Cervical cancer clinical target region segmentation method guided by anatomical prior position information

Abstract

The invention provides a cervical cancer clinical target segmentation method guided by anatomical prior position information, and the method comprises the steps: employing the CT image data of a cervical cancer patient and a reference CTV as basic data, and calculating a group of corresponding distance information image data Dimg according to the basic data; a multi-task double-branch feature fusion network MTF-Net based on a UNet network is constructed; the reference CTV and the Dimg are used as references for training the MTF-Net; in addition, CT image data is used as input information of MTF-Net intelligent segmentation; wherein the reference CTV is the DICOM format data of the cervical cancer CTV, which is drawn and determined manually according to a professional guide and a process; the distance information image data Dimg refers to a distance information graph between a surrounding adjacent organ and the cervical cancer CTV. The method has the advantages that the anatomical structure position information of the cervical cancer CTV and the surrounding adjacent organs is fully utilized, the anatomical prior position information is obtained, and the accuracy and efficiency of cervical cancer segmentation are improved.

Description

Technical Field

[0001] The invention relates to a method for segmenting a clinical target area of ​​cervical cancer, and in particular to a method for segmenting a clinical target area of ​​cervical cancer guided by anatomical priori position information. Background Art

[0002] Radiotherapy (abbreviated as RT) is one of the important methods in the treatment of cervical cancer. The precise radiotherapy process requires doctors to manually segment the clinical target volume (CTV) of cervical cancer. The process is time-consuming and labor-intensive, highly subjective, and related to the level of experience. Obviously, the automatic segmentation of cervical cancer CTV is crucial for the precise radiotherapy of malignant tumors. However, cervical cancer CTV is the microscopic spread of tumor cells, which is not visible on the positioning CT required for radiotherapy, and is still unclear even under magnetic resonance imaging with soft tissue advantages. At present, there is no unified expert consensus on the segmentation of cervical cancer CTV, and there are multiple relevant guidelines, and different people may have different understandings of the guidelines. Therefore, it is extremely challenging for radiotherapists to accurately outline the contours of cervical cancer CTV on CT. Traditional automatic segmentation methods, such as the automatic segmentation method based on the atlas library ABAS, are not accurate, require manual editing by doctors, and the process is very time-consuming.

[0003] Because the cervix is ​​located in the pelvic cavity and its regional anatomical structure is complex, traditional segmentation methods often fail to achieve satisfactory accuracy. In recent years, deep learning technology has made significant progress in the field of cervical cancer CTV segmentation, but existing methods that apply deep learning technology to cervical cancer CTV segmentation often ignore the use of anatomical structure information, resulting in limited segmentation accuracy. In actual clinical practice, doctors also need to refer to the location information of the surrounding adjacent anatomical structures when outlining the cervical cancer CTV. Therefore, making full use of prior information is crucial to improving the accuracy of cervical cancer CTV segmentation. Summary of the invention

[0004] Aiming at the problems that the existing technology ignores the anatomical position information of the target area, resulting in limited segmentation accuracy, etc., the present invention proposes a cervical cancer clinical target area segmentation method guided by anatomical prior position information.

[0005] The cervical cancer clinical target segmentation method guided by anatomical prior position information of the present invention adopts CT image data and benchmark CTV of cervical cancer patients as basic data, and calculates a corresponding set of distance information image data Dimg according to the basic data; constructs a multi-task dual-branch feature fusion network MTF-Net based on UNet network; uses benchmark CTV and Dimg as references for training MTF-Net; and uses CT image data as input information for MTF-Net intelligent segmentation; wherein, the benchmark CTV refers to DICOM format data of cervical cancer CTV manually outlined and determined according to professional guidelines and processes; the distance information image data Dimg refers to the distance information map between the surrounding adjacent organs and the cervical cancer CTV.

[0006] Preferably, the use of CT image data of cervical cancer patients and reference CTV as basic data includes converting the CT image data of cervical cancer patients into a set of two-dimensional slice image data; and generating corresponding label data from the reference CTV.

[0007] Preferably, the step of calculating a corresponding set of distance information image data Dimg according to the basic data includes calculating a corresponding set of distance information image data Dimg using Python computer language according to the label data generated by the reference CTV; wherein the distance between the surrounding adjacent organs and the cervical cancer CTV is calculated using the Euclidean distance to obtain the distance information image data Dimg; as shown in the following formula (1):

[0008]

[0009] In the formula, Ω out is the external space of the cervical cancer CTV surface, Ω in is the internal space of cervical cancer CTV, (x1, y1) is the point of the external space of cervical cancer CTV, (x2, y2) is the point on the boundary of cervical cancer CTV, the distance value of the voxel inside cervical cancer CTV is set to 0, and D(x, y) is the Euclidean distance.

[0010] Preferably, the multi-task dual-branch feature fusion network MTF-Net based on the UNet network is constructed; it includes a shared encoder and two parallel decoders with the same structure, and the two parallel decoders are respectively a distance information decoder and a target CTV segmentation decoder; the distance information decoder and the target CTV segmentation decoder are respectively connected to the shared encoder through their own jump connections; and an information attention module DIA is set on each jump connection between the target CTV segmentation decoder and the shared encoder; the decoders all use bilinear interpolation upsampling to restore to the size of the input image; wherein the shared encoder extracts multi-level features of the input image to provide basic feature representation for the subsequent two parallel decoders; the distance information decoder reconstructs the distance information between the target area and the surrounding tissue; the target CTV segmentation decoder restores the multi-level features extracted by the encoder to the original image size by upsampling and feature fusion, and accurately classifies each pixel to achieve cervical cancer target area segmentation; the information attention module DIA introduces the edge position information of the distance information decoder as a strong clue to enhance the edge part of the segmentation feature.

[0011] Preferably, the shared encoder extracts multi-level features of the input image and provides basic feature representation for the subsequent two parallel decoders, including using ResNet50 as the backbone network in the shared encoder. The shared encoder includes five stages, and its output features are recorded as Ei (i=1, 2, 3, 4, 5); the size of Ei is 1 / 2, 1 / 4, 1 / 8, 1 / 16 and 1 / 32 of the input image respectively; from stage 1 to stage 5, as the network deepens, the feature information becomes richer.

[0012] Preferably, the distance information decoder reconstructs the distance information between the target area and the surrounding tissue, including obtaining the shared encoder output features E1, E2, E3 and E4, performing 1×1 convolution on the output features, and reducing the number of channels to 1 to generate a distance information prediction map Dimg'.

[0013] Preferably, the information attention module DIA introduces the edge position information of the distance information decoder as a strong clue to enhance the edge part of the segmentation feature, including using 4 DIA modules to process the features E1, E2, E3, and E4 of stages 1 to 4 obtained by the shared encoder; thereby generating corresponding distance information enhanced features A1, A2, A3, and A4 of 4 different stages; including,

[0014] Each DIA module has two inputs, namely the distance information prediction map Dimg' and the features Ei (i∈{1, 2, 3, 4}) from stage 1 to 4 obtained by the shared encoder; Dimg' is max-pooled and downsampled to the same spatial size as Ei, and an element-wise product is performed between Dimg' and Ei, and then fused with Ei through a feature parallel operation Concat to generate the initial distance enhancement feature, which is used by E' i express;

[0015] E' i =C(M(Dimg')*Ei),Ei) (2)

[0016] In the formula, M and * are Maxpool and element-wise product operations respectively, and C is the concatenation operation Concat;

[0017] The spatial attention mechanism is introduced to improve the recognition ability of cervical cancer target areas in images. That is, features are extracted through a 1×1 convolution layer, and then the ReLU activation function is applied to introduce nonlinear transformation, and then the local spatial features are further refined through a 3×3 convolution layer. Finally, the spatial attention map Es is output through the Sigmoid function, which reflects the importance weight of each spatial position. The spatial attention map Es is element-wise multiplied with the input feature Ei to enhance the spatial features and obtain the enhanced feature A:

[0018] A=E s *E' i (3)

[0019] The feature A is equally divided into two parts in the channel dimension, denoted as Ad and As; Ad and As correspond to the distance information feature and the segmentation key area feature respectively; the feature Ei is added element by element with Ad and As to obtain the final enhanced feature Ai:

[0020] Ai=A d +A s +E i (4)

[0021] At this point, the features E1, E2, E3, and E4 of stages 1 to 4 obtained by the shared encoder are used by the DIA module to generate the corresponding distance information enhanced features A1, A2, A3, and A4 of four different stages.

[0022] Preferably, the target CTV segmentation decoder restores the multi-level features extracted by the encoder to the original image size by upsampling and feature fusion, and accurately classifies each pixel to achieve cervical cancer target segmentation, including that the segmentation decoder obtains distance information enhancement features A1, A2, A3 and A4, performs 1×1 convolution on the output features, and reduces the number of channels to 2 to generate a target CTV segmentation prediction map.

[0023] Preferably, the reference CTV and Dimg are used as references for training MTF-Net, including collecting CT image data and reference CTV of 418 cervical cancer patients, and randomly dividing the data set into a training set, a validation set and a test set in a ratio of 7:1:2; the training data set and the validation data set include CT image data, reference CTV and distance information image data Dimg, which are used to train and verify the dual-branch network model; the test data set includes CT image data and reference CTV, which are used to test the segmentation effect of the network; and during training, the target CTV segmentation prediction map and the distance information prediction map Dimg' generated by MTF-Net are respectively subjected to Loss loss operation with the reference CTV data and Dimg, so that the target CTV segmentation prediction map and the distance information prediction map Dimg' gradually approach the reference CTV data and Dimg; including using Dice loss L dice , cross entropy loss L ce And the mean absolute error loss L mae Train MTF-Net and combine the three loss functions to get the fusion loss function as follows:

[0024] Loss = L dice (S ctv ,G ctv )+0.8L ce (S ctv ,G ctv )+0.2L mae (S dia ,G dia ) (4)

[0025] Where S ctv , G ctv are the segmentation prediction value of CTV and the benchmark CTV, Sdia and Gdia are the distance information prediction value and Dimg respectively; the definitions of these three loss functions are as follows:

[0026]

[0027] Where n is the number of voxels in the input CT image data, sj represents the predicted probability that the jth voxel belongs to the CTV, and gj represents the voxel value of the corresponding benchmark CTV. At the same time, the validation dataset is used to evaluate, optimize and select the best parameters for MTF-Net to ensure the generalization of the model.

[0028] The beneficial technical effect of the cervical cancer clinical target segmentation method guided by anatomical prior position information of the present invention is to fully utilize the anatomical structure position information of the cervical cancer CTV and the surrounding adjacent organs, effectively improving the accuracy and efficiency of cervical cancer segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Attached Figure 1 This is a schematic diagram of the MTF-Net architecture of the multi-task dual-branch feature fusion network of the present invention;

[0030] Attached Figure 2 It is a schematic diagram of the structure of the information attention module DIA of the present invention;

[0031] Attached Figure 3 is the distance information image data Dimg of the anatomical priori position information of the present invention;

[0032] Attached Figure 4 It is a schematic diagram of the visualization results of the segmentation results of the method of the present invention and other currently most advanced methods.

[0033] The cervical cancer CTV segmentation method guided by anatomical priori position information of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments. DETAILED DESCRIPTION

[0034] The cervical cancer clinical target segmentation method guided by anatomical prior position information of the present invention uses CT image data and reference CTV of cervical cancer patients as basic data, and calculates a corresponding set of distance information image data Dimg according to the basic data; constructs a multi-task dual-branch feature fusion network MTF-Net based on the UNet network; uses reference CTV and Dimg as references for training MTF-Net; and uses CT image data as input information for MTF-Net intelligent segmentation; wherein the reference CTV refers to the DICOM format data of the cervical cancer CTV manually outlined and determined according to professional guidelines and processes; the distance information image data Dimg refers to the distance information map between the surrounding adjacent organs and the cervical cancer CTV. The cervical cancer CTV segmentation method of the present invention calculates a corresponding set of distance information image data Dimg according to the basic data, and uses reference CTV and Dimg as references for training MTF-Net, so that MTF-Net makes full use of the anatomical structure position information of the cervical cancer CTV and the surrounding adjacent organs when segmenting the cervical cancer CTV, effectively improving the accuracy and efficiency of cervical cancer CTV segmentation. Because the cervical cancer CTV segmentation method of the present invention is guided by the anatomical structure position information of the cervical cancer CTV and the surrounding adjacent organs, and fully utilizes the anatomical structure position information of the cervical cancer CTV and the surrounding adjacent organs during segmentation, it is called a cervical cancer CTV segmentation method guided by anatomical prior position information.

[0035] As one of the preferred technical solutions, the CT image data of cervical cancer patients and the reference CTV are used as basic data, including converting the CT image data of cervical cancer patients into a set of two-dimensional slice image data; and generating corresponding label data from the reference CTV. The conversion of the CT image data of cervical cancer patients and the reference CTV is beneficial for MTF-Net recognition on the one hand, and is beneficial for the calculation of distance information image data Dimg on the other hand.

[0036] As one of the preferred technical solutions, the method of calculating a corresponding set of distance information image data Dimg according to the basic data includes: using Python computer language to calculate a corresponding set of distance information image data Dimg according to the label data generated by the reference CTV; wherein the distance between the surrounding adjacent organs and the cervical cancer CTV is calculated using the Euclidean distance to obtain the distance information image data Dimg; as shown in the following formula (1):

[0037]

[0038] In the formula, Ω out is the external space of the cervical cancer CTV surface, Ω inis the internal space of cervical cancer CTV, (x1, y1) is the point of the external space of cervical cancer CTV, (x2, y2) is the point on the boundary of cervical cancer CTV, the distance value of the voxel inside cervical cancer CTV is set to 0, and D(x, y) is the Euclidean distance. Euclidean distance, also known as Euclidean metric and Euclidean distance, is a widely used distance definition. It refers to the true distance between two points in m-dimensional space, or the natural length of a vector (that is, the distance from the point to the origin). In clustering algorithms, Euclidean distance is often used to calculate the distance between sample points, thereby realizing clustering of samples. The method of the present invention uses Euclidean distance to construct distance information image data Dimg, which can not only accurately reflect the position information between the surrounding adjacent organs and cervical cancer CTV, but also make the calculation results more representative (see Appendix). Figure 3 ,In the figure, a is a CT image with a CTV label, and b is a distance position information map Dimg).

[0039] See attached Figure 1 As one of the preferred technical solutions, the multi-task dual-branch feature fusion network MTF-Net based on the UNet network is constructed; it includes a shared encoder and two parallel decoders with the same structure, and the two parallel decoders are respectively a distance information decoder and a target CTV segmentation decoder; the distance information decoder and the target CTV segmentation decoder are respectively connected to the shared encoder through their own jump connections; and an information attention module DIA is set on each jump connection between the target CTV segmentation decoder and the shared encoder; the decoders all use bilinear interpolation upsampling to restore to the size of the input image; wherein the shared encoder extracts multi-level features of the input image to provide basic feature representation for the subsequent two parallel decoders; the distance information decoder reconstructs the distance information between the target area and the surrounding tissue; the target CTV segmentation decoder restores the multi-level features extracted by the encoder to the original image size by upsampling and feature fusion, and accurately classifies each pixel to achieve cervical cancer target segmentation; the information attention module DIA introduces the edge position information of the distance information decoder as a strong clue to enhance the edge part of the segmentation feature. The cervical cancer CTV segmentation method guided by anatomical prior position information of the present invention sets a distance information decoder in the MTF-Net network architecture to decode the distance information image data Dimg, which can help MTF-Net better understand the spatial relationship between different tissue structures, thereby improving the accuracy of segmentation; setting an information attention module DIA to introduce the edge position information of the distance information decoder as a strong clue to enhance the edge part of the segmentation feature, and making full use of the anatomical structure position information of the cervical cancer CTV and the surrounding adjacent organs, so as to obtain the anatomical prior position information more accurately and comprehensively.

[0040] As one of the preferred technical solutions, the shared encoder extracts multi-level features of the input image and provides basic feature representation for the subsequent two parallel decoders, including using ResNet50 as the backbone network in the shared encoder. The shared encoder includes five stages, and its output features are recorded as Ei (i = 1, 2, 3, 4, 5); the size of Ei is 1 / 2, 1 / 4, 1 / 8, 1 / 16 and 1 / 32 of the input image respectively; from stage 1 to stage 5, as the network deepens, the feature information becomes richer. The key information in the image is captured through the encoding process, which is crucial for the subsequent cervical cancer CTV segmentation.

[0041] As one of the preferred technical solutions, the distance information decoder reconstructs the distance information between the target area and the surrounding tissue, including obtaining the shared encoder output features E1, E2, E3 and E4, performing 1×1 convolution on the output features, and reducing the number of channels to 1 respectively, to generate a distance information prediction map Dimg'. The distance information prediction map Dimg' provides important edge position information for subsequent segmentation.

[0042] See attached Figure 2 As one of the preferred technical solutions, the information attention module DIA introduces the edge position information of the distance information decoder as a strong clue to enhance the edge part of the segmentation feature, including using 4 DIA modules to process the features E1, E2, E3, and E4 of stages 1 to 4 obtained by the shared encoder; thereby generating corresponding distance information enhanced features A1, A2, A3, and A4 of 4 different stages; including,

[0043] Each DIA module has two inputs, namely the distance information prediction map Dimg' and the features Ei (i∈{1, 2, 3, 4}) from stage 1 to 4 obtained by the shared encoder; Dimg' is max-pooled and downsampled to the same spatial size as Ei, and an element-wise product is performed between Dimg' and Ei, and then fused with Ei through a feature parallel operation Concat to generate the initial distance enhancement feature, which is used by E' i express;

[0044] E' i =C(M(Dimg')*Ei),Ei) (2)

[0045] In the formula, M and * are Maxpool and element-wise product operations respectively, and C is the concatenation operation Concat;

[0046] The spatial attention mechanism is introduced to improve the recognition ability of cervical cancer target areas in images. That is, features are extracted through a 1×1 convolution layer, and then the ReLU activation function is applied to introduce nonlinear transformation, and then the local spatial features are further refined through a 3×3 convolution layer. Finally, the spatial attention map Es is output through the Sigmoid function, which reflects the importance weight of each spatial position. The spatial attention map Es is element-wise multiplied with the input feature Ei to enhance the spatial features and obtain the enhanced feature A:

[0047] A=E s *E' i (3)

[0048] The feature A is equally divided into two parts in the channel dimension, denoted as Ad and As; Ad and As correspond to the distance information feature and the segmentation key area feature respectively; the feature Ei is added element by element with Ad and As to obtain the final enhanced feature Ai:

[0049] Ai=A d +A s +E i (4)

[0050] At this point, the features E1, E2, E3, and E4 of stages 1 to 4 obtained by the shared encoder generate the corresponding distance information enhancement features A1, A2, A3, and A4 of four different stages through the DIA module. The method of the present invention performs an element-by-element product operation on the spatial attention map Es and the input feature Ei, thereby achieving the first enhancement of the spatial feature; then, the input feature Ei is again element-by-element added to the features Ad and As, thereby achieving the second enhancement of the spatial feature, further effectively integrating the distance information and the segmentation key area information. In addition, in order to achieve refined feature processing, the feature A is equally divided into two parts in the channel dimension, namely the distance information feature and the segmentation key area feature, laying the foundation for achieving spatial feature enhancement again.

[0051] As one of the preferred technical solutions, the target CTV segmentation decoder restores the multi-level features extracted by the encoder to the original image size by upsampling and feature fusion, and accurately classifies each pixel to achieve cervical cancer target segmentation, including: the segmentation decoder obtains distance information enhancement features A1, A2, A3 and A4, performs 1×1 convolution on the output features, and reduces the number of channels to 2 respectively, to generate a target CTV segmentation prediction map. Since the distance information enhancement features A1, A2, A3 and A4 fuse the anatomical structure position information of the cervical cancer CTV and the surrounding adjacent organs, that is, the anatomical prior position information, the accuracy of the segmentation is improved.

[0052] As one of the preferred technical solutions, the method uses the benchmark CTV and Dimg as a reference for training MTF-Net, including collecting CT image data and benchmark CTV of 418 cervical cancer patients, and randomly dividing the data set into a training set, a validation set and a test set in a ratio of 7:1:2; the training data set and the validation data set include CT image data, benchmark CTV and distance information image data Dimg, which are used to train and verify the dual-branch network model; the test data set includes CT image data and benchmark CTV, which are used to test the segmentation effect of the network; and during training, the target CTV segmentation prediction map and the distance information prediction map Dimg' generated by MTF-Net are respectively subjected to Loss loss operation with the benchmark CTV data and Dimg, so that the target CTV segmentation prediction map and the distance information prediction map Dimg' gradually approach the benchmark CTV data and Dimg; including using Dice loss L dice , cross entropy loss L ce And the mean absolute error loss L mae Train MTF-Net and combine the three loss functions to get the fusion loss function as follows:

[0053] Loss = L dice (S ctv ,G ctv )+0.8L ce (S ctv ,G ctv )+0.2L mae (S dia ,G dia ) (4)

[0054] Among them, S ctv , G ctv are the segmentation prediction value of CTV and the benchmark CTV, Sdia and Gdia are the distance information prediction value and Dimg respectively; the definitions of these three loss functions are as follows:

[0055]

[0056] Where n is the number of voxels in the input CT image data, sj represents the predicted probability that the jth voxel belongs to the CTV, and gj represents the voxel value of the corresponding benchmark CTV. At the same time, the validation dataset is used to evaluate, optimize and select the best parameters for MTF-Net to ensure the generalization of the model.

[0057] The clinical target segmentation method for cervical cancer guided by anatomical prior position information of the present invention uses CT image data and reference CTV of cervical cancer patients as basic data, and converts them into CT image data and label data, and then calculates distance information image data Dimg according to the label data; the reference CTV and Dimg are used as references for training MTF-Net, so that the target CTV segmentation prediction map and distance information prediction map Dimg' predicted by MTF-Net gradually approach the reference CTV and Dimg. In addition, the method of the present invention uses the DIA module to reversely input the Dimg' feature map generated by the CT image data into the encoding, and combines it with the original CT encoding features, effectively improving the segmentation accuracy of the target CTV.

[0058] See Table 1. In order to evaluate the segmentation performance of MTF-Net, the method MTF-Net of the present invention and the most advanced method at present are used to segment the same data set, and four quantitative evaluation indicators are used to evaluate the segmentation results, including Dice similarity coefficient DSC, sensitivity SEN, positive prediction value PPV and 95% Hausdorff distance HD95. Table 1 lists the evaluation results. From the data in Table 1, it can be seen that the MTF-Net of the present invention is 2.55%, 2.84% and 2.22% higher than PPAF_Net in DSC, SEN and PPV respectively; and it is reduced by 1.45mm in 95HD; which fully demonstrates that the cervical cancer CTV segmentation method guided by anatomical prior position information has the advantages of more accurate segmentation and higher efficiency.

[0059] Table 1 Test set comparison experimental results

[0060]

[0061]

[0062] Attached Figure 4 This is a visualization diagram of the segmentation results of the method of the present invention and other currently most advanced methods. In the figure, the first column is the CT slice of the cervical cancer CTV, and the red part in the middle is the benchmark CTV; the green part in the second to eighth columns is the benchmark CTV, and the red part is the prediction result of various methods. In fact, the red part should also be blocky, but a large part is covered by green and looks like a line. Therefore, the less red part, the closer the prediction result is to the benchmark CTV. It is not difficult to see that MTF-Net shows the closest overall similarity to the benchmark CTV.

[0063] See Table 2, ablation experiments are used to verify the effectiveness of MTF-Net. A basic model baseline is built based on the ResU-Net model, while keeping its architecture the same as MTF-Net. Distance prediction and DIA attention modules are added as auxiliary tasks, and quantitative performance comparisons are performed. Table 2 shows the comparison results of different combinations in the test set. Adding distance prediction as an auxiliary task improves the DSC level on the test set by 1.37%; adding the DIA module improves the processing performance of DSC by another 1.84% in the test set. These comparisons show that adding distance prediction as an auxiliary task enhances the model's ability to locate the boundaries of CTV, while the addition of the DIA attention module further improves the segmentation performance.

[0064] Table 2: MTF-Net ablation experiment results

[0065]

[0066] Obviously, the beneficial technical effect of the cervical cancer CTV segmentation method guided by anatomical prior position information of the present invention is to make full use of the anatomical structure position information of the cervical cancer CTV and the surrounding adjacent organs, obtain anatomical prior position information, and improve the accuracy and efficiency of cervical cancer segmentation.

Claims

1. A method for segmenting a clinical target area of ​​cervical cancer guided by anatomical prior position information, characterized in that: Using CT image data of cervical cancer patients and benchmark CTV as basic data, a corresponding set of distance information image data Dimg is calculated according to the basic data; A multi-task dual-branch feature fusion network MTF-Net based on the UNet network is constructed; the benchmark CTV and Dimg are used as references for training MTF-Net; and CT image data is used as input information for MTF-Net intelligent segmentation; wherein the benchmark CTV refers to the DICOM format data of the cervical cancer CTV manually outlined and determined according to professional guidelines and processes; the distance information image data Dimg refers to the distance information map between the surrounding adjacent organs and the cervical cancer CTV.

2. The method for clinical target segmentation of cervical cancer guided by anatomical prior position information according to claim 1, characterized in that: The method adopts CT image data of cervical cancer patients and reference CTV as basic data, including converting the CT image data of cervical cancer patients into a set of two-dimensional slice image data; and generating corresponding label data from the reference CTV.

3. The method for clinical target segmentation of cervical cancer guided by anatomical prior position information according to claim 1, characterized in that: The corresponding set of distance information image data Dimg is calculated according to the basic data, The method includes calculating a corresponding set of distance information image data Dimg using Python computer language according to the label data generated by the reference CTV; wherein the distance between the surrounding adjacent organs and the cervical cancer CTV is calculated using the Euclidean distance to obtain the distance information image data Dimg; as shown in the following formula (1): In the formula, Ω out is the external space of the cervical cancer CTV surface, Ω in is the internal space of cervical cancer CTV, (x1, y1) is the point of the external space of cervical cancer CTV, (x2, y2) is the point on the boundary of cervical cancer CTV, the distance value of the voxel inside cervical cancer CTV is set to 0, and D(x, y) is the Euclidean distance.

4. The method for clinical target segmentation of cervical cancer guided by anatomical prior position information according to claim 1, characterized in that: The invention discloses a multi-task dual-branch feature fusion network MTF-Net based on UNet network, which includes a shared encoder and two parallel decoders with the same structure, and the two parallel decoders are respectively a distance information decoder and a target CTV segmentation decoder; the distance information decoder and the target CTV segmentation decoder are respectively connected to the shared encoder through their own jump connections; and an information attention module DIA is set on each jump connection between the target CTV segmentation decoder and the shared encoder; the decoders all adopt bilinear interpolation upsampling to restore to the size of the input image; wherein the shared encoder extracts multi-level features of the input image to provide basic feature representation for the subsequent two parallel decoders; the distance information decoder reconstructs the distance information between the target area and the surrounding tissue; the target CTV segmentation decoder restores the multi-level features extracted by the encoder to the original image size by upsampling and feature fusion, and accurately classifies each pixel to achieve cervical cancer target area segmentation; the information attention module DIA introduces the edge position information of the distance information decoder as a strong clue to enhance the edge part of the segmentation feature.

5. The method for clinical target segmentation of cervical cancer guided by anatomical prior position information according to claim 4, characterized in that: The shared encoder extracts multi-level features of the input image and provides basic feature representation for the subsequent two parallel decoders. Including using ResNet50 as the backbone network in the shared encoder, the shared encoder consists of five stages, and its output features are denoted as Ei (i=1, 2, 3, 4, 5); the sizes of Ei are 1 / 2, 1 / 4, 1 / 8, 1 / 16 and 1 / 32 of the input image respectively; from stage 1 to stage 5, as the network deepens, the feature information becomes richer.

6. The method for clinical target segmentation of cervical cancer guided by anatomical prior position information according to claim 5, characterized in that: The distance information decoder reconstructs the distance information between the target area and the surrounding tissue, Including obtaining the shared encoder output features E1, E2, E3 and E4, performing 1×1 convolution on the output features, and reducing the number of channels to 1 respectively to generate the distance information prediction map Dimg'.

7. The method for clinical target segmentation of cervical cancer guided by anatomical prior position information according to claim 6, characterized in that: The information attention module DIA introduces the edge position information of the distance information decoder as a strong clue to enhance the edge part of the segmentation feature, including using 4 DIA modules to process the features E1, E2, E3, and E4 of stages 1 to 4 obtained by the shared encoder; thereby generating corresponding distance information enhanced features A1, A2, A3, and A4 of 4 different stages; include, Each DIA module has two inputs, namely the distance information prediction map Dimg' and the features Ei (i∈{1, 2, 3, 4}) from stage 1 to 4 obtained by the shared encoder; Dimg' is max-pooled and downsampled to the same spatial size as Ei, and an element-wise product is performed between Dimg' and Ei, and then fused with Ei through a feature parallel operation Concat to generate the initial distance enhancement feature, which is used by E' i express; E' i =C(M(Dimg')*Ei),Ei) (2) In the formula, M and * are Maxpool and element-wise product operations respectively, and C is the concatenation operation Concat; The spatial attention mechanism is introduced to improve the recognition ability of cervical cancer target areas in images. That is, features are extracted through a 1×1 convolution layer, and then the ReLU activation function is applied to introduce nonlinear transformation, and then the local spatial features are further refined through a 3×3 convolution layer. Finally, the spatial attention map Es is output through the Sigmoid function, which reflects the importance weight of each spatial position. The spatial attention map Es is element-wise multiplied with the input feature Ei to enhance the spatial features and obtain the enhanced feature A: A=E s *AND' i (3) The feature A is equally divided into two parts in the channel dimension, denoted as Ad and As; Ad and As correspond to the distance information feature and the segmentation key area feature respectively; the feature Ei is added element by element with Ad and As to obtain the final enhanced feature Ai: Oh=A d +A s +E i (4) At this point, the features E1, E2, E3, and E4 of stages 1 to 4 obtained by the shared encoder are used by the DIA module to generate the corresponding distance information enhanced features A1, A2, A3, and A4 of four different stages.

8. The method for clinical target segmentation of cervical cancer guided by anatomical prior position information according to claim 7, characterized in that: The target CTV segmentation decoder restores the multi-level features extracted by the encoder to the original image size through upsampling and feature fusion, and accurately classifies each pixel to achieve cervical cancer target area segmentation. Including, the segmentation decoder obtains the distance information enhanced features A1, A2, A3 and A4, performs 1×1 convolution on the output features, and reduces the number of channels to 2 to generate the target CTV segmentation prediction map.

9. The method for clinical target segmentation of cervical cancer guided by anatomical prior position information according to claim 8, characterized in that: The method of using the benchmark CTV and Dimg as references for training MTF-Net includes collecting CT image data and benchmark CTV of 418 cervical cancer patients, and randomly dividing the data set into a training set, a validation set and a test set at a ratio of 7:1:2; the training data set and the validation data set include CT image data, benchmark CTV and distance information image data Dimg, which are used to train and verify the dual-branch network model; the test data set includes CT image data and benchmark CTV, which are used to test the segmentation effect of the network; and during the training, the target CTV segmentation prediction map and the distance information prediction map Dimg' generated by MTF-Net are respectively subjected to Loss loss operation with the benchmark CTV data and Dimg, so that the target CTV segmentation prediction map and the distance information prediction map Dimg' gradually approach the benchmark CTV data and Dimg; Including, using Dice loss L dice , cross entropy loss L ce And the mean absolute error loss L mae Train MTF-Net and combine the three loss functions to get the fusion loss function as follows: Loss=L dice (S ctv ,G ctv )+0.8L ce (S ctv ,G ctv )+0.2L mae (S dia ,G dia ) (4) Where S ctv , G ctv are the segmentation prediction value of CTV and the benchmark CTV, Sdia and Gdia are the distance information prediction value and Dimg respectively; the definitions of these three loss functions are as follows: Where n is the number of voxels in the input CT image data, sj represents the predicted probability that the jth voxel belongs to the CTV, and gj represents the voxel value of the corresponding benchmark CTV. At the same time, the validation dataset is used to evaluate, optimize and select the best parameters for MTF-Net to ensure the generalization of the model.

Images