Deep learning dose bias-based lung mrgart stratification decision system

Abstract

This invention relates to the field of pulmonary magnetic resonance-guided adaptive radiotherapy (MRART) technology, and provides a pulmonary MRgART hierarchical decision-making system based on deep learning dose bias. The system includes a data preprocessing module, a physics prior guidance module, a deep learning prediction module, a risk stratification decision-making module, and a model training module. The data preprocessing module standardizes the initial data; the physics prior guidance module generates a physics prior enhanced feature map and a dual-channel input tensor based on the physical mechanism of dose bias; the deep learning prediction module uses 3D-ResNet to output predicted probabilities; the risk stratification decision-making module triggers the corresponding dose calculation process based on the comparison result of the predicted probability and the classification threshold; and the model training module generates risk labels and trains the risk prediction model. This invention significantly improves the efficiency of online adaptive radiotherapy while ensuring dose calculation accuracy, achieving personalized and precise radiotherapy.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a lung MRgART hierarchical decision system based on deep learning dose bias. Background Technology

[0002] Magnetic resonance imaging (MRI) cannot directly provide electron density information; therefore, the bulk density assignment method is currently the main method for generating pseudo-CT scans in online workflows. This method calculates dose by pre-setting electron density, and has advantages such as simplicity, efficiency, and clear physical meaning, meeting the timeliness requirements of online adaptive radiotherapy. However, lung tissue is highly heterogeneous, and the overall average assignment is difficult to reflect the true density distribution of individuals, easily leading to dose deviation.

[0003] Therefore, a lung MRgART hierarchical decision system based on deep learning dose bias is needed to address the problems of large dose calculation errors and low online treatment efficiency. Summary of the Invention

[0004] The purpose of this invention is to solve the technical problem that there are significant individual differences in the dose calculation of the bulk density assignment method in lung MRgART, while uniform implementation of enhanced assignment affects online efficiency. The invention provides a lung MRgART hierarchical decision system based on deep learning dose bias, which includes a data preprocessing module, a physical prior guidance module, a deep learning prediction module, and a risk stratification decision module.

[0005] The data preprocessing module is used to standardize the input initial data to obtain standardized data; the standardized data includes standardized CT images and standardized GTV masks.

[0006] The physics prior guidance module is used to perform prior enhancement on standardized CT images based on the physical mechanism formed by dose deviation, generate a physical prior enhancement feature map; and fuse and stitch standardized data to obtain a dual-channel input tensor.

[0007] The deep learning prediction module is used to output the prediction probability using the risk prediction model; the risk prediction model receives the physical prior enhanced feature map or dual-channel input tensor output by the physical prior guidance module and outputs the prediction probability.

[0008] The risk stratification decision module is used to determine the risk type based on the predicted probability and a preset classification threshold; and to trigger the corresponding dose calculation process based on the risk type.

[0009] Existing technologies generally treat dose deviation in the pulmonary dose assignment method as a "systemic problem" and adopt a uniform enhancement strategy. This invention constructs a two-path prediction model using a dual-channel input tensor based on enhancement features and multimodal fusion features, and establishes a hierarchical decision path to predict individual dose deviation risk. It reveals that pulmonary dose deviation in the pulmonary dose assignment method has structural dependence and predictability, and based on this, proposes the technical ideas of "risk stratification" and "selective enhancement," breaking through the cognitive limitations of existing technologies.

[0010] Furthermore, the data preprocessing module includes a resampling unit, a normalization unit, and a spatial alignment and cropping unit;

[0011] The initial data includes initial CT images, GTV masks, and sets of masks for organs at risk;

[0012] The resampling unit is used to convert the initial CT image from the original acquired voxel grid size to a uniform voxel size to obtain a resampled CT image.

[0013] The normalization unit is used to mathematically transform the gray values ​​of the resampled CT images and adjust them to a uniform numerical distribution standard to obtain normalized CT images.

[0014] The spatial alignment and cropping unit is used to obtain lung masks based on the endangered organ mask set. Using the centroid of the lung mask as a spatial reference point, the normalized CT image and GTV mask are cropped to a uniform size to obtain standardized data.

[0015] Furthermore, the physics prior guidance module includes a spatial proximity calculation unit, a local density heterogeneity calculation unit, a beam path cumulative error calculation unit, a prior fusion and enhancement unit, a feature fusion unit, and a multimodal input unit;

[0016] The GTV mask is a binary representation of the gross tumor volume;

[0017] The spatial proximity calculation unit is used to calculate the spatial proximity factor, which is used to quantify the spatial proximity between the gross tumor volume and the surrounding high-density structures.

[0018] The local density heterogeneity is used to calculate the local density heterogeneity factor, which is used to quantify the degree of density inhomogeneity of lung tissue within a preset area surrounding the gross tumor volume.

[0019] The beam path cumulative error calculation unit is used to calculate the beam path cumulative error factor, which is used to quantify the cumulative electron density assignment error along the beam path under SBRT multi-beam irradiation.

[0020] The prior fusion and enhancement unit is used to generate a prior weight map by weighting and fusing the spatial proximity factor, local density heterogeneity factor, and beam path cumulative error factor, and to perform adaptive enhancement on the normalized CT image based on the prior weight map, and output a physical prior enhancement feature map.

[0021] The feature fusion unit is used to fuse standardized data to generate a GTV feature fusion map;

[0022] The multimodal input unit is used to stitch standardized CT images and GTV feature fusion maps together in the channel dimension to construct a dual-channel input tensor.

[0023] This invention starts from the physical mechanism of dose deviation formation, quantifies three core factors—spatial proximity, local density heterogeneity, and cumulative error of multi-beam paths—into mathematical expressions, and inputs them into a deep learning network to achieve prediction of individual dose deviation risk.

[0024] Furthermore, the formula for calculating the spatial proximity factor is as follows:

[0025] ;

[0026] Where i1 represents the voxel index, which traverses the outer region of the gross tumor volume; Represents the spatial proximity factor of voxel i1; Represents a high-density structure set; Represents the set of voxels i1 to high-density structures. The minimum Euclidean distance; Indicates the scale parameter; Indicates the outer region representing the gross tumor volume; For indicator functions;

[0027] The formula for calculating the local density heterogeneity factor is as follows:

[0028] ;

[0029] Among them, voxel i2 traverses the lung tissue region. The local density heterogeneity factor representing voxel i2; Indicates Gaussian smoothing operation; Represents standardized CT images; This represents the local Gaussian weighted mean of the squared normalized CT image at voxel i2; This represents the local Gaussian weighted mean of the standardized CT image at voxel i2;

[0030] The formula for calculating the cumulative error factor of the beam path is:

[0031] ;

[0032] Among them, voxel i3 traverses all regions traversed by the beam path. B is the cumulative error factor of the beam path for voxel i3; B is the number of beams, b=1,2,...,B; L b The path length of beam b within the patient's body; path b represents the path traversed by the b-th beam; s represents the spatial coordinates of a point along the beam path; Assign error to the electron density at spatial coordinate point s. ,in, This represents the high-frequency component of the normalized CT image at point s; This represents a normalized image.

[0033] Furthermore, the calculation formula for the prior fusion and enhancement unit includes:

[0034] ;

[0035] Among them, voxel i traverses the outer region of the gross tumor volume, the lung tissue region, and all regions traversed by the beam path. This represents the prior weight of voxel i. The first fusion coefficient, For the second fusion coefficient, The third fusion coefficient; This represents the spatial proximity factor corresponding to voxel i; Indicates the local density heterogeneity factor; This represents the cumulative error factor of the beam path;

[0036] The prior weights of all voxels are arranged according to their spatial positions to obtain the prior weight map. ;

[0037] ;

[0038] in, Enhanced feature maps for physical priors; This represents element-wise multiplication; To enhance strength parameters.

[0039] Furthermore, the calculation formula for the GTV feature fusion map is as follows:

[0040] ;

[0041] in, This is a GTV feature fusion map. For standardized CT images; For standardized GTV masks, within the standardized GTV mask area Outside the standardized GTV mask area .

[0042] The physics-guided prior learning module proposed in this invention quantifies three core physical factors—spatial proximity, local density heterogeneity, and beam path cumulative error—into mathematical expressions and designs them as fully differentiable network components, achieving end-to-end collaborative learning driven by physics knowledge and data. No similar design exists in the prior art, and this innovation enables risk prediction models to achieve superior prediction performance and generalization ability under small sample conditions.

[0043] Furthermore, the risk stratification decision-making module includes a threshold setting unit, a risk judgment unit, a process triggering unit, and a plan review interface unit;

[0044] The threshold setting unit is used to set the classification threshold;

[0045] The risk determination unit is used to receive the predicted probability output by the deep learning prediction module, compare it with the classification threshold, and output the risk type, which includes high-risk cases and low-risk cases.

[0046] The process triggering unit is used to automatically trigger the corresponding dose calculation process according to the risk type;

[0047] The planned review interface unit is used to provide an access point for planned review after the dose calculation process is triggered, and to determine whether the dose index meets the safety tolerance.

[0048] Furthermore, the system also includes a model training module; the model training module is used to generate risk labels required for training the risk prediction model and to train the risk prediction model.

[0049] The model training module includes a reference CT design unit, a pseudo-CT generation unit, a dose deviation calculation and labeling unit, a data augmentation unit, and a model building unit;

[0050] The reference CT design unit is used to generate a reference dose distribution based on the initial CT images of several historical patients and in accordance with a reference treatment plan.

[0051] The pseudo-CT generation unit is used to generate pseudo-CT on the same set of initial CT images using the block density assignment method, and to calculate the corresponding dose distribution.

[0052] The dose deviation calculation and labeling unit is used to compare the dose distribution with the reference dose distribution, determine the risk type based on preset conditions, and label the risk.

[0053] The data augmentation unit is used to augment the physical prior enhanced feature map and the dual-channel input tensor.

[0054] The model building unit uses a 3D-ResNet network to build and train a risk prediction model.

[0055] This invention proposes a model training module that can effectively eliminate confounding factors. Using the same initial set of data, only the electron density representation method is changed. By comparing the dose differences between the reference CT plan and the pseudo-CT plan using the mass density assignment method, binary labels for high-risk and low-risk cases are generated. This module eliminates confounding factors such as anatomical differences, registration errors, and delineation variations, ensuring that the risk labels purely reflect the dose deviation caused by errors in the electron density representation method using the mass density assignment method, providing a high-quality gold standard for model training.

[0056] This invention employs three-fold cross-validation to select the optimal model, and then averages the prediction probabilities of the three optimal models during independent testing. Performance is evaluated using Bootstrap repeated sampling (200 times). This strategy effectively reduces the risk of overfitting to a single model and improves the stability and reliability of the prediction results.

[0057] Technical effects:

[0058] This invention directly transforms the output of the risk prediction model into a basis for process decision-making. Based on the principle of maximizing the Youden Index, a classification threshold is set, diverting patients to two parallel paths: an "enhanced value assignment process" and a "routine value assignment process." This mechanism realizes a complete decision chain of "pre-planning prediction—high-risk case screening—selective enhancement," solving the efficiency problem caused by a uniform enhancement strategy. Attached Figure Description

[0059] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 This is a system structure diagram of the present invention;

[0061] Figure 2 This is a diagram of the data preprocessing module of the present invention;

[0062] Figure 3 This is a diagram of the physics prior guidance module of the present invention;

[0063] Figure 4 This is a diagram of the risk stratification decision-making module of the present invention;

[0064] Figure 5 This is a diagram of the model training module of the present invention;

[0065] Figure 6 This is a data processing flow diagram in this invention. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0067] It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance, or suggesting any such actual relationship or order between these entities or operations. Additionally, the terms "connected," "linked," etc., can refer to a direct connection between elements, components, modules, etc., or an indirect connection via other elements, components, modules, etc.

[0068] Example 1:

[0069] This invention is achieved through the following technical solutions, such as... Figure 1 , Figure 6 As shown, the present invention provides a lung MRgART hierarchical decision system based on deep learning dose bias. The system includes: a data preprocessing module, a physical prior guidance module, a deep learning prediction module, a risk hierarchical decision module, and a model training module.

[0070] The data preprocessing module is used to standardize the input initial data to obtain standardized data.

[0071] like Figure 2 As shown, the data preprocessing module includes a resampling unit, a normalization unit, and a spatial alignment and clipping unit.

[0072] like Figure 6 As shown, the initial data includes initial CT images, GTV masks, and sets of masks for organs at risk.

[0073] The GTV mask is a binary representation of the gross tumor volume (GTV), which is manually drawn by the physician on the initial CT images. The organ at risk mask set refers to a set of binary images used to represent the entire area of ​​normal tissue where radiation dose needs to be limited during radiotherapy.

[0074] The resampling unit is used to convert the initial CT image from the original acquired voxel grid size to a uniform voxel size to obtain a resampled CT image.

[0075] Furthermore, since the initial CT images acquired by different patients, different devices, or different scanning parameters have different spatial resolutions, the resampling unit resamples the initial CT images using the Plastimatch linear interpolation method, so that all input data have the same voxel size and isotropy.

[0076] Specifically, the resampling unit uses a size of 1.5×1.5×1.5mm. 3 The initial CT image is resampled using a voxel grid to obtain a resampled CT image. It should be noted that using this sampling scale can preserve image detail while maintaining computational efficiency, and avoid feature distribution shifts caused by voxel size differences.

[0077] The normalization unit is used to mathematically transform the gray values ​​of resampled CT images and adjust them to a uniform numerical distribution standard to obtain normalized CT images.

[0078] Specifically, the normalization unit uses the z-score method to normalize the intensity of the resampled CT images, and the calculation formula is as follows:

[0079] ;

[0080] in, For normalized CT images, it represents the standardized value of each voxel in the resampled CT image after intensity normalization processing. The resampled CT value represents the value stored in each voxel of the resampled CT image. The mean of the resampled CT images represents the mean value within the lung region of the resampled CT images. The standard deviation of resampled CT images represents the standard deviation of the lung region in resampled CT images.

[0081] It should be noted that normalization units can eliminate inconsistencies in intensity caused by different scanning devices, scanning parameters, or individual patient differences.

[0082] The spatial alignment and cropping unit is used to obtain lung masks based on the organ at risk mask set. Using the centroid of the lung mask as a spatial reference point, the normalized CT image and GTV mask are cropped to a uniform size to obtain standardized data, which includes standardized CT images and standardized GTV masks.

[0083] Specifically, the spatial alignment and cropping unit first parses the endangered organ mask set, which includes multiple binary masks that are spatially aligned with the initial CT image and have structural name labels; then, it performs normalization mapping based on predefined structural names, filters out masks labeled "left lung" and "right lung" from the endangered organ mask set, merges them into a complete lung mask, and finally outputs the lung mask.

[0084] Furthermore, the spatial alignment and cropping unit calculates the centroid coordinates of the lung mask, uses these centroid coordinates as a spatial reference point, and uses a voxel of size 256×256×256 as the center to crop the normalized CT image and GTV mask to a uniform size, thus obtaining the standardized CT image and the standardized GTV mask.

[0085] Spatial alignment and cropping units ensure consistency of all input samples in anatomical space, enabling the system to learn features related to anatomical locations.

[0086] The physics-guided prior module is used to perform prior enhancement on standardized CT images based on the physical mechanism formed by dose bias, generate physical prior enhancement feature maps, and fuse and stitch standardized data to obtain dual-channel input tensors.

[0087] like Figure 3 As shown, the physics prior guidance module includes a spatial proximity calculation unit, a local density heterogeneity calculation unit, a beam path cumulative error calculation unit, a prior fusion and enhancement unit, a feature fusion unit, and a multimodal input unit.

[0088] Dose deviation refers to the dose distribution obtained by recalculating the dose using a pseudo-CT generated by the volume density assignment method; the physical mechanism of dose deviation refers to the fundamental physical cause of the dose calculation error caused by the volume density assignment method.

[0089] The spatial proximity calculation unit is used to calculate the spatial proximity factor, which quantifies the spatial proximity between the GTV and surrounding high-density structures. The spatial proximity factor ranges from (0,1). A larger spatial proximity factor indicates that the GTV is closer to the surrounding high-density structures within a preset area, i.e., the higher the spatial proximity. On the other hand, it also indicates a higher risk of dose deviation due to electron density averaging in the bulk density assignment method.

[0090] Furthermore, the formula for calculating the spatial proximity factor is as follows:

[0091] ;

[0092] Where i1 represents the voxel index, traversing the outer region of GTV; Represents the spatial proximity factor of voxel i1; It refers to a high-density collection of structures, including but not limited to the chest wall, mediastinum, and major blood vessels; Represents the set of voxels i1 to high-density structures. The minimum Euclidean distance; The dimension parameter is preferably 5 mm. This indicates the outer area of ​​the GTV (i.e., the preset area surrounding the GTV). Specifically, This indicates the area within 10mm of the GTV. As an indicator function, when voxel i1 is within the extended region of GTV, Conversely, The spatial proximity factor is non-zero only within a preset area surrounding the GTV.

[0093] The local density heterogeneity calculation unit is used to calculate the local density heterogeneity factor. The local density heterogeneity factor is used to quantify the degree of density non-uniformity of lung tissue in a preset area around GTV. The larger the local density heterogeneity factor, the higher the heterogeneity, that is, the higher the degree of density non-uniformity of lung tissue (i.e., the more intense the density fluctuation). On the other hand, it also indicates that the dose deviation risk generated by the bulk density assignment method is higher.

[0094] Furthermore, the formula for calculating the local density heterogeneity factor is as follows:

[0095] ;

[0096] Among them, voxel i2 traverses the lung tissue region. The local density heterogeneity factor representing voxel i2; This indicates a Gaussian smoothing operation, and the kernel size for the Gaussian smoothing operation is preferably 7×7×7; Represents standardized CT images; This represents the local Gaussian weighted mean of the squared normalized CT image at voxel i2; This represents the local Gaussian weighted mean of the standardized CT image at voxel i2.

[0097] The beam path cumulative error calculation unit is used to calculate the beam path cumulative error factor, which quantifies the cumulative electron density assignment error along the beam path under SBRT multi-beam irradiation. A larger beam path cumulative error factor indicates a larger electron density assignment error at voxel i, and also indicates a higher risk of dose deviation caused by the assignment method. It should be noted that SBRT multi-beam irradiation typically uses multiple coplanar or non-coplanar beams; the more high-density structures each beam path passes through, the greater the cumulative electron density assignment error.

[0098] Furthermore, the formula for calculating the cumulative error factor of the beam path is as follows:

[0099] ;

[0100] Among them, voxel i3 traverses all regions traversed by the beam path. is the cumulative error factor for the beam path of voxel i3, which is a pre-calculated beam path depth map representing the average projection depth of voxel i3 in each beam direction. This beam path depth map is generated offline based on the beam set parameters exported from the planning system; B is the number of beams, preset according to the clinical SBRT multi-beam irradiation plan, b=1,2,...,B; path b L represents the path traversed by the b-th beam; b is the path length of beam b within the patient's body; s represents the spatial coordinates of a point along the beam path; The error in assigning the electron density value to the spatial coordinate point s is approximated by the high-frequency components of the normalized CT image. Specifically, ,in, This represents the high-frequency component of the normalized CT image at point s.

[0101] The prior fusion and enhancement unit is used to generate a prior weight map by weighting and fusing the spatial proximity factor, local density heterogeneity factor, and beam path cumulative error factor. Based on the prior weight map, the normalized CT image is adaptively enhanced, thereby outputting a physical prior enhancement feature map.

[0102] It should be noted that the spatial proximity factor, local density heterogeneity factor, and beam path cumulative error factor are all normalized before being input into the prior fusion and enhancement units, so that the value ranges of the spatial proximity factor, local density heterogeneity factor, and beam path cumulative error factor are in a similar range.

[0103] Specifically, for each voxel i, voxel i traverses the outer region of the GTV, the lung tissue region, and all regions traversed by the beam path, and calculates the spatial proximity factor corresponding to voxel i for each region. Local density heterogeneity factor Beam path cumulative error factor (If a voxel is not within the traversal range of a factor, then the contribution of that factor to that voxel is 0), and then prior weights are generated through weighted summation, including:

[0104] ;

[0105] Among them, voxel i3 traverses all regions traversed by the beam path. This represents the prior weight of voxel i. The first fusion coefficient, For the second fusion coefficient, The third fusion coefficient, and the first fusion coefficient Second fusion coefficient and the third fusion coefficient The initial value is 1 / 3, and the fusion coefficient is automatically optimized through backpropagation.

[0106] If voxel i is not within the traversal range of the spatial proximity factor, then the contribution of the spatial proximity factor to voxel i is 0. Assuming voxel i is not within the ergodic range of the local density heterogeneity factor, then the contribution of the local density heterogeneity factor to voxel i is 0. Assuming voxel i is not within the traversal range of the cumulative error factor of the beam path, then the contribution of the cumulative error factor of the beam path to voxel i is 0. .

[0107] Assuming voxel i exists only within the ergodic range of the spatial proximity factor, the contributions of the local density heterogeneity factor and the beam cumulative error factor to voxel i are 0. and Assuming voxel i exists only within the ergodic range of the local density heterogeneity factor, then the contributions of the spatial proximity factor and the beam cumulative error factor to voxel i are 0. and Assuming voxel i exists only within the traversal range of the cumulative error factor of the beam path, then the contributions of the spatial proximity factor and the local density heterogeneity factor to voxel i are 0. and .

[0108] Arranging the prior weights of all voxels according to their spatial location creates a prior weight map with the same size as the normalized CT image. Each element represents the physical prior fusion weight of the corresponding voxel.

[0109] Then, the prior weight map is compared with the normalized CT images. Performing volumewise multiplication yields the physical prior enhanced feature map. And introduce residual joins to preserve the original information, specifically including:

[0110] ;

[0111] in, This represents element-wise multiplication; To enhance the intensity parameter, the initial value is 0.5, with a range of [0,1]. Physical prior enhancement feature map. The enhancement strength can be adaptively determined based on the physical prior; specifically, when the enhancement strength parameter... When the value approaches 0, the physics prior guidance module degenerates into an identity mapping; when the enhancement intensity parameter When the value approaches 1, the prior weights are fully applied to the original image.

[0112] The feature fusion unit is used to fuse standardized data to generate a GTV feature fusion map.

[0113] Specifically, the feature fusion unit performs voxel-by-voxel multiplication between the standardized CT image and the standardized GTV mask to generate an enhanced input that fuses the spatial information of the GTV, defined as the GTV feature fusion map. :

[0114] ;

[0115] in, For standardized CT images; For standardized GTV masks, within the standardized GTV mask area Outside the standardized GTV mask area .

[0116] The multimodal input unit is used to stitch standardized CT images with GTV feature fusion maps in the channel dimension to construct a dual-channel input tensor. A dual-channel input tensor is a three-dimensional tensor formed by stitching image data from two different sources or different modalities in the channel dimension.

[0117] The deep learning prediction module is used to output predicted probabilities using a risk prediction model. The risk prediction model receives either a physical prior enhanced feature map or a dual-channel input tensor output by the physical prior guidance module and outputs the predicted probabilities.

[0118] The risk prediction model supports two input modes: a first input mode using a physical prior guided feature map, and a second input mode using a dual-channel input tensor. Preferably, the second input mode can fully utilize GTV spatial information to improve prediction accuracy.

[0119] The risk stratification decision module is used to determine the risk type based on the predicted probability output by the deep learning prediction module and a preset classification threshold; and to trigger the corresponding dose calculation process based on the risk type.

[0120] like Figure 4 As shown, the risk stratification decision-making module includes a threshold setting unit, a risk judgment unit, a process triggering unit, and a plan review interface unit.

[0121] The threshold setting unit is used to set the classification threshold.

[0122] The threshold setting unit calculates the classification threshold using the Youden Index maximization principle. The specific calculation formula is as follows:

[0123] ;

[0124] Where T is the classification threshold; t is the candidate threshold variable, with a value range of [0,1]; The sensitivity of the candidate threshold variable t; represents the specificity of the candidate threshold variable t.

[0125] The risk assessment unit receives the predicted probability p from the deep learning prediction module, compares it with the classification threshold T, and outputs the risk type. When p ≥ T, the current case is classified as a high-risk case; when p < T, the current case is classified as a low-risk case.

[0126] The process triggering unit automatically triggers the corresponding dose calculation process based on the risk type. Specifically, for high-risk cases, the process triggering unit automatically triggers the lung sub-organ segmentation and enhancement assignment process. First, high-density structures such as pulmonary vessels and bronchial walls are automatically segmented. The block density assignment method is used to assign corresponding electron density values ​​to the segmented high-density structures, generating a pseudo-CT. Subsequently, dose recalculation and planning evaluation are performed based on the pseudo-CT to ensure dose accuracy. For low-risk cases, the conventional block assignment method is used to preset the electron density to generate a pseudo-CT, and dose calculation is completed to maintain the efficiency of the online adaptive radiotherapy process.

[0127] Through the process triggering unit, the lung sub-organ segmentation and enhancement assignment process is automatically triggered only for high-risk cases, while the block assignment method is used for low-risk cases. This avoids the online efficiency loss caused by uniform enhancement for all patients while ensuring dosage accuracy, and achieves selective enhancement.

[0128] The planned review interface unit is used to provide an access point for planned review after the dose calculation process is triggered, and to determine whether the key dose indicators meet the safety tolerance.

[0129] Specifically, for high-risk cases, after completing the enhancement assignment and dose calculation process, the planned review interface unit automatically compares the dose distribution before and after enhancement assignment. When the difference in key dose indicators exceeds the safety tolerance, manual review is triggered. Key metrological indicators include the D90 of the PTV region and the dose Gamma pass rate. In this embodiment, the safety tolerance includes a D90 of the PTV region lower than the prescribed dose of the PTV, or a dose Gamma pass rate (3mm / 3%) lower than 90%.

[0130] For low-risk cases, the rapid planning review step is retained to ensure that the difference between the volume density-assigned dose and the actual dose is within a safe tolerance. Rapid planning review refers to a process of quickly verifying only key metrological indicators and does not involve manual review.

[0131] In this embodiment, the system further includes a model training module, which is used to generate risk labels required for training the risk prediction model and to train the risk prediction model. It should be noted that the model training module is only activated during the offline data preparation phase before model training and does not participate in the online prediction process. Figure 5 As shown, the model training module includes a reference CT design unit, a pseudo-CT generation unit, a dose deviation calculation and labeling unit, a dataset partitioning unit, a data augmentation unit, and a model building unit.

[0132] The reference CT design unit is used to design reference treatment plans and generate reference dose distributions based on initial CT images of several historical patients, following standardized clinical guidelines. Specifically, based on the patient's initial CT images, a senior physicist designs a reference treatment plan according to standardized clinical guidelines, with the preferred prescription dose being 50 Gy / 5F, adhering to TG101 reporting, Timmerman target volume requirements, and OAR dose limits, to obtain a CT-based reference dose distribution. .

[0133] The pseudo-CT generation unit is used to generate pseudo-CT images from the same set of initial CT images using a volume density assignment method, and to calculate the corresponding dose distribution. Dose distribution refers to the numerical distribution of radiation dose within the three-dimensional space of the patient's body during radiotherapy, denoted as... It should be noted that the pseudo-CT generation unit only changes the electron density representation method, maintaining complete consistency in anatomical structure, spatial geometry, and contour relationships. Therefore, it can eliminate confounding factors such as anatomical differences, registration errors, and delineation variations, ensuring that dose deviation only reflects the error in the assignment-based electron density representation.

[0134] The dose deviation calculation and labeling unit is used to compare dose distributions. Compared with reference dose distribution The risk type is determined and a risk label is assigned based on preset conditions.

[0135] Specifically, the preset conditions are:

[0136] (1) PTV area deviation ;

[0137] (2) Overall pass rate (3% / 3mm standard) < 90%.

[0138] When any condition is met, the corresponding case is marked as a high-risk case (risk label = 1); otherwise, it is marked as a low-risk case (risk label = 0).

[0139] The PTV region refers to the core target area for treatment planning and dosage evaluation. This represents the percentage of the PTV volume that received a dose of ≥50 Gy, calculated using the following formula:

[0140] ;

[0141] in, The absolute volume of PTV with a dose ≥50 Gy; This indicates the total volume of the PTV.

[0142] Furthermore, PTV area deviation The calculation formula is:

[0143] ;

[0144] in, This indicates the volume of PTV receiving a 50 Gy dose in the reference treatment plan; This represents the volume of PTV receiving a 50 Gy dose in the bulk density assignment method.

[0145] The dataset partitioning unit is used to divide all labeled cases into training and test sets according to a preset ratio. The preset ratio is preferably 4:1 (training set:test set), where the training set is used for subsequent three-fold cross-validation.

[0146] The data augmentation unit is used to augment the physical prior enhanced feature maps and the dual-channel input tensor during the training phase. Data augmentation methods include:

[0147] (1) Random cropping: Using the center of the input image as a reference, random cropping with a maximum offset of 20% is performed, and the size of the cropped image is 128×128×128 voxels.

[0148] (2) Random mirroring: Performs horizontal or vertical mirroring operation on the cropped image with a 50% probability.

[0149] (3) Random rotation: The cropped image is randomly rotated within the range of 0° to 60° with a probability of 20%.

[0150] The above data augmentation strategy is executed dynamically in each iteration, which is equivalent to expanding the training sample space.

[0151] The model building unit uses a 3D-ResNet network to construct and train a risk prediction model. The risk prediction model includes an encoder and fully connected blocks.

[0152] The encoder consists of four residual block layers, each containing six 3D convolutional units. Each 3D convolutional unit uses a 3×3×3 convolutional kernel with a stride of 1. After each residual block layer, a max pooling layer is applied, with a 2×2×2 kernel size and a stride of 2, used to progressively reduce the spatial dimension of the image while expanding the receptive field.

[0153] Fully connected block: The feature map output by the encoder is sequentially processed by Flatten, Dropout, and fully connected layer operations, and finally output as the predicted probability after passing through the Sigmoid activation function. .

[0154] The binary cross-entropy loss function is used as the loss function for the risk prediction model. The specific formula is as follows:

[0155] ;

[0156] Where L is the binary crossover loss function; N is the number of samples in a batch, n=1,2,...,N; For the true risk label of sample n, when sample n is a high-risk case When sample n consists of low-risk cases ; Let n be the predicted probability of sample n.

[0157] In summary, this invention achieves a balance between dose accuracy and online efficiency in pulmonary MRgART through an innovative technical path of "pre-planning prediction—risk stratification—selective enhancement—plan review." The direct benefits to patients include a more precise, safer, and more convenient radiotherapy experience; the benefits to medical staff include reduced workload, improved decision-making efficiency, and smoother processes; and the benefits to the healthcare system and society include optimized resource allocation, reduced medical costs, and promotion of equitable access to and intelligent development of advanced technologies. These social benefits fully demonstrate the clinical application value and promotional significance of this invention.

[0158] Compared to existing conventional volume density assignment methods, this system requires the addition of a feature fusion unit in the data preprocessing module, a deep learning prediction module, and a risk stratification decision-making module. The main costs are concentrated in software algorithm development and maintenance, and a small amount of computing resource upgrades, requiring no additional hardware investment. Due to the standardized software deployment model, the marginal production cost is extremely low; the licensed production cost of a single software unit can be controlled within 5% of the existing process cost. Simultaneously, by improving the efficiency of online adaptive radiotherapy and reducing planned rework rates, the product can significantly increase the utilization rate of magnetic resonance accelerators, indirectly driving the sales of equipment consumables and related services. Furthermore, this product has the potential for international market expansion; with the global promotion of MRgART, its economic benefits will be further amplified.

[0159] 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 variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A hierarchical decision-making system for lung MRgART based on deep learning dose bias, characterized in that, The system includes a data preprocessing module, a physics prior guidance module, a deep learning prediction module, and a risk stratification decision-making module. The data preprocessing module is used to standardize the input initial data to obtain standardized data; the standardized data includes standardized CT images and standardized GTV masks. The data preprocessing module includes a resampling unit, a normalization unit, and a spatial alignment and cropping unit. The initial data includes initial CT images, GTV masks, and sets of masks for organs at risk; The resampling unit is used to convert the initial CT image from the original acquired voxel grid size to a uniform voxel size to obtain a resampled CT image. The normalization unit is used to mathematically transform the gray values ​​of the resampled CT images and adjust them to a uniform numerical distribution standard to obtain normalized CT images. The spatial alignment and cropping unit is used to obtain lung masks based on the endangered organ mask set. Using the centroid of the lung mask as a spatial reference point, the normalized CT image and GTV mask are cropped to a uniform size to obtain standardized data. The physics prior guidance module is used to perform prior enhancement on standardized CT images based on the physical mechanism formed by dose deviation, generate a physical prior enhancement feature map; and fuse and stitch standardized data to obtain a dual-channel input tensor. The physics prior guidance module includes a spatial proximity calculation unit, a local density heterogeneity calculation unit, a beam path cumulative error calculation unit, a prior fusion and enhancement unit, a feature fusion unit, and a multimodal input unit; The GTV mask is a binary representation of the gross tumor volume; The spatial proximity calculation unit is used to calculate the spatial proximity factor, which is used to quantify the spatial proximity between the gross tumor volume and the surrounding high-density structures. The local density heterogeneity is used to calculate the local density heterogeneity factor, which is used to quantify the degree of density inhomogeneity of lung tissue within a preset area surrounding the gross tumor volume. The beam path cumulative error calculation unit is used to calculate the beam path cumulative error factor, which is used to quantify the cumulative electron density assignment error along the beam path under SBRT multi-beam irradiation. The prior fusion and enhancement unit is used to generate a prior weight map by weighting and fusing the spatial proximity factor, local density heterogeneity factor, and beam path cumulative error factor, and to perform adaptive enhancement on the normalized CT image based on the prior weight map, and output a physical prior enhancement feature map. The feature fusion unit is used to fuse standardized data to generate a GTV feature fusion map; The multimodal input unit is used to stitch standardized CT images and GTV feature fusion maps in the channel dimension to construct a dual-channel input tensor; The deep learning prediction module is used to output the prediction probability using the risk prediction model; the risk prediction model receives the physical prior enhanced feature map or dual-channel input tensor output by the physical prior guidance module and outputs the prediction probability. The risk stratification decision module is used to determine the risk type based on the predicted probability and a preset classification threshold. And trigger the corresponding dose calculation process based on the risk type.

2. The lung MRgART hierarchical decision-making system based on deep learning dose bias according to claim 1, characterized in that, The formula for calculating the spatial proximity factor is as follows: ； Where i1 represents the voxel index, which traverses the outer region of the gross tumor volume; Represents the spatial proximity factor of voxel i1; Represents a high-density structure set; Represents the set of voxels i1 to high-density structures. The minimum Euclidean distance; Indicates the scale parameter; The outer region representing the gross tumor volume; For indicator functions; The formula for calculating the local density heterogeneity factor is as follows: ； Among them, voxel i2 traverses the lung tissue region. The local density heterogeneity factor representing voxel i2; Indicates Gaussian smoothing operation; Represents standardized CT images; This represents the local Gaussian weighted mean of the squared normalized CT image at voxel i2; This represents the local Gaussian weighted mean of the standardized CT image at voxel i2; The formula for calculating the cumulative error factor of the beam path is: ； Among them, voxel i3 traverses all regions traversed by the beam path. B is the cumulative error factor of the beam path for voxel i3; B is the number of beams, b=1,2,...,B; path b L represents the path traversed by the b-th beam; b is the path length of beam b within the patient's body; s represents the spatial coordinates of a point along the beam path; Assign error to the electron density at spatial coordinate point s. ,in, This represents the high-frequency component of the normalized CT image at point s; This represents a normalized image.

3. The lung MRgART hierarchical decision-making system based on deep learning dose bias according to claim 2, characterized in that, The calculation formula for the prior fusion and enhancement unit includes: ； Among them, voxel i traverses the outer region of the gross tumor volume, the lung tissue region, and all regions traversed by the beam path. This represents the prior weight of voxel i. The first fusion coefficient, For the second fusion coefficient, The third fusion coefficient; This represents the spatial proximity factor corresponding to voxel i; Indicates the local density heterogeneity factor; This represents the cumulative error factor of the beam path; The prior weights of all voxels are arranged according to their spatial positions to obtain the prior weight map. ; ； in, Enhanced feature maps for physical priors; This represents element-wise multiplication; To enhance strength parameters.

4. The lung MRgART hierarchical decision-making system based on deep learning dose bias according to claim 1, characterized in that, The calculation formula for the GTV feature fusion map is as follows: ； in, This is a GTV feature fusion map. For standardized CT images; For standardized GTV masks, within the standardized GTV mask area Outside the standardized GTV mask area .

5. The lung MRgART hierarchical decision-making system based on deep learning dose bias according to claim 1, characterized in that, The risk stratification decision-making module includes a threshold setting unit, a risk judgment unit, a process triggering unit, and a plan review interface unit. The threshold setting unit is used to set the classification threshold; The risk determination unit is used to receive the predicted probability output by the deep learning prediction module, compare it with the classification threshold, and output the risk type, which includes high-risk cases and low-risk cases. The process triggering unit is used to automatically trigger the corresponding dose calculation process according to the risk type; The planned review interface unit is used to provide an access point for planned review after the dose calculation process is triggered, and to determine whether the key dose indicators meet the safety tolerance.

6. The lung MRgART hierarchical decision-making system based on deep learning dose bias according to claim 1, characterized in that, The system also includes a model training module; the model training module is used to generate risk labels required for training the risk prediction model and to train the risk prediction model. The model training module includes a reference CT design unit, a pseudo-CT generation unit, a dose deviation calculation and labeling unit, a data augmentation unit, and a model building unit; The reference CT design unit is used to generate a reference dose distribution based on the initial CT images of several historical patients and in accordance with a reference treatment plan. The pseudo-CT generation unit is used to generate pseudo-CT on the same set of initial CT images using the block density assignment method, and to calculate the corresponding dose distribution. The dose deviation calculation and labeling unit is used to compare the dose distribution with the reference dose distribution, determine the risk type based on preset conditions, and label the risk. The data augmentation unit is used to augment the physical prior enhanced feature map and the dual-channel input tensor. The model building unit uses a 3D-ResNet network to build and train a risk prediction model.

Images