An interactive 3D medical image segmentation method based on bidirectional gated memory network

Abstract

The present invention discloses an interactive three-dimensional medical image segmentation method based on a bidirectional gated memory network, comprising: using an interactive network with internal and external guidance to process expert-labeled slices S i The current query slice is matched against the memory slices in the memory storage using a memory network that follows a key-value embedding mechanism. A bidirectional gated recurrent unit (Bi-GRU) is added to the memory network to obtain contextual information about the query slice. Atrous spatial convolutional pooling (ASPP) is used to expand the receptive field and achieve multi-scale feature integration of the query slice. This invention bidirectionally transfers user-provided heuristic information to other slices, enabling the network to obtain richer feature information and thus improving lesion segmentation accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image segmentation, and in particular to an interactive three-dimensional medical image segmentation method based on a bidirectional gated memory network. Background Art

[0002] In recent years, deep learning techniques have achieved state-of-the-art results in the field of medical image semantic segmentation. For example, the traditional UNet framework uses a concatenation operation to fuse feature maps. It was proposed to address the poor robustness of medical image segmentation tasks. Its U-shaped network structure alleviates the inability of fully convolutional neural networks to capture contextual information and positional relationships. The UNet encoder extracts features by progressively downsampling to learn a global contextual representation, while the decoder progressively upsamples the extracted features to the input resolution for feature fusion. While various network models based on the UNet framework have been widely used in medical image processing, their performance in learning long-term dependencies is limited by their localized receptive fields. Consequently, these models have limitations in capturing multi-scale information, resulting in low segmentation accuracy for irregularly shaped pathological regions. While recent studies have shown that these techniques have demonstrated excellent performance on various segmentation tasks, their performance in clinical applications remains limited due to inherent limitations of medical images, such as low contrast, varying imaging and segmentation rules, and inter-patient variability. The main reasons are as follows: when the distribution of training data and test data is inconsistent, the deep learning network model is more sensitive to its distribution deviation, and the automatic segmentation network based on deep learning has poor generalization ability for unknown new samples, which makes it unsuitable for special groups; the manual segmentation method requires labeling each target object, and this process requires a lot of manual operation. If the target area is not labeled, the network cannot proceed to the next step of segmentation, so it is not suitable for large and complex medical image datasets.

[0003] At present, some studies use interactive segmentation technology to solve the above problems, adding user annotations during the training process of the automatic segmentation network to make the segmentation results more accurate and scientific. Many traditional interactive methods, such as Graph Cuts and GeoS, use low-quality features (such as grayscale or color distribution) to perform image segmentation. Since low-quality features cannot effectively segment the target area in many low-contrast situations, and these methods often require a lot of user interaction to obtain reliable results, they are not applicable to complex medical image segmentation tasks. In existing three-dimensional medical image segmentation research, interactive segmentation methods combined with deep learning can effectively solve the shortcomings of traditional segmentation methods in processing medical images. Deep learning technology is used to reduce user interaction, allowing the network to mine richer lesion feature information and better handle complex targets, weak edges and other problems. It not only improves the efficiency of model segmentation, but also requires only a small amount of user intervention to obtain accurate segmentation results.

[0004] Chinese patent document CN101819679B discloses a 3D medical image segmentation method that utilizes 3D image reconstruction and Gaussian filtering to smooth gradient information. Although this method can generate 3D images of multiple regions of interest, its processing capabilities for segmenting small targets and specific lesions are limited.

[0005] Chinese patent document CN111798458A discloses an interactive medical image segmentation method based on uncertainty guidance, which uses a grouped convolutional neural network to perform initial segmentation and uncertainty estimation on three-dimensional medical images. Although it improves segmentation accuracy, since the network uses high-confidence specific areas to continuously supplement the currently labeled samples to reduce incorrect segmentation caused by low-confidence uncertain areas, a large amount of calculation is required to represent low-level features as high-level features, which makes it unsuitable for applications with limited computing resources.

[0006] "DeepIGeoS: A Deep Interactive Geodesic Framework for Medical Image Segmentation" proposes combining user interaction with convolutional neural networks through geodesic distance transforms. However, DeepIGeoS can only process known objects in the training set and lacks adaptability to unknown objects.

[0007] "MIDeepSeg: Minimally Interactive Segmentation of Unseen Objects from Medical Images Using Deep Learning" This annotation algorithm has superior performance in a large number of new and unseen organs and image modalities, but it does not consider the long-term dependencies and remote spatial context information of three-dimensional medical images, and has inaccurate segmentation problems for narrow and long targets and irregular lesions, and the network is limited in recognizing user annotation types. Summary of the Invention

[0008] The purpose of the present invention is to provide an interactive three-dimensional medical image segmentation method based on a bidirectional gated memory network (Bi-MEM), which can identify difficult cases by using heuristic interactive information and achieve more accurate segmentation.

[0009] To achieve the above objectives, this application proposes an interactive three-dimensional medical image segmentation method based on a bidirectional gated memory network, comprising:

[0010] An interactive network of internal and external guidance (IOG) is used to process expert-labeled slices S i ;

[0011] The current query slice is feature matched with the memory slice in the memory storage through the memory network following the key-value embedding mechanism;

[0012] Add a bidirectional gated recurrent unit (Bi-GRU) to the memory network to obtain contextual information of the query slice;

[0013] The atrous spatial convolutional pooling pyramid (ASPP) is used to expand the receptive field and achieve multi-scale feature integration of query slices.

[0014] Furthermore, an interactive network of internal and external guidance (IOG) is used to process the expert-labeled slices S i , specifically: the expert selects a slice S i The lesion area is marked and the slice S is annotated using an interactive network guided by internal and external i The lesion area is initially segmented to obtain a two-dimensional segmentation mask M i , where i represents the i-th slice of the entire group of cases.

[0015] Furthermore, the internal and external guided interaction network models the segmentation task as learning a bidirectional sequential image patch representation; the expert clicks on the slice S iThe rough center point of the target lesion is used to preliminarily determine the approximate location of the target lesion. Then multiple rays are extended from the inside to the outside of the lesion area, and the feature representation of these sequential image blocks is learned to capture the shape and appearance changes from the target to the background and from the background to the target.

[0016] Furthermore, the features are retrieved and matched by the memory network following the key-value embedding mechanism. Specifically, the slice that needs to be segmented in the input case is used as the query slice S q , the query slice S q and memory slice S m Encoded into a key-value pair mapping through a dedicated encoder, that is, the query encoder only accepts query slices S q As input, the memory encoder also accepts the memory slice S m and its segmentation mask M m as input.

[0017] Furthermore, the memory network is sliced ​​by querying S q Key mapping K q and memory slice S in memory storage m Key mapping K m The similarity between them is then used to perform feature matching, and then the relative matching score is used to compare the memory slice S in the memory storage m Memory value mapping V m , then weighted and retrieve memory value map V m , mapping it to the query value V q Combined output; if there are multiple memory slices S m , then embed them into key-value mappings respectively, and then different memory slices S m The key and value mappings are stacked to output a set of three-dimensional key-value mappings.

[0018] Furthermore, in the gated recurrent unit, the input of each unit includes the feature information x of the slice t , the hidden state information h of the previous unit t-1 , the newly generated hidden information h t And the candidate state h', each gated recurrent unit GRU includes an update gate z and a reset gate r.

[0019] Furthermore, in a gated recurrent unit, the state information of the previous moment is input into the state of the current moment. The update gate determines the degree of information transmission. The closer the gate signal is to 1, the more information is selected; and the closer the gate signal is to 0, the more information is forgotten. The formula is as follows:

[0020] z t =σ(x t w z +ht-1 w z +b z ) (1)

[0021] The reset gate controls the hidden state information h passed down from the previous node t-1 The impact on the candidate memory h', that is, the reset gate determines the information to be forgotten; if the reset signal is close to 0, the gated recurrent unit forgets the previously acquired hidden state information h t-1 ; Its expression is as follows:

[0022] r t =σ(x t w r +h t-1 w r +b r ) (2)

[0023] Next, “selectively” remember the current state and bring h’ into the hidden state at the current moment. The candidate memory h’ contains the data x t , which is expressed as follows:

[0024] h'=tanh(x t w h' +r t ⊙h t-1 w h' +b h' ) (3)

[0025] In the memory update phase, the tanh activation function is used for weighted assignment to determine the importance of the passed value, ranging from -1 to 1; the gated recurrent unit GRU performs both "selection" and "forgetting" at this step, and its expression is as follows:

[0026] h t =(1-z t )⊙h t-1 +z t ⊙h' (4)

[0027] Among them, ⊙ represents the element-wise multiplication operation, σ represents the sigmoid activation function, w represents the learned transformation, and b is the bias parameter.

[0028] Furthermore, the bidirectional gated recurrent unit Bi-GRU converts 3D medical image slices into sequence representations, where each query slice S q The forward information and backward information are respectively passed to the two hidden layers connected to the same output layer; therefore, the state information of the two gated recurrent units GRU are fused to obtain the output result, which is expressed as follows:

[0029]

[0030] in, are the outputs of forward propagation and backward propagation respectively, are the weights for forward and backward propagation, respectively.

[0031] Furthermore, the atrous spatial convolution pooling pyramid ASPP includes a 1×1 convolution layer, a pooling pyramid and a pooling module. The pooling pyramid includes a 3×3 convolution layer with expansion rates of 2, 4, and 8, respectively. The input query slice feature map is sampled in parallel by atrous convolutions at different sampling rates to obtain a reconstructed multi-scale feature map. The feature map is decoded using a decoder to obtain a lesion segmentation result containing contextual spatial information.

[0032] Compared with the existing technology, the above technical solution adopted by the present invention has the following advantages: 1. It uses the knowledge and experience of experts to mark lesions, allowing the entire network to learn richer lesion feature information and make real-time and rapid adjustments to improve the prediction and segmentation accuracy.

[0033] 2. The slices are processed using an interactive network with internal and external guidance to improve the efficiency of human-computer interaction and solve the problems of time-consuming and high cost of annotating 3D medical image data.

[0034] 3. A memory network based on bidirectional gated recurrent units can effectively capture subtle feature changes between forward and backward slices based on the slice information in the memory, and interactively distinguish the foreground and background. The "gate" is a learnable parameter, making the proposed mechanism applicable to datasets of any size and working well even on smaller medical image datasets.

[0035] 4. The dilated spatial convolutional pooling pyramid uses three parallel dilated convolutional layers to expand the receptive field without adding new parameters or more computational effort. It achieves the acquisition and integration of feature maps of different scales and improves the robustness of complex lesion area segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flowchart of an interactive three-dimensional medical image segmentation method;

[0037] Figure 2 Graph showing the key-value feature matching process in Bi-MEM;

[0038] Figure 3 It is the structure diagram of the bidirectional gated recurrent unit Bi-GRU;

[0039] Figure 4 Flowchart for expanding the receptive field of Atrous Spatial Convolution Pooling Pyramid ASPP. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solutions and advantages of this application more clearly understood, this application is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application. That is, the embodiments described are only part of the embodiments of this application, not all of them.

[0041] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application for protection, but merely represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present application.

[0042] At present, medical image data collection and annotation are very difficult, and building a sufficiently large data set is challenging. For training complex medical images, deep learning methods use segmentation mask annotation data, and this process requires a large amount of computing resources. In addition, due to the inherent limitations of medical images (such as low contrast, uneven lesion edges, differences between patients, etc.) and their complex pathological areas (such as occluded areas, fuzzy boundary areas, narrow areas, etc.), fully automatic segmentation networks often show problems such as insufficient feature learning ability and lack of clinically acceptable segmentation accuracy and robustness. Recent studies have shown that adding a sufficient number of high-quality manual annotations to the network can improve the above problems. However, while increasing prior knowledge, more manual interaction is also required, which increases labor costs and reduces network interaction efficiency. Therefore, the present invention combines interactive segmentation technology based on internal and external guidance (IOG) with memory networks, and uses the knowledge and experience of experts to provide a small amount of manual annotations for difficult lesion areas, so that the network can learn more unrecognized lesion feature information, thereby achieving the purpose of solving the above problems. At the same time, a bidirectional gated recurrent unit (Bi-GRU) is added to the memory network. The bidirectional design structure can not only utilize information in the forward sequence, but also capture reverse information. The memory network transmits the inspiration information provided by the user to other slices in a bidirectional manner, obtaining richer feature information, thereby improving the accuracy of lesion segmentation.

[0043] like Figure 1As shown, an interactive three-dimensional medical image segmentation method based on a bidirectional gated memory network, first, the expert selects a slice from the entire group of cases based on diagnostic experience to mark its difficult lesion area, that is, the interactive slice is initially segmented through the IOG interactive network. Subsequently, the initial segmentation result and the entire group of cases are input into Bi-MEM for feature extraction. Bi-MEM's external memory storage stores memory slices with target lesion segmentation masks, which can provide different types of lesion feature information. Bi-MEM encodes features according to the key-value embedding mechanism. The memory network learns the important features of the target lesion from the interactive slice, and combines the memory slice in the memory storage and its corresponding two-dimensional segmentation mask to improve the segmentation result of the query slice, and further expands to other slices to complete the segmentation of the target lesion. Finally, Bi-MEM outputs the segmentation results of all slices of the entire group of cases. Therefore, the present invention can identify difficult cases and achieve more accurate segmentation by using heuristic interactive information. The specific implementation process is as follows:

[0044] S1. Using the interactive network of internal and external guidance IOG to process expert-annotated slices S i ;

[0045] Specifically, the expert selects a slice S i The lesion area is marked and the interactive slice S is analyzed using the IOG interaction network. i The lesion area is initially segmented to obtain a rough two-dimensional segmentation image M i , where i represents the i-th slice of the entire case (a set of 3D medical images containing k slices). The IOG network models the segmentation task as learning a bidirectional representation of sequential image patches. Experts click on interactive slices S i The rough center point of the target lesion is used to preliminarily determine the approximate location of the target lesion. Multiple rays extend from the inside of the lesion area to the outside, and the feature representation of these sequential image blocks is learned to capture the shape and appearance changes from the target to the background and from the background to the target to obtain accurate segmentation results. This interactive network improves the segmentation performance while reducing the segmentation time. i The entire set of cases was entered into Bi-MEM.

[0046] S2. Feature matching between the current query slice and the memory slice in the memory storage through the memory network following the key-value embedding mechanism;

[0047] Specifically, the slice that needs to be segmented in the input case is used as the query slice S q , query slice S q and memory slice S m Encoded into a key-value pair mapping through a dedicated encoder, that is, the query encoder only accepts query slices S qAs input, the memory encoder also accepts the memory slice S m and its segmentation mask M m As input. In the memory read operation, Bi-MEM queries the slice S q Key mapping K q and memory slice S in memory storage m Key mapping K m The similarity between them is then used to perform feature matching, and then the relative matching score is used to compare the memory slice S in the memory storage m Memory value mapping V m , then weighted and retrieve memory value map V m , mapping it to the query value V q Combined output; if there are multiple memory slices S m , they are embedded in the key-value mapping respectively, and then the key and value mappings of different memory slices are stacked to output a set of three-dimensional key-value mappings; the Bi-MEM key-value feature matching process is shown in the figure Figure 2 shown.

[0048] S3. Add a bidirectional gated recurrent unit (Bi-GRU) to the memory network to obtain contextual information of the query slice.

[0049] Specifically, the ordinary long short-term memory network LSTM requires multiple gate interactions to "select" the retained memory. However, the gated recurrent unit GRU is a recurrent neural network with an optimized and improved structure based on the long short-term memory network LSTM. It regularizes the memory through only one gate and can "select" and "forget" at the same time, thereby dynamically modifying the stored memory and effectively alleviating the problems of gradient disappearance and gradient explosion during training. The input of each unit includes the feature information x of the slice t , the hidden state information h of the previous unit t-1 , the newly generated hidden information h t And the candidate state h', each gated recurrent unit GRU includes an update gate z and a reset gate r.

[0050] The state information of the previous moment is input into the state of the current moment. The update gate determines the degree of information transmission. The closer the gate signal is to 1, the more information is selected; and the closer the gate signal is to 0, the more information is forgotten. The formula is as follows:

[0051] z t =σ(x t w z +h t-1 w z +b z ) (1)

[0052] The reset gate controls the hidden state information h passed down from the previous node t-1 The impact on the candidate memory h', that is, the reset gate determines the information to be forgotten; if the reset signal is close to 0, the gated recurrent unit forgets the previously acquired hidden state information h t-1 ; Its expression is as follows:

[0053] r t =σ(x t w r +h t-1 w r +b r ) (2)

[0054] Next, “selectively” remember the current state and bring h’ into the hidden state at the current moment. The candidate memory h’ contains the data x t , which is expressed as follows:

[0055] h'=tanh(x t w h' +r t ⊙h t-1 w h' +b h' ) (3)

[0056] In the memory update phase, the tanh activation function is used for weighted assignment to determine the importance of the passed value, ranging from -1 to 1; the gated recurrent unit GRU performs both "selection" and "forgetting" at this step, and its expression is as follows:

[0057] h t =(1-z t )⊙h t-1 +z t ⊙h' (4)

[0058] Among them, ⊙ represents the element-wise multiplication operation, σ represents the sigmoid activation function, w represents the learned transformation, and b is the bias parameter.

[0059] The bidirectional gated recurrent unit Bi-GRU converts 3D medical image slices into sequence representations, where each query slice S q The forward information and backward information of are passed to the two hidden layers connected to the same output layer. Thus, the output layer has the query slice S q Complete contextual information ensures that when multiple lesion areas are present in a group of cases, other lesions are not mistakenly identified. The gated design facilitates dynamic modification of stored memory, and the bidirectional structure effectively captures long-term dependencies between 3D medical image slice sequences, reducing the weak edge discrimination of multiple regions with certain connectivity, which can lead to over-segmentation or under-segmentation.

[0060] The unidirectional GRU state always outputs a unidirectional propagation from the front to the back, which can only extract the forward information of the current state and cannot use the backward sequence information. The bidirectional gated recurrent unit Bi-GRU allows the network to obtain information from both forward and backward propagation, fully extracting the relevant features of the lesion area. The state information of the two GRUs is fused to obtain the output result, which is expressed as follows:

[0061]

[0062] in, are the outputs of forward propagation and backward propagation respectively, are the weights of forward and backward propagation respectively. By fusing the feature information of forward and backward, the query slice S q The spatial context information of the target lesion is utilized to the greatest extent, so the network can extract the important features of the target lesion. Figure 3 shown.

[0063] S4. Use the Atrous Spatial Convolution Pooling Pyramid (ASPP) to expand the receptive field and achieve multi-scale feature integration of query slices;

[0064] Specifically, the general pooling pyramid structure is not conducive to the increase of model depth and will reduce computing performance when the target object is very small. The present invention uses ASPP to improve this problem. The void space convolution pooling pyramid ASPP includes a 1×1 convolution layer, a pooling pyramid and a pooling module. The pooling pyramid includes a 3×3 convolution layer with expansion rates of 2, 4, and 8 respectively. The main function of its void design is to expand the receptive field, and to improve the network's segmentation accuracy for lesions with obvious scale changes (for example, lesions with irregular shapes and low contrast with surrounding normal tissues) without losing resolution. After the memory reading operation, ASPP samples the input query slice feature map through void convolutions of different sampling rates in parallel, and sets different sampling rates to change the size of the field of view, thereby realizing free multi-scale feature information integration. Finally, the feature map is decoded by a decoder to obtain the output of the memory reading operation, and a lesion segmentation result containing contextual spatial information is obtained. The process diagram of ASPP integrating multi-scale features is shown in FIG. Figure 4 shown.

[0065] The foregoing descriptions of specific exemplary embodiments of the present invention are for purposes of illustration and description. These descriptions are not intended to limit the invention to the precise forms disclosed, and it is apparent that many variations and modifications are possible in light of the foregoing teachings. The exemplary embodiments have been selected and described for the purpose of explaining the specific principles of the invention and their practical application, thereby enabling those skilled in the art to realize and utilize a variety of exemplary embodiments of the invention and various options and modifications. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. An interactive 3D medical image segmentation method based on a bidirectional gated memory network, characterized in that: include: An interactive network with internal and external guidance is used to process expert-labeled slices S i ; The current query slice is feature matched with the memory slice in the memory storage through the memory network following the key-value embedding mechanism; Add a bidirectional gated recurrent unit (Bi-GRU) to the memory network to obtain contextual information of the query slice; Use the atrous spatial convolutional pooling pyramid (ASPP) to expand the receptive field and achieve multi-scale feature integration of query slices; An interactive network with internal and external guidance is used to process expert-labeled slices S i , specifically: the expert selects a slice S i The lesion area is marked and the slice S is annotated using an interactive network guided by internal and external i The lesion area is initially segmented to obtain a two-dimensional segmentation mask M i , where i represents the i-th slice of the entire group of cases; The internal-external guided interaction network models the segmentation task as learning a bidirectional representation of sequential image patches; the expert clicks on the slice S i The rough center point of the target lesion is used to preliminarily determine the approximate location of the target lesion. Then, multiple rays are extended from the inside of the lesion area to the outside, and the feature representation of these sequential image blocks is learned to capture the shape and appearance changes from the target to the background and from the background to the target. The features are retrieved and matched through the memory network following the key-value embedding mechanism. Specifically, the slice that needs to be segmented in the input case is used as the query slice S q , the query slice S q and memory slice S m Encoded into a key-value pair mapping through a dedicated encoder, that is, the query encoder only accepts query slices S q As input, the memory encoder also accepts the memory slice S m and its segmentation mask M m As input; The memory network is passed through the query slice S q Key mapping K q and memory slice S in memory storage m Key mapping K m The similarity between them is then used to perform feature matching, and then the relative matching score is used to compare the memory slice S in the memory storage m Memory value mapping V m , then weighted and retrieve memory value map V m , mapping it to the query value V q Combined output; if there are multiple memory slices S m , then embed them into key-value mappings respectively, and then different memory slices S m The key and value mappings are stacked to output a set of three-dimensional key-value mappings.

2. The interactive three-dimensional medical image segmentation method based on a bidirectional gated memory network according to claim 1, characterized in that: In the gated recurrent unit, the input of each unit includes the feature information x of the slice t , the hidden state information h of the previous unit t-1 , the newly generated hidden information h t And the candidate state h', each gated recurrent unit GRU includes an update gate z and a reset gate r.

3. The interactive three-dimensional medical image segmentation method based on a bidirectional gated memory network according to claim 2, characterized in that: In a gated recurrent unit, the state information of the previous moment is input into the state of the current moment. The update gate determines the degree of information transmission. The closer the gate signal is to 1, the more information is selected; and the closer the gate signal is to 0, the more information is forgotten. The formula is as follows: z t =σ(x t ·w z +h t-1 ·w z +b z ) (1) The reset gate controls the hidden state information h passed down from the previous node t-1 The impact on the candidate memory h', that is, the reset gate determines the information to be forgotten; if the reset signal is close to 0, the gated recurrent unit forgets the previously acquired hidden state information h t-1 ; Its expression is as follows: r t =σ(x t ·w r +h t-1 ·w r +b r ) (2) Next, "selectively" remember the current state and bring h' into the hidden state at the current moment. The candidate memory h' contains the data x t , which is expressed as follows: h'=tanh(x t ·w h' +r t ⊙h t-1 ·w h' +b h' ) (3) In the memory update phase, the tanh activation function is used for weighted assignment to determine the importance of the passed value, ranging from -1 to 1; the gated recurrent unit GRU performs both "selection" and "forgetting" at this step, and its expression is as follows: h t =(1-z t )⊙h t-1 +z t ⊙h' (4) Among them, ⊙ represents the element-wise multiplication operation, σ represents the sigmoid activation function, w represents the learned transformation, and b is the bias parameter.

4. The interactive three-dimensional medical image segmentation method based on a bidirectional gated memory network according to claim 3, characterized in that: The bidirectional gated recurrent unit Bi-GRU converts 3D medical image slices into sequence representations, where each query slice S q The forward information and backward information are respectively passed to the two hidden layers connected to the same output layer; therefore, the state information of the two gated recurrent units GRU are fused to obtain the output result, which is expressed as follows: in, are the outputs of forward propagation and backward propagation respectively, are the weights for forward and backward propagation, respectively.

5. The interactive three-dimensional medical image segmentation method based on a bidirectional gated memory network according to claim 1, characterized in that: The atrous spatial convolution pooling pyramid (ASPP) includes a 1×1 convolution layer, a pooling pyramid, and a pooling module. The pooling pyramid includes a 3×3 convolution layer with expansion rates of 2, 4, and 8, respectively. The input query slice feature map is sampled in parallel through atrous convolutions at different sampling rates to obtain a reconstructed multi-scale feature map. The feature map is decoded using a decoder to obtain a lesion segmentation result containing contextual spatial information.

Images