Method for establishing and correcting radiotherapy target volume based on dose distribution preview system

Abstract

An embodiment of the present invention discloses a method for establishing and correcting a radiotherapy target area based on a dose distribution preview system, comprising: reading CT image data, PTV contours, and OARs contours; processing the CT image data and PTV contours to obtain processing results; the processing results include a 3D dose distribution map, a DVH map, a dose index of interest, and a HI / CI indicator; based on the processing results, performing a "sculpted" correction on the PTV contour so that the corrected PTV contour strikes a balance between ensuring the dose received by the lesion and reducing the risk of critical organs. The present invention designs a dose distribution preview system based on deep learning, which simplifies the target area modification and plan formulation process, improves the rationality of radiotherapy target area establishment and the quality of radiotherapy plans, shortens the treatment delay time of cancer patients, reduces the burden on staff, and improves the patient's radiotherapy efficacy.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method for establishing and correcting a radiotherapy target area based on a dose distribution preview system. Background Art

[0002] With the development of technologies such as radiobiology and computers, radiotherapy (RT) is playing an increasingly irreplaceable role in cancer treatment. The rational formulation of radiotherapy plans (RT plans) is the basis for ensuring the accuracy and good therapeutic effect of radiotherapy. In this process, the determination of radiotherapy target volumes is an important prerequisite for plan formulation. Target volumes are usually divided into three levels: GTV (Gross Tumor Volume), CTV (Clinical Target Volume) and PTV (Planning Target Volume). Among them, PTV is a key reference object for field layout, dose calculation and plan evaluation. In addition to the clinically diagnosed tumor site represented by GTV and the range of subclinical lesions included in CTV, oncologists will also expand CTV based on factors such as volume changes, organ motion and imaging errors when determining the PTV area.

[0003] In reality, PTV determination is an iterative process that is difficult to replace with machines. First, oncologists need to consider individual differences in patients, such as age and condition. For example, considering the growth of a child's spine, the dose needs to be as symmetrical as possible. Second, anatomical differences around the patient's target volume can lead to irregularities in the dose received by organs at risk (OARs). Third, physicians' skills and work experience vary, leading to differences in the rationality of the PTV determination process. In addition, in intensity-modulated radiotherapy techniques such as simultaneous integrated boost (SIB), it is necessary to consider the multi-level three-dimensional dose distribution in different defined areas, which further complicates the determination and modification of the target volume.

[0004] Currently, the process of target volume determination and revision in radiotherapy relies primarily on the clinical experience and skills of oncologists. Furthermore, in cases involving large or uniquely located primary lesions and SIB (simultaneous boost), repeated adjustments to the target volume are necessary in conjunction with the physicist's radiotherapy plan optimization. The current detailed process from radiotherapy plan designation to implementation is as follows:

[0005] (1) Oncologists determine the location and area of ​​the GTV based on CT or multimodal (MRI, PET-CT, etc.) fused medical imaging data, combined with examination or test results and their own clinical experience, and then expand the GTV area to obtain the CTV area.

[0006] (2) Based on the established CTV area, the initial PTV outline is obtained by expanding the area outside the CTV. The initial PTV outline is also obtained by outlining the OARs. The physician modifies the initial PTV outline based on their clinical experience and the actual situation of the tumor patient.

[0007] (3) Under the guidance of experienced or senior physicians, a comprehensive prediction of the therapeutic or preventive dose of the lesion and surrounding tissues and the dose distribution of the critical organs is made to seek a balance between ensuring the ideal therapeutic dose and protecting the critical organs, and the PTV is modified based on this.

[0008] (4) PTV, OARs, and related imaging data are uploaded to TPS, and the physicist performs equipment selection, field layout, constraint setting, dose calculation, plan optimization, and preliminary evaluation. After the plan design is completed, it is fed back to the physician for evaluation.

[0009] (5) The oncologist evaluates the plan based on the three-dimensional dose distribution, the dose index (or DVH map) of the PTV and OARs, and other indicators such as HI\CI provided by the physicist. Based on clinical regulations, the oncologist is asked to further optimize the plan. The plan needs to be repeatedly adjusted until it reaches the optimal state.

[0010] (6) For radiotherapy plans that have been repeatedly optimized but still cannot meet clinical requirements (sufficiently high target dose, good uniformity and conformality, and sufficiently low risk to critical organs), the oncologist will reasonably modify the PTV contour and provide feedback to the physicist to re-formulate the radiotherapy plan.

[0011] (7) The physicist re-plans the radiotherapy plan and repeats steps (5)-(6) until the radiotherapy plan meets the clinical requirements and ensures that the lesion (and the surrounding necessary areas) receive sufficient treatment doses while fully protecting the critical organs.

[0012] (8) The radiotherapy plan is evaluated and reviewed by physicians and senior physicists before being delivered for clinical verification and implementation.

[0013] The above solution has the following main disadvantages:

[0014] (1) In the process of modifying the radiotherapy target volume, oncologists lack the prediction of the overall dose distribution corresponding to the current PTV, dose index (or DVH map), HI\CI and other estimation methods, which makes it difficult to find a balance between ensuring the prescribed treatment dose (for the lesion and the necessary area around it) and reducing the risk of critical organs (normal tissues and organs). On the one hand, in order to ensure the dose to the lesion, while considering eliminating the uncertainty of the radiation dose caused by factors such as organ movement, shape change, positioning error and imaging error, the PTV boundary may be over-expanded; on the other hand, for cases with large original lesions, adjacent to or overlapping critical organs or SIB (simultaneous boost), the PTV boundary may be under-expanded due to the consideration of minimizing the dose to critical organs.

[0015] (2) Due to differences in clinical experience and professional level, different clinical workers have different speeds and effects in correcting the PTV area, which will accordingly affect the cycle and quality of plan formulation.

[0016] (3) For radiotherapy plans that have been repeatedly optimized but still fail to obtain the ideal dose distribution, dose index (or DVH map), and HI\CI indicators that still do not meet clinical requirements, it is difficult for physicists to determine whether the reason is due to unreasonable PTV or errors in the plan design process.

[0017] (4) After each modification of the PTV, the oncologist can only estimate the corresponding changes in the dose distribution, dose index (or DVH map), and HI\CI indicators based on clinical experience. This requires the physicist to re-plan the design through TPS to verify it. If it still does not meet the clinical requirements, it needs to be fed back to the physician to modify the PTV again, and then repeat this process. Therefore, for some special (lesion volume or location) cases, the PTV correction and re-planning process will consume a lot of time and energy. The delay in patient treatment time and the change in lesion volume caused during this period will reduce the accuracy of radiotherapy, resulting in insufficient actual lesion dose or damage to normal tissues and organs. Summary of the Invention

[0018] The purpose of the embodiments of the present invention is to provide a method for establishing and correcting a radiotherapy target area based on a dose distribution preview system, and a dose distribution preview system, so as to simplify the target area modification and plan formulation process, improve the rationality of radiotherapy target area establishment and the quality of radiotherapy plans, shorten the treatment delay time of tumor patients, reduce the burden on staff and improve the radiotherapy efficacy of patients.

[0019] To achieve the above objectives, in a first aspect, an embodiment of the present invention provides a method for establishing and correcting a radiotherapy target area based on a dose distribution preview system. The dose distribution system includes a data reading and processing module, a case setting and dose index selection module, a 3D dose distribution map preview module, a calculation module, and a communication module. The method includes:

[0020] The data reading and processing module reads the CT image data, the PTV contour and the OARs contour; the contour data is drawn by an oncologist based on the CT image data;

[0021] receiving the oncologist's operation on the case setting and dose index selection module, implementing case setting and selecting the dose index indicator corresponding to the set case, and obtaining the patient's condition;

[0022] The CT image data and PTV contour are processed using the 3D dose distribution map preview module and the calculation module to obtain processing results; the processing results include a 3D dose distribution map, a DVH map, a dose index of interest, and a HI / CI indicator;

[0023] receiving, based on the processing result, a modification of the PTV contour by an oncologist, so that the modified PTV contour strikes a balance between ensuring the dose received by the lesion and reducing the risk of critical organs;

[0024] The calculation module calculates the corrected processing results to obtain a multi-objective Pareto optimal solution set for the target area contour, so that the oncologist can make auxiliary decisions based on the multi-objective Pareto optimal solution set for the target area contour, combined with the patient's condition and clinical requirements, and obtain the target area delineation result;

[0025] The communication module uploads CT image data, PTV contours, and OARs contours to the radiotherapy planning system, allowing the physicist to select equipment, arrange the field of view, set constraints, calculate doses, optimize the plan, and conduct preliminary evaluations to obtain a radiotherapy plan. The oncologist and the physicist make repeated adjustments to the radiotherapy plan before delivering it for clinical verification and implementation.

[0026] In a second aspect, an embodiment of the present invention provides a dose distribution preview system, comprising:

[0027] A data reading and processing module is used to read CT image data, PTV contours and OARs contours, and to train a dose prediction model; the contour data is drawn by an oncologist based on the CT image data;

[0028] A case setting and dose index selection module is configured to receive an operation of the oncologist on the case setting and dose index selection module, thereby implementing case setting and selecting a dose index indicator corresponding to the set case;

[0029] A 3D dose distribution map preview module is used to process the CT image data and PTV data using the trained dose prediction model to obtain a 3D dose distribution map for preview by oncologists;

[0030] A calculation module is used to calculate and draw the DVH curves of the target area and each critical organ based on the 3D dose distribution map, and to analyze the dose index and HI / CI indicators of interest for oncologists to make prejudgment and evaluation;

[0031] A communication module is used to realize communication between the dose distribution system and the external radiotherapy planning system.

[0032] As a specific embodiment of the present application, the data reading and processing module is used to train the dose prediction model. The specific process includes:

[0033] Collect case data; the case data includes radiotherapy simulation positioning CT medical imaging data, PTV outline data, OARs outline data, and clinical dose distribution data;

[0034] The case data were format converted, initially converted to 2D / 3D, and image preprocessed, and the processed case data were randomly divided into a training set and a validation set at a ratio of 5:1;

[0035] Building a deep learning network model architecture; the deep learning network model architecture includes a generator network and a discriminator network;

[0036] Setting different objective function combinations for the generative network and the discriminative network respectively;

[0037] The deep learning network model is trained and tested using the training set and validation set to obtain a dose prediction model.

[0038] As a specific embodiment of the present application, the case setting and dose index selection module is specifically used to:

[0039] Receive medical condition data set by an oncologist, wherein the case data includes the affected area and specific disease type;

[0040] A default dosage index is selected according to the disease condition data.

[0041] The implementation of the embodiments of the present invention simplifies the target area modification and plan formulation process, improves the rationality of radiotherapy target area establishment and the quality of radiotherapy plans, shortens the treatment delay time of tumor patients, reduces the burden on staff and improves the radiotherapy efficacy of patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the specific implementation of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the specific implementation or the description of the prior art.

[0043] Figure 1 is a structural diagram of a dose distribution preview system provided by an embodiment of the present invention;

[0044] Figure 2 This is a data preprocessing flow chart in the development of the dose distribution preview system;

[0045] Figure 3 This is a schematic diagram of the overall deep network architecture for dose prediction;

[0046] Figure 4 This is a schematic diagram of the dose prediction and discrimination network structure;

[0047] Figure 5 This is a schematic diagram of the network structure for dose prediction generation;

[0048] Figure 6 This is a diagram of the data enhancement method during model training;

[0049] Figure 7 This is a flow chart of the process of establishing and revising radiotherapy target areas, and specifying and implementing radiotherapy plans based on the dose distribution preview system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0050] 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 them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0051] The definitions of abbreviations and key terms involved in the embodiments of the present invention are as follows:

[0052]

[0053]

[0054] The inventive concept of this invention is that deep learning technology has been increasingly applied in radiotherapy in recent years, primarily in areas such as automated diagnosis, automated segmentation, and dose prediction. This invention first develops a dose distribution preview system based on the collection of extensive clinical case data, deep learning methods, and related software engineering techniques. This system then guides oncologists in establishing and revising radiotherapy target volumes. The general process is as follows:

[0055] First, a dose distribution preview system is developed based on a self-built deep learning (DL) network structure. Then, physicians can preview and obtain evaluation indicators such as dose distribution, HI\CI, and dose index (or DVH map) based on the current target contour in real time to perform "sculpting" modifications to the radiotherapy target area. The system automatically stores the target contour and its corresponding evaluation indicators after each modification, and provides physicians with a "Pareto Optimality solution set" of the target contour based on multiple indicators, ultimately seeking a balance between ensuring the ideal treatment dose and protecting critical organs.

[0056] Based on the aforementioned inventive concept, embodiments of the present invention provide a dose distribution preview system. The system's development phase can be broadly divided into the following stages: case data collection, data preprocessing, deep learning network model architecture development, objective function setting, model training, validation, and test evaluation, and dose preview system process design.

[0057] like Figure 1 As shown, the dose distribution preview system mainly includes:

[0058] A data reading and processing module is used to read CT image data, PTV contours and OARs contours, and to train a dose prediction model; the contour data is drawn by an oncologist based on the CT image data;

[0059] The case setting and dose index selection module is used to receive the oncologist's operation on the case setting and dose index selection module, realize case setting and select the dose index indicator corresponding to the set case; the oncologist sets the diseased site (head, chest, abdomen, pelvis or limbs) and the specific disease type (brain tumor, nasopharyngeal cancer, esophageal cancer, lung cancer, rectal cancer, cervical cancer and prostate cancer), and the system automatically selects the default dose index based on the case situation; the physician can also set the dose index indicator of the target area or different critical organs according to clinical requirements or personal preferences;

[0060] A 3D dose distribution map preview module is used to process the CT image data and PTV data using the trained dose prediction model to obtain a 3D dose distribution map for preview by oncologists;

[0061] A calculation module is used to calculate and draw the DVH curves of the target area and each critical organ based on the 3D dose distribution map, and to analyze the dose index and HI / CI indicators of interest for oncologists to make prejudgment and evaluation;

[0062] A communication module is used to realize communication between the dose distribution system and the external radiotherapy planning system.

[0063] The specific process of training the dose prediction model includes:

[0064] (1) Case data collection

[0065] The original case data collected in this invention are all paired data, that is, for a patient case, the corresponding DICOM (Digital Imaging and Communications in Medicine) communication protocol medical data should be fully collected, including:

[0066] Medical imaging data, mainly radiotherapy simulation positioning CT medical imaging data, in the format of RT Image

[0067] PTV, OARs outline information data, format is RT Structure

[0068] Clinical dose distribution map data, in the format of RT Dose

[0069] Among them, all original medical imaging data are from the simulation positioning equipment actually used in the clinical practice of the radiotherapy department, and are acquired by scanning with parameters consistent with the clinical scenario; all contouring is completed by at least two experienced oncologists and evaluated and reviewed by senior physicians; the radiotherapy plans obtained based on the above contouring data are completed by at least two experienced physicists, and all plans are reviewed by senior physicists and physicians and have been delivered for clinical use. The diseases involved in the present invention include brain tumors, nasopharyngeal cancer, esophageal cancer, lung cancer, rectal cancer, cervical cancer and prostate cancer. For all the above diseases, the number of original case collections is 100 to 120, which are randomly divided into training-validation sets (Train-Valid Dataset) and test sets (Test Dataset) at a ratio of about 5:1.

[0070] (2) Data preprocessing

[0071] Raw data preprocessing is the prerequisite for deep learning network training and verification. All data are exported through TPS and converted into .dcm / .nii format and 2D / 3D format, and then image preprocessing is performed. The steps of medical image preprocessing in this invention are as follows: Figure 2 As shown, specifically including:

[0072] Medical image resampling (Image Resampling): 3D image voxel spacing (Voxel spacing) of CT, Structures and Dose distribution map after initial conversion Suppose that the 3D image resolutions are (DCT_X, DCT_Y, DCT_Z), (DStruc_X, DStruc_Y, DStruc_Z), (DDose_X, DDose_Y, DDose_Z), and the 3D image sizes are (FCT_X, FCT_Y, FCT_Z), (FStruc_X, FStruc_Y, FStruc_Z), (FDose_X, FDose_Y, FDose_Z). To ensure that the real physical space sizes mapped by different types of 3D image pixels are consistent, the voxel spacing needs to be kept uniform at S0. According to the image size (mm) in a single dimension, F = S*D and F is a constant, all CT, Structures and Dose distributions after resampling can be obtained. In this invention, the nearest neighbor interpolation is used for resampling the Structures binary image (0 / 255), and B-spline interpolation is used for CT and Dose distribution maps.

[0073] Structures image labeling: Different label values ​​(L1, L2, …, Ln) are set for the PTV, different OARs, and body contours. The air portion is kept at 0, forming corresponding binary images. The upper and lower bounds of the label values ​​are roughly consistent with the CT grayscale range.

[0074] Dose distribution map pixel value to dose value conversion: Based on Python-pydicom or ITK and other related tools, the DoseGridScaling parameters of the initial Dose distribution map are read and the dose image pixel value matrix is ​​converted into a dose value matrix.

[0075] Image Cropping and Padding: Cropping and unifying the 3D resolution of all CT, Structures, and Dose distribution map images to the same size. That is, the number of rows and columns of all image matrices remains the same. To facilitate network training and validation, the values ​​are generally set to 128 or 256. If the actual resolution is smaller than this value, it is padded with zeros.

[0076] Image normalization. Normalize the CT, Structures, and Dose distribution map image matrices to the range [-1, 1] according to their corresponding intervals.

[0077] Multi-Channel Data Stacking: In this invention, neural network input data is multi-channel data to improve network parameterization and prediction performance. The overall multi-channel data (Channel_1, Channel_2, Channel_3, Channel_4…Channel_n) respectively includes CT imaging data, 3D dose data, target volume data, and binarized image data of different OARs.

[0078] (3) Construction of deep learning network model structure

[0079] The dose prediction involved in this invention is mainly based on the deep learning image generation method. The network model structure adopted in the overall architecture integrates two deep learning ideas, one is the Conditional Generative Adversarial Network (CGAN) idea, and the other is the multi-scale / multi-channel feature fusion (Multi-scale / channels feature fusion) encoding-decoding network idea. The overall network structure consists of two parts, such as Figure 3 As shown in , the generating network and the discriminative network have opposite optimization goals, so there is a process of mutual confrontation and mutual promotion in each training cycle. Figure 3 In the above figure, X is the multi-channel input data consisting of CT and contour data, G(X) is the output of the generative network, i.e., the predicted dose distribution map, and “Fake” and “Real” represent the classification results of the discriminator as “fake” and “real”.

[0080] Discriminator network, such as Figure 4As shown in the figure, the ideal optimization goal is to accurately distinguish the clinical dose distribution map from the predicted dose distribution map, that is, when the input is the above two, the network output is "real" and "fake" respectively. The network structure adopts the PatchGAN discriminant network composed of convolution / normalization / activation layers, in which the normalization layer adopts batch normalization (Batch Normalize) and the activation layer uses LeakyReLU (Leaky Rectified Linear Unit, with leaky linear rectification function) as the activation function. Based on the binary output probability value matrix of the overlapping blocks of the input image, let F n is the receptive field size of the nth layer, KS is the convolution kernel size (Kernel Size), S is the step size, then the receptive field size of the nth layer is:

[0081] F n =KS*F n+1 -(KS-S)*(F n+1 -1) (1)

[0082] From this, we can obtain the receptive fields from back to front as follows: 4*4*4, 7*7*7, 16*16*16, 34*34*34, 70*70*70…, and the number of convolutional layers of the discriminator can be obtained according to the required receptive field size.

[0083] exist Figure 4 In the equation (5), X is the multi-channel input data consisting of CT and contour data, G(X) is the output of the generative network, i.e., the predicted dose distribution map, Cli is the clinical dose distribution map, and MCli and MPre are the probability output matrices of the discriminator for the clinical dose distribution and the predicted dose distribution, respectively.

[0084] Generate network (Generator), as shown in the attached Figure 5 As shown, the ideal optimization goal is to output a predicted dose distribution that is indistinguishable from the clinical dose distribution map, so that the output result is "real" when the prediction result is input into the discriminator. The encoding part uses a coding block consisting of multiple convolution / normalization / activation layers and a series of downsampling operations to extract image features from multiple channels at multiple scales. The normalization layer uses batch normalization, and the activation layer uses ReLU (Rectified Linear Unit) as the activation function. For each decoding layer, its input includes feature fusion of different scales and channels, and the feature map is convolved / normalized / activated. The decoding part uses upsampling to gradually transform the high-dimensional feature map into a 3D predicted dose distribution map.

[0085] exist Figure 5 middle, and Represents different levels of encoding layer and decoding layer, N@S 3 Indicates that the current number of feature maps is N and the three-dimensional size is S 3 .

[0086] (4) Objective function setting

[0087] In the present invention, different objective function combinations are set for the discriminant network and the generative network respectively.

[0088] The discriminant network objective function mainly adopts the combination of conventional binary cross entropy (BCE) and Sigmoid (σ) function, as shown below:

[0089]

[0090]

[0091]

[0092] Where X represents the multi-channel input data, Cli and Pre represent the clinical dose distribution and predicted dose distribution respectively, and L D_Cli and L D_Pre They represent the loss terms of the discriminant network for clinical dose distribution and predicted dose distribution respectively; L D Represents the total loss function of the discriminator; M Cli and M Pre These are the classification output probability matrices of the discriminant network for clinical dose distribution and predicted dose distribution, respectively.

[0093] The objective function of the generated network consists of the following multiple loss terms with different weights:

[0094] A. Adversarial loss (with the discriminator)

[0095]

[0096] Where G(X) represents the predicted dose map output by the generator, M G(X) Represents a matrix consisting of probability values, L G_CGAN represents the adversarial loss term of the generative network.

[0097] B. Adjacent Voxels Difference Loss

[0098]

[0099] where m T 、m C 、m S Respectively represent the cross-sectional, coronal and sagittal slices of the 3D dose distribution image, Cli i 、Cli j 、、Cli k represent the dose value matrix of a certain layer in the clinical dose map, and G(X) i 、G(X) j 、G(X) k The dose matrix corresponding to the dose distribution map predicted by the generator, L G_AVD It represents the loss term that tries to preserve the edge features of the image by minimizing the difference in dose values ​​of adjacent voxels.

[0100] C.L1 norm loss term

[0101] L G_L1 =||Cli-G(X)||1 (7)

[0102] Among them, L G_L1 Represents the L1 norm loss term of the generative network.

[0103] The overall objective function of the generated network is:

[0104] L G =λ CGAN ·L G_CGAN +λ G_AVD ·L G_AVD +λ L1 ·L G_L1 (8)

[0105] Among them, λ CGAN ,λ G_AVD ,λ L1 The hyperparameters represent the weights of the adversarial loss term, the neighboring voxel difference loss term, and the L1 norm loss term, respectively.

[0106] (5) Model training, validation, and test evaluation

[0107] All cases were randomly divided into training-validation set and test set in a ratio of about 5:1. The model was trained and validated by 5-fold cross validation. The input was enhanced by data augmentation method (such as Figure 6 As shown in the figure, D0 represents the original image size, D X (where represents the size of the randomly cropped image). Each iteration randomly rotates and crops the multi-channel image. During training, model hyperparameters are continuously adjusted, and the model is stored every five training cycles (epochs) for later testing. The evaluation metrics after testing mainly include the following:

[0108] Dose distribution difference map and difference statistical histogram

[0109] MAE (mean absolute error) or MSE (mean squared error)

[0110] DVH (Dose-Volume Histograms)

[0111] Dose index statistics: D99%, D98%, D95%, D50%, D2% of PTV, Dmax and Dmean of different OARs, etc.

[0112] Homogeneity index (HI) of PTV dose distribution:

[0113] Conformity index (CI) of PTV dose distribution: Where V T represents the PTV volume, V P Indicates the prescription dose coverage volume, V TP Indicates the volume of the PTV area covered by the prescribed dose

[0114] DICE coefficient: Where Pre and Cli represent the predicted dose distribution map and clinical dose distribution map respectively, and Prescription_Dose represents the prescribed dose

[0115] Furthermore, the communication module is specifically used for:

[0116] CT image data, PTV contours, and OARs contours are uploaded to the radiotherapy planning system, allowing physicists to select equipment, arrange radiation fields, set constraints, calculate doses, optimize plans, and conduct preliminary assessments to obtain a radiotherapy plan. The oncologist and physicist will make repeated adjustments to the radiotherapy plan before delivering it for clinical verification and implementation.

[0117] It should be noted that the dose distribution preview system in this embodiment uses the UDP / IP transport layer protocol and fully integrates the DICOM protocol to implement network communication functions. The client is implemented based on the ASP.NET Core application development framework. This dose distribution preview system primarily provides oncologists with functions such as dose distribution preview, dose index (or DVH map), and HI / CI indicator calculation.

[0118] Based on the above dose distribution preview system, please refer to Figure 7 The process of establishing and correcting the radiotherapy target area provided by the embodiment of the present invention is as follows:

[0119] (1) Oncologists determine the location and area of ​​the GTV based on CT or multimodal (MRI, PET-CT, etc.) fused medical imaging data, combined with examination or test results and their own clinical experience, and then expand the GTV area to obtain the CTV area.

[0120] (2) Based on the determined CTV area, the initial PTV contour is expanded and the OARs contours are outlined. Oncologists can perform case setting and dose index selection through the aforementioned case setting and dose index selection module.

[0121] (3) The CT and contour data are read into the dose preview system. The system processes the raw data and provides the oncologist with a 3D dose distribution map, DVH map, and the dose index and HI / CI calculation results corresponding to the current PTV. That is, the CT image data and PTV data are input into the 3D dose distribution map preview module for processing to obtain a 3D dose distribution map. The calculation module calculates and plots the DVH curves of the target area and each critical organ based on the 3D dose distribution map, and parses out the dose index and HI / CI indicators of interest.

[0122] (4) The physician refers to the preview system and, based on the results obtained from the current contour data, performs a "sculpted" correction on the PTV. After each minor correction, the physician can obtain a three-dimensional dose distribution map, DVH, and various dose indices in real time, and strive to achieve a balance between ensuring the dose received by the lesion (and the surrounding necessary areas) and reducing the risk of critical organs. That is, based on the processing results (including the 3D dose distribution map, DVH map, the dose index of interest, and the HI / CI indicator), the physician receives the correction of the PTV contour, so that the corrected PTV contour achieves a balance between ensuring the dose received by the lesion and reducing the risk of critical organs.

[0123] (5) The system saves each modified contour and its corresponding three-dimensional dose distribution map, DVH, PTV\OARs dose index, and automatically provides a multi-objective "Pareto Optimality solution set" for the target contour based on the above indicators. Oncologists make decisions based on the patient's condition and clinical requirements. That is, the calculation module calculates the modified processing results to obtain a multi-objective Pareto optimal solution set for the target contour, allowing oncologists to make auxiliary decisions based on the multi-objective Pareto optimal solution set for the target contour, combined with the patient's condition and clinical requirements, and obtain the target delineation result.

[0124] (6) Submit the target area delineation results to senior physicians for review.

[0125] (7) PTV, OARs, and related imaging data are uploaded to TPS, and the physicist performs equipment selection, field layout, constraint setting, dose calculation, plan optimization, and preliminary evaluation. After the plan design is completed, it is fed back to the physician for evaluation.

[0126] (8) The oncologist evaluates the plan based on the three-dimensional dose distribution, PTV and OARs dose index (or DVH map), HI\CI, etc. provided by the physicist, and requires the physicist to further optimize the plan according to clinical regulations. The plan needs to be repeatedly adjusted until it is optimal.

[0127] (9) The radiotherapy plan is evaluated and reviewed by physicians and senior physicists before being delivered for clinical verification and implementation.

[0128] From the above description, it can be seen that compared with the current method of establishing and correcting radiotherapy target volumes, the dose preview system based on deep learning and the method of establishing and correcting radiotherapy target volumes under auxiliary guidance proposed in this invention have the following advantages:

[0129] (1) During the process of establishing and revising the target area by the oncologist, the method proposed in the present invention can provide a real-time preview of the three-dimensional dose distribution, the relevant dose index (or DVH map) and the advance calculation and evaluation of indicators such as HI\CI for the target area contour being modified without formulating a radiotherapy plan, so as to guide the oncologist to find a balance between ensuring the prescribed treatment dose (for the lesion and its surrounding necessary areas) and reducing the risk of critical organs (normal tissues and organs), so that the oncologist can make "sculpted" modifications to the target area contour.

[0130] (2) The method proposed in the present invention enables oncologists to use the dose distribution and dose index (or DVH map) provided by the system as the standard in the process of establishing and correcting the target area (especially for cases with large lesions, special locations or SIB cases), thereby compensating for the lack of clinical experience and differences in professional level.

[0131] (3) For radiotherapy physicists, on the one hand, since doctors can obtain indicators such as dose distribution, dose index, HI\CI based on the method proposed by the present invention, there is no need for physicists to make plans to feedback various evaluation indicators to doctors after each target area modification; on the other hand, physicists make radiotherapy plans based on the ideal target area contour data provided by doctors. When the planning results are not ideal, the irrationality of the target area delineation can be eliminated, so that only the errors in the planning process can be focused on; therefore, the method provided by the present invention can reduce the workload of physicists to a large extent.

[0132] (4) The method provided by the present invention reduces the redundant time brought about by the target area establishment and correction process coordinated by physicians and physicists, which greatly reduces the useless work in the planning process, and is conducive to shortening the delay time for patients to receive treatment, thereby avoiding the reduced efficacy and the risk of toxic side effects of radiotherapy caused by delayed treatment.

[0133] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A dose distribution preview system, characterized in that: include: A data reading and processing module is used to read CT image data, PTV contours and OARs contours, and to train a dose prediction model; the contour data is drawn by an oncologist based on the CT image data; A case setting and dose index selection module is configured to receive an operation of the oncologist on the case setting and dose index selection module, thereby implementing case setting and selecting a dose index indicator corresponding to the set case; A 3D dose distribution map preview module is used to process the CT image data and PTV data using the trained dose prediction model to obtain a 3D dose distribution map for preview by oncologists; a calculation module, configured to calculate and plot DVH curves of the target volume and organs at risk based on the 3D dose distribution map, and parse out the dose index and HI / CI index of interest for prejudgment and evaluation by the oncologist, and to calculate the processing results corresponding to the PTV contour revised by the oncologist to obtain a multi-objective Pareto optimal solution set for the target contour, so that the oncologist can make auxiliary decisions based on the multi-objective Pareto optimal solution set for the target contour, in combination with the patient's condition and clinical requirements, and obtain a target delineation result; the processing results include the 3D dose distribution map, the DVH curve, the dose index of interest, and HI / CI index; A communication module is used to realize communication between the dose distribution system and the external radiotherapy planning system.

2. The system according to claim 1, wherein The data reading and processing module is used to train the dose prediction model. The specific process includes: Collect case data; the case data includes radiotherapy simulation positioning CT medical imaging data, PTV outline data, OARs outline data, and clinical dose distribution data; The case data were format converted, initially converted to 2D / 3D, and image preprocessed, and the processed case data were randomly divided into a training set and a validation set at a ratio of 5:1; Building a deep learning network model architecture; the deep learning network model architecture includes a generator network and a discriminator network; Setting different objective function combinations for the generative network and the discriminative network respectively; The deep learning network model is trained and tested using the training set and validation set to obtain a dose prediction model.

3. The system according to claim 2, wherein: The image preprocessing of the case data specifically includes: medical image resampling, binary image label value setting, dose distribution map pixel value-dose value conversion, image cropping and filling, image normalization and multi-channel data stacking.

4. The system according to claim 2, wherein: The discriminant network adopts a PatchGAN discriminant network consisting of convolution / normalization / activation layers.

5. The system according to claim 2, wherein: The objective function of the discriminant network adopts a combination of value cross entropy and Sigmoid (σ) function; the objective function of the generative network is composed of multiple loss terms according to different weights, and the multiple loss terms include: adversarial loss term, neighbor voxel difference loss term and L1 norm loss term.

6. The system according to claim 2, wherein: The deep learning network model is trained and tested using the training set and validation set to obtain a dose prediction model, specifically including: The model is trained and validated using a five-fold cross-validation approach. Data augmentation is used for the input, and multi-channel images are randomly rotated and cropped in each iteration. The model hyperparameters are continuously adjusted during training, and the model is stored every five training cycles for later testing.

7. The system according to claim 6, wherein: The evaluation indicators after the test include dose distribution difference map and difference statistical histogram, mean absolute error or mean square error, DVH, dose index statistics, PTV dose distribution uniformity index, PTV dose distribution conformity index and DICE coefficient.

8. The system according to claim 1, wherein: The case setting and dose index selection module is specifically used for: Receive medical condition data set by an oncologist, wherein the case data includes the affected area and specific disease type; A default dosage index is selected according to the disease condition data.

9. The system according to claim 1, wherein: The modified PTV contour achieves a balance between ensuring the dose received by the lesion and reducing the risk of critical organs.

10. The system according to claim 1, wherein: The communication module is specifically used for: CT image data, PTV contours, and OARs contours are uploaded to the radiotherapy planning system, allowing physicists to select equipment, arrange radiation fields, set constraints, calculate doses, optimize plans, and conduct preliminary assessments to obtain a radiotherapy plan. The oncologist and physicist will make repeated adjustments to the radiotherapy plan before delivering it for clinical verification and implementation.

Images