Automatic sketching method and device for radiotherapy structure based on 2D data

Abstract

The invention discloses an automatic sketching method and device for a radiotherapy structure based on 2D data, and belongs to the field of medical image processing, and the method comprises the steps: S1, collecting a patient data set comprising an X-ray image and a paired CT image; s2, a spatial dimension expansion generation model from the x-ray image to the CT image is constructed, and the spatial dimension expansion generation model comprises an encoder, a spatial expansion module and a decoder; s3, training the spatial dimension expansion generation model in the step S2 by using the X-ray image acquired in the step S1 and the paired CT image, and adjusting the weight through iteration to minimize CT reconstruction loss and CT projection loss; and S4, training the basic sketching model by using the X-ray image acquired in the step S1 and the paired CT image. The problems that the current adaptive radiotherapy technology depends on a 3D image, the cost is high, and the time of fractional treatment is long can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a method and device for automatically outlining an adaptive radiotherapy structure based on 2D online data of a patient. Background Art

[0002] In modern radiotherapy, accurate structural delineation is crucial to improving treatment efficacy. Adaptive radiotherapy structural delineation refers to the need to contour various target structures in radiotherapy, including adaptive radiotherapy, including various organs at risk (OARs), tumor target volumes (GTVs), clinical target volumes (CTVs), target volumes (PTVs), and various auxiliary structures, such as cavities, machine tools, etc.

[0003] Fractionation in adaptive radiotherapy: Traditional radiotherapy usually adopts fractionated treatment, that is, the total dose is divided into multiple smaller doses, with a certain interval between each treatment. In adaptive radiotherapy, the concept of fractionated treatment has been further developed. Specifically, before each fractionated treatment, the doctor will re-evaluate the patient's anatomy through imaging technology (such as CT, CBCT, MRI or PET) to determine whether the location of the tumor has changed. If changes are detected, the treatment plan will be adjusted to adapt to the new situation to ensure that each fractionated dose can accurately irradiate the target area.

[0004] Traditional methods usually rely on pre-treatment planning CT for structural outlining. However, during adaptive radiotherapy, the anatomical changes of various organs or tissues in the patient's body need to be considered, and the outlining results need to be updated in each treatment fraction. Existing solutions often rely on MR, CT, CBCT or PET scans. These images are not only expensive to capture, but also take a long time to scan, which seriously limits the promotion and application of adaptive radiotherapy.

[0005] Automatic structural delineation is a key link in adaptive radiotherapy, and some solutions are currently available. For example, Shandong Cancer Hospital proposed an MRI-based deep learning automatic segmentation model in 2022 ([1] Li Z, Zhang W, Li B. Patient-specific daily updated deep learning auto-segmentation for MRI-guided adaptive radiotherapy. Radiother Oncol. 2022 Dec; 177: 222-230.), which first uses diagnostic MRI to train a basic automatic delineation model, and then uses each patient's completed fractionated MRI and delineation results to fine-tune the model, obtaining a customized delineation model for the patient that is updated daily.

[0006] In addition, Huiruan Technology has proposed an automatic outlining method for registration guidance in adaptive radiotherapy (patent publication number CN117911462A). This method performs image registration on the original planned CT image and the fractionated images, and uses the registered CT image, the outlining information after registration mapping, and the fractionated images as inputs to the automatic outlining model to complete the training of the automatic outlining model and obtain the predicted contour. Although there is no restriction on whether the fractionated images can be MR, CT, or CBCT, the technical solution limits the fractionated images to three-dimensional images.

[0007] In 2023, Peking Union Medical College Hospital proposed an effective segmentation method based on domain adaptation ([3]Liu Y, Yang B, Chen X. Efficient segmentation using domain adaptation for MRI-guidedand CBCT-guided online adaptive radiotherapy. Radiother Oncol. 2023Nov; 188: 109871.), which can be used in MRI-guided or CBCT-guided online adaptive radiotherapy. For CBCT-guided online adaptive radiotherapy, this method first uses all planning CTs to train a general segmentation model, and then uses the planning CT of each patient to fine-tune the model to obtain a customized model for each patient. During fractionated treatment, CycleGAN is used to generate sCT from CBCT, and then sCT is input into the customized model to predict the delineation results of each fraction. The same idea is used in MR-guided adaptive radiotherapy, except that the fractionated images are replaced from MR to CBCT.

[0008] In adaptive radiotherapy, the time spent on each fraction has a great impact on the final treatment effect. The shorter the time, the more obvious the advantage of adaptive radiotherapy technology. The automatic structure delineation methods realized by the above technologies all rely on three-dimensional images such as MR or CT, CBCT, etc. The scanning of these three-dimensional images is very time-consuming, which seriously restricts the improvement of the efficiency of fractionated treatment. In addition, the problem of expensive shooting costs is magnified under multiple fractions in one course of treatment, which makes it unaffordable for many patients.

[0009] The information disclosed in this background technology section is only intended to enhance the understanding of the overall background of the invention and should not be regarded as an acknowledgment or any form of suggestion that the information constitutes the prior art already known to a person skilled in the art. Summary of the invention

[0010] The purpose of the present invention is to provide a method and device for automatically outlining radiotherapy structures based on 2D data, which can solve the problems of current adaptive radiotherapy technology relying on 3D images, high cost and long time for fractionated treatment. 2D x-ray images with low shooting cost and high speed can be used to reconstruct sCT images, which can then be used for automatic outlining of treatment structures and fractionated treatment.

[0011] To achieve the above object, the present invention provides a method for automatically outlining a radiotherapy structure based on 2D data, characterized in that it comprises the following steps:

[0012] S1: Acquire a patient dataset including X-ray images and paired CT images;

[0013] S2: constructing a spatial dimension expansion generation model from an X-ray image to a CT image, wherein the spatial dimension expansion generation model includes an encoder, a spatial expansion module, and a decoder;

[0014] S3: using the X-ray image and paired CT image acquired in step S1, the spatial dimension expansion generative model in step S2 is trained, and the weights are iteratively adjusted to minimize the CT reconstruction loss and the CT projection loss;

[0015] S4: Use the X-ray image and paired CT image acquired in step S1 to train the basic delineation model.

[0016] In one embodiment of the present invention, the encoder is composed of 5 2D convolution modules; each 2D convolution module includes 2 2D convolution layers, 1 normalization layer and 1 Relu activation layer; each 2D convolution module is connected to a 2D convolution layer with a stride of 2;

[0017] The decoder is composed of 4 3D convolution modules; each 3D convolution module includes 2 3D convolution layers, 1 normalization layer and 1 Relu activation layer; each 3D convolution module is connected to a 3D deconvolution layer;

[0018] The spatial expansion module is composed of a series of 1 2D convolution layer, 1 normalization layer, 1 dimension expansion layer, 1 3D convolution layer, 1 normalization layer and 1 Relu activation layer.

[0019] In one embodiment of the present invention, the calculation formulas for the CT reconstruction loss and the CT projection loss are respectively the following formula (1) and formula (2):

[0020]

[0021] Where y is the CT image acquired in step S1; G(x) is the sCT image generated by the spatial dimension expansion generative model in step S2; x is a pixel in the image; Ω represents the CT image spatial domain; p ax (y), p co (y), p sa (y) are the projections of the CT images in the cross-sectional, coronal and sagittal planes respectively; p ax (G(x)),p co (G(x)),p sa (G(x)) are the projections of the sCT image in the cross-sectional, coronal and sagittal planes respectively; loss re is the CT reconstruction loss, loss pr It is CT projection loss.

[0022] In one embodiment of the present invention, the following steps are also included:

[0023] S5: In each fraction, take an X-ray image, input the spatial dimension expansion generation model trained in step S4, and generate an sCT image; input the generated sCT image into the basic delineation model trained in step S5, and output the automatic delineation result of the current fraction.

[0024] In one embodiment of the present invention, the following steps are also included:

[0025] S6: After the fractionated treatment is completed, the sCT images obtained in the current fraction and the automatic delineation results are added to the online database of the patient, and the basic delineation model is trained using the online data to obtain a customized segmentation model for the patient.

[0026] The present invention also provides an automatic delineation device for radiotherapy structures based on 2D data, comprising:

[0027] A data acquisition module for acquiring a patient data set including an X-ray image and a paired CT image;

[0028] A dimension expansion model construction module, used to construct a spatial dimension expansion generation model from an X-ray image to a CT image, wherein the spatial dimension expansion generation model includes an encoder, a spatial expansion module and a decoder;

[0029] A dimension expansion model training module, used to train a spatial dimension expansion generation model using the acquired X-ray images and paired CT images, and iteratively adjust weights to minimize CT reconstruction loss and CT projection loss;

[0030] The delineation model training module is used to train the basic delineation model using the acquired X-ray images and paired CT images.

[0031] In one embodiment of the present invention, the encoder is composed of 5 2D convolution modules; each 2D convolution module includes 2 2D convolution layers, 1 normalization layer and 1 Relu activation layer; each 2D convolution module is connected to a 2D convolution layer with a stride of 2;

[0032] The decoder is composed of 4 3D convolution modules; each 3D convolution module includes 2 3D convolution layers, 1 normalization layer and 1 Relu activation layer; each 3D convolution module is connected to a 3D deconvolution layer;

[0033] The spatial expansion module is composed of a series of 1 2D convolution layer, 1 normalization layer, 1 dimension expansion layer, 1 3D convolution layer, 1 normalization layer and 1 Relu activation layer.

[0034] In one embodiment of the present invention, the calculation formulas for the CT reconstruction loss and the CT projection loss are respectively the following formula (1) and formula (2):

[0035]

[0036] Where y is the acquired CT image; G(x) is the sCT image generated by the spatial dimension expansion generative model; x is a pixel in the image; Ω represents the CT image spatial domain; p ax (y), p co (y), p sa (y) are the projections of the CT images in the cross-sectional, coronal and sagittal planes respectively; p ax (G(x)),p co (G(x)),p sa (G(x)) are the projections of the sCT image in the cross-sectional, coronal and sagittal planes respectively; loss re is the CT reconstruction loss, loss pr It is CT projection loss.

[0037] In one embodiment of the present invention, it also includes:

[0038] The automatic delineation module is used to take X-ray images in each fraction, input the trained spatial dimension expansion generation model, and generate sCT images; input the generated sCT images into the trained basic delineation model, and output the automatic delineation results of the current fraction.

[0039] In one embodiment of the present invention, it also includes:

[0040] The customized segmentation module is used to add the sCT images obtained in the current fraction and the automatic delineation results to the patient's online database after the fractionated treatment is completed, and use the online data to train the basic delineation model to obtain a customized segmentation model for the patient.

[0041] Compared with the prior art, the method and device for automatically outlining radiotherapy structures based on 2D data according to the present invention reconstructs 3D CT images based on 2D X-ray images taken by the patient in each fraction of adaptive radiotherapy, and then uses the trained outlining model to automatically outline the structure, which reduces the scanning time of the fractionated images and greatly reduces the cost of adaptive radiotherapy technology. In addition, after each fraction of treatment, the outlining model is fine-tuned and trained using the patient's online data, so that the outlining model continues to learn online and continuously improves the accuracy of structure outlining. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is a flow chart of a method for automatically outlining a radiotherapy structure based on 2D data according to an embodiment of the present invention;

[0043] Figure 2 2 is a schematic diagram of the structure of a dimensionally extended generative network DEGNet according to an embodiment of the present invention;

[0044] Figure 3 It is a schematic structural diagram of an automatic delineation device for radiotherapy structures based on 2D data according to one embodiment of the present invention. DETAILED DESCRIPTION

[0045] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific implementation modes.

[0046] Unless explicitly stated otherwise, throughout the specification and claims, the term “comprise” or variations such as “include” or “comprising”, etc., will be understood to include the stated elements or components but not to exclude other elements or components.

[0047] like Figure 1 to Figure 2 As shown, according to a preferred embodiment of the present invention, a method for automatically outlining a radiotherapy structure based on 2D online data of a patient specifically comprises the following steps:

[0048] S1: Acquire a patient dataset including X-ray images and paired CT images.

[0049] Specifically, before starting adaptive radiotherapy, patient data with paired X-ray images and CT images are collected from various hospitals. If there is no corresponding structural contour on the CT, the CT needs to be structurally outlined first for training the outline model.

[0050] S2: Construct a spatial dimension expansion generation model from X-ray images to CT images, wherein the spatial dimension expansion generation model includes an encoder, a spatial expansion module and a decoder.

[0051] In this paper, a spatial dimension expansion generation model DEGNet (Dimension Expansion Generation Network) is proposed, and the network structure is as follows Figure 2 As shown in Figure 2, the spatial dimension expansion generation model DEGNet adds a spatial expansion module DEx to the basic encoding-decoding network structure to realize the conversion of 2D features to 3D features.

[0052] In common 2D-2D or 3D-3D generation models, the spatial dimensions between the encoding module and the corresponding decoding module at each level are consistent, and feature fusion can be performed directly by addition or channel splicing. However, the feature dimensions of x-ray to CT generation are inconsistent. Therefore, in the present invention, a spatial expansion module is introduced between the 2D encoding module and the corresponding 3D encoding module at each level to convert 2D features into 3D features.

[0053] Specifically, the input of the spatial dimension expansion generation model DEGNet is a 2D x-ray image, and the output is a 3D CT image.

[0054] The encoder part of the network DEGNet is mainly composed of 5 2D convolution modules, which are used to extract features of different scales of x-rays. Among them, each 2D convolution module includes 2 2D convolution layers, 1 normalization layer and 1 Relu activation layer. Each 2D convolution module is followed by a 2D convolution layer with a stride of 2, which is used to compress the feature size and increase the number of feature channels, that is, for downsampling.

[0055] The decoder part of the network DEGNet is mainly composed of 4 3D convolution modules, each of which includes 2 3D convolution layers, 1 normalization layer and 1 Relu activation layer. Each 3D convolution module is followed by a 3D deconvolution layer to upsample the features and compress the number of feature channels.

[0056] The spatial expansion module DEx (Dimension Expand module) of the network DEGNet is mainly composed of a series of 1 2D convolution layer, 1 normalization layer, 1 dimension expansion layer, 1 3D convolution layer, 1 normalization layer and 1 Relu activation layer. The dimension expansion layer is implemented by feature replication. The main function of the spatial expansion module DEx is to better utilize the features extracted from the x-ray. Before entering the 3D convolution module, the deconvolution upsampling result is feature spliced ​​with the encoder output at the same scale.

[0057] The encoder, spatial extension module DEx and decoder are connected in sequence to form a complete spatial dimension extension generation model DEGNet from input 2D X-ray images to output 3D CT images. Using the spatial dimension extension generation model DEGNet to convert 2D x-ray images into 3D CT images saves the scanning time of three-dimensional images such as MR\CT\CBCT\PET, greatly improves the efficiency of each fraction in adaptive radiotherapy, and alleviates the high cost of adaptive radiotherapy technology.

[0058] S3: Use the x-ray images and paired CT images collected in step S1 to train the spatial dimension expansion generation model in step S2, and iteratively adjust the weights to minimize the CT reconstruction loss and CT projection loss.

[0059] In this step, the adjusted weights refer to the trainable parameters of each layer of the entire model. The condition for stopping training is that the model converges and reaches a stable state, that is, the loss function does not decrease within several consecutive epochs (or iterations) (usually 3-5), then training can be stopped.

[0060] In this step S3, the general training method of the generative adversarial network GAN is used to train the spatial dimension expansion generative model DEGNet, in which the discriminator network Discriminator can be various common classification models, and its main task is to judge whether the input image is a real CT or an sCT image generated by DEGNet. Based on the commonly used generative adversarial loss, for the mapping from x-ray to CT, this step constrains the sCT generated by the model from multiple angles in order to enable DEGNet to learn the mapping from x-ray images to CT images more stably and efficiently. Based on the commonly used generative adversarial loss, the CT reconstruction loss loss is calculated by re and CT projection loss pr To improve the convergence speed of DEGNet, the specific calculation formulas are as follows:

[0061]

[0062] Where y is the CT image collected in step S1; G(x) is the sCT image generated by the spatial dimension expansion generative model DEGNet in step S2; x is a pixel in the image; Ω represents the CT image spatial domain, which refers to the entire CT image; p ax (y), p co (y), p sa (y) are the projections of the CT images in the cross-sectional, coronal and sagittal planes respectively; p ax (G(x)),p co (G(x)),p sa (G(x)) are the projections of the sCT image in the cross-sectional, coronal and sagittal planes, respectively.

[0063] During the neural network training process, loss re and loss pr They are used to calculate the mean square error between the generated sCT image and the real CT image in 2D projection and 3D volume, respectively, which greatly improves the convergence speed of DEGNet.

[0064] During the neural network training process, the loss function is used to converge the model through back propagation. The loss function measures the difference between the network's predicted output and the true label. Through iterative optimization, the network's weights are continuously adjusted to minimize this difference, so that the network can learn the mapping relationship between input data and output labels.

[0065] S4: Use the X-ray image and paired CT image acquired in step S1 to train the basic delineation model.

[0066] Among them, the basic contouring model ContourNet can adopt various common 3D segmentation network structures.

[0067] In this step, similar to the spatial dimension expansion generation model, the condition for stopping the training of the delineation model is that the model converges and reaches a stable state, that is, the model delineation accuracy does not improve within several consecutive (generally 3-5) epochs or iterations, then the training can be stopped. Alternatively, a convergence threshold can be set, that is, if the delineation model has no improvement in accuracy on the validation set within several consecutive (generally 3-5) epochs or iterations, then the training can be stopped.

[0068] S5: In each fraction, take an X-ray image, input it into the spatial dimension expansion generation model trained in step S4, and generate an sCT image; input the generated sCT image into the basic delineation model trained in step S5, and output the automatic delineation result of the current fraction for this treatment.

[0069] S6: After the fractionated treatment is completed, the sCT images obtained in the current fraction and the automatic delineation results are added to the online database of the patient, and the basic delineation model is trained using the online data to obtain a customized segmentation model for the patient.

[0070] In step S6, after each treatment session, the basic delineation model can be fine-tuned and trained using the online data of each patient to obtain a customized delineation model for each patient. As the entire course of treatment progresses, the model continues to learn online, and the accuracy of automatic delineation becomes higher and higher.

[0071] Each patient only uses the universal contouring model for the first treatment. Then, an online database of the patient can be established, and the contouring model can be continuously fine-tuned using the patient's online data. As the course of treatment progresses, the patient's personal online database becomes more and more complete, and the automatic contouring accuracy of the customized model for the individual patient based on this continuous learning becomes higher and higher.

[0072] like Figure 3 As shown, according to a preferred embodiment of the present invention, an automatic delineation device for radiotherapy structures based on patient 2D online data specifically includes: a data acquisition module 1, a dimensionality expansion model construction module 2, a dimensionality expansion model training module 3, a delineation model training module 4, an automatic delineation module 5 and a customized segmentation module 6.

[0073] The data acquisition module 1 is used to acquire a patient data set including an X-ray image and a paired CT image.

[0074] Specifically, before starting adaptive radiotherapy, patient data with paired X-ray images and CT images are collected from various hospitals. If there is no corresponding structural contour on the CT, the CT needs to be structurally outlined first for training the outline model.

[0075] The dimension expansion model construction module 2 is used to construct a spatial dimension expansion generation model from an X-ray image to a CT image, wherein the spatial dimension expansion generation model includes an encoder, a spatial expansion module and a decoder.

[0076] Specifically, Figure 2 As shown in Figure 3, the input of the spatial dimension expansion generation model DEGNet is a 2D x-ray image and the output is a 3D CT image.

[0077] The encoder part of the network DEGNet is mainly composed of 5 2D convolution modules, which are used to extract features of different scales of x-rays. Among them, each 2D convolution module includes 2 2D convolution layers, 1 normalization layer and 1 Relu activation layer. Each 2D convolution module is followed by a 2D convolution layer with a stride of 2, which is used to compress the feature size and increase the number of feature channels, that is, for downsampling.

[0078] The decoder part of the network DEGNet is mainly composed of 4 3D convolution modules, each of which includes 2 3D convolution layers, 1 normalization layer and 1 Relu activation layer. Each 3D convolution module is followed by a 3D deconvolution layer to upsample the features and compress the number of feature channels.

[0079] The spatial expansion module DEx (Dimension Expand module) of the network DEGNet is mainly composed of a series of 1 2D convolution layer, 1 normalization layer, 1 dimension expansion layer, 1 3D convolution layer, 1 normalization layer and 1 Relu activation layer. The dimension expansion layer is implemented by feature replication. The main function of the spatial expansion module DEx is to better utilize the features extracted from the x-ray. Before entering the 3D convolution module, the deconvolution upsampling result is feature spliced ​​with the encoder output at the same scale.

[0080] The dimension expansion model training module 3 is used to train the spatial dimension expansion generation model using the acquired X-ray images and paired CT images, and iteratively adjust the weights to minimize the CT reconstruction loss and CT projection loss.

[0081] The adjusted weights refer to the trainable parameters of each layer of the entire model. The condition for stopping training is that the model converges and reaches a stable state, that is, the loss function does not decrease for several consecutive (usually 3-5) epochs or iterations, then training can be stopped.

[0082] Specifically, the dimension expansion model training module 3 uses the general training method of the generative adversarial network GAN to train the spatial dimension expansion generative model DEGNet, in which the discriminator network Discriminator can be various common classification models, and its main task is to discriminate whether the input image is a real CT or an sCT image generated by DEGNet. On the basis of the commonly used generative adversarial loss, for the mapping from x-rays to CT, in order to enable DEGNet to learn the mapping from x-ray images to CT images more stably and efficiently, the sCT generated by the model is constrained from multiple angles. On the basis of the commonly used generative adversarial loss, the CT reconstruction loss loss is calculated by re and CT projection loss pr To improve the convergence speed of DEGNet, the specific calculation formulas are as follows:

[0083]

[0084] Where y is the acquired CT image; G(x) is the sCT image generated by the spatial dimension extension generation model DEGNet; x is a pixel in the image; Ω represents the CT image spatial domain, which refers to the entire CT image; p ax (y), p co (y), p sa (y) are the projections of the CT images in the cross-sectional, coronal and sagittal planes respectively; p ax (G(x)),p co (G(x)),p sa (G(x)) are the projections of the sCT image in the cross-sectional, coronal and sagittal planes, respectively.

[0085] During the neural network training process, loss re and loss pr They are used to calculate the mean square error between the generated sCT image and the real CT image in 2D projection and 3D volume, respectively, which greatly improves the convergence speed of DEGNet.

[0086] The delineation model training module 4 is used to train the basic delineation model using the acquired x-ray images and paired CT images.

[0087] Among them, the basic contouring model ContourNet can adopt various common 3D segmentation network structures.

[0088] Similar to the spatial dimension expansion generative model, the condition for stopping the training of the delineation model is that the model converges and reaches a stable state, that is, the model delineation accuracy does not improve within several consecutive (generally 3-5) epochs or iterations, then the training can be stopped. Alternatively, a convergence threshold can be set, that is, if the accuracy of the delineation model on the validation set does not improve within several consecutive (generally 3-5) epochs or iterations, then the training can be stopped.

[0089] The automatic delineation module 5 is used to take X-ray images in each fraction, input the trained spatial dimension expansion generation model, and generate sCT images; input the generated sCT images into the trained basic delineation model, and output the automatic delineation results of the current fraction for this treatment.

[0090] The customized segmentation module 6 is used to add the sCT image obtained in the current fractionation and the automatic delineation result to the online database of the patient after the fractionation treatment is completed, and train the basic delineation model using the online data to obtain a customized segmentation model for the patient.

[0091] After each treatment session, the customized segmentation module 6 can use the online data of each patient to fine-tune the basic delineation model to obtain a customized delineation model for each patient. As the entire course of treatment progresses, the model continues to learn online, and the accuracy of automatic delineation becomes higher and higher.

[0092] Each patient only uses the universal contouring model for the first treatment. Then, an online database of the patient can be established, and the contouring model can be continuously fine-tuned using the patient's online data. As the course of treatment progresses, the patient's personal online database becomes more and more complete, and the automatic contouring accuracy of the customized model for the individual patient based on this continuous learning becomes higher and higher.

[0093] The foregoing description of specific exemplary embodiments of the present invention is for the purpose of illustration and demonstration. These descriptions are not intended to limit the present invention to the precise form disclosed, and it is clear that many changes and variations can be made based on the above teachings. The purpose of selecting and describing the exemplary embodiments is to explain the specific principles of the present invention and its practical application, so that those skilled in the art can realize and utilize various different exemplary embodiments of the present invention and various different selections and changes. The scope of the present invention is intended to be limited by the claims and their equivalents.

Claims

1. A method for automatically outlining a radiotherapy structure based on 2D data, characterized in that: The following steps are involved: S1: Acquire a patient dataset including X-ray images and paired CT images; S2: constructing a spatial dimension expansion generation model from an X-ray image to a CT image, wherein the spatial dimension expansion generation model includes an encoder, a spatial expansion module, and a decoder; S3: using the X-ray image and paired CT image acquired in step S1, the spatial dimension expansion generative model in step S2 is trained, and the weights are iteratively adjusted to minimize the CT reconstruction loss and the CT projection loss; S4: Use the X-ray image and paired CT image acquired in step S1 to train the basic delineation model.

2. The method for automatically outlining a radiotherapy structure based on 2D data according to claim 1, characterized in that: The encoder is composed of 5 2D convolution modules; each 2D convolution module includes 2 2D convolution layers, 1 normalization layer and 1 Relu activation layer; each 2D convolution module is connected to a 2D convolution layer with a step size of 2; The decoder is composed of 4 3D convolution modules; each 3D convolution module includes 2 3D convolution layers, 1 normalization layer and 1 Relu activation layer; each 3D convolution module is connected to a 3D deconvolution layer; The spatial expansion module is composed of a series of 1 2D convolution layer, 1 normalization layer, 1 dimension expansion layer, 1 3D convolution layer, 1 normalization layer and 1 Relu activation layer.

3. The method for automatically outlining a radiotherapy structure based on 2D data according to claim 1, characterized in that: The calculation formulas of the CT reconstruction loss and the CT projection loss are respectively the following formula (1) and formula (2): Where y is the CT image acquired in step S1; G(x) is the sCT image generated by the spatial dimension expansion generative model in step S2; x is a pixel in the image; Ω represents the CT image spatial domain; p ax (y), p co (y), p sa (y) are the projections of the CT images in the cross-sectional, coronal and sagittal planes respectively; p ax (G(x)),p co (G(x)),p sa (G(x)) are the projections of the sCT image in the cross-sectional, coronal and sagittal planes respectively; loss re is the CT reconstruction loss, loss pr It is CT projection loss.

4. The method for automatically outlining a radiotherapy structure based on 2D data according to claim 1, characterized in that: The following steps are also included: S5: In each fraction, take an X-ray image, input the spatial dimension expansion generation model trained in step S4, and generate an sCT image; input the generated sCT image into the basic delineation model trained in step S5, and output the automatic delineation result of the current fraction.

5. The method for automatically outlining a radiotherapy structure based on 2D data according to claim 4, characterized in that: The following steps are also included: S6: After the fractionated treatment is completed, the sCT images obtained in the current fraction and the automatic delineation results are added to the online database of the patient, and the basic delineation model is trained using the online data to obtain a customized segmentation model for the patient.

6. An automatic delineation device for radiotherapy structures based on 2D data, characterized in that: include: A data acquisition module for acquiring a patient data set including an X-ray image and a paired CT image; A dimension expansion model construction module, used to construct a spatial dimension expansion generation model from an X-ray image to a CT image, wherein the spatial dimension expansion generation model includes an encoder, a spatial expansion module and a decoder; A dimension expansion model training module, used to train a spatial dimension expansion generation model using the acquired X-ray images and paired CT images, and iteratively adjust weights to minimize CT reconstruction loss and CT projection loss; The delineation model training module is used to train the basic delineation model using the acquired X-ray images and paired CT images.

7. The automatic delineation device for radiotherapy structure based on 2D data according to claim 1, characterized in that: The encoder is composed of 5 2D convolution modules; each 2D convolution module includes 2 2D convolution layers, 1 normalization layer and 1 Relu activation layer; each 2D convolution module is connected to a 2D convolution layer with a step size of 2; The decoder is composed of 4 3D convolution modules; each 3D convolution module includes 2 3D convolution layers, 1 normalization layer and 1 Relu activation layer; each 3D convolution module is connected to a 3D deconvolution layer; The spatial expansion module is composed of a series of 1 2D convolution layer, 1 normalization layer, 1 dimension expansion layer, 1 3D convolution layer, 1 normalization layer and 1 Relu activation layer.

8. The automatic delineation device for radiotherapy structure based on 2D data according to claim 1, characterized in that: The calculation formulas of the CT reconstruction loss and the CT projection loss are respectively the following formula (1) and formula (2): Where y is the acquired CT image; G(x) is the sCT image generated by the spatial dimension expansion generative model; x is a pixel in the image; Ω represents the CT image spatial domain; p ax (y), p co (y), p sa (y) are the projections of the CT images in the cross-sectional, coronal and sagittal planes respectively; p ax (G(x)),p co (G(x)),p sa (G(x)) are the projections of the sCT image in the cross-sectional, coronal and sagittal planes respectively; loss re is the CT reconstruction loss, loss pr It is CT projection loss.

9. The automatic delineation device for radiotherapy structure based on 2D data according to claim 1, characterized in that: Also includes: The automatic delineation module is used to take X-ray images in each fraction, input the trained spatial dimension expansion generation model, and generate sCT images; input the generated sCT images into the trained basic delineation model, and output the automatic delineation results of the current fraction.

10. The automatic delineation device for radiotherapy structure based on 2D data according to claim 9, characterized in that: Also includes: The customized segmentation module is used to add the sCT images obtained in the current fraction and the automatic delineation results to the patient's online database after the fractionated treatment is completed, and use the online data to train the basic delineation model to obtain a customized segmentation model for the patient.

Images