invention, wherein: the decoder includes, the decoder imitates the upsampling structure of the U-net network, connects the features passing through the graph attention module with the underlying and high-level query features saved during decoding through jump connections, and then decodes the features to predict the segmentation results.

[0034] The beneficial effects of the present invention are: (1) directly acting on 3D images to reduce overfitting;

[0035] (2) Tumor images can be accurately segmented under small sample conditions, with an average segmentation accuracy of . BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them:

[0037] Figure 1 A flowchart of a small sample learning tumor segmentation method driven by a graph attention model according to the first embodiment of the present invention;

[0038] Figure 2 A data partitioning diagram of a small sample learning tumor segmentation method driven by a graph attention model according to the first embodiment of the present invention;

[0039] Figure 3 A network structure diagram of a small sample learning tumor segmentation method driven by a graph attention model according to the first embodiment of the present invention;

[0040] Figure 4 A network structure diagram of a k-shot segmentation task of a small sample learning tumor segmentation method driven by a graph attention model according to the first embodiment of the present invention;

[0041] Figure 5 The magnetic resonance image and molecular typing report diagram of the small sample learning tumor segmentation method driven by the graph attention model described in the first embodiment of the present invention;

[0042] Figure 6 This is a comparison chart of the segmentation effects of the small sample learning tumor segmentation method driven by the graph attention model described in the second embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0044] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0045] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0046] The present invention is described in detail with reference to schematic diagrams. When describing the embodiments of the present invention, for the sake of convenience, the cross-sectional diagrams showing the device structure will not be partially enlarged according to the general scale, and the schematic diagrams are only examples, which should not limit the scope of protection of the present invention. In addition, in actual production, the three-dimensional dimensions of length, width and depth should be included.

[0047] At the same time, in the description of the present invention, it should be noted that the directions or positional relationships indicated by the terms "upper, lower, inner and outer" are based on the directions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as limiting the present invention. In addition, the terms "first, second or third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0048] In the present invention, unless otherwise clearly specified and limited, the terms "install, connect, connect" should be understood in a broad sense, for example: it can be a fixed connection, a detachable connection or an integral connection; it can also be a mechanical connection, an electrical connection or a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0049] Example 1

[0050] Reference Figures 1 to 6 , which is the first embodiment of the present invention, and provides a small sample learning tumor segmentation method driven by a graph attention model, comprising:

[0051] S1: Reference Figure 2 Divide the data into a support set and a query set, input the support set and the query set into the encoder for encoding, and obtain support features and query features;

[0052] (1) Obtain sagittal MRI images of 3D breast cancer tumors; obtain binary images corresponding to the 3D tumor images; obtain molecular classification reports corresponding to each 3D tumor, including: normal type, luminal type, HER2 overexpression type, and triple-negative breast cancer.

[0053] (2) Classify the tumor data using the molecular typing report; divide each type of tumor data into a training set and a test set, and use the strategy of training three types of tumors to test the remaining type of tumor to train the network; the support set and query set are randomly selected from the other three types of training data.

[0054] (3) After encoding the tth slice of the support and query images, the support features and query features obtained are:

[0055] X q，t =E(I q，t )

[0056] Among them, E is the encoder, K is the K samples in the support set, is the t-th slice image of the k-th sample in the support set, I q,t is a slice of the query image, X s,t To support the feature, X q,t For query features.

[0057] S2: Obtain feature A of the tumor region of interest based on the supporting features, and perform de-pooling on feature A so that the supporting features have the same resolution as the query features;

[0058] (1) Use bilinear interpolation technology to scale the size of the support set segmentation labels;

[0059] (2) Performing Hadamard product between the segmentation labels of the support set and the encoded support features to calculate the features of the tumor region of interest;

[0060] (3) Global average pooling technology is used to pool the features of the tumor region of interest:

[0061]

[0062] in, is the support feature of the t-th slice of the k-th sample after encoding, for The corresponding segmentation label, ⊙ is the Hadamard product, is the supporting feature of the t-th slice of the representative k-th sample after pooling, and GAP() is the global average pooling technique.

[0063] S3: Superimpose the unpooled support features and query features by channel, put the superimposed features into the graph attention module, and learn the spatial information between feature slices through the graph attention module;

[0064] (1) The unpooled support features and query features are superimposed by channel as follows:

[0065]

[0066] in, is the feature of the tumor region of interest of the support set of the kth sample and the tth slice after processing, For channel superposition operation, h t is the slice feature;

[0067] (2) The graph attention module learns the spatial information between feature slices, including treating a slice of a 3D tumor image as a vertex on the graph, and the feature information of the slice As vertex features, a series of new slice features are needed Using the new slice features, we build a shared attention mechanism: Calculate the adjacent slice h i 、h j The attention weights are:

[0068] a ij = att(Wh i ,Wh j )

[0069] Among them, a ij is the attention weight of vertex i and vertex j, att() is the attention mechanism, Wh i ,Wh j Represents matrix multiplication by features;

[0070] Use the normalized exponential function to slice h i 、h j The attention weight a ij To standardize:

[0071]

[0072] Among them, σ refers to the activation function, || is the concatenation operation, and a T represents the transpose of the attention vector a, and exp() is the exponential function e x ,Wh i The randomly generated matrix is ​​multiplied by the feature of the i-th slice, Wh j The randomly generated matrix is ​​multiplied by the feature of the jth slice, Wh k is a randomly generated matrix multiplied by the feature of an adjacent slice of the i-th slice, N i are all adjacent slices of slice i, k is the sequence number of a slice among all adjacent slices of slice i, α ij The slice h is normalized by the normalized exponential function. i and all its adjacent slices N i The attention value, is a collection of slices of dimension F, is a set of slices of D;

[0073] Slice h i The features of can be obtained by summing the features of adjacent slices with corresponding attention weights:

[0074]

[0075] Among them, h′ i h i The slice feature after a graph attention layer, j is the slice number of all adjacent slices of slice i.

[0076] S4: Put the output of the graph attention module into the decoder and combine it with the jump connection to obtain the segmentation result.

[0077] (1) Extend the attention layer to multiple attention levels to increase training stability:

[0078]

[0079] Where r is the number of times the graph attention layer is reused.

[0080] (2) When extended to the k-shot segmentation task, given K supporting data: The graph attention module will act on each support and query image, and the features obtained are:

[0081]

[0082] (3) The decoder imitates the upsampling structure of the U-net network and connects the features passed through the graph attention module with the bottom-level and high-level query features saved during decoding through skip connections, and then decodes the features to predict the segmentation results.

[0083] Example 2

[0084] To verify and illustrate the technical effects adopted in the present method, this embodiment selects a U-shaped network, a prototype alignment network, a two-way gate recurrent unit small sample method and the present method for comparative testing, and compares the test results by means of scientific demonstration to verify the real effect of the present method.

[0085] In order to verify that this method has a high segmentation accuracy compared with the U-net, prototype alignment net, and bidirectional gate recurrent unit small sample method, in this embodiment, the traditional method and this method will be used to perform real-time measurement and comparison on tumor image data.

[0086] Test environment: The hardware environment of both the traditional method and this method is: the computer CPU is i9-9900k, the memory is 64G, the graphics card is Nvidia TITAN, and the video memory is 24G; the software environment is: the program is run in the pytorch environment, and the same number of samples is used for measurement. The U-net cannot recognize small samples and requires a large number of experimental samples. The test results are shown in the following table.

[0087] Table 1: Segmentation accuracy comparison table.

[0088]

[0089]

[0090] From the above table, we can see that for the same data, the accuracy of this method is significantly higher than that of the other three methods. The number of samples required by the U-net is much larger than that of this method, but its segmentation accuracy is much lower than that of this method.

[0091] It should be appreciated that embodiments of the present invention may be implemented or enforced by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable memory. The method may be implemented in a computer program using standard programming techniques - including a non-transitory computer-readable storage medium configured with a computer program, wherein the storage medium so configured causes the computer to operate in a specific and predefined manner - according to the methods and drawings described in the specific embodiments. Each program may be implemented in a high-level procedural or object-oriented programming language to communicate with a computer system. However, if desired, the program may be implemented in an assembly or machine language. In any case, the language may be a compiled or interpreted language. In addition, the program may be run on a programmed ASIC for this purpose.

[0092] Furthermore, the operations of the processes described herein may be performed in any suitable order unless otherwise indicated herein or otherwise clearly contradicted by context. The processes described herein (or variations and / or combinations thereof) may be performed under the control of one or more computer systems configured with executable instructions, and may be implemented as code (e.g., executable instructions, one or more computer programs, or one or more applications) that is executed collectively on one or more processors, by hardware, or a combination thereof. The computer program includes a plurality of instructions that may be executed by one or more processors.

[0093] Further, the method can be implemented in any type of computing platform that is operably connected to a suitable computer, including but not limited to a personal computer, a minicomputer, a mainframe, a workstation, a network or distributed computing environment, a separate or integrated computer platform, or in communication with a charged particle tool or other imaging device, etc. Aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether removable or integrated into a computing platform, such as a hard disk, an optical read and / or write storage medium, a RAM, a ROM, etc., so that it can be read by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the process described herein. In addition, the machine-readable code, or part thereof, can be transmitted via a wired or wireless network. When such media includes instructions or programs that implement the steps described above in conjunction with a microprocessor or other data processor, the invention described herein includes these and other different types of non-transitory computer-readable storage media. When programmed according to the methods and techniques described in the present invention, the present invention also includes the computer itself. The computer program can be applied to input data to perform the functions described herein, thereby converting the input data to generate output data stored in a non-volatile memory. The output information can also be applied to one or more output devices such as a display. In a preferred embodiment of the present invention, the converted data represents a physical and tangible object, including a specific visual depiction of the physical and tangible object produced on a display.

[0094] As used in this application, the terms "component", "module", "system", etc. are intended to refer to a computer-related entity, which can be hardware, firmware, a combination of hardware and software, software, or software in operation. For example, a component can be, but is not limited to: a process running on a processor, a processor, an object, an executable file, a thread in execution, a program, and / or a computer. As an example, an application running on a computing device and the computing device can both be components. One or more components can exist in a process and / or thread in execution, and a component can be located in a computer and / or distributed between two or more computers. In addition, these components can be executed from various computer-readable media having various data structures thereon. These components can communicate in a local and / or remote process manner, such as according to a signal having one or more data packets (e.g., data from a component that interacts with another component in a local system, a distributed system, and / or interacts with other systems in a signal manner through a network such as the Internet).

[0095] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A small sample learning tumor segmentation method driven by a graph attention model, characterized by: include, Divide the data into a support set and a query set, input the support set and the query set into the encoder for encoding, and obtain support features and query features; Acquire a feature A of the tumor region of interest according to the supporting feature, and perform de-pooling on the feature A so that the supporting feature has the same resolution as the query feature; The unpooled support features and query features are superimposed by channel, and the superimposed features are put into a graph attention module, and the spatial information between feature slices is learned through the graph attention module; The output of the graph attention module is put into the decoder, and the segmentation result is obtained by combining the jump connection method; The graph attention module learns the spatial information between feature slices including, A slice of a 3D tumor image is considered as a vertex on the graph, and the features of the slice As vertex features, a series of new slice features are needed Using the new slice features, a shared attention mechanism is constructed: Calculate the adjacent slice h i 、h j The attention weights are: a ij =att(Wh i ,Wh j ) Among them, a ij is the attention weight of vertex i and vertex j, att() is the attention mechanism, Wh i ,Wh j Represents matrix multiplication by features; The slice h is normalized using the exponential function i 、h j The attention weight a ij To standardize: Among them, σ refers to the activation function, || is the concatenation operation, and a T represents the transpose of the attention vector a, and exp() is the exponential function e x ,Wh i The randomly generated matrix is ​​multiplied by the feature of the i-th slice, Wh j The randomly generated matrix is ​​multiplied by the feature of the jth slice, Wh k is a randomly generated matrix multiplied by the feature of an adjacent slice of the i-th slice, N i are all adjacent slices of slice i, k is the slice number of all adjacent slices of slice i, α ij The slice h is normalized by the normalized exponential function. i and all its adjacent slices N i The attention value, is a collection of slices of dimension F, is a set of slices of D; Slice h i The features of can be obtained by summing the features of adjacent slices with corresponding attention weights: Among them, h' i h i The slice feature after a graph attention layer, j is the slice number of all adjacent slices of slice i.

2. The small sample learning tumor segmentation method driven by the graph attention model as claimed in claim 1, characterized in that: The data include, Obtain sagittal MRI images of 3D breast cancer tumors; Obtain a binary image corresponding to the 3D tumor image; Obtain molecular typing reports for each 3D tumor, including normal type, luminal type, HER2 overexpression type and triple-negative breast cancer.

3. The small sample learning tumor segmentation method driven by the graph attention model as claimed in claim 2, characterized in that: The data is divided into a support set and a query set including: Classifying tumor data using the molecular typing report; Divide each type of tumor data into training sets and test sets, and train the network using the strategy of training three types of tumors to test the remaining type of tumor; The support set and query set are randomly selected from the other three types of training data.

4. The small sample learning tumor segmentation method driven by the graph attention model as claimed in claim 3, characterized in that: The obtaining of supporting features and query features includes: The support features and query features obtained after encoding the tth slice of the support and query images are: X q，t =E(I q，t ) Among them, E is the encoder, K is the K samples in the support set, is the t-th slice image of the k-th sample in the support set, I q,t is a slice of the query image, X s,t To support the feature, X q,t For query features.

5. The small sample learning tumor segmentation method driven by the graph attention model as claimed in claim 4, characterized in that: Obtaining a feature A of the tumor region of interest from the supporting features, and performing reverse pooling on the feature A includes: Bilinear interpolation technique is used to scale the size of support set segmentation labels; Perform Hadamard product between the segmentation labels of the support set and the encoded support features to calculate the features of the tumor region of interest; The global average pooling technique is used to pool the features of the tumor region of interest: in, is the support feature of the t-th slice of the k-th sample after encoding, for The corresponding segmentation label, ⊙ is the Hadamard product, is the supporting feature of the t-th slice of the representative k-th sample after pooling, and GAP() is the global average pooling technique.

6. The small sample learning tumor segmentation method driven by the graph attention model as claimed in claim 5, characterized in that: The stacking of the unpooled support features and query features by channel includes: in, is the feature of the tumor region of interest of the support set of the kth sample and the tth slice after processing, ⊕ is the channel superposition operation, and h t Slice feature.

7. The small sample learning tumor segmentation method driven by the graph attention model as claimed in claim 6, characterized in that: Also includes, Extend the attention layer to multiple attention heads to increase the stability of training: Where r is the number of times the graph attention layer is reused.

8. The small sample learning tumor segmentation method driven by the graph attention model as claimed in claim 7, characterized in that: Also includes, When extended to the k-shot segmentation task, given K support data: The graph attention module will act on each support and query image, and the features obtained are:

9. The small sample learning tumor segmentation method driven by the graph attention model as claimed in claim 8, characterized in that: The decoder comprises, The decoder imitates the upsampling structure of the U-net network, and connects the features passed through the graph attention module with the bottom-level and high-level query features saved during decoding through skip connections, and then decodes the features to predict the segmentation results.

Images