A DVV curve generation method, device, equipment and readable storage medium

Abstract

A DVV curve generation method, device, equipment and readable storage medium, including obtaining a planned radiation dose corresponding to a target risk organ, generating multiple target radiation doses based on a dose gradient coefficient and the planned radiation dose; performing feature extraction on a pre-radiotherapy CT image and a risk organ dose image corresponding to the target radiation dose to generate a first deformation field; performing feature extraction on a pre-radiotherapy CT image and a non-risk organ dose image corresponding to the target radiation dose to generate a second deformation field; generating a target deformation field after fusing the first deformation field and the second deformation field, performing spatial transformation on a pre-radiotherapy target risk organ mask corresponding to the target radiation dose based on the target deformation field to output a post-radiotherapy target risk organ mask; determining a target atrophy amount based on the post-radiotherapy and pre-radiotherapy target risk organ masks; generating a DVV curve based on the correspondence between the target atrophy amount and the dose gradient coefficient, so as to objectively and accurately obtain the correspondence between the dose and the risk organ atrophy.

Description

Technical Field

[0001] The present application relates to the technical field of medical radiotherapy, and in particular to a DVV curve generation method, apparatus, device and readable storage medium. Background Art

[0002] Nasopharyngeal carcinoma (NPC) is one of the most common cancers. Among them, radiotherapy is the main method for treating NPC. However, since radiation can cause deformation of the gross tumor volume (GTV) and risk organs (OARs) including the parotid gland, liver, and lungs, it is crucial to accurately understand the correspondence between radiation dose and risk organ atrophy. In related technologies, the correspondence between radiation dose and risk organ atrophy is often subjectively judged based on the clinical experience of doctors. This not only places high demands on and is highly dependent on the clinical experience of doctors, but also because different doctors have different levels of experience, it is impossible to ensure the accuracy of the judgment results.

[0003] It can be seen that how to objectively and accurately depict the correspondence between radiation dose and risk organ atrophy on the basis of reducing dependence on high-level clinical experience is an issue that needs to be urgently addressed. Summary of the Invention

[0004] The present application provides a DVV curve generation method, device, equipment and readable storage medium, which can solve the technical problems existing in the prior art of strong dependence on high-level clinical experience and poor accuracy of judgment results.

[0005] In a first aspect, an embodiment of the present application provides a method for generating a DVV curve, the method comprising:

[0006] Acquiring a planned radiation dose corresponding to a target risk organ, and generating a plurality of target radiation doses based on a preset dose gradient coefficient and the planned radiation dose, wherein a mapping relationship exists between the dose gradient coefficient and the target radiation dose;

[0007] For each target radiation dose, the pre-radiotherapy CT image and the risk organ dose image corresponding to the target radiation dose are input into a preset target prediction network for feature extraction to generate a first deformation field;

[0008] Inputting the pre-radiotherapy CT image and the non-risk organ dose image corresponding to the target radiation dose into the target prediction network for feature extraction to generate a second deformation field;

[0009] The first deformation field and the second deformation field are fused through the target prediction network to generate a target deformation field, and a pre-radiotherapy target risk organ mask corresponding to the target radiation dose is spatially transformed based on the target deformation field to predict and output a post-radiotherapy target risk organ mask;

[0010] The target atrophy amount corresponding to the target radiation dose is determined based on the target risk organ mask after radiotherapy and the target risk organ mask before radiotherapy;

[0011] After the target atrophy amounts corresponding to all target radiation doses are output, a dose-volume variation curve DVV corresponding to the target risk organ is generated based on the correspondence between all target atrophy amounts and dose gradient coefficients.

[0012] In conjunction with the first aspect, in one embodiment, the risk organ dose image includes a left risk organ dose image and a right risk organ dose image, and the pre-radiotherapy CT image and the risk organ dose image corresponding to the target radiation dose are input into a preset target prediction network for feature extraction to generate a first deformation field, including:

[0013] Inputting the pre-radiotherapy CT image and the left risk organ dose image into the target prediction network for feature extraction to generate a left deformation field;

[0014] Inputting the pre-radiotherapy CT image and the right-side risk organ dose image into the target prediction network for feature extraction to generate a right-side deformation field;

[0015] The left deformation field and the right deformation field are used as the first deformation field.

[0016] In conjunction with the first aspect, in one embodiment, fusing the first deformation field and the second deformation field through the target prediction network to generate a target deformation field includes:

[0017] The left deformation field, the right deformation field, and the second deformation field are fused through the target prediction network to generate a target deformation field.

[0018] In conjunction with the first aspect, in one embodiment, before the step of inputting the risk organ dose image corresponding to the target radiation dose and the pre-radiotherapy CT image into a preset target prediction network for feature extraction, the method further includes:

[0019] Acquiring an image training set, the image training set including a pre-radiotherapy CT image and its corresponding radiation dose image, and a pre-radiotherapy risk organ mask, the radiation dose image including a radiation dose image with a zero dose and a radiation dose image with a non-zero dose, and the radiation dose image including a risk organ dose image and a non-risk organ dose image;

[0020] Constructing a fully convolutional neural network, the fully convolutional neural network includes an input layer, a backbone network, a spatial transformation layer, and an output layer. The backbone network is used to extract features from the input image of the input layer to generate a deformation field. The spatial transformation layer is used to perform spatial transformation processing on the deformation field output by the backbone network. The output layer is used to output the processing results of the spatial transformation layer.

[0021] The fully convolutional neural network is trained using the image training set to obtain a target prediction network.

[0022] In a second aspect, an embodiment of the present application provides a DVV curve generation device, the DVV curve generation device comprising: a processing module, a prediction module, a determination module, and a generation module;

[0023] The processing module is used to obtain a planned radiation dose corresponding to a target risk organ, and generate a plurality of target radiation doses based on a preset dose gradient coefficient and the planned radiation dose, wherein a mapping relationship exists between the dose gradient coefficient and the target radiation dose;

[0024] For each target radiation dose, the prediction module is used to input the pre-radiotherapy CT image and the risk organ dose image corresponding to the target radiation dose into a preset target prediction network for feature extraction to generate a first deformation field; input the pre-radiotherapy CT image and the non-risk organ dose image corresponding to the target radiation dose into the target prediction network for feature extraction to generate a second deformation field; fuse the first deformation field and the second deformation field through the target prediction network to generate a target deformation field, and perform spatial transformation on the pre-radiotherapy target risk organ mask corresponding to the target radiation dose based on the target deformation field, and predict and output the post-radiotherapy target risk organ mask;

[0025] The determination module is used to determine the target atrophy amount corresponding to the target radiation dose based on the target risk organ mask after radiotherapy and the target risk organ mask before radiotherapy;

[0026] The generation module is used to generate a dose volume change curve DVV corresponding to the target risk organ based on the corresponding relationship between all target atrophy amounts and dose gradient coefficients after outputting the target atrophy amounts corresponding to all target radiation doses.

[0027] In conjunction with the second aspect, in one embodiment, the risk organ dose image includes a left risk organ dose image and a right risk organ dose image, and the prediction module is specifically configured to:

[0028] Inputting the pre-radiotherapy CT image and the left risk organ dose image into the target prediction network for feature extraction to generate a left deformation field;

[0029] Inputting the pre-radiotherapy CT image and the right-side risk organ dose image into the target prediction network for feature extraction to generate a right-side deformation field;

[0030] The left deformation field and the right deformation field are used as the first deformation field.

[0031] In conjunction with the second aspect, in one embodiment, the prediction module is further configured to:

[0032] The left deformation field, the right deformation field, and the second deformation field are fused through the target prediction network to generate a target deformation field.

[0033] In conjunction with the second aspect, in one embodiment, the apparatus further includes a training module, which is configured to:

[0034] Acquiring an image training set, the image training set including a pre-radiotherapy CT image and its corresponding radiation dose image, and a pre-radiotherapy risk organ mask, the radiation dose image including a radiation dose image with a zero dose and a radiation dose image with a non-zero dose, and the radiation dose image including a risk organ dose image and a non-risk organ dose image;

[0035] Constructing a fully convolutional neural network, the fully convolutional neural network includes an input layer, a backbone network, a spatial transformation layer, and an output layer. The backbone network is used to extract features from the input image of the input layer to generate a deformation field. The spatial transformation layer is used to perform spatial transformation processing on the deformation field output by the backbone network. The output layer is used to output the processing results of the spatial transformation layer.

[0036] The fully convolutional neural network is trained using the image training set to obtain a target prediction network.

[0037] In a third aspect, an embodiment of the present application provides a DVV curve generation device, comprising a processor, a memory, and a DVV curve generation program stored in the memory and executable by the processor, wherein when the DVV curve generation program is executed by the processor, the steps of the aforementioned DVV curve generation method are implemented.

[0038] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, on which a DVV curve generation program is stored. When the DVV curve generation program is executed by a processor, the steps of the aforementioned DVV curve generation method are implemented.

[0039] The beneficial effects of the technical solutions provided in the embodiments of the present application include:

[0040] By obtaining the planned radiation dose corresponding to the target risk organ, and generating multiple target radiation doses based on the preset dose gradient coefficient and the planned radiation dose; extracting features from the pre-radiotherapy CT image and the risk organ dose image through the target prediction network to generate a first deformation field; then extracting features from the pre-radiotherapy CT image and the non-risk organ dose image to generate a second deformation field; then fusing the first deformation field and the second deformation field to generate a target deformation field to break the false correlation between the non-risk organ dose and the risk organ atrophy, so that the finally calculated risk organ atrophy amount is only related to the risk organ dose, so as to improve the accuracy of the atrophy amount calculation; based on the target deformation field, the radiotherapy corresponding to the target radiation dose is performed. The mask of the target risk organ before radiotherapy is spatially transformed to predict and output the mask of the target risk organ after radiotherapy; then, the target atrophy amount corresponding to the target radiation dose is determined based on the target risk organ mask after radiotherapy and the target risk organ mask before radiotherapy; after the target atrophy amounts corresponding to all target radiation doses are output, a dose-volume change curve DVV corresponding to the target risk organ is generated based on the correspondence between all target atrophy amounts and the dose gradient coefficient, so that doctors can objectively and accurately learn the correspondence between radiation dose and risk organ atrophy from the DVV curve without relying on high-level clinical experience, thereby solving the technical problems in the existing technology that are highly dependent on high-level clinical experience and have poor accuracy in judgment results. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a flow chart of an embodiment of the DVV curve generation method of the present application;

[0042] Figure 2 A schematic diagram of a DVV curve according to an embodiment of the present application;

[0043] Figure 3 For this application Figure 1 Detailed flow chart of step S20;

[0044] Figure 4 Schematic diagram of the hardware structure of the DVV curve generation device involved in the embodiment of the present application. DETAILED DESCRIPTION

[0045] In order to enable those skilled in the art to better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0046] First, some technical terms in this application are explained to facilitate those skilled in the art to understand this application.

[0047] Voxel: It is the smallest unit of a 3D image. For example, a 100×100×100 CT image has 1,000,000 voxels.

[0048] Vector: A vector consists of two factors: direction and length. The coordinate system (x, y, z) can be regarded as a 3D vector.

[0049] Vector field: Many vectors clustered together form a vector field. A 3D vector field is formed by many 3D vectors clustered together. Each point in the 3D vector field has its own 3D vector.

[0050] Deformation Field (DVF): A 3D DVF is a vector field. The size of a 3D DVF is often the same as the size of the corresponding CT image, and the number of channels of the 3D deformation field is 3. For example, the vector field corresponding to a CT vector is 100×100×100×3. This means that each voxel on the CT image has a 3D vector that guides the voxel to move in the direction of the vector by the length of the vector. After the voxel moves, the shape of the object in the image changes.

[0051] Spurious correlation: When two factors are statistically correlated but not causally related, it is a spurious correlation.

[0052] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0053] In a first aspect, an embodiment of the present application provides a method for generating a DVV curve.

[0054] In one embodiment, referring to Figure 1 , Figure 1 This is a flow chart of an embodiment of the DVV curve generation method of this application. Figure 1 As shown, the DVV curve generation method includes:

[0055] Step S10: obtaining a planned radiation dose corresponding to a target risk organ, and generating a plurality of target radiation doses based on a preset dose gradient coefficient and the planned radiation dose, wherein a mapping relationship exists between the dose gradient coefficient and the target radiation dose.

[0056] For illustrative purposes, it should be noted that in this embodiment, risk organs include, but are not limited to, the parotid gland, liver, and lungs, which are prone to deformation. This embodiment uses the parotid gland as an example to illustrate the process and principles of DVV curve generation, with the target risk organ being the parotid gland. Furthermore, a preset dose gradient coefficient serves as the abscissa of the DVV curve. Its specific value can be determined based on actual needs, as long as it lies within the [0, 1] interval. For example, with a gradient of 0.2, the dose gradient coefficients are 0.0, 0.2, 0.4, 0.6, 0.8, and 1.0.

[0057] For each target parotid gland, there is a corresponding planned radiation dose, which is provided by the physician. When the relationship between the radiation dose and risk organ atrophy for the target parotid gland needs to be objectively and accurately depicted, the planned radiation dose is processed according to the dose gradient coefficient to obtain multiple target radiation doses that correspond to the dose gradient coefficient. For example, assuming the dose gradient coefficients are 0.0, 0.2, 0.4, 0.6, 0.8, and 1.0, and the planned radiation dose is x, the target radiation doses are 0, 0.2x, 0.4x, 0.6x, 0.8x, and x, respectively.

[0058] Step S20: For each target radiation dose, the pre-radiotherapy CT image and the risk organ dose image corresponding to the target radiation dose are input into a preset target prediction network for feature extraction to generate a first deformation field.

[0059] For example, in this embodiment, before generating the DDV curve, a neural network model that can be used to predict the post-radiotherapy CT image and the post-radiotherapy risk organ mask (i.e., the post-radiotherapy risk organ Mask) will be constructed, and the neural network model will be iteratively trained. The loss value is determined based on the qualified true value of the corresponding predicted value after training, and the parameters in the model are updated according to the loss value to obtain the final target prediction network.

[0060] Since the target prediction network uses the same principles in predicting the post-radiotherapy CT image and the post-radiotherapy risk organ mask corresponding to each target radiation dose, for the sake of simplicity, the following embodiment will explain the prediction process and principle of the post-radiotherapy risk organ mask for one of the target radiation doses.

[0061] Understandably, in deep learning, spurious correlations often exist between input and labels. For example, in image recognition tasks, images of camels are often accompanied by deserts. The camel in the image (part of the input) is statistically correlated with the camel label, and there's also a causal relationship. However, the desert in the image (another part of the input) is also statistically correlated with the camel label, but there's no causal relationship. Therefore, the correlation between the desert and camel labels is spurious, which leads to the problem that the model can't correctly identify a camel standing on grass.

[0062] There is often a problem of false correlation when predicting parotid atrophy. Specifically, the radiation dose can usually be divided into two parts: parotid dose and extraparotid dose (i.e., non-parotid dose); it should be understood that the parotid dose and parotid atrophy are not only statistically correlated, but also causally related, while the extraparotid dose and parotid atrophy are statistically correlated, but not causally related, so the extraparotid dose and parotid atrophy are falsely correlated. Among them, the reason for the statistical correlation between extraparotid dose and parotid atrophy is that the larger the tumor area, the greater the parotid dose tends to be, and the parotid gland will be affected by a larger dose, so the parotid dose and extraparotid dose are statistically correlated, and the parotid dose and parotid atrophy are statistically correlated, so the extraparotid dose is also statistically correlated with parotid atrophy.

[0063] It should be noted that false correlation can lead to the following problem for parotid gland atrophy: when only the parotid gland dose is reduced, that is, the extraparotid gland dose is not reduced, the predicted parotid gland atrophy does not decrease due to the statistical correlation between parotid gland atrophy and extraparotid gland dose. Therefore, it is very important to address the false correlation between extraparotid gland dose and parotid gland atrophy when predicting parotid gland atrophy, thereby effectively improving the accuracy of parotid gland atrophy prediction.

[0064] To eliminate the spurious correlation between extra-parotid dose and parotid atrophy, this embodiment uses a segmented dose transmission method to eliminate the statistical correlation between the parotid dose and the extra-parotid dose, and further eliminates the statistical correlation between the extra-parotid dose and parotid atrophy. This eliminates the necessary conditions for eliminating the spurious correlation, and thus eliminates the spurious correlation. Specifically, the dose image corresponding to the target radiation dose is divided into a parotid dose image and a non-parotid dose image. These images are then fed into the target prediction network for feature extraction. This ensures that the deformation field generated by the target prediction network is only related to the parotid dose or the non-parotid dose, thus preventing the non-parotid dose from influencing the generation of the parotid dose deformation field.

[0065] The extra-parotid dose portion in the parotid dose image is padded with 0, and then the parotid dose image and the pre-radiotherapy CT image are feature extracted through the target prediction network to generate a first deformation field, which is independent of the extra-parotid dose.

[0066] Step S30: inputting the pre-radiotherapy CT image and the non-risk organ dose image corresponding to the target radiation dose into the target prediction network for feature extraction to generate a second deformation field.

[0067] For example, in this embodiment, the parotid gland dose portion in the non-parotid gland dose image is padded with 0, and then the non-parotid gland dose image and the pre-radiotherapy CT image are feature extracted through the target prediction network to generate a second deformation field, so that the second deformation field is independent of the parotid gland dose.

[0068] Since the parotid gland dose portion in the non-parotid gland dose image is padded with 0, the parotid gland dose is always 0 regardless of how the non-parotid gland dose changes; similarly, the non-parotid gland dose portion in the parotid gland dose image is padded with 0, so that the extra-parotid gland dose is always 0 regardless of how the parotid gland dose changes, thereby breaking the statistical correlation between the parotid gland dose and the extra-parotid gland dose, thereby breaking the statistical correlation between the extra-parotid gland dose and parotid gland atrophy, and thus breaking the necessary conditions for false correlation.

[0069] Step S40: The first deformation field and the second deformation field are fused through the target prediction network to generate a target deformation field, and a pre-radiotherapy target risk organ mask corresponding to the target radiation dose is spatially transformed based on the target deformation field to predict and output a post-radiotherapy target risk organ mask.

[0070] For example, in this embodiment, since the false correlation between the extra-parotid dose and parotid atrophy is resolved, the first deformation field and the second deformation field are fused to generate a target deformation field through the target prediction network, and then the pre-radiotherapy target parotid mask corresponding to the target radiation dose is spatially transformed through the target deformation field, so that the post-radiotherapy target parotid mask that has broken the false correlation and is accurate can be predicted, so that parotid atrophy is only related to the parotid dose, and the predicted parotid atrophy degree is more sensitive to the change of the parotid dose.

[0071] It should be noted that the physical meaning of the spatial transformation in this embodiment is to move each voxel on the Mask or CT in the direction of the vector by the length of the vector, thereby obtaining a deformed Mask or image; the specific implementation method of the spatial transformation can be determined according to actual needs, as long as the condition of being differentiable (that is, the gradient in the back-propagation mechanism can pass through the spatial transformation matrix) is met, and is not limited here. For example, bilinear interpolation, trilinear interpolation and other methods can be selected as specific implementation methods of the spatial transformation.

[0072] Step S50: determining a target atrophy amount corresponding to a target radiation dose according to the target risk organ mask after radiotherapy and the target risk organ mask before radiotherapy.

[0073] For example, in this embodiment, after the target parotid gland mask after radiotherapy is predicted and output through spatial transformation, the target amount of parotid gland atrophy at the target radiation dose can be determined based on the predicted target parotid gland mask after radiotherapy and the target parotid gland mask before radiotherapy. Similarly, the amount of target parotid gland atrophy at various target radiation doses can also be predicted using the above methods and principles.

[0074] Step S60: After the target atrophy amounts corresponding to all target radiation doses are output, a dose-volume variation curve DVV corresponding to the target risk organ is generated based on the correspondence between all target atrophy amounts and dose gradient coefficients.

[0075] For example, in this embodiment, since there is a mapping relationship between the dose gradient coefficient and the target radiation dose, and a mapping relationship between the target radiation dose and the target atrophy amount, there is also a mapping relationship between the dose gradient coefficient and the target atrophy amount. Therefore, once the target prediction network completes the prediction of the target atrophy amount for the target parotid gland at each target radiation dose, i.e., after outputting the target atrophy amounts corresponding to 0, 0.2x, 0.4x, 0.6x, 0.8x, and x, respectively, a dose-volume variation curve DVV corresponding to the target organ at risk can be generated based on the correspondence between the target atrophy amount and the dose gradient coefficient.

[0076] For example, the target atrophy amounts corresponding to 0, 0.2x, 0.4x, 0.6x, 0.8x and x are y1, y2, y3, y4, y5 and y6 respectively, and there is a mapping relationship between the dose gradient coefficient 0 and y1, 0.2 and y2, 0.4 and y3, 0.6 and y4, 0.8 and y51, and 1.0 and y6; using the dose gradient coefficient as the horizontal coordinate and the residual organ volume after treatment corresponding to the target atrophy amount as the vertical coordinate, the following can be drawn: Figure 2 As can be understood, the DVV curve objectively and accurately depicts the relationship between radiation dose and risk organ atrophy, that is, the relationship between radiation dose and the remaining organ volume after treatment. This allows doctors to accurately determine the remaining organ volume after treatment corresponding to each radiation dose based on the DVV curve, thereby enabling them to formulate a more accurate radiotherapy plan.

[0077] It's important to note that the vertical axis of the DVV curve represents the residual organ volume after treatment, expressed as a ratio compared to the pre-treatment volume. For example, if the target parotid gland shrinks by p%, the vertical axis is (1-p%). Generally speaking, organs at risk (OARs) shrink after radiation exposure, so the vertical axis ranges from 0 to 1. The horizontal axis of the DVV curve represents a multiplier (OAR Dose Multiplier, also known as the dose gradient coefficient) between 0 and 1. Given a complete radiotherapy dose map, the dose to each organ at risk is proportionally reduced by the number on the horizontal axis until the dose to the organ at risk reaches 0, resulting in multiple target radiation doses. For example, a horizontal axis of 1 indicates that the dose to the organ at risk (i.e., the target radiation dose) is the planned radiation dose; a horizontal axis of 0.5 indicates that the dose to the organ at risk is 0.5 times the planned radiation dose.

[0078] It should be understood that for the DVV curve, on the horizontal axis, the physical meaning of only reducing the dose of the risk organs without reducing the dose outside the risk organs (including the target area) is to protect the risk organs while curing cancer, which corresponds to the doctors' commonly used strategies; and the physical meaning of the entire DVV curve is to predict the impact of the doctors' commonly used strategies for protecting risk organs on the risk organs after radiotherapy, so that doctors can objectively and accurately understand the correspondence between radiation dose and risk organ atrophy from the DVV curve without relying on high-level clinical experience, thereby solving the technical problems in the existing technology that are highly dependent on high-level clinical experience and have poor accuracy in judgment results.

[0079] Further, in one embodiment, see Figure 3 As shown, the risk organ dose image includes a left risk organ dose image and a right risk organ dose image. The pre-radiotherapy CT image and the risk organ dose image corresponding to the target radiation dose are input into a preset target prediction network for feature extraction to generate a first deformation field, including:

[0080] Step S201: inputting the pre-radiotherapy CT image and the left risk organ dose image into the target prediction network for feature extraction to generate a left deformation field;

[0081] Step S202: inputting the pre-radiotherapy CT image and the right risk organ dose image into the target prediction network for feature extraction to generate a right deformation field;

[0082] Step S203: taking the left deformation field and the right deformation field as the first deformation field.

[0083] For example, in this embodiment, not only are the dose images corresponding to the target radiation dose divided into parotid gland dose images and non-parotid gland dose images, but the parotid gland dose images are also divided into left and right parotid gland dose images. This prevents mutual influence between the left and right parotid gland doses, enabling the target prediction network to output a more accurate deformation field. Therefore, the target prediction network can first perform feature extraction on the pre-radiotherapy CT image and the left parotid gland dose image to generate the left deformation field. Feature extraction is then performed on the pre-radiotherapy CT image and the right parotid gland dose image to generate the right deformation field. The left and right deformation fields are then used as the first deformation field for deformation field fusion.

[0084] Furthermore, in one embodiment, fusing the first deformation field and the second deformation field through the target prediction network to generate a target deformation field includes:

[0085] The left deformation field, the right deformation field, and the second deformation field are fused through the target prediction network to generate a target deformation field.

[0086] Exemplarily, in this embodiment, since the formation of the left, right, and second deformation fields is solely related to the parotid gland dose at the corresponding location, they can accurately express the relationship between the radiation dose at the corresponding location and parotid gland atrophy. Therefore, the left, right, and second deformation fields can be fused via a target prediction network to generate a target deformation field. This target deformation field can then be used to spatially transform the pre-radiotherapy target parotid gland mask corresponding to the target radiation dose. This allows the prediction of an accurate post-radiotherapy target parotid gland mask that eliminates spurious correlations, ensuring that parotid gland atrophy is solely related to the parotid gland dose.

[0087] Furthermore, in one embodiment, before the step of inputting the risk organ dose image corresponding to the target radiation dose and the pre-radiotherapy CT image into a preset target prediction network for feature extraction, the method further includes:

[0088] Acquiring an image training set, the image training set including a pre-radiotherapy CT image and its corresponding radiation dose image, and a pre-radiotherapy risk organ mask, the radiation dose image including a radiation dose image with a zero dose and a radiation dose image with a non-zero dose, and the radiation dose image including a risk organ dose image and a non-risk organ dose image;

[0089] Constructing a fully convolutional neural network, the fully convolutional neural network includes an input layer, a backbone network, a spatial transformation layer, and an output layer. The backbone network is used to extract features from the input image of the input layer to generate a deformation field. The spatial transformation layer is used to perform spatial transformation processing on the deformation field output by the backbone network. The output layer is used to output the processing results of the spatial transformation layer.

[0090] The fully convolutional neural network is trained using the image training set to obtain a target prediction network.

[0091] For example, it's understandable that in existing nasopharyngeal carcinoma radiotherapy datasets, the average parotid gland dose is mostly between 20gy and 30gy. Therefore, predicting only data with an average parotid gland dose between 20gy and 30gy is not a problem. However, during the test of reducing the parotid gland dose, the average parotid gland dose gradually decreases to more than ten gy, then to single digits, and finally to zero. Therefore, since the current training dataset contains no data with single-digit or zero average parotid gland doses, the model performs relatively poorly when predicting data with lower average parotid gland doses, making it impossible to accurately predict the parotid gland mask after radiotherapy.

[0092] In this embodiment, to address the above-mentioned issues, a training strategy of segmented dose conduction and dose data enhancement is proposed. Specifically, a pre-radiotherapy parotid gland mask, a pre-radiotherapy CT image, and its corresponding radiation dose image are obtained; wherein the radiation dose image includes a parotid gland dose image and a non-parotid gland dose image after segmented conduction processing, and the radiation dose image also includes a radiation dose image corresponding to the original dose and a radiation dose image corresponding to a dose of 0 after data enhancement processing. That is, a data image with an average parotid gland dose of 0 is added, so that the original N data images become 2N data images to form an image training set.

[0093] For example, there are pre-radiotherapy CT image 1 and pre-radiotherapy CT image 2, and the radiation doses corresponding to pre-radiotherapy CT image 1 and pre-radiotherapy CT image 2 are x1 and x2 respectively, then a radiation dose image with a radiation dose of x1 and a radiation dose image with a radiation dose of 0 are generated for pre-radiotherapy CT image 1; similarly, a radiation dose image with a radiation dose of x2 and a radiation dose image with a radiation dose of 0 are generated for pre-radiotherapy CT image 2; and then each radiation dose image is mapped into a parotid gland dose image and a non-parotid gland dose image.

[0094] It can be understood that on the DVV curve, the original N data are all at the position of 1 on the horizontal axis, and the enhanced N data are all at the position of 0 on the horizontal axis, so that there is no problem of dose data imbalance at the two positions 0 and 1 on the DVV curve, so as to achieve the purpose of making the degree of parotid atrophy more sensitive to changes in parotid gland dose, so that the model can more accurately predict the parotid gland Mask after radiotherapy.

[0095] In this embodiment, a neural network model is constructed to predict post-radiotherapy CT images and post-radiotherapy parotid gland masks. This neural network model can be a fully convolutional neural network (i.e., DoseNet) based on the U-net structure (U-net is a network model suitable for medical image segmentation). The model aims to predict voxel displacement based on the two images, effectively utilizing the dependency between radiation dose and parotid gland volume change to generate images corresponding to different expected parotid gland deformations that adapt to the varying dose. It should be noted that the specific structure of the neural network model can be determined according to actual needs and is not limited here.

[0096] The fully convolutional neural network includes an input layer, a backbone network, a spatial transformation layer, and an output layer. The input layer is used for image input; the backbone network is used to extract contextual features from the input radiation dose image and pre-radiotherapy CT image to output a deformation field. It should be noted that the backbone network can be any segmented backbone network, such as a U-net, which uses the dependency between radiation dose and parotid gland volume change for deformation prediction; the spatial transformation layer is used to perform spatial transformation processing on the pre-radiotherapy parotid gland mask and pre-radiotherapy CT image based on the deformation field output by the backbone network to predict and output the post-radiotherapy parotid gland mask and post-radiotherapy CT image; the output layer is used to output the post-radiotherapy parotid gland mask and post-radiotherapy CT image.

[0097] When training a fully convolutional neural network using an image training set, for a pre-radiotherapy CT image, assuming its radiation dose is x1, its corresponding radiation dose images include the first left parotid gland dose image, the first right parotid gland dose image, and the first non-parotid gland dose image, all with a radiation dose of x1, as well as the second left parotid gland dose image, the second right parotid gland dose image, and the second non-parotid gland dose image, all with a radiation dose of 0. In this case, the first left parotid gland dose image must first be input into the target prediction network for feature extraction to obtain a deformation field. Similarly, the first right parotid gland dose image and the first non-parotid gland dose image are sequentially input into the target prediction network for feature extraction to obtain two deformation fields. The three deformation fields are then fused to obtain the final deformation field, which is then used to spatially transform the pre-radiotherapy parotid gland mask to predict the first post-radiotherapy parotid gland mask.

[0098] Similarly, the second left parotid gland dose image, the second right parotid gland dose image, and the second non-parotid gland dose image are sequentially input into the target prediction network for feature extraction to obtain three deformation fields. These three deformation fields are then fused to obtain the final deformation field. The pre-radiotherapy parotid gland mask is then spatially transformed using this final deformation field to predict the second post-radiotherapy parotid gland mask. Thus, this embodiment generates a post-radiotherapy parotid gland mask corresponding to the original dose and a post-radiotherapy parotid gland mask corresponding to zero dose for each pre-radiotherapy CT image.

[0099] It should be understood that in model training, when the parotid gland mask before radiotherapy is deformed by the deformation field to obtain the predicted parotid gland mask after radiotherapy, the predicted parotid gland mask after radiotherapy can be compared with the actual parotid gland mask after radiotherapy to determine the loss between the predicted parotid gland mask after radiotherapy and the actual parotid gland mask after radiotherapy, and the model parameters are updated according to the determined loss, and then iterative training is performed to minimize the loss of the model; it should be noted that the loss calculation can also be performed on the output result of the deformation field, and then the model parameters are updated and iterative training is performed according to the loss result to finally generate the target prediction network.

[0100] It should be understood that if dose segmentation and dose data enhancement are not performed, the deformation field output by the traditional fully convolutional neural network is: represents the deformation field, g represents the backbone network, θ represents the parameters in the backbone network, m represents the CT image before radiotherapy, and d represents the radiation dose image before radiotherapy. The loss is calculated based on the CT image before and after radiotherapy and the deformation field: f represents the true value of the CT image after radiotherapy.

[0101] After the dose segmentation transmission and dose data enhancement processing are performed in this embodiment, the deformation field output by the full convolutional neural network is: is the deformation field after data enhancement, K represents the number of blocks obtained after the dose is segmented (for example, if the radiation dose image is divided into the left parotid dose image, the right parotid dose image and the non-parotid dose image, then K = 3), s represents the mask, s m Represents the risk organ mask before radiotherapy, represents the kth block in the risk organ mask before radiotherapy; and its loss is calculated based on the CT image before radiotherapy and the deformation field: It can be seen that when calculating the loss of the enhanced data, this embodiment uses the pre-radiotherapy CT image (m) instead of the post-radiotherapy CT image (f) as the true value. That is, the loss in this embodiment is calculated by the pre-radiotherapy CT image, the pre-radiotherapy CT image and the deformation field, which effectively improves the prediction accuracy of the model.

[0102] In a second aspect, an embodiment of the present application further provides a DVV curve generating device.

[0103] In one embodiment, the DVV curve generating device includes: a processing module, a prediction module, a determination module, and a generation module;

[0104] The processing module is used to obtain a planned radiation dose corresponding to a target risk organ, and generate a plurality of target radiation doses based on a preset dose gradient coefficient and the planned radiation dose, wherein a mapping relationship exists between the dose gradient coefficient and the target radiation dose;

[0105] For each target radiation dose, the prediction module is used to input the pre-radiotherapy CT image and the risk organ dose image corresponding to the target radiation dose into a preset target prediction network for feature extraction to generate a first deformation field; input the pre-radiotherapy CT image and the non-risk organ dose image corresponding to the target radiation dose into the target prediction network for feature extraction to generate a second deformation field; fuse the first deformation field and the second deformation field through the target prediction network to generate a target deformation field, and perform spatial transformation on the pre-radiotherapy target risk organ mask corresponding to the target radiation dose based on the target deformation field, and predict and output the post-radiotherapy target risk organ mask;

[0106] The determination module is used to determine the target atrophy amount corresponding to the target radiation dose based on the target risk organ mask after radiotherapy and the target risk organ mask before radiotherapy;

[0107] The generation module is used to generate a dose volume change curve DVV corresponding to the target risk organ based on the corresponding relationship between all target atrophy amounts and dose gradient coefficients after outputting the target atrophy amounts corresponding to all target radiation doses.

[0108] Through this embodiment, doctors can objectively and accurately learn the correspondence between radiation dose and risk organ atrophy from the DVV curve without relying on high-level clinical experience, thereby solving the technical problems in the existing technology of strong dependence on high-level clinical experience and poor accuracy of judgment results.

[0109] Furthermore, in one embodiment, the risk organ dose image includes a left risk organ dose image and a right risk organ dose image, and the prediction module is specifically configured to:

[0110] Inputting the pre-radiotherapy CT image and the left risk organ dose image into the target prediction network for feature extraction to generate a left deformation field;

[0111] Inputting the pre-radiotherapy CT image and the right-side risk organ dose image into the target prediction network for feature extraction to generate a right-side deformation field;

[0112] The left deformation field and the right deformation field are used as the first deformation field.

[0113] Furthermore, in one embodiment, the prediction module is further configured to:

[0114] The left deformation field, the right deformation field, and the second deformation field are fused through the target prediction network to generate a target deformation field.

[0115] Furthermore, in one embodiment, the apparatus further includes a training module, which is configured to:

[0116] Acquiring an image training set, the image training set including a pre-radiotherapy CT image and its corresponding radiation dose image, and a pre-radiotherapy risk organ mask, the radiation dose image including a radiation dose image with a zero dose and a radiation dose image with a non-zero dose, and the radiation dose image including a risk organ dose image and a non-risk organ dose image;

[0117] Constructing a fully convolutional neural network, the fully convolutional neural network includes an input layer, a backbone network, a spatial transformation layer, and an output layer. The backbone network is used to extract features from the input image of the input layer to generate a deformation field. The spatial transformation layer is used to perform spatial transformation processing on the deformation field output by the backbone network. The output layer is used to output the processing results of the spatial transformation layer.

[0118] The fully convolutional neural network is trained using the image training set to obtain a target prediction network.

[0119] The functional implementation of each module in the above-mentioned DVV curve generation device corresponds to each step in the above-mentioned DVV curve generation method embodiment, and their functions and implementation processes are not repeated here one by one.

[0120] In a third aspect, an embodiment of the present application provides a DVV curve generating device, which may be a device with data processing capabilities, such as a personal computer (PC), a laptop computer, or a server.

[0121] Reference Figure 4 , Figure 4Schematic diagram of the hardware structure of the DVV curve generation device involved in the embodiment of the present application. In the embodiment of the present application, the DVV curve generation device may include a processor, a memory, a communication interface and a communication bus.

[0122] The communication bus may be of any type and is used to interconnect the processor, memory, and communication interface.

[0123] Communication interfaces include input / output (I / O) interfaces, physical interfaces, and logical interfaces, which interconnect components within the DVV curve generation device and connect the device to other devices (such as other computing devices or user devices). Physical interfaces can include Ethernet, fiber, or ATM interfaces; user devices can include displays and keyboards.

[0124] The memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0125] The processor may be a general-purpose processor that can call a DVV curve generation program stored in a memory and execute the DVV curve generation method provided in the embodiments of the present application. For example, the general-purpose processor may be a central processing unit (CPU). The method executed when the DVV curve generation program is called can be referenced in the various embodiments of the DVV curve generation method of the present application and will not be further described here.

[0126] Those skilled in the art will understand that Figure 4 The hardware structure shown in the figure does not constitute a limitation to the present application and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0127] In a fourth aspect, an embodiment of the present application also provides a computer-readable storage medium.

[0128] The readable storage medium of the present application stores a DVV curve generation program, wherein when the DVV curve generation program is executed by a processor, the steps of the DVV curve generation method described above are implemented.

[0129] The method implemented when the DVV curve generation program is executed can refer to the various embodiments of the DVV curve generation method of the present application, and will not be repeated here.

[0130] It should be noted that the serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0131] The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices. The terms "first", "second" and "third" are used to distinguish different objects, etc., and do not represent a sequence, nor do they limit the "first", "second" and "third" to different types.

[0132] In the description of the embodiments of this application, the words "exemplary," "for example," or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary," "for example," or "for example" in the embodiments of this application should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary," "for example," or "for example" is intended to present the relevant concepts in a concrete manner.

[0133] In the description of the embodiments of the present application, unless otherwise specified, “ / ” means or, for example, A / B can mean A or B; “and / or” in the text is merely a description of the association relationship of associated objects, indicating that three relationships may exist, for example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, in the description of the embodiments of the present application, “multiple” refers to two or more than two.

[0134] In some processes described in the embodiments of the present application, multiple operations or steps are included that appear in a specific order. However, it should be understood that these operations or steps may not be performed in the order in which they appear in the embodiments of the present application or may be performed in parallel. The sequence numbers of the operations are only used to distinguish between different operations, and the sequence numbers themselves do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations or steps may be performed in sequence or in parallel, and these operations or steps may be combined.

[0135] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, of course, it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for enabling a terminal device to execute the methods described in each embodiment of the present application.

[0136] The above are only preferred embodiments of the present application and do not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A DVV curve generation method, characterized in that: The DVV curve generation method includes: Acquiring a planned radiation dose corresponding to a target risk organ, and generating a plurality of target radiation doses based on a preset dose gradient coefficient and the planned radiation dose, wherein a mapping relationship exists between the dose gradient coefficient and the target radiation dose; For each target radiation dose, the pre-radiotherapy CT image and the risk organ dose image corresponding to the target radiation dose are input into a preset target prediction network for feature extraction to generate a first deformation field; Inputting the pre-radiotherapy CT image and the non-risk organ dose image corresponding to the target radiation dose into the target prediction network for feature extraction to generate a second deformation field; The first deformation field and the second deformation field are fused through the target prediction network to generate a target deformation field, and a pre-radiotherapy target risk organ mask corresponding to the target radiation dose is spatially transformed based on the target deformation field to predict and output a post-radiotherapy target risk organ mask; The target atrophy amount corresponding to the target radiation dose is determined based on the target risk organ mask after radiotherapy and the target risk organ mask before radiotherapy; After the target atrophy amounts corresponding to all target radiation doses are output, a dose-volume variation curve DVV corresponding to the target risk organ is generated based on the correspondence between all target atrophy amounts and dose gradient coefficients.

2. The DVV curve generation method according to claim 1, wherein: The risk organ dose image includes a left risk organ dose image and a right risk organ dose image. The pre-radiotherapy CT image and the risk organ dose image corresponding to the target radiation dose are input into a preset target prediction network for feature extraction to generate a first deformation field, including: Inputting the pre-radiotherapy CT image and the left risk organ dose image into the target prediction network for feature extraction to generate a left deformation field; Inputting the pre-radiotherapy CT image and the right-side risk organ dose image into the target prediction network for feature extraction to generate a right-side deformation field; The left deformation field and the right deformation field are used as the first deformation field.

3. The DVV curve generation method according to claim 2, wherein: The fusing the first deformation field and the second deformation field through the target prediction network to generate a target deformation field includes: The left deformation field, the right deformation field, and the second deformation field are fused through the target prediction network to generate a target deformation field.

4. The DVV curve generation method according to claim 1, wherein: Before the step of inputting the risk organ dose image corresponding to the target radiation dose and the pre-radiotherapy CT image into a preset target prediction network for feature extraction, the method further includes: Acquiring an image training set, the image training set including a pre-radiotherapy CT image and its corresponding radiation dose image, and a pre-radiotherapy risk organ mask, the radiation dose image including a radiation dose image with a zero dose and a radiation dose image with a non-zero dose, and the radiation dose image including a risk organ dose image and a non-risk organ dose image; Constructing a fully convolutional neural network, the fully convolutional neural network includes an input layer, a backbone network, a spatial transformation layer, and an output layer. The backbone network is used to extract features from the input image of the input layer to generate a deformation field. The spatial transformation layer is used to perform spatial transformation processing on the deformation field output by the backbone network. The output layer is used to output the processing results of the spatial transformation layer. The fully convolutional neural network is trained using the image training set to obtain a target prediction network.

5. A DVV curve generating device, characterized in that: The DVV curve generation device includes: a processing module, a prediction module, a determination module and a generation module; The processing module is used to obtain a planned radiation dose corresponding to a target risk organ, and generate a plurality of target radiation doses based on a preset dose gradient coefficient and the planned radiation dose, wherein a mapping relationship exists between the dose gradient coefficient and the target radiation dose; For each target radiation dose, the prediction module is used to input the pre-radiotherapy CT image and the risk organ dose image corresponding to the target radiation dose into a preset target prediction network for feature extraction to generate a first deformation field; input the pre-radiotherapy CT image and the non-risk organ dose image corresponding to the target radiation dose into the target prediction network for feature extraction to generate a second deformation field; fuse the first deformation field and the second deformation field through the target prediction network to generate a target deformation field, and perform spatial transformation on the pre-radiotherapy target risk organ mask corresponding to the target radiation dose based on the target deformation field, and predict and output the post-radiotherapy target risk organ mask; The determination module is used to determine the target atrophy amount corresponding to the target radiation dose based on the target risk organ mask after radiotherapy and the target risk organ mask before radiotherapy; The generation module is used to generate a dose volume change curve DVV corresponding to the target risk organ based on the corresponding relationship between all target atrophy amounts and dose gradient coefficients after outputting the target atrophy amounts corresponding to all target radiation doses.

6. The DVV curve generating device according to claim 5, characterized in that: The risk organ dose image includes a left risk organ dose image and a right risk organ dose image, and the prediction module is specifically used to: Inputting the pre-radiotherapy CT image and the left risk organ dose image into the target prediction network for feature extraction to generate a left deformation field; Inputting the pre-radiotherapy CT image and the right-side risk organ dose image into the target prediction network for feature extraction to generate a right-side deformation field; The left deformation field and the right deformation field are used as the first deformation field.

7. The DVV curve generating device according to claim 6, wherein: The prediction module is further configured to: The left deformation field, the right deformation field, and the second deformation field are fused through the target prediction network to generate a target deformation field.

8. The DVV curve generating device according to claim 5, wherein: The apparatus further comprises a training module for: Acquiring an image training set, the image training set including a pre-radiotherapy CT image and its corresponding radiation dose image, and a pre-radiotherapy risk organ mask, the radiation dose image including a radiation dose image with a zero dose and a radiation dose image with a non-zero dose, and the radiation dose image including a risk organ dose image and a non-risk organ dose image; Constructing a fully convolutional neural network, the fully convolutional neural network includes an input layer, a backbone network, a spatial transformation layer, and an output layer. The backbone network is used to extract features from the input image of the input layer to generate a deformation field. The spatial transformation layer is used to perform spatial transformation processing on the deformation field output by the backbone network. The output layer is used to output the processing results of the spatial transformation layer. The fully convolutional neural network is trained using the image training set to obtain a target prediction network.

9. A DVV curve generating device, characterized in that: The DVV curve generation device includes a processor, a memory, and a DVV curve generation program stored in the memory and executable by the processor, wherein when the DVV curve generation program is executed by the processor, the steps of the DVV curve generation method according to any one of claims 1 to 4 are implemented.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a DVV curve generation program, wherein when the DVV curve generation program is executed by a processor, the steps of the DVV curve generation method according to any one of claims 1 to 4 are implemented.

Images