An Automated Cervical Cancer Target Region Segmentation Method Based on Dual Attention Mechanism

Abstract

This invention discloses an automatic cervical cancer target region segmentation method based on a dual attention mechanism, consisting of two parts: a spatial constraint network and a width-driven attention network. The spatial constraint network includes a CTV segmentation network and an OAR segmentation network. The CTV and OAR segmentation networks extract features related to CTV and OAR segmentation from cervical cancer CT scans. A cross-attention mechanism is used to learn the attention weights in cervical cancer CTV segmentation, and then CTV features are used as values ​​V for weighted aggregation to learn the spatial constraints between OAR and CTV. After restoring the shape of the obtained features, they are concatenated with the output feature map of the CTV segmentation network to obtain the CTV segmentation result. A width-driven attention network is constructed and inserted into the decoder stage of the CTV segmentation network to fine-tune the CTV segmentation result. This invention integrates the relative positional relationship between OAR and CTV into the neural network through network structure design, providing prior knowledge to the neural network and improving the accuracy of cervical cancer CTV segmentation.

Description

Technical Field

[0001] This invention relates to the field of cancer target region segmentation technology, and in particular to an automatic segmentation method for cervical cancer target regions based on a dual attention mechanism. Background Technology

[0002] Cervical cancer is one of the most common malignant tumors in women, and radiotherapy plays a crucial role in its treatment. For early-stage cervical cancer, radiotherapy is usually administered as adjuvant therapy after surgery. For locally advanced or metastatic cervical cancer, external beam radiation therapy (EBRT) and brachytherapy (BT) are also key treatment modalities. Intensity-modulated radiation therapy (IMRT) is the preferred radiotherapy method for treating locally advanced cervical cancer because it provides a highly precise therapeutic dose to the tumor and reduces the dose to organs at risk. Accurate delineation of the clinical target volume (CTV) and organs at risk (OAR) is central to the radiotherapy planning process. The late toxicity rate associated with pelvic chemoradiotherapy for cervical cancer is approximately 6%–23%. Accurate CTV segmentation is particularly important in reducing late toxicities associated with pelvic chemoradiotherapy, as late toxicities such as urinary incontinence, fistulas, and malabsorption can persist for many years, causing significant harm, especially to younger patients.

[0003] Typically, this delineation is a laborious manual process performed by radiation oncologists based on gynecological examinations, surgical reports, and assessments from computed tomography (CT), magnetic resonance imaging (MRI), and other imaging modalities. This delineation is extremely time-consuming, and the definition of the target area depends on the physician's understanding of clinical guidelines, consensus, and experience. Despite the existence of standard guidelines, inter-observer and intra-observer variability persists in the quality, efficiency, and reproducibility of segmentation, remaining one of the major challenges in radiotherapy planning. Therefore, computational tools capable of automatically and accurately segmenting anatomical structures within a reasonable timeframe can significantly reduce the manual work of radiation oncologists. Computer-aided automated image segmentation can alleviate physician workload while improving the accuracy, consistency, and reproducibility of target area delineation.

[0004] CTV segmentation of cervical cancer on CT and MRI is a challenging task due to soft tissue deformation, variations in shape and appearance between patients, and anatomical changes during treatment. Traditional automated segmentation methods include atlas-based auto-segmentation (ABAS) algorithms. ABAS algorithms primarily use deformable image registration to generate contours, which involves a transformation between two images where voxels in the moving image set are deformed to match voxels in the target image set, creating a deformable vector field during this process. In ABAS, the segmentation structure from the atlas library is propagated to the target volume image using a deformable image registration algorithm. Because multiple ABAS uses a voting scheme to determine whether voxels are inside or outside the structure, it is more susceptible to topological artifacts compared to single ABAS. However, multiple ABAS can overcome the problems encountered by single ABAS, such as significant volume and positional differences between the atlas library and the target data. Although not advantageous in segmentation time, atlas segmentation combined with manual correction yields better segmentation results than purely atlas-based segmentation methods, and both methods offer significant efficiency improvements over purely manual segmentation.

[0005] In recent years, ABAS-based software has been frequently used for automatic segmentation of CTV and OAR in radiotherapy. However, ABAS has its limitations. First, constructing a "universal atlas" is difficult due to differences in patient age and organ shape and size. Second, image registration is an extremely time-consuming process due to its variability. ABAS cannot handle significant differences between different images, such as those caused by organ motion or deformation. Third, for organs and tumors lacking clear boundaries or exhibiting complex shapes, especially target areas like CTV with unclear boundaries and irregular shapes, the segmentation results of ABAS methods are often unsatisfactory. Furthermore, image processing for atlas-based automatic segmentation is time-consuming. ABAS methods limit the final segmentation results to anatomically reasonable but less flexible shapes or locations described by the model, making it difficult to meet the requirements for target area delineation in cervical cancer radiotherapy.

[0006] Accurate delineation of the clinical target volume for radiotherapy is of great practical significance for improving physician efficiency and reducing harm to patients. However, to date, automatic CTV segmentation methods for cervical cancer based on Convolutional Neural Networks (CNNs) have mainly focused on U-Net and its variants. The general process for CTV delineation is as follows: guided by clinical information, cancer pathology, and diagnostic images, the visible cancer volume is delineated as the Gross Tumor Volume (GTV); then, clinicians, considering the possibility of cancer spread, expand the volume by adding edges, thus defining the CTV. This results in CTV delineation differing from the usual clear OAR delineation, with the difficulty lying in the unclear boundaries of the CTV and its significant dependence on surrounding organs. Existing technologies do not consider the special characteristics of CTV delineation and its dependence on surrounding organs, directly applying deep learning segmentation algorithms to the CTV task, which is extremely unreasonable and leads to low accuracy in the segmentation models. Summary of the Invention

[0007] To address the aforementioned issues, this invention provides an automatic cervical cancer target region segmentation method based on a dual attention mechanism. This method integrates the relative positional relationship between the OAR and CTV into a neural network, providing prior knowledge to the neural network and improving the accuracy of cervical cancer CTV segmentation.

[0008] Therefore, the technical solution of the present invention is: an automatic segmentation method for cervical cancer target area based on a dual attention mechanism, which is a two-stage training method, including a spatially constrained network and a width-driven attention network finely tuned based on this.

[0009] Specifically, the following steps are included:

[0010] 1) Construct a spatial constraint network, which consists of two convolutional neural networks with the same architecture: a CTV segmentation network and an OAR segmentation network.

[0011] 2) The CTV segmentation network and the OAR segmentation network learn to delineate CTV and OAR from cervical cancer CT scans, respectively, and extract features related to CTV segmentation and OAR segmentation.

[0012] 3) Using CTV features as the key K and OAR features as the query Q, a cross-attention mechanism is used to learn the attention weights in cervical cancer CTV segmentation. Then, CTV features are used as the value V for weighted aggregation to learn the spatial constraints between OAR and CTV.

[0013] 4) After restoring the shape of the features obtained in step 3), concatenate them with the output feature map of the CTV segmentation network, and then pass them through a convolutional layer to obtain the CTV segmentation result;

[0014] 5) Construct a width-driven attention network and insert it into the decoder stage of the CTV segmentation network to fine-tune the CTV segmentation network.

[0015] Based on the above scheme and as a preferred embodiment of the above scheme: both the CTV segmentation network and the OAR segmentation network are convolutional neural networks that include encoder and decoder structures. The CTV segmentation network and the OAR segmentation network share a single encoder, while the decoders operate independently.

[0016] Based on the above scheme and as a preferred embodiment of the above scheme: the spatial constraint network is a 3D-Unet network, including a CTV segmentation network and an OAR segmentation network;

[0017] M = F E (I)

[0018] X c =F D-CTV-2 (M)

[0019] X O =F D-OAR-2 (M)

[0020] Where I represents the input image, F E M represents the encoder of the network, and F represents the intermediate layer features obtained after the encoder. D-CTV-2 The first two modules of the CTV decoder, representing the network, consist of two consecutive convolutional blocks and an upsampling operation; X C This represents the CTV feature map obtained after passing through this module; F D-OAR-2 The first two modules of the OAR decoder represent the network; X O This represents the OAR feature map obtained after passing through this module.

[0021] Based on the above scheme and as a preferred scheme: in step 3), the output feature maps of the CTV segmentation network and the OAR segmentation network are expanded into sequence vectors and position encoding is added; the key K and value V corresponding to the CTV feature and the query Q corresponding to the OAR feature are obtained; the positional relationship between CTV and OAR in the features is learned using the cross-attention mechanism.

[0022] Based on the above scheme and as a preferred option: In the cross-attention mechanism, after multiplying the key K and value V matrices, the result is processed by the softmax function, then multiplied by the query Q matrix. The calculated result is concatenated with the output feature map of the CTV segmentation network, and then passed through a convolutional layer to obtain the CTV segmentation result. The output result is calculated in the form of a weighted sum, where the weight assigned to each value is calculated based on the compatibility function between the query and the corresponding key.

[0023]

[0024] Where: dk represents the dimension of K.

[0025] Based on the above scheme and as a preferred embodiment of the above scheme: the width-driven attention network is used to extract contextual information in terms of width, generate a scaling factor for each channel for each column; calculate the width-driven attention weights to utilize the context to represent the importance of each column feature.

[0026] Based on the above scheme and as a preferred embodiment: the width-driven attention network is merged into a specific layer to form a semantic segmentation network, including: height-oriented pooling operation, interpolation to a coarse attention map, and computation of width-driven attention. Figure 3 part;

[0027] Low-level feature mapping in semantic segmentation networks Advanced Feature Mapping Where C is the number of channels, and D, H and W are the spatial dimensions of the input tensor, namely depth, height and width, respectively;

[0028] Given a low-level feature map X l After that, F HANet This will generate an attention map along the channel direction. It consists of a scaling factor in the height direction of each channel, and is associated with the high-level feature map X. h The channels and depth / height dimensions are matched;

[0029] After computing the attention map, the given high-level feature map X is... h Convert to a new feature This new feature is composed of A and X h The product of the elements is obtained by multiplying them together, and the formula is:

[0030]

[0031] I) Height-direction pooling operation:

[0032] Using height-oriented pooling operation G pool C l ×D l ×H l ×W l Input size X l Aggregate into C l ×D l ×1×W l A matrix Z of size, from each column, extracts contextual information in the height direction, i.e.:

[0033] Z = G pool (X l );

[0034] Formally, the w-th column vector of depth d is calculated as follows:

[0035]

[0036] II) Interpolation is a coarse attention map:

[0037] C is downsampled l ×D l ×1×W l The matrix Z interpolation is matrix

[0038] III) Calculate the width-driven attention map:

[0039] The width-driven channel orientation attention map A is obtained through a convolutional layer, whose input is the pooled and interpolated feature map in the width direction.

[0040] Attention map A shows the key channels for each column, and the operation consisting of N convolutional layers can be written as:

[0041]

[0042] Where σ is the sigmoid function and δ is the ReLU activation. This represents the i-th two-dimensional convolutional layer;

[0043] IV) The attention map is transformed by upsampling so that its width dimension is equal to that of the given high-level feature map X. h .

[0044] This invention introduces a dual attention mechanism to spatially constrain CTV segmentation. The first attention mechanism is the Spatial Constraint Net (SC-Net), which creates a CTV segmentation network and an OAR segmentation network. The CTV segmentation network uses an encoder-decoder architecture to learn CTV segmentation, and the OAR segmentation network uses the same architecture to learn OAR segmentation. The encoder parameters of the two segmentation networks are shared. The decoders of the two segmentation networks are not shared, allowing them to obtain different feature maps. After patching and embedding, the features extracted by the CTV decoder are used as the key K and value V, and the features extracted by the OAR decoder are used as the query Q. A cross-attention mechanism is used to learn the spatial constraints between OAR and CTV, and the OAR features are used to guide CTV segmentation.

[0045] The second attention mechanism is a width-driven attention network, which is inserted into the decoder stage of the CTV segmentation network to fine-tune the CTV segmentation results. The width-driven attention network annotates a small portion of the lumbar muscle data. Based on the relatively strong anatomical priors of CTV and lumbar muscles, the left and right boundaries of the upper half of the CTV are located on the medial edge of the lumbar muscles. Using the parameters trained on the first architecture, the width-driven attention network is adjusted to selectively emphasize information categories according to the horizontal position of the pixel category distribution, thus fine-tuning the CTV segmentation obtained in the first step.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] 1. This invention proposes an automatic segmentation technology for cervical cancer target areas based on a dual attention mechanism with spatial constraints. By designing the network structure, the relative positional relationship between OAR and CTV is integrated into the neural network, providing prior knowledge to the neural network and improving the accuracy of cervical cancer CTV segmentation.

[0048] 2. Accurate delineation of the clinical target area for radiotherapy is crucial for subsequent treatment steps such as adaptive radiotherapy for cervical cancer, and has many practical implications for improving doctors' work efficiency and reducing harm to patients. Attached Figure Description

[0049] The following detailed description, in conjunction with the accompanying drawings and embodiments of the present invention, will provide further information.

[0050] Figure 1 This is a network structure diagram of the spatial constraint network SC-Net of the present invention;

[0051] Figure 2 This is a flowchart of the OAR feature-guided CTV segmentation process of the present invention;

[0052] Figure 3 This is a diagram showing the composition of the width-driven attention network of the present invention;

[0053] Figure 4 This is a flowchart illustrating the fine-tuning training process of inserting a CTV segmentation network into the width-driven attention network of this invention. Detailed Implementation

[0054] See the attached figures. The automatic cervical cancer target region segmentation method described in this embodiment introduces a dual attention mechanism to spatially constrain CTV segmentation, proposing two networks, SC-Net and WA-Net. The first network is the spatial constraint network SC-Net, which includes a CTV segmentation network and an OAR segmentation network. The CTV segmentation network uses an encoder-decoder architecture to learn CTV segmentation, and the OAR segmentation network uses the same architecture to learn OAR segmentation. The encoder parameters of the two segmentation networks are shared. The decoders of the two segmentation networks are not shared, allowing them to obtain different feature maps. After patch embedding, the features extracted by the CTV decoder are used as the key K and value V, and the features extracted by the OAR decoder are used as the query Q. A cross-attention mechanism is used to learn the relative positional relationship between OAR and CTV, and the OAR features are used to guide CTV segmentation.

[0055] A second network, WA-Net (width-driven attention), is inserted into the decoder stage of the CTV segmentation network to fine-tune the CTV segmentation results. The width-driven attention network annotates a small subset of the lumbar muscle data. Based on the relatively strong anatomical priors of the CTV and the lumbar muscles, the left and right boundaries of the upper half of the CTV lie on the medial edge of the lumbar muscles. Using parameters trained on the first architecture, the width-driven attention network is adjusted to selectively emphasize information categories based on the horizontal position of the pixel category distribution, thus fine-tuning the CTV segmentation network obtained in the first step.

[0056] The specific steps are as follows:

[0057] 1. The first network, SC-Net, includes the CTV segmentation network and the OAR segmentation network, both of which are common encoder-decoder structures in CNNs.

[0058] like Figure 1 As shown, the basic framework of the SC-Net network, taking 3D-Unet as an example, consists of two parts: the CTV segmentation network and the OAR segmentation network.

[0059] M = F E (I) (1)

[0060] X C =F D-CTV-2 (M) (2)

[0061] X O =F D-OAR-2 (M) (3)

[0062] Where I represents the input image, F EM represents the encoder of the network, and F represents the intermediate layer features obtained after the encoder. D-CTV-2 The first two modules of the CTV decoder, representing the network, consist of two consecutive convolutional blocks and an upsampling operation; X C This represents the CTV feature map obtained after processing this module. F D-OAR-2 Similarly, the first two modules of the OAR decoder represent the network; X O This represents the OAR feature map obtained after passing through this module.

[0063] The CTV segmentation network and the OAR segmentation network share a single encoder, but their decoders do not. The query retrieves feature maps from the OAR decoder, with keys and values ​​taken from feature maps in the CTV decoder.

[0064] The OAR feature module learns from the OAR delineated in cervical cancer CT scans, including OAR segmentation of five parts: bladder, pelvis, pelvic bone marrow, rectum, and small intestine.

[0065] Then, the correlation between CTV segmentation and OAR segmentation is calculated through a cross-attention mechanism. Specifically, the feature maps of CTV and OAR extracted by CNN are unfolded into sequence vectors, positional encoding is added, and the positional relationship between the parts of the features is learned using a cross-attention mechanism. The attention function can be described as mapping a query Q and a set of key-value (KV) pairs to an output, where the query Q, key K, value V and output are all vectors, as described in formula (4), where d k The dimension representing K. For example... Figure 2 As shown, the output is calculated as a weighted sum, where the weight assigned to each value is calculated based on a compatibility function between the query and the corresponding key.

[0066]

[0067] 2. Medical images often involve common structural priors dependent on spatial location. Based on the relative position of the lumbar muscles and the CTV (Central Television Channel), each column of the CTV image has a different category distribution statistics. In this sense, capturing contextual information solely from the width (representing the global context of each column) can be used to estimate how channels should be weighted during pixel-level classification in medical image segmentation. Therefore, this invention proposes a WA (Width-Driven Attention) network, the purpose of which is to: i) extract contextual information in terms of width; ii) compute width-driven attention weights to utilize context to represent the importance of each column's features (in intermediate layers). In this section, WA-Net is first described as a general plug-in module, and then WA-Net is merged into specific layers to form a semantic segmentation network specifically for medical image segmentation. This embodiment only annotates a small portion of the lumbar muscle data and fine-tunes the CTV segmentation network obtained in the first step.

[0068] like Figure 3 As shown, WA-Net, which fine-tunes the network, generates a scaling factor for each channel for each column based on contextual information about the column width.

[0069] and Let X represent the low-level and high-level feature maps in the semantic segmentation network, where C is the number of channels, and D, H, and W are the spatial dimensions of the input tensor, namely depth, height, and width, respectively. Given a low-level feature map X... l After that, F HANet This will generate an attention map along the channel direction. It consists of a scaling factor in the height direction of each channel, and is associated with the high-level feature map X. h The channels are matched with the depth and height dimensions. This is accomplished through a series of steps: height-oriented pooling operations ( Figure 3 (a)), interpolation is a coarse attention map ( Figure 3 (b,d)) and computation width-driven attention map ( Figure 3 (c)).

[0070] After computing the attention map, the given high-level feature map X can be... h Convert to a new feature This new feature is composed of A and X h The elements are multiplied together to obtain the result. Note that a single per-channel scaling vector is derived for each set of columns or multiple consecutive columns, thus replicating the vector horizontally, as shown in the formula:

[0071]

[0072] Specifically, first, a height-oriented pooling operation is performed. Figure 3 (a)). To obtain the attention map along the channel direction, we first use the height-oriented pooling operation G. pool C l ×D l ×H l ×W l Input size X l Aggregate into C l ×D l ×1×W l A matrix Z of size, from each column, extracts contextual information in the height direction, i.e.

[0073] Z = G pool (X l (6)

[0074] The pooling operation along the height direction is set to average pooling. Formally, the w-th column vector of depth d is calculated as follows:

[0075]

[0076] The second step is to calculate the coarse attention interpolation. Figure 3 (b, d)). After the pooling operation, the model generates a matrix. However, only some columns in matrix Z are necessary for computing an effective attention map. Some columns have only one class (i.e., CTV), while others have two classes (i.e., CTV and lumbar muscle). Downsampling C... l ×D l ×1×W l The matrix Z interpolation is A matrix ( Figure 3 (b) H is a hyperparameter set to 16. To enable multiplication with the feature map, the attention map is upsampled so that its width dimension equals the given high-level feature map X. h ( Figure 3 (d)).

[0077] Next, the width-driven attention map is calculated. Figure 3 (c) The width-driven channel orientation attention map A is obtained through a convolutional layer, whose input is the pooled and interpolated feature map in the width direction. The convolutional layers we use take into account the relationships between adjacent columns when estimating the attention map, because each column is related to its neighboring columns.

[0078] Attention map A shows the key channels for each column. For example, in the last layer, each column can be associated with one or two labels (CTV or CTV+Psoas). These operations, consisting of N convolutional layers, can be written as follows:

[0079]

[0080] Where σ is the sigmoid function and δ is the ReLU activation. Let represent the i-th 2D convolutional layer. We used three convolutional layers. Decreasing the ratio r can reduce the parameter overhead of WANet and also provide potential regularization.

[0081] 3. For example Figure 4 As shown, WA-Net is inserted into the decoder stage of the CTV segmentation network to fine-tune the segmentation results. To save manpower, only a portion of the data is labeled for training in this stage.

[0082] This invention is based on the approach taken by human experts when learning to delineate the cervical cancer target area (CTV). A certain level of understanding of human anatomy is necessary for CTV delineation. This invention proposes an automatic cervical cancer target area delineation technology based on a dual-attention mechanism with spatial constraints. By designing the network structure, the relative positional relationship between the target area (OAR) and the CTV is integrated into the neural network, providing prior knowledge and improving the accuracy of cervical cancer CTV segmentation. Accurate delineation of the clinical target area for radiotherapy is crucial for subsequent treatment steps such as adaptive radiotherapy for cervical cancer, and has significant practical implications for improving physician efficiency and reducing harm to patients.

[0083] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for automatic segmentation of cervical cancer target volume based on a double-attention mechanism, characterized in that: This is a two-stage training method, including a spatial constraint-based network and a width-driven attention network based on fine-tuning; Specifically comprising the following steps: 1) constructing a spatial constraint network, the spatial constraint network comprising two convolutional neural networks with the same architecture, respectively a CTV segmentation network and an OAR segmentation network; 2) the CTV segmentation network and the OAR segmentation network respectively learn to delineate CTV and OAR from cervical cancer CT, and respectively extract features related to CTV segmentation and OAR segmentation; 3) using the CTV features as the key K, using the OAR features as the query Q, using the cross-attention mechanism to learn the attention weight in the cervical cancer CTV segmentation, and then using the CTV features as the value V to perform weighted aggregation, to learn the spatial constraint between the OAR and the CTV; 4) restoring the features obtained in step 3) to shape and splicing with the output feature map of the CTV segmentation network, and then performing convolution to obtain the CTV segmentation result; constructing a width-driven attention network and inserting it into the decoder stage of the CTV segmentation network to fine-tune the CTV segmentation network; The CTV segmentation network and the OAR segmentation network are both convolutional neural networks comprising an encoder and a decoder structure, the CTV segmentation network and the OAR segmentation network share an encoder, and the decoders work independently; The width-driven attention network is merged into a specific layer to form a semantic segmentation network, including: height direction pooling operation, interpolation to coarse attention map and calculation of width-driven attention map; Low-level feature mapping in semantic segmentation networks High-level feature mapping where C is the number of channels, and D, H, and W are the spatial dimensions of the input tensor, depth, height, and width, respectively; Given a low-level feature map X l After F HANet A channel-wise attention map is generated Composed of a scaling factor for each channel height direction and matches the channel and depth, height dimensions of the high-level feature map X h ​ After computing the attention map, the given high-level feature map X h is converted to a new feature The new feature is obtained by multiplying the elements of A and X h , whose formula is: I) height direction pooling operation: A pooling operation G in the height direction is used pool C l ×D l ×H l ×W l An input X of size C l is aggregated into a matrix Z of size C l ×D l ×1×W l from which the context information in the height direction is extracted from each column, i.e.: Z = G pool (X l ); Formally, the wth column vector of depth d is calculated as: II) interpolation to coarse attention map: C l ×D l ×1×W l is interpolated to a matrix Z of size ​ III) calculation of width-driven attention map: The channel direction attention map A driven by the width is obtained through a convolution layer, and an input of the convolution layer is the feature map after the width direction pooling and interpolation The attention map A shows the key channels of each column, and the operation consisting of N convolutional layers can be written as: where σ is a sigmoid function, and δ is a ReLU activation, represents the i-th two-dimensional convolutional layer; The attention map is converted by upsampling so that its width dimension is equal to the given high-level feature map X h .

2. The method of claim 1, wherein the method is based on a dual attention mechanism. The spatial constraint network is a 3D-Unet network, comprising a CTV segmentation network and an OAR segmentation network; M = F E (I) X C = F D-CTV-2 (M) X O = F D-OAR-2 (M) where I represents the input image, F E represents the encoder of the network, M represents the intermediate layer feature obtained after the encoder, F D-CTV-2 represents the first two modules of the CTV decoder of the network, which includes two consecutive convolution blocks and an up-sampling operation; X C represents the CTV feature map obtained after the module; F D-OAR-2 represents the first two modules of the OAR decoder of the network; X O represents the OAR feature map obtained after the module.

3. The method of claim 1, wherein the method comprises: obtaining a cervical cancer image; inputting the cervical cancer image into the dual attention mechanism based model to obtain a cervical cancer target volume segmentation result. The output feature maps of the CTV segmentation network and the OAR segmentation network are respectively unfolded into sequence vectors and added with position encoding; the key K and the value V corresponding to the CTV features, and the query Q corresponding to the OAR features are obtained; the cross-attention mechanism is used to learn the position relationship between the CTV and the OAR in the features.

4. The method of claim 3, wherein the method comprises: In the cross-attention mechanism, after the key K and the value V matrix are multiplied, they are processed by the softmax function, and then multiplied with the query Q matrix, the calculation result is spliced with the output feature map of the CTV segmentation network, and then a convolution layer is used to obtain the CTV segmentation result; the output result is calculated in the form of weighted sum, wherein the weight assigned to each value is calculated according to the compatibility function of the query and the corresponding key: Where: dk represents the dimension of K.

5. The method of claim 1, wherein the method comprises: The width-driven attention network is used to extract the context information in the width aspect, and generate a scaling factor for each channel of each column; The width-driven attention weight is calculated to use the context to represent the importance of each column feature.

Images